research article improving causality induction with...
TRANSCRIPT
Research ArticleImproving Causality Induction with Category Learning
Yi Guo12 Zhihong Wang1 and Zhiqing Shao1
1 Department of Computer Science and Engineering East China University of Science and TechnologyPO Box 408 Shanghai 200237 China
2 School of Information Science and Technology Shihezi University Shihezi 832003 China
Correspondence should be addressed to Yi Guo yguo71625foxmailcom
Received 7 April 2014 Accepted 14 April 2014 Published 30 April 2014
Academic Editor Shan Zhao
Copyright copy 2014 Yi Guo et al This is an open access article distributed under the Creative Commons Attribution License whichpermits unrestricted use distribution and reproduction in any medium provided the original work is properly cited
Causal relations are of fundamental importance for human perception and reasoning According to the nature of causality causalityhas explicit and implicit forms In the case of explicit form causal-effect relations exist at either clausal or discourse levels Theimplicit causal-effect relations heavily rely on empirical analysis and evidence accumulationThis paper proposes a comprehensivecausality extraction system (CL-CIS) integrated with the means of category-learning CL-CIS considers cause-effect relations inboth explicit and implicit forms and especially practices the relation between category and causality in computation In elaboratelydesigned experiments CL-CIS is evaluated together with general causality analysis system (GCAS) and general causality analysissystemwith learning (GCAS-L) and it testified to its own capability and performance in construction of cause-effect relationsThispaper confirms the expectation that the precision and coverage of causality induction can be remarkably improved by means ofcausal and category learning
1 Introduction
A general philosophical definition of causality states thatthe philosophical concept of causality refers to the set of allparticular causal or cause-and-effect relations [1] Causal rela-tions are of fundamental importance for human perceptionand reasoning Since ignoring causal relationships may havefatal consequences their knowledge plays a crucial role indaily life to ensure survival in an ever changing environment
In many research works the causality generally refersto the existence of causality in mathematics and physicsCausalities are often investigated in situations influencedby uncertainty involving several variables Thus causalitieswhich can be presented in terms of flows among processesor events are generally expressed in mathematical languagesand analyzed in a mathematical manner Therefore statisticsand probability theories seem to be the most popular math-ematical languages for modeling causality in most scientificdisciplines There is an extensive range of literature oncausality modeling applying and combining mathematicallogic graph theoryMarkovmodels Bayesian probability andso forth [2] But they seem not to be able to predominate allrelevant issues or questions
In recent years clarification and extraction of cause-effect relationships among texts (eg objects or events)causality extraction is elevated to a prominent researchtopic in text mining knowledge engineering and knowledgemanagement
A variety of research works testify that causalities intexts can be extracted at three technology levels The firstis the clausal level (CL) which includes cue phrases andlexical clues [3] and semantic similarity [4]The second is thediscourse level (DL) which implements connective markers[5] and constructs discourse relations [6] The third is themode level (ML) which extracts causal relations from a QAsystem [7 8] applies commonsense rules [9] associativememory [10] and Chain Event Graphs [11]
According to the nature of causality in texts causalityhas explicit and implicit forms In the case of explicit formcausal-effect relations exist at either clausal or discourselevels The implicit causal-effect relations heavily rely onempirical analysis and evidence accumulation
Moreover Waldmann and Hagmayer [12] in theirresearch work of cognitive psychology state that ldquocategoriesthat have been acquired in previous learning contexts mayinfluence subsequent causal learningrdquo which indicate that
Hindawi Publishing Corporatione Scientific World JournalVolume 2014 Article ID 650147 8 pageshttpdxdoiorg1011552014650147
2 The Scientific World Journal
(1) the category information about objects or events in textsis a necessary supplement for causality extraction and (2)the category information has impact on subsequent learning-based cause-effect identification
Based on the facts the task of causality extraction eveninduction in texts could not be accomplished in an arbitrarymanner This paper proposes a comprehensive causalityextraction system (CL-CIS) integrated with the means ofcategory-learning CL-CIS considers cause-effect relations inboth explicit and implicit forms and especially practices therelation between category and causality in computation
The rest of this paper is organized as follows Section 2states the causality category information and the relation-ships between them in texts and constructs the theoreticalfoundation of our research work Section 3 expatiates themethodology and the technical details of system structureand kernel algorithms Section 4 focuses on experimentillustration and result analysis of the experimental resultsSection 5 concludes this paper and provides future researchworks
2 Causality and Category
Traditionally research about the representation of causalrelations and research about the representation of categorieswere separatedThis research strategy rests on the assumptionthat categories summarize objects or events on the basis oftheir similarity structure whereas causality refers to relationsbetween causal objects or events Literature [12] proves thatthe relationship between causality and categorization is moredynamic than previously thought
21 Causality The standard view guiding research on causal-ity presupposes the existence of objective networks of causesand effects that cognitive systems try to mirror Regardlessof whether causal learning is viewed as the attempt to inducecausality on the basis of statistical information or on the basisof mechanism information it is generally assumed that thegoal of causal learning is to form adequate representations ofthe texture of the causal world
22 Causality Rests on Fixed Categories Studies on causallearning typically investigate trial-by-trial learning taskswhich involve learning the contingencies between causes andeffects In a large number of studies which focus on causalcontingency learning A characteristic feature of these tasksis that they present categorized events representing causesand effects which are statistically related Cause and effectcategories are viewed as fixed entities that are already presentprior to the learning task The goal of learning is to estimatecausal strength of individual causal links or to induce causalmodels on the basis of observed covariations The role ofcause and effect categories in the learning process is not thefocus of interest in these approaches they are simply viewedas given
A similar approach underlies research on the relationshipbetween categories and causality According to the view thatcategorization is theory-based traditional similarity-based
accounts of categorization are deficient because they ignorethe fact that many categories are grounded in knowledgeabout causal structures [13] In natural concepts featuresoften represent causes or effects with the category labelreferring to a complex causalmodel For example disease cat-egories frequently refer to common-cause models of diseaseswith the category features representing causes (eg virus) andeffects (eg symptoms) within this causal model A numberof studies using these and similar materials have shown thatthe type of causalmodel connecting otherwise identical causeand effect features influences learning typicality judgmentsor generalization [14ndash17] The main goal of these studieswas to investigate the effect of different causal relationsconnecting the causal features As in contingency learningstudies the cause and effect features within the causal modelswere treated as fixed categorized entities which alreadyexisted prior to the learning context
23 Categories Shape Causality It is certainly true that manyinteresting insights can be gained from investigating howpeople learn about causal models on the basis of preexistingcause and effect categories However there is also a linkbetween categories and causality in the opposite directionThe categories that have been acquired in previous learningcontexts may have a crucial influence on subsequent causallearning
The basis of the potential influence of categories on causalinduction lies in the fact that the acquisition and use of causalknowledge is based on categorized events Causal laws suchas the fact that smoking causes heart disease can only benoticed on the basis of events that are categorized (eg eventsof smoking and cases of heart disease)
Without such categories causal laws neither could bedetected nor could causal knowledge be applied to newcases Thus causal knowledge not only affects the creation ofcategories it also presupposes already existing categories forthe description of causes and effects The potential influenceof categories is due to the fact that one of the most importantcues to causality is statistical covariation between causes andeffects
To study the relation between categories and causalinduction [12] have developed a new paradigm that consistsof three phases the category learning phase the causallearning phase and the third test phase The main goal isto answer the question that under what condition learnerswill tend to activate the categorical information (Figure 1)from the earlier category learning phase when learning aboutcausal contingencies in the later phase This effect is entailedby the fact that the alternative categories form differentreference classes
In category learning phase causes will be classified intothe distinct categories (upper left arrow in Figure 1) Incausal learning phase causes in collection are paired withthe presence or absence of an effect (lower arrow) In thesubsequent test phase the test causes are rated with thelikelihood of the effect The crucial question is when peoplewould go through the upper route in Figure 1 and assign thetest causes to the categories in the category learning phase or
The Scientific World Journal 3
Categories
EffectsCauses
Figure 1 Possible routes of category learning between causes andeffects
Categories
EffectsCausesCauses
Category learning
learning
Figure 2 Enhanced strategy of category learning between causesand effects
whether they would stick to the lower route and induce newcategories within causal learning phase
If the learning strategy opted for the lower route onepossible solution may be to induce new categories that aremaximally predictive of the effects This solution wouldobviously generate maximally predictive categories Lien andCheng [18] reported a research that is consistent with themaximal-contrast hypothesis Their experiments show thatthe substances are categorized according to the feature and tothe hierarchical level that were maximally predictive for theeffect Thus the induced substance category was determinedby its suitability for predicting the effect Lien and Cheng[18] interpreted this as evidence for their maximal-contrasthypothesis In another word people tend to induce categoriesthat maximize their causal predictability
In sum our research addresses the question of whichroute learners will go Will they routinely go through theupper road and activate category knowledge when learningabout novel effects A simple connectionist one-layer net-work that may be used to understand the task we are going toexplore in our experiments or will they go the lower road andlearn a new set of categories on the basis of causal informationin causal learning phase that is maximally predictive withinthis learning phase Thus Figure 1 is enhanced with twolearning phases as shown in Figure 2
3 Research Methodology andSystem Structure
31 Research Methodology Causality is a fundamental con-cept in human thinking and reasoning It is not surprisingthat most if not all languages in the world have a range of
Table 1 Statistics for backward causality connectives
Connectives For objectivereason
For subjectivereason
Total percentage inall connectives
Because 83 17 50For 29 71 18As 60 40 13Since 14 86 6While 14 86 6Misc 0 100 7Total 56 44 100
lexical expressions specifically designed for communicatingcausal relations
This paper focuses on explicit causality markers andimplicit causal relations in text The explicit causality mark-ers include two grammatically different types of causalitymarkers in English We investigate the semantic contrastsexpressed by different causal auxiliary verbs marking causalrelations expressed within one clause and those expressedby different causal connectives marking causal relationsbetween clauses
311 Causality in Verbs Some instances of causality in verbsare listed below These verbs include but are not limitedto ldquomakerdquo ldquoletrdquo ldquohaverdquo ldquocauserdquo and their synonyms fromWordNet [19] which is generally referred to as an onlinelexical database
[The extreme cold] cause madecaused even [the rivers(to) freeze] effect[She] cause madehad [her son empty his plate] effectdespite his complaints
312 Causality in Discourse Connectives For a complexsentence the predicative or relative clauses of the nounsynonyms of ldquocauserdquo are labeled as ldquopotential causerdquo ofantecedent sentence or main clause Additionally each clauseinducted with ldquosincerdquo ldquoasrdquo or ldquobecauserdquo is also labeled asldquopotential causerdquo of its main clause which is labeled asldquopotential effectrdquo
Table 1 lists the statistics for above backward causalitycollected upon Reuters-21578 currently the most widelyused test collection for text processing research and BBC-News2000 collected by a self-developed Web Crawler fromhttpwwwbbccouk The forward causality where in pre-sentation order the cause precedes the effect is the mostfrequently used ones for example ldquothereforerdquo ldquobecause ofthatrdquo ldquothat is whyrdquo and ldquosordquo As the forward causality con-nectives are one hundred percent strong causality indicatorsthe backward causality connectives in Table 1 need morespecific notation For example in all ldquobecauserdquo connectivesin collections 83 of them indicate objective reasons whilethe other 17 indicate subjective reasons Meanwhile theldquobecauserdquo connectives persist 50 of all the connectivesincluding ldquobecauserdquo ldquoforrdquo ldquoasrdquo ldquosincerdquo ldquowhilerdquo and othermiscellaneous connectivesThe ldquoforrdquo ldquoasrdquo ldquosincerdquo ldquowhilerdquo and
4 The Scientific World Journal
Table 2 Types of implicit causality sentences
Types Subtypes Exemplar sentences
Compoundsentences
Cause-effect sentencesconnected with ldquoandrdquo
Cause-effect(1)This was the first time I made an international call and my heart was beating fast(2) The filter is much more efficient than the primary filter and it removes allremaining solid particles from the fuel
Effect-cause(3) The crops had failed and there had not been any rain for months(4) Aluminum is used as engineering material for planes and spaceships and it islight and tough
Cause-effect sentenceswithout connectives
Effect-cause(5) My heart sank Some of them did not weigh more than 135 pounds and theothers looked too young to be in trouble(6) but the Voyagerrsquos captain and three crewmen had stayed in board They hadhoped the storm would die out and they could save the shipCause-effect(7) The red distress flare shot up from the sinking ship the Voyager Everymanaboard our Coast Guard cutter knew that time had run out
Relative clauses SV SVO SVC SVOC SVOOSVA SVOA
(8) To make an atom we have to use uranium in which the atoms are available forfission(9) We know that a cat whose eyes can take in more rays of light than our eyes cansee clearly in the night
If clauses
(10) This system of subsidies must be maintained if the farmer will sufferconsiderable losses if it is abolished(11) If the water will rise above this level we must warn everybody in theneighborhood
That clauses (12) The act was even the bolder that he stood utterly alone(13) The crying was all the more racking that she was not hysterical but hopeless
SVO-SVOC(14) Her falling ill spoiled everything(15) Timidity and shamefacedness caused her to stand back and looked indifferentlyaway
other miscellaneous connectives correspondingly hold 1813 6 6 and 7 of all the connectives in collections
313 Implicit Causality in Texts Causal effect relations aregeneral connections in the objective world and the causalitysentences exist in languages in a pervasive manner Theexplicit causality sentences are those conducted by backwardand forward causality connectives listed in Section 312while the implicit causality sentences are those connectedwith other connectives even without any one Our researchworks have testified that there are five types of implicitcausality sentences shown in Table 2
In the practical English texts there exist a few inter-personal verbs like ldquopraiserdquo and ldquoapologizerdquo thus supplyinginformation about whose behavior or state is the more likelyimmediate cause of the event at hand Because it is conveyedimplicitly as part of themeaning of the verb this probabilisticcue is usually referred to as implicit causality
Such exemplar implicit causality verbs [20] adopted inthis paper are listed in Table 3 together with their biasindicating the probabilities as causal cues for example 100is the causal baseline the higher the bias value is the morelikely is the cause
32 System Description Our causality induction system withassistance of category learning named CL-CIS is composedwith the following functional modules (shown in Figure 3)category learning classify exemplars into categories categoryand causal mapping causal learning building causal-effectrelations and the final testing module CL-CIS also includesthree reference libraries (databases) causality in verbscausality in discourse connectives and implicit causalityfor query and assistance Table 4 compares the compositiondifference among general causality analysis system (GCAS)general causality analysis system with learning (GCAS-L)and CL-CIS As the construction of three libraries is elabo-rated in Section 31 Methodology this section concentrateson the six functional modules
321 Category Learning (CL) The category learning modulebuilds up different distinct and exhaustive categories accord-ing to semantic contents (eg noun phrases verb phrases)of sentences in our text collections This module is a basicpreprocessing step and the target is to construct a collectionof distinct and exhaustive categories with the text learningmethods
The Scientific World Journal 5
Table 3 Exemplar implicit causality verbs
NP1 verbs Bias NP2 verbs BiasAmazed 119 Admire 200Annoy 119 Adore 186Apologize 100 Appreciate 200Be in the way 108 Comfort 191Beg 117 Compliment on something 196Bore 105 Congratulate 195Call 119 Criticize 200Confess 103 Envy 200Disappoint 103 Fear 200Disturb 114 Fire 196Fascinate 100 Hate 196Hurt 113 Hold in contempt 200Inspire 123 Hold responsible 191Intimidate 123 Loathe 186Irritate 122 Love 186Lie to 122 Praise 196Mislead 122 Press charges against 191Swindle 113 Punish 195Win 119 Respect 195Worry 119 Thank 182
Table 4 System composition comparison of GCAS GCAS-L andCL-CIS
System composition GCAS GCAS-L CL-CISModules
Category learning radic
Classify exemplars intocategories radic
Category and causal mapping radic
Causal learning radic radic
Building causal-effectrelations radic radic radic
Testing causal-effectrelations radic radic radic
LibrariesCausality in verbs radic radic radic
Causality in discourseconnectives Optional radic radic
Implicit causality Optional radic
322 Classify Exemplars into Categories (CEC) In the clas-sify exemplars into categories module each sentence istreated as an individual independent event and parsed intophrases The classification is based on the comparison ofa sentence with features of each category so as to set upthe category background knowledge for each sentence Forexample the concept bank can be clarified as a financeorganization or a river body boundary with correspondingcategory background knowledge of its sentence and contexts
Category learning
Classify exemplars into
categories
Category and causal
mapping
Causal learning
Building causal-effect
relations
Testing causal-effect
relations
Causality in verbs
Causality in discourse
connectivesImplicit causality
Figure 3 System framework of CL-CIS
Inputlayer (IL)
Recurrentlayer (RL)
Outputlayer (OL)
Contextlayer (CL)
Time delay
OU1 OU2
RU1RU2
WOR
IU1 IU2
WRI W
RC
CU1 CU2
middot middot middot
middot middot middot
middot middot middot middot middot middot
OU|O|
RU|R|
IU|I| CU|C|
Figure 4 Architecture of SRNN
323 Category and Causal Mapping (CCM) This moduleconstructs the category-causal-effect mapping relations forexample a virus from a predefined category causes specificdisease-related symptoms such as a swelling of the spleen(splenomegaly) The construction mechanism to a greatextent is based on collection storage and indexing ofmassive cases
324 Causal Learning (CauL) As analyzed before a sim-ple connectionist one-layer network that may be used tounderstand the task of when to activate category knowledgewhen learning about novel effects This module adopts asimple recurrent neural network (SRNN) to simulate theconnectionist model
Symbols in Figure 4 are defined in Table 5 the first orderweight matrices WRI and WOR fully connect the units ofthe input layer (IL) the recurrent layer (RL) and the outputlayer (OL) respectively as in the feed forward multilayerperceptron (MLP) The current activities of recurrent unitsRU(119905) are fed back through time delay connections to thecontext layer which is presented as CU(119905+1) = RU(119905)
Therefore each unit in recurrent layer is fed by activitiesof all recurrent units from previous time step through
6 The Scientific World Journal
Table 5 Definition of SRNN Symbols
Symbols DefinitionIU A unit of input layerRU A unit of recurrent layerCU A unit of context layerOU A unit of output layer|119868| The number of units in IL|119877| The number of units in RL|119862| The number of units in CL|119874| The number of units in OLWRI The weight vector from IL to RLWRC The weight vector from CL to RLWOR The weight vector from RL to OL
recurrent weight matrix WRC The context layer which iscomposed of activities of recurrent units from previous timestep can be viewed as an extension of input layer to therecurrent layer Above working procedure represents thememory of the network via holding contextual informationfrom previous time steps
The weight matrices 119882RI 119882RC and 119882OR are presentedas follows
119882RI= [(119908
RI1 )119879 (119908
RI2 )119879 (119908
RI|119877|)119879]
=
[[[[[
[
119908ri11 119908
ri12 sdot sdot sdot 119908
ri1|119877|
119908ri21 119908ri22 sdot sdot sdot 119908
ri1|119877|
d
119908ri|119868|1 119908
ri|119868|2 sdot sdot sdot 119908
ri|119868||119877|
]]]]]
]
119882RC= [(119908
RC1 )119879 (119908
RC2 )119879 (119908
RC|119877| )119879]
=
[[[[
[
119908rc11 119908rc12 sdot sdot sdot 119908
rc1|119877|
119908rc21 119908rc22 sdot sdot sdot 119908
rc1|119877|
d
119908rc|119862|1 119908
rc|119862|2 sdot sdot sdot 119908
rc|119862||119877|
]]]]
]
119882OR= [(119908
OR1 )119879 (119908
OR2 )119879 (119908
OR|119874| )119879]
=
[[[[
[
119908or11 119908or12 sdot sdot sdot 119908
or1|119874|
119908or21 119908or22 sdot sdot sdot 119908
or1|119874|
d
119908or|119877|1 119908
or|119877|2 sdot sdot sdot 119908
or|119877||119874|
]]]]
]
(1)
In the above formulations (119908RI119896 )119879 is the transpose of 119908RI119896
for the instance of119882RI where 119908RI119896 is a row vector and (119908RI119896 )119879
is the column vector of the same elements The vector 119908RI119896 =(119908ri1119896 119908
ri2119896 119908
ri|119868|119896) represents the weights from all the input
layer units to the recurrent (hidden) layer unit RU119896The sameconclusion applies with119882RC and119882OR
Given an input pattern in time t IU(t) = (IU(119905)1 IU(119905)2
IU(119905)|119868|) and recurrent activities RU(t) = (RU(119905)1 RU
(119905)2
RU(119905)|119877|) for the 119894th recurrent unit the net input RU1015840(119905)
119894and
output activity RU(119905)119894
are calculated as follows
RU1015840(119905)119894 = IU(t)sdot (119908
RI119894 )119879+ RU(tminus1) sdot (119908RC119894 )
119879
=
|119868|
sum119895=1
IU(119905)119895 119908ri119895119894 +
|119877|
sum119895=1
RU(119905minus1)119895 119908rc119895119894
RU(119905)119894 = 119891 (RU1015840(119905)
119894 )
(2)
For the kth output unit its net inputOU1015840(119905)119896
and output activityOU(119905)119896
are calculated as (3) Consider the following
OU1015840(119905)119896= RU(t) sdot (119908OR119896 )
119879=
|119877|
sum119895=1
RU(119905)119895 119908or119895119896
OU(119905)119896= 119891 (OU1015840(119905)
119896)
(3)
Here the activation function 119891 applies the logistic sigmoidfunction (4) in this paper Consider the following
119891 (119909) =1
1 + 119890minus119909=119890119909
1 + 119890119909 (4)
325 Building Causal-Effect Relations (BCER) This modulebuilds up the ldquocause-effectrdquo pairs and connections with theinputs of SRNN and corresponding outputs The causalitydetection results could include four possible types (1) singlecause-effect pairs in which any two pairs are independentfrom each other (2) cause-effect chains which are formedwithmore than one cause-effect pairs connected together (aneffect is a cause of another effect) (3) one-cause-multiple-effect pairs (4) multiple-cause-one-effect pairs All of thecausality connections are archived in a database Both CauLandBCERmodules exploit three libraries (databases) causal-ity in verbs causality in discourse connectives and implicitcausality for reference
326 Testing Causal-Effect Relations (TCER) This moduletests and verifies a new processed causal-effect relation withour existing and expanding collection ofmassive causal-effectrelations The concrete technology includes comparison oftriples ⟨CategoryCauseEffect⟩
4 Experiments and Results
Our experiments test general causality analysis system(GCAS) general causality analysis system with learning(GCAS-L) and CL-CIS together in order to examine andreveal the assertion that category information is a necessarysupplement for causality extraction and has impact on subse-quent learning-based cause-effect identification
In our experiments we have used two text collections (1)Reuters-21578 currently the most widely used test collectionfor text processing research (2) BBC-News2000 collected bya self-developed Web Crawler from httpwwwbbccouk
The Scientific World Journal 7
00
01
02
03
04
05
06
07
08
09
10
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
0681 0706 069807320719
0755 074207930776
0815 0809
0897
(a) Precision
00
01
02
03
04
05
06
07
08
09
10
0593 0581 0602
0706
0621 0609 0616
0734
0685 0661 0683
0805
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
(b) Recall
00
01
02
03
04
05
06
07
08
09
10
0634 0637 0646 0719
0666 0674 0673 0762
0728 0730 0741
0849
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
GCASGCAS-L
CL-CIS
(c) F-measure (balanced)
Figure 5 Evaluation results of GCAS GCAS-L and CL-CIS
BBC-News2000 collected 2000 news articles which havebeen pruned off unnecessary Web page elements such ashtml tags images URLs and so forth BBC-News2000 is notcategorized and is treated as a whole hybrid
30 judges are involved to manually search cause-effectpairs and construct standard causality references (SCR) SCR-Reuters for Reuters-21578 and SCR-BBC for BBC-News2000Due to limited human resources in current stage SCR-Reuters only covers 2000 documents newid ranges from 1 to2000 contained in files ldquoreut2-000sgmrdquo and ldquoreut2-001sgmrdquoThe rest of the files of Reuters-21578 are still involved in thetraining phase
The performances of GCAS GCAS-L and CL-CIS areevaluated with precision recall and F-measure [21] thetraditional measures that have been widely applied by mostinformation retrieval systems to analyze and evaluate theirperformance The F-measure is a harmonic combination ofthe precision and recall values used in information retrievalAs shown in Tables 6 7 and 8 the experimental results statethat (1) GCAS scores from 0681 to 0732 on precision andfrom 0581 to 0706 on recall (2) GCAS-L scores from 0719to 0793 on precision and from 0609 to 0734 on recall (3)CL-CIS scores from 0776 to 0897 on precision and from0661 to 0805 on recall Figure 5 explicitly states that (1)
Table 6 Experimental results of GCAS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0681 0593 0634reut2-001sgm (newid 1001ndash2000) 0706 0581 0637Both (reut2) (newid 1ndash2000) 0698 0602 0646BBC-News2000 0732 0706 0719
Table 7 Experimental results of GCAS-L
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0719 0621 0666reut2-001sgm (newid 1001ndash2000) 0755 0609 0674Both (reut2) (newid 1ndash2000) 0742 0616 0673BBC-News2000 0793 0734 0762
the general causality analysis system with learning (GCAS-L) performances better than the general causality analysissystem (GCAS) on all evaluation measures (2) the causalityanalysis system strengthened with causal and category learn-ing (CL-CIS) exceeds GCAS-L in the meantime
8 The Scientific World Journal
Table 8 Experimental results of CL-CIS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0776 0685 0728reut2-001sgm (newid 1001ndash2000) 0815 0661 0730Both (reut2) (newid 1ndash2000) 0809 0683 0741BBC-News2000 0897 0805 0849
5 Concluding Remarks
In recent years detection and clarification of cause-effectrelationships among texts events or objects has been elevatedto a prominent research topic of natural and social sciencesover the human knowledge development history
This paper demonstrates a novel comprehensive causalityextraction system (CL-CIS) integrated with the means ofcategory-learning CL-CIS considers cause-effect relations inboth explicit and implicit forms and especially practices therelation between category and causality in computation
CL-CIS is inspired with cognitive philosophy in categoryand causality In causality extraction and induction tasks CL-CIS implements a simple recurrent neural network (SRNN)to simulate human associative memory which has the abilityto associate different types of inputs when processing infor-mation
In elaborately designed experimental tasks CL-CIS hasbeen examined in full with two text collections Reuters-21578and BBC-News2000 The experimental results have testifiedthe capability and performance of CL-CIS in constructionof cause-effect relations and also confirmed the expectationthat the means of causal and category learning will improvethe precision and coverage of causality induction in a notablemanner
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work is financially supported by the National NaturalScience Foundation of China (Grant no 61003126)
References
[1] httpenwikipediaorgwikiCausality[2] J Pearl Causality Models Reasoning and Inference Cambridge
University Press New York NY USA 2000[3] D-S Chang and K-S Choi ldquoIncremental cue phrase learning
and bootstrapping method for causality extraction using cuephrase and word pair probabilitiesrdquo Information Processing andManagement vol 42 no 3 pp 662ndash678 2006
[4] S Kim R H Bracewell and K M Wallace ldquoA frameworkfor automatic causality extraction using semantic similarityrdquoin Proceedings of the ASME International Design EngineeringTechnical Conferences and Computers and Information in Engi-neering Conference (IDETCCIE rsquo07) pp 831ndash840 Las VegasNev USA September 2007
[5] T Inui K Inui andYMatsumoto ldquoAcquiring causal knowledgefrom text using the connective markersrdquo Transactions of theInformation Processing Society of Japan vol 45 no 3 pp 919ndash932 2004
[6] D Marcu and A Echihabi ldquoAn unsupervised approach torecognizing discourse relationsrdquo in Proceedings of the 40thAnnual Meeting of the Association for Computational LinguisticsConference (ACL rsquo02) pp 368ndash375 University of PennsylvaniaPhiladelphia Pa USA 2002
[7] C Pechsiri and A Kawtraku ldquoMining causality from texts forquestion answering systemrdquo IEICE Transactions on Informationand Systems vol 90 no 10 pp 1523ndash1533 2007
[8] R Girju ldquoAutomatic detection of causal relations for questionansweringrdquo in Proceedings of the 41st Annual Meeting of theAssociation for Computational Linguistics 2003
[9] K Torisawa ldquoAutomatic extraction of commonsense inferencerules from corporardquo in Proceedings of the 9th Annual Meeting ofthe Association for Natural Language Processing 2003
[10] Y Guo N Hua and Z Shao ldquoCognitive causality detectionwith associative memory in textual eventsrdquo in Proceedings ofthe IEEE International Symposium on Information Engineeringand Electronic Commerce (IEEC rsquo09) pp 140ndash144 TernopilUkraine May 2009
[11] P Thwaites J Q Smith and E Riccomagno ldquoCausal analysiswith Chain Event GraphsrdquoArtificial Intelligence vol 174 no 12-13 pp 889ndash909 2010
[12] M R Waldmann and Y Hagmayer ldquoCategories and causalitythe neglected directionrdquo Cognitive Psychology vol 53 no 1 pp27ndash58 2006
[13] G L MurphyTheBig Book of Concepts MIT Press CambridgeMass USA 2002
[14] B Rehder ldquoA causal-model theory of conceptual representationand categorizationrdquo Journal of Experimental Psychology Learn-ing Memory and Cognition vol 29 no 6 pp 1141ndash1159 2003
[15] B Rehder ldquoCategorization as casual reasoningrdquo Cognitive Sci-ence vol 27 no 5 pp 709ndash748 2003
[16] B Rehder and R Hastie ldquoCausal knowledge and categoriesthe effects of causal beliefs on categorization induction andsimilarityrdquo Journal of Experimental Psychology General vol 130no 3 pp 323ndash360 2001
[17] B Rehder and R Hastie ldquoCategory coherence and category-based property inductionrdquo Cognition vol 91 no 2 pp 113ndash1532004
[18] Y Lien and PWCheng ldquoDistinguishing genuine from spuriouscauses a coherence hypothesisrdquo Cognitive Psychology vol 40pp 87ndash137 2000
[19] G A Miller R Beckwith C Fellbaum D Gross and K JMiller ldquoIntroduction to WordNet an on-line lexical databaserdquoInternational Journal of Lexicography vol 3 no 4 pp 235ndash3121990
[20] A W Koornneef and J J A Van Berkum ldquoOn the use of verb-based implicit causality in sentence comprehension evidencefrom self-paced reading and eye trackingrdquo Journal of Memoryand Language vol 54 no 4 pp 445ndash465 2006
[21] C D Manning and H Schutze Foundations of StatisticalNatural Language Processing TheMIT Press Cambridge MassUSA 4th edition 2001
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
2 The Scientific World Journal
(1) the category information about objects or events in textsis a necessary supplement for causality extraction and (2)the category information has impact on subsequent learning-based cause-effect identification
Based on the facts the task of causality extraction eveninduction in texts could not be accomplished in an arbitrarymanner This paper proposes a comprehensive causalityextraction system (CL-CIS) integrated with the means ofcategory-learning CL-CIS considers cause-effect relations inboth explicit and implicit forms and especially practices therelation between category and causality in computation
The rest of this paper is organized as follows Section 2states the causality category information and the relation-ships between them in texts and constructs the theoreticalfoundation of our research work Section 3 expatiates themethodology and the technical details of system structureand kernel algorithms Section 4 focuses on experimentillustration and result analysis of the experimental resultsSection 5 concludes this paper and provides future researchworks
2 Causality and Category
Traditionally research about the representation of causalrelations and research about the representation of categorieswere separatedThis research strategy rests on the assumptionthat categories summarize objects or events on the basis oftheir similarity structure whereas causality refers to relationsbetween causal objects or events Literature [12] proves thatthe relationship between causality and categorization is moredynamic than previously thought
21 Causality The standard view guiding research on causal-ity presupposes the existence of objective networks of causesand effects that cognitive systems try to mirror Regardlessof whether causal learning is viewed as the attempt to inducecausality on the basis of statistical information or on the basisof mechanism information it is generally assumed that thegoal of causal learning is to form adequate representations ofthe texture of the causal world
22 Causality Rests on Fixed Categories Studies on causallearning typically investigate trial-by-trial learning taskswhich involve learning the contingencies between causes andeffects In a large number of studies which focus on causalcontingency learning A characteristic feature of these tasksis that they present categorized events representing causesand effects which are statistically related Cause and effectcategories are viewed as fixed entities that are already presentprior to the learning task The goal of learning is to estimatecausal strength of individual causal links or to induce causalmodels on the basis of observed covariations The role ofcause and effect categories in the learning process is not thefocus of interest in these approaches they are simply viewedas given
A similar approach underlies research on the relationshipbetween categories and causality According to the view thatcategorization is theory-based traditional similarity-based
accounts of categorization are deficient because they ignorethe fact that many categories are grounded in knowledgeabout causal structures [13] In natural concepts featuresoften represent causes or effects with the category labelreferring to a complex causalmodel For example disease cat-egories frequently refer to common-cause models of diseaseswith the category features representing causes (eg virus) andeffects (eg symptoms) within this causal model A numberof studies using these and similar materials have shown thatthe type of causalmodel connecting otherwise identical causeand effect features influences learning typicality judgmentsor generalization [14ndash17] The main goal of these studieswas to investigate the effect of different causal relationsconnecting the causal features As in contingency learningstudies the cause and effect features within the causal modelswere treated as fixed categorized entities which alreadyexisted prior to the learning context
23 Categories Shape Causality It is certainly true that manyinteresting insights can be gained from investigating howpeople learn about causal models on the basis of preexistingcause and effect categories However there is also a linkbetween categories and causality in the opposite directionThe categories that have been acquired in previous learningcontexts may have a crucial influence on subsequent causallearning
The basis of the potential influence of categories on causalinduction lies in the fact that the acquisition and use of causalknowledge is based on categorized events Causal laws suchas the fact that smoking causes heart disease can only benoticed on the basis of events that are categorized (eg eventsof smoking and cases of heart disease)
Without such categories causal laws neither could bedetected nor could causal knowledge be applied to newcases Thus causal knowledge not only affects the creation ofcategories it also presupposes already existing categories forthe description of causes and effects The potential influenceof categories is due to the fact that one of the most importantcues to causality is statistical covariation between causes andeffects
To study the relation between categories and causalinduction [12] have developed a new paradigm that consistsof three phases the category learning phase the causallearning phase and the third test phase The main goal isto answer the question that under what condition learnerswill tend to activate the categorical information (Figure 1)from the earlier category learning phase when learning aboutcausal contingencies in the later phase This effect is entailedby the fact that the alternative categories form differentreference classes
In category learning phase causes will be classified intothe distinct categories (upper left arrow in Figure 1) Incausal learning phase causes in collection are paired withthe presence or absence of an effect (lower arrow) In thesubsequent test phase the test causes are rated with thelikelihood of the effect The crucial question is when peoplewould go through the upper route in Figure 1 and assign thetest causes to the categories in the category learning phase or
The Scientific World Journal 3
Categories
EffectsCauses
Figure 1 Possible routes of category learning between causes andeffects
Categories
EffectsCausesCauses
Category learning
learning
Figure 2 Enhanced strategy of category learning between causesand effects
whether they would stick to the lower route and induce newcategories within causal learning phase
If the learning strategy opted for the lower route onepossible solution may be to induce new categories that aremaximally predictive of the effects This solution wouldobviously generate maximally predictive categories Lien andCheng [18] reported a research that is consistent with themaximal-contrast hypothesis Their experiments show thatthe substances are categorized according to the feature and tothe hierarchical level that were maximally predictive for theeffect Thus the induced substance category was determinedby its suitability for predicting the effect Lien and Cheng[18] interpreted this as evidence for their maximal-contrasthypothesis In another word people tend to induce categoriesthat maximize their causal predictability
In sum our research addresses the question of whichroute learners will go Will they routinely go through theupper road and activate category knowledge when learningabout novel effects A simple connectionist one-layer net-work that may be used to understand the task we are going toexplore in our experiments or will they go the lower road andlearn a new set of categories on the basis of causal informationin causal learning phase that is maximally predictive withinthis learning phase Thus Figure 1 is enhanced with twolearning phases as shown in Figure 2
3 Research Methodology andSystem Structure
31 Research Methodology Causality is a fundamental con-cept in human thinking and reasoning It is not surprisingthat most if not all languages in the world have a range of
Table 1 Statistics for backward causality connectives
Connectives For objectivereason
For subjectivereason
Total percentage inall connectives
Because 83 17 50For 29 71 18As 60 40 13Since 14 86 6While 14 86 6Misc 0 100 7Total 56 44 100
lexical expressions specifically designed for communicatingcausal relations
This paper focuses on explicit causality markers andimplicit causal relations in text The explicit causality mark-ers include two grammatically different types of causalitymarkers in English We investigate the semantic contrastsexpressed by different causal auxiliary verbs marking causalrelations expressed within one clause and those expressedby different causal connectives marking causal relationsbetween clauses
311 Causality in Verbs Some instances of causality in verbsare listed below These verbs include but are not limitedto ldquomakerdquo ldquoletrdquo ldquohaverdquo ldquocauserdquo and their synonyms fromWordNet [19] which is generally referred to as an onlinelexical database
[The extreme cold] cause madecaused even [the rivers(to) freeze] effect[She] cause madehad [her son empty his plate] effectdespite his complaints
312 Causality in Discourse Connectives For a complexsentence the predicative or relative clauses of the nounsynonyms of ldquocauserdquo are labeled as ldquopotential causerdquo ofantecedent sentence or main clause Additionally each clauseinducted with ldquosincerdquo ldquoasrdquo or ldquobecauserdquo is also labeled asldquopotential causerdquo of its main clause which is labeled asldquopotential effectrdquo
Table 1 lists the statistics for above backward causalitycollected upon Reuters-21578 currently the most widelyused test collection for text processing research and BBC-News2000 collected by a self-developed Web Crawler fromhttpwwwbbccouk The forward causality where in pre-sentation order the cause precedes the effect is the mostfrequently used ones for example ldquothereforerdquo ldquobecause ofthatrdquo ldquothat is whyrdquo and ldquosordquo As the forward causality con-nectives are one hundred percent strong causality indicatorsthe backward causality connectives in Table 1 need morespecific notation For example in all ldquobecauserdquo connectivesin collections 83 of them indicate objective reasons whilethe other 17 indicate subjective reasons Meanwhile theldquobecauserdquo connectives persist 50 of all the connectivesincluding ldquobecauserdquo ldquoforrdquo ldquoasrdquo ldquosincerdquo ldquowhilerdquo and othermiscellaneous connectivesThe ldquoforrdquo ldquoasrdquo ldquosincerdquo ldquowhilerdquo and
4 The Scientific World Journal
Table 2 Types of implicit causality sentences
Types Subtypes Exemplar sentences
Compoundsentences
Cause-effect sentencesconnected with ldquoandrdquo
Cause-effect(1)This was the first time I made an international call and my heart was beating fast(2) The filter is much more efficient than the primary filter and it removes allremaining solid particles from the fuel
Effect-cause(3) The crops had failed and there had not been any rain for months(4) Aluminum is used as engineering material for planes and spaceships and it islight and tough
Cause-effect sentenceswithout connectives
Effect-cause(5) My heart sank Some of them did not weigh more than 135 pounds and theothers looked too young to be in trouble(6) but the Voyagerrsquos captain and three crewmen had stayed in board They hadhoped the storm would die out and they could save the shipCause-effect(7) The red distress flare shot up from the sinking ship the Voyager Everymanaboard our Coast Guard cutter knew that time had run out
Relative clauses SV SVO SVC SVOC SVOOSVA SVOA
(8) To make an atom we have to use uranium in which the atoms are available forfission(9) We know that a cat whose eyes can take in more rays of light than our eyes cansee clearly in the night
If clauses
(10) This system of subsidies must be maintained if the farmer will sufferconsiderable losses if it is abolished(11) If the water will rise above this level we must warn everybody in theneighborhood
That clauses (12) The act was even the bolder that he stood utterly alone(13) The crying was all the more racking that she was not hysterical but hopeless
SVO-SVOC(14) Her falling ill spoiled everything(15) Timidity and shamefacedness caused her to stand back and looked indifferentlyaway
other miscellaneous connectives correspondingly hold 1813 6 6 and 7 of all the connectives in collections
313 Implicit Causality in Texts Causal effect relations aregeneral connections in the objective world and the causalitysentences exist in languages in a pervasive manner Theexplicit causality sentences are those conducted by backwardand forward causality connectives listed in Section 312while the implicit causality sentences are those connectedwith other connectives even without any one Our researchworks have testified that there are five types of implicitcausality sentences shown in Table 2
In the practical English texts there exist a few inter-personal verbs like ldquopraiserdquo and ldquoapologizerdquo thus supplyinginformation about whose behavior or state is the more likelyimmediate cause of the event at hand Because it is conveyedimplicitly as part of themeaning of the verb this probabilisticcue is usually referred to as implicit causality
Such exemplar implicit causality verbs [20] adopted inthis paper are listed in Table 3 together with their biasindicating the probabilities as causal cues for example 100is the causal baseline the higher the bias value is the morelikely is the cause
32 System Description Our causality induction system withassistance of category learning named CL-CIS is composedwith the following functional modules (shown in Figure 3)category learning classify exemplars into categories categoryand causal mapping causal learning building causal-effectrelations and the final testing module CL-CIS also includesthree reference libraries (databases) causality in verbscausality in discourse connectives and implicit causalityfor query and assistance Table 4 compares the compositiondifference among general causality analysis system (GCAS)general causality analysis system with learning (GCAS-L)and CL-CIS As the construction of three libraries is elabo-rated in Section 31 Methodology this section concentrateson the six functional modules
321 Category Learning (CL) The category learning modulebuilds up different distinct and exhaustive categories accord-ing to semantic contents (eg noun phrases verb phrases)of sentences in our text collections This module is a basicpreprocessing step and the target is to construct a collectionof distinct and exhaustive categories with the text learningmethods
The Scientific World Journal 5
Table 3 Exemplar implicit causality verbs
NP1 verbs Bias NP2 verbs BiasAmazed 119 Admire 200Annoy 119 Adore 186Apologize 100 Appreciate 200Be in the way 108 Comfort 191Beg 117 Compliment on something 196Bore 105 Congratulate 195Call 119 Criticize 200Confess 103 Envy 200Disappoint 103 Fear 200Disturb 114 Fire 196Fascinate 100 Hate 196Hurt 113 Hold in contempt 200Inspire 123 Hold responsible 191Intimidate 123 Loathe 186Irritate 122 Love 186Lie to 122 Praise 196Mislead 122 Press charges against 191Swindle 113 Punish 195Win 119 Respect 195Worry 119 Thank 182
Table 4 System composition comparison of GCAS GCAS-L andCL-CIS
System composition GCAS GCAS-L CL-CISModules
Category learning radic
Classify exemplars intocategories radic
Category and causal mapping radic
Causal learning radic radic
Building causal-effectrelations radic radic radic
Testing causal-effectrelations radic radic radic
LibrariesCausality in verbs radic radic radic
Causality in discourseconnectives Optional radic radic
Implicit causality Optional radic
322 Classify Exemplars into Categories (CEC) In the clas-sify exemplars into categories module each sentence istreated as an individual independent event and parsed intophrases The classification is based on the comparison ofa sentence with features of each category so as to set upthe category background knowledge for each sentence Forexample the concept bank can be clarified as a financeorganization or a river body boundary with correspondingcategory background knowledge of its sentence and contexts
Category learning
Classify exemplars into
categories
Category and causal
mapping
Causal learning
Building causal-effect
relations
Testing causal-effect
relations
Causality in verbs
Causality in discourse
connectivesImplicit causality
Figure 3 System framework of CL-CIS
Inputlayer (IL)
Recurrentlayer (RL)
Outputlayer (OL)
Contextlayer (CL)
Time delay
OU1 OU2
RU1RU2
WOR
IU1 IU2
WRI W
RC
CU1 CU2
middot middot middot
middot middot middot
middot middot middot middot middot middot
OU|O|
RU|R|
IU|I| CU|C|
Figure 4 Architecture of SRNN
323 Category and Causal Mapping (CCM) This moduleconstructs the category-causal-effect mapping relations forexample a virus from a predefined category causes specificdisease-related symptoms such as a swelling of the spleen(splenomegaly) The construction mechanism to a greatextent is based on collection storage and indexing ofmassive cases
324 Causal Learning (CauL) As analyzed before a sim-ple connectionist one-layer network that may be used tounderstand the task of when to activate category knowledgewhen learning about novel effects This module adopts asimple recurrent neural network (SRNN) to simulate theconnectionist model
Symbols in Figure 4 are defined in Table 5 the first orderweight matrices WRI and WOR fully connect the units ofthe input layer (IL) the recurrent layer (RL) and the outputlayer (OL) respectively as in the feed forward multilayerperceptron (MLP) The current activities of recurrent unitsRU(119905) are fed back through time delay connections to thecontext layer which is presented as CU(119905+1) = RU(119905)
Therefore each unit in recurrent layer is fed by activitiesof all recurrent units from previous time step through
6 The Scientific World Journal
Table 5 Definition of SRNN Symbols
Symbols DefinitionIU A unit of input layerRU A unit of recurrent layerCU A unit of context layerOU A unit of output layer|119868| The number of units in IL|119877| The number of units in RL|119862| The number of units in CL|119874| The number of units in OLWRI The weight vector from IL to RLWRC The weight vector from CL to RLWOR The weight vector from RL to OL
recurrent weight matrix WRC The context layer which iscomposed of activities of recurrent units from previous timestep can be viewed as an extension of input layer to therecurrent layer Above working procedure represents thememory of the network via holding contextual informationfrom previous time steps
The weight matrices 119882RI 119882RC and 119882OR are presentedas follows
119882RI= [(119908
RI1 )119879 (119908
RI2 )119879 (119908
RI|119877|)119879]
=
[[[[[
[
119908ri11 119908
ri12 sdot sdot sdot 119908
ri1|119877|
119908ri21 119908ri22 sdot sdot sdot 119908
ri1|119877|
d
119908ri|119868|1 119908
ri|119868|2 sdot sdot sdot 119908
ri|119868||119877|
]]]]]
]
119882RC= [(119908
RC1 )119879 (119908
RC2 )119879 (119908
RC|119877| )119879]
=
[[[[
[
119908rc11 119908rc12 sdot sdot sdot 119908
rc1|119877|
119908rc21 119908rc22 sdot sdot sdot 119908
rc1|119877|
d
119908rc|119862|1 119908
rc|119862|2 sdot sdot sdot 119908
rc|119862||119877|
]]]]
]
119882OR= [(119908
OR1 )119879 (119908
OR2 )119879 (119908
OR|119874| )119879]
=
[[[[
[
119908or11 119908or12 sdot sdot sdot 119908
or1|119874|
119908or21 119908or22 sdot sdot sdot 119908
or1|119874|
d
119908or|119877|1 119908
or|119877|2 sdot sdot sdot 119908
or|119877||119874|
]]]]
]
(1)
In the above formulations (119908RI119896 )119879 is the transpose of 119908RI119896
for the instance of119882RI where 119908RI119896 is a row vector and (119908RI119896 )119879
is the column vector of the same elements The vector 119908RI119896 =(119908ri1119896 119908
ri2119896 119908
ri|119868|119896) represents the weights from all the input
layer units to the recurrent (hidden) layer unit RU119896The sameconclusion applies with119882RC and119882OR
Given an input pattern in time t IU(t) = (IU(119905)1 IU(119905)2
IU(119905)|119868|) and recurrent activities RU(t) = (RU(119905)1 RU
(119905)2
RU(119905)|119877|) for the 119894th recurrent unit the net input RU1015840(119905)
119894and
output activity RU(119905)119894
are calculated as follows
RU1015840(119905)119894 = IU(t)sdot (119908
RI119894 )119879+ RU(tminus1) sdot (119908RC119894 )
119879
=
|119868|
sum119895=1
IU(119905)119895 119908ri119895119894 +
|119877|
sum119895=1
RU(119905minus1)119895 119908rc119895119894
RU(119905)119894 = 119891 (RU1015840(119905)
119894 )
(2)
For the kth output unit its net inputOU1015840(119905)119896
and output activityOU(119905)119896
are calculated as (3) Consider the following
OU1015840(119905)119896= RU(t) sdot (119908OR119896 )
119879=
|119877|
sum119895=1
RU(119905)119895 119908or119895119896
OU(119905)119896= 119891 (OU1015840(119905)
119896)
(3)
Here the activation function 119891 applies the logistic sigmoidfunction (4) in this paper Consider the following
119891 (119909) =1
1 + 119890minus119909=119890119909
1 + 119890119909 (4)
325 Building Causal-Effect Relations (BCER) This modulebuilds up the ldquocause-effectrdquo pairs and connections with theinputs of SRNN and corresponding outputs The causalitydetection results could include four possible types (1) singlecause-effect pairs in which any two pairs are independentfrom each other (2) cause-effect chains which are formedwithmore than one cause-effect pairs connected together (aneffect is a cause of another effect) (3) one-cause-multiple-effect pairs (4) multiple-cause-one-effect pairs All of thecausality connections are archived in a database Both CauLandBCERmodules exploit three libraries (databases) causal-ity in verbs causality in discourse connectives and implicitcausality for reference
326 Testing Causal-Effect Relations (TCER) This moduletests and verifies a new processed causal-effect relation withour existing and expanding collection ofmassive causal-effectrelations The concrete technology includes comparison oftriples ⟨CategoryCauseEffect⟩
4 Experiments and Results
Our experiments test general causality analysis system(GCAS) general causality analysis system with learning(GCAS-L) and CL-CIS together in order to examine andreveal the assertion that category information is a necessarysupplement for causality extraction and has impact on subse-quent learning-based cause-effect identification
In our experiments we have used two text collections (1)Reuters-21578 currently the most widely used test collectionfor text processing research (2) BBC-News2000 collected bya self-developed Web Crawler from httpwwwbbccouk
The Scientific World Journal 7
00
01
02
03
04
05
06
07
08
09
10
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
0681 0706 069807320719
0755 074207930776
0815 0809
0897
(a) Precision
00
01
02
03
04
05
06
07
08
09
10
0593 0581 0602
0706
0621 0609 0616
0734
0685 0661 0683
0805
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
(b) Recall
00
01
02
03
04
05
06
07
08
09
10
0634 0637 0646 0719
0666 0674 0673 0762
0728 0730 0741
0849
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
GCASGCAS-L
CL-CIS
(c) F-measure (balanced)
Figure 5 Evaluation results of GCAS GCAS-L and CL-CIS
BBC-News2000 collected 2000 news articles which havebeen pruned off unnecessary Web page elements such ashtml tags images URLs and so forth BBC-News2000 is notcategorized and is treated as a whole hybrid
30 judges are involved to manually search cause-effectpairs and construct standard causality references (SCR) SCR-Reuters for Reuters-21578 and SCR-BBC for BBC-News2000Due to limited human resources in current stage SCR-Reuters only covers 2000 documents newid ranges from 1 to2000 contained in files ldquoreut2-000sgmrdquo and ldquoreut2-001sgmrdquoThe rest of the files of Reuters-21578 are still involved in thetraining phase
The performances of GCAS GCAS-L and CL-CIS areevaluated with precision recall and F-measure [21] thetraditional measures that have been widely applied by mostinformation retrieval systems to analyze and evaluate theirperformance The F-measure is a harmonic combination ofthe precision and recall values used in information retrievalAs shown in Tables 6 7 and 8 the experimental results statethat (1) GCAS scores from 0681 to 0732 on precision andfrom 0581 to 0706 on recall (2) GCAS-L scores from 0719to 0793 on precision and from 0609 to 0734 on recall (3)CL-CIS scores from 0776 to 0897 on precision and from0661 to 0805 on recall Figure 5 explicitly states that (1)
Table 6 Experimental results of GCAS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0681 0593 0634reut2-001sgm (newid 1001ndash2000) 0706 0581 0637Both (reut2) (newid 1ndash2000) 0698 0602 0646BBC-News2000 0732 0706 0719
Table 7 Experimental results of GCAS-L
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0719 0621 0666reut2-001sgm (newid 1001ndash2000) 0755 0609 0674Both (reut2) (newid 1ndash2000) 0742 0616 0673BBC-News2000 0793 0734 0762
the general causality analysis system with learning (GCAS-L) performances better than the general causality analysissystem (GCAS) on all evaluation measures (2) the causalityanalysis system strengthened with causal and category learn-ing (CL-CIS) exceeds GCAS-L in the meantime
8 The Scientific World Journal
Table 8 Experimental results of CL-CIS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0776 0685 0728reut2-001sgm (newid 1001ndash2000) 0815 0661 0730Both (reut2) (newid 1ndash2000) 0809 0683 0741BBC-News2000 0897 0805 0849
5 Concluding Remarks
In recent years detection and clarification of cause-effectrelationships among texts events or objects has been elevatedto a prominent research topic of natural and social sciencesover the human knowledge development history
This paper demonstrates a novel comprehensive causalityextraction system (CL-CIS) integrated with the means ofcategory-learning CL-CIS considers cause-effect relations inboth explicit and implicit forms and especially practices therelation between category and causality in computation
CL-CIS is inspired with cognitive philosophy in categoryand causality In causality extraction and induction tasks CL-CIS implements a simple recurrent neural network (SRNN)to simulate human associative memory which has the abilityto associate different types of inputs when processing infor-mation
In elaborately designed experimental tasks CL-CIS hasbeen examined in full with two text collections Reuters-21578and BBC-News2000 The experimental results have testifiedthe capability and performance of CL-CIS in constructionof cause-effect relations and also confirmed the expectationthat the means of causal and category learning will improvethe precision and coverage of causality induction in a notablemanner
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work is financially supported by the National NaturalScience Foundation of China (Grant no 61003126)
References
[1] httpenwikipediaorgwikiCausality[2] J Pearl Causality Models Reasoning and Inference Cambridge
University Press New York NY USA 2000[3] D-S Chang and K-S Choi ldquoIncremental cue phrase learning
and bootstrapping method for causality extraction using cuephrase and word pair probabilitiesrdquo Information Processing andManagement vol 42 no 3 pp 662ndash678 2006
[4] S Kim R H Bracewell and K M Wallace ldquoA frameworkfor automatic causality extraction using semantic similarityrdquoin Proceedings of the ASME International Design EngineeringTechnical Conferences and Computers and Information in Engi-neering Conference (IDETCCIE rsquo07) pp 831ndash840 Las VegasNev USA September 2007
[5] T Inui K Inui andYMatsumoto ldquoAcquiring causal knowledgefrom text using the connective markersrdquo Transactions of theInformation Processing Society of Japan vol 45 no 3 pp 919ndash932 2004
[6] D Marcu and A Echihabi ldquoAn unsupervised approach torecognizing discourse relationsrdquo in Proceedings of the 40thAnnual Meeting of the Association for Computational LinguisticsConference (ACL rsquo02) pp 368ndash375 University of PennsylvaniaPhiladelphia Pa USA 2002
[7] C Pechsiri and A Kawtraku ldquoMining causality from texts forquestion answering systemrdquo IEICE Transactions on Informationand Systems vol 90 no 10 pp 1523ndash1533 2007
[8] R Girju ldquoAutomatic detection of causal relations for questionansweringrdquo in Proceedings of the 41st Annual Meeting of theAssociation for Computational Linguistics 2003
[9] K Torisawa ldquoAutomatic extraction of commonsense inferencerules from corporardquo in Proceedings of the 9th Annual Meeting ofthe Association for Natural Language Processing 2003
[10] Y Guo N Hua and Z Shao ldquoCognitive causality detectionwith associative memory in textual eventsrdquo in Proceedings ofthe IEEE International Symposium on Information Engineeringand Electronic Commerce (IEEC rsquo09) pp 140ndash144 TernopilUkraine May 2009
[11] P Thwaites J Q Smith and E Riccomagno ldquoCausal analysiswith Chain Event GraphsrdquoArtificial Intelligence vol 174 no 12-13 pp 889ndash909 2010
[12] M R Waldmann and Y Hagmayer ldquoCategories and causalitythe neglected directionrdquo Cognitive Psychology vol 53 no 1 pp27ndash58 2006
[13] G L MurphyTheBig Book of Concepts MIT Press CambridgeMass USA 2002
[14] B Rehder ldquoA causal-model theory of conceptual representationand categorizationrdquo Journal of Experimental Psychology Learn-ing Memory and Cognition vol 29 no 6 pp 1141ndash1159 2003
[15] B Rehder ldquoCategorization as casual reasoningrdquo Cognitive Sci-ence vol 27 no 5 pp 709ndash748 2003
[16] B Rehder and R Hastie ldquoCausal knowledge and categoriesthe effects of causal beliefs on categorization induction andsimilarityrdquo Journal of Experimental Psychology General vol 130no 3 pp 323ndash360 2001
[17] B Rehder and R Hastie ldquoCategory coherence and category-based property inductionrdquo Cognition vol 91 no 2 pp 113ndash1532004
[18] Y Lien and PWCheng ldquoDistinguishing genuine from spuriouscauses a coherence hypothesisrdquo Cognitive Psychology vol 40pp 87ndash137 2000
[19] G A Miller R Beckwith C Fellbaum D Gross and K JMiller ldquoIntroduction to WordNet an on-line lexical databaserdquoInternational Journal of Lexicography vol 3 no 4 pp 235ndash3121990
[20] A W Koornneef and J J A Van Berkum ldquoOn the use of verb-based implicit causality in sentence comprehension evidencefrom self-paced reading and eye trackingrdquo Journal of Memoryand Language vol 54 no 4 pp 445ndash465 2006
[21] C D Manning and H Schutze Foundations of StatisticalNatural Language Processing TheMIT Press Cambridge MassUSA 4th edition 2001
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World Journal 3
Categories
EffectsCauses
Figure 1 Possible routes of category learning between causes andeffects
Categories
EffectsCausesCauses
Category learning
learning
Figure 2 Enhanced strategy of category learning between causesand effects
whether they would stick to the lower route and induce newcategories within causal learning phase
If the learning strategy opted for the lower route onepossible solution may be to induce new categories that aremaximally predictive of the effects This solution wouldobviously generate maximally predictive categories Lien andCheng [18] reported a research that is consistent with themaximal-contrast hypothesis Their experiments show thatthe substances are categorized according to the feature and tothe hierarchical level that were maximally predictive for theeffect Thus the induced substance category was determinedby its suitability for predicting the effect Lien and Cheng[18] interpreted this as evidence for their maximal-contrasthypothesis In another word people tend to induce categoriesthat maximize their causal predictability
In sum our research addresses the question of whichroute learners will go Will they routinely go through theupper road and activate category knowledge when learningabout novel effects A simple connectionist one-layer net-work that may be used to understand the task we are going toexplore in our experiments or will they go the lower road andlearn a new set of categories on the basis of causal informationin causal learning phase that is maximally predictive withinthis learning phase Thus Figure 1 is enhanced with twolearning phases as shown in Figure 2
3 Research Methodology andSystem Structure
31 Research Methodology Causality is a fundamental con-cept in human thinking and reasoning It is not surprisingthat most if not all languages in the world have a range of
Table 1 Statistics for backward causality connectives
Connectives For objectivereason
For subjectivereason
Total percentage inall connectives
Because 83 17 50For 29 71 18As 60 40 13Since 14 86 6While 14 86 6Misc 0 100 7Total 56 44 100
lexical expressions specifically designed for communicatingcausal relations
This paper focuses on explicit causality markers andimplicit causal relations in text The explicit causality mark-ers include two grammatically different types of causalitymarkers in English We investigate the semantic contrastsexpressed by different causal auxiliary verbs marking causalrelations expressed within one clause and those expressedby different causal connectives marking causal relationsbetween clauses
311 Causality in Verbs Some instances of causality in verbsare listed below These verbs include but are not limitedto ldquomakerdquo ldquoletrdquo ldquohaverdquo ldquocauserdquo and their synonyms fromWordNet [19] which is generally referred to as an onlinelexical database
[The extreme cold] cause madecaused even [the rivers(to) freeze] effect[She] cause madehad [her son empty his plate] effectdespite his complaints
312 Causality in Discourse Connectives For a complexsentence the predicative or relative clauses of the nounsynonyms of ldquocauserdquo are labeled as ldquopotential causerdquo ofantecedent sentence or main clause Additionally each clauseinducted with ldquosincerdquo ldquoasrdquo or ldquobecauserdquo is also labeled asldquopotential causerdquo of its main clause which is labeled asldquopotential effectrdquo
Table 1 lists the statistics for above backward causalitycollected upon Reuters-21578 currently the most widelyused test collection for text processing research and BBC-News2000 collected by a self-developed Web Crawler fromhttpwwwbbccouk The forward causality where in pre-sentation order the cause precedes the effect is the mostfrequently used ones for example ldquothereforerdquo ldquobecause ofthatrdquo ldquothat is whyrdquo and ldquosordquo As the forward causality con-nectives are one hundred percent strong causality indicatorsthe backward causality connectives in Table 1 need morespecific notation For example in all ldquobecauserdquo connectivesin collections 83 of them indicate objective reasons whilethe other 17 indicate subjective reasons Meanwhile theldquobecauserdquo connectives persist 50 of all the connectivesincluding ldquobecauserdquo ldquoforrdquo ldquoasrdquo ldquosincerdquo ldquowhilerdquo and othermiscellaneous connectivesThe ldquoforrdquo ldquoasrdquo ldquosincerdquo ldquowhilerdquo and
4 The Scientific World Journal
Table 2 Types of implicit causality sentences
Types Subtypes Exemplar sentences
Compoundsentences
Cause-effect sentencesconnected with ldquoandrdquo
Cause-effect(1)This was the first time I made an international call and my heart was beating fast(2) The filter is much more efficient than the primary filter and it removes allremaining solid particles from the fuel
Effect-cause(3) The crops had failed and there had not been any rain for months(4) Aluminum is used as engineering material for planes and spaceships and it islight and tough
Cause-effect sentenceswithout connectives
Effect-cause(5) My heart sank Some of them did not weigh more than 135 pounds and theothers looked too young to be in trouble(6) but the Voyagerrsquos captain and three crewmen had stayed in board They hadhoped the storm would die out and they could save the shipCause-effect(7) The red distress flare shot up from the sinking ship the Voyager Everymanaboard our Coast Guard cutter knew that time had run out
Relative clauses SV SVO SVC SVOC SVOOSVA SVOA
(8) To make an atom we have to use uranium in which the atoms are available forfission(9) We know that a cat whose eyes can take in more rays of light than our eyes cansee clearly in the night
If clauses
(10) This system of subsidies must be maintained if the farmer will sufferconsiderable losses if it is abolished(11) If the water will rise above this level we must warn everybody in theneighborhood
That clauses (12) The act was even the bolder that he stood utterly alone(13) The crying was all the more racking that she was not hysterical but hopeless
SVO-SVOC(14) Her falling ill spoiled everything(15) Timidity and shamefacedness caused her to stand back and looked indifferentlyaway
other miscellaneous connectives correspondingly hold 1813 6 6 and 7 of all the connectives in collections
313 Implicit Causality in Texts Causal effect relations aregeneral connections in the objective world and the causalitysentences exist in languages in a pervasive manner Theexplicit causality sentences are those conducted by backwardand forward causality connectives listed in Section 312while the implicit causality sentences are those connectedwith other connectives even without any one Our researchworks have testified that there are five types of implicitcausality sentences shown in Table 2
In the practical English texts there exist a few inter-personal verbs like ldquopraiserdquo and ldquoapologizerdquo thus supplyinginformation about whose behavior or state is the more likelyimmediate cause of the event at hand Because it is conveyedimplicitly as part of themeaning of the verb this probabilisticcue is usually referred to as implicit causality
Such exemplar implicit causality verbs [20] adopted inthis paper are listed in Table 3 together with their biasindicating the probabilities as causal cues for example 100is the causal baseline the higher the bias value is the morelikely is the cause
32 System Description Our causality induction system withassistance of category learning named CL-CIS is composedwith the following functional modules (shown in Figure 3)category learning classify exemplars into categories categoryand causal mapping causal learning building causal-effectrelations and the final testing module CL-CIS also includesthree reference libraries (databases) causality in verbscausality in discourse connectives and implicit causalityfor query and assistance Table 4 compares the compositiondifference among general causality analysis system (GCAS)general causality analysis system with learning (GCAS-L)and CL-CIS As the construction of three libraries is elabo-rated in Section 31 Methodology this section concentrateson the six functional modules
321 Category Learning (CL) The category learning modulebuilds up different distinct and exhaustive categories accord-ing to semantic contents (eg noun phrases verb phrases)of sentences in our text collections This module is a basicpreprocessing step and the target is to construct a collectionof distinct and exhaustive categories with the text learningmethods
The Scientific World Journal 5
Table 3 Exemplar implicit causality verbs
NP1 verbs Bias NP2 verbs BiasAmazed 119 Admire 200Annoy 119 Adore 186Apologize 100 Appreciate 200Be in the way 108 Comfort 191Beg 117 Compliment on something 196Bore 105 Congratulate 195Call 119 Criticize 200Confess 103 Envy 200Disappoint 103 Fear 200Disturb 114 Fire 196Fascinate 100 Hate 196Hurt 113 Hold in contempt 200Inspire 123 Hold responsible 191Intimidate 123 Loathe 186Irritate 122 Love 186Lie to 122 Praise 196Mislead 122 Press charges against 191Swindle 113 Punish 195Win 119 Respect 195Worry 119 Thank 182
Table 4 System composition comparison of GCAS GCAS-L andCL-CIS
System composition GCAS GCAS-L CL-CISModules
Category learning radic
Classify exemplars intocategories radic
Category and causal mapping radic
Causal learning radic radic
Building causal-effectrelations radic radic radic
Testing causal-effectrelations radic radic radic
LibrariesCausality in verbs radic radic radic
Causality in discourseconnectives Optional radic radic
Implicit causality Optional radic
322 Classify Exemplars into Categories (CEC) In the clas-sify exemplars into categories module each sentence istreated as an individual independent event and parsed intophrases The classification is based on the comparison ofa sentence with features of each category so as to set upthe category background knowledge for each sentence Forexample the concept bank can be clarified as a financeorganization or a river body boundary with correspondingcategory background knowledge of its sentence and contexts
Category learning
Classify exemplars into
categories
Category and causal
mapping
Causal learning
Building causal-effect
relations
Testing causal-effect
relations
Causality in verbs
Causality in discourse
connectivesImplicit causality
Figure 3 System framework of CL-CIS
Inputlayer (IL)
Recurrentlayer (RL)
Outputlayer (OL)
Contextlayer (CL)
Time delay
OU1 OU2
RU1RU2
WOR
IU1 IU2
WRI W
RC
CU1 CU2
middot middot middot
middot middot middot
middot middot middot middot middot middot
OU|O|
RU|R|
IU|I| CU|C|
Figure 4 Architecture of SRNN
323 Category and Causal Mapping (CCM) This moduleconstructs the category-causal-effect mapping relations forexample a virus from a predefined category causes specificdisease-related symptoms such as a swelling of the spleen(splenomegaly) The construction mechanism to a greatextent is based on collection storage and indexing ofmassive cases
324 Causal Learning (CauL) As analyzed before a sim-ple connectionist one-layer network that may be used tounderstand the task of when to activate category knowledgewhen learning about novel effects This module adopts asimple recurrent neural network (SRNN) to simulate theconnectionist model
Symbols in Figure 4 are defined in Table 5 the first orderweight matrices WRI and WOR fully connect the units ofthe input layer (IL) the recurrent layer (RL) and the outputlayer (OL) respectively as in the feed forward multilayerperceptron (MLP) The current activities of recurrent unitsRU(119905) are fed back through time delay connections to thecontext layer which is presented as CU(119905+1) = RU(119905)
Therefore each unit in recurrent layer is fed by activitiesof all recurrent units from previous time step through
6 The Scientific World Journal
Table 5 Definition of SRNN Symbols
Symbols DefinitionIU A unit of input layerRU A unit of recurrent layerCU A unit of context layerOU A unit of output layer|119868| The number of units in IL|119877| The number of units in RL|119862| The number of units in CL|119874| The number of units in OLWRI The weight vector from IL to RLWRC The weight vector from CL to RLWOR The weight vector from RL to OL
recurrent weight matrix WRC The context layer which iscomposed of activities of recurrent units from previous timestep can be viewed as an extension of input layer to therecurrent layer Above working procedure represents thememory of the network via holding contextual informationfrom previous time steps
The weight matrices 119882RI 119882RC and 119882OR are presentedas follows
119882RI= [(119908
RI1 )119879 (119908
RI2 )119879 (119908
RI|119877|)119879]
=
[[[[[
[
119908ri11 119908
ri12 sdot sdot sdot 119908
ri1|119877|
119908ri21 119908ri22 sdot sdot sdot 119908
ri1|119877|
d
119908ri|119868|1 119908
ri|119868|2 sdot sdot sdot 119908
ri|119868||119877|
]]]]]
]
119882RC= [(119908
RC1 )119879 (119908
RC2 )119879 (119908
RC|119877| )119879]
=
[[[[
[
119908rc11 119908rc12 sdot sdot sdot 119908
rc1|119877|
119908rc21 119908rc22 sdot sdot sdot 119908
rc1|119877|
d
119908rc|119862|1 119908
rc|119862|2 sdot sdot sdot 119908
rc|119862||119877|
]]]]
]
119882OR= [(119908
OR1 )119879 (119908
OR2 )119879 (119908
OR|119874| )119879]
=
[[[[
[
119908or11 119908or12 sdot sdot sdot 119908
or1|119874|
119908or21 119908or22 sdot sdot sdot 119908
or1|119874|
d
119908or|119877|1 119908
or|119877|2 sdot sdot sdot 119908
or|119877||119874|
]]]]
]
(1)
In the above formulations (119908RI119896 )119879 is the transpose of 119908RI119896
for the instance of119882RI where 119908RI119896 is a row vector and (119908RI119896 )119879
is the column vector of the same elements The vector 119908RI119896 =(119908ri1119896 119908
ri2119896 119908
ri|119868|119896) represents the weights from all the input
layer units to the recurrent (hidden) layer unit RU119896The sameconclusion applies with119882RC and119882OR
Given an input pattern in time t IU(t) = (IU(119905)1 IU(119905)2
IU(119905)|119868|) and recurrent activities RU(t) = (RU(119905)1 RU
(119905)2
RU(119905)|119877|) for the 119894th recurrent unit the net input RU1015840(119905)
119894and
output activity RU(119905)119894
are calculated as follows
RU1015840(119905)119894 = IU(t)sdot (119908
RI119894 )119879+ RU(tminus1) sdot (119908RC119894 )
119879
=
|119868|
sum119895=1
IU(119905)119895 119908ri119895119894 +
|119877|
sum119895=1
RU(119905minus1)119895 119908rc119895119894
RU(119905)119894 = 119891 (RU1015840(119905)
119894 )
(2)
For the kth output unit its net inputOU1015840(119905)119896
and output activityOU(119905)119896
are calculated as (3) Consider the following
OU1015840(119905)119896= RU(t) sdot (119908OR119896 )
119879=
|119877|
sum119895=1
RU(119905)119895 119908or119895119896
OU(119905)119896= 119891 (OU1015840(119905)
119896)
(3)
Here the activation function 119891 applies the logistic sigmoidfunction (4) in this paper Consider the following
119891 (119909) =1
1 + 119890minus119909=119890119909
1 + 119890119909 (4)
325 Building Causal-Effect Relations (BCER) This modulebuilds up the ldquocause-effectrdquo pairs and connections with theinputs of SRNN and corresponding outputs The causalitydetection results could include four possible types (1) singlecause-effect pairs in which any two pairs are independentfrom each other (2) cause-effect chains which are formedwithmore than one cause-effect pairs connected together (aneffect is a cause of another effect) (3) one-cause-multiple-effect pairs (4) multiple-cause-one-effect pairs All of thecausality connections are archived in a database Both CauLandBCERmodules exploit three libraries (databases) causal-ity in verbs causality in discourse connectives and implicitcausality for reference
326 Testing Causal-Effect Relations (TCER) This moduletests and verifies a new processed causal-effect relation withour existing and expanding collection ofmassive causal-effectrelations The concrete technology includes comparison oftriples ⟨CategoryCauseEffect⟩
4 Experiments and Results
Our experiments test general causality analysis system(GCAS) general causality analysis system with learning(GCAS-L) and CL-CIS together in order to examine andreveal the assertion that category information is a necessarysupplement for causality extraction and has impact on subse-quent learning-based cause-effect identification
In our experiments we have used two text collections (1)Reuters-21578 currently the most widely used test collectionfor text processing research (2) BBC-News2000 collected bya self-developed Web Crawler from httpwwwbbccouk
The Scientific World Journal 7
00
01
02
03
04
05
06
07
08
09
10
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
0681 0706 069807320719
0755 074207930776
0815 0809
0897
(a) Precision
00
01
02
03
04
05
06
07
08
09
10
0593 0581 0602
0706
0621 0609 0616
0734
0685 0661 0683
0805
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
(b) Recall
00
01
02
03
04
05
06
07
08
09
10
0634 0637 0646 0719
0666 0674 0673 0762
0728 0730 0741
0849
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
GCASGCAS-L
CL-CIS
(c) F-measure (balanced)
Figure 5 Evaluation results of GCAS GCAS-L and CL-CIS
BBC-News2000 collected 2000 news articles which havebeen pruned off unnecessary Web page elements such ashtml tags images URLs and so forth BBC-News2000 is notcategorized and is treated as a whole hybrid
30 judges are involved to manually search cause-effectpairs and construct standard causality references (SCR) SCR-Reuters for Reuters-21578 and SCR-BBC for BBC-News2000Due to limited human resources in current stage SCR-Reuters only covers 2000 documents newid ranges from 1 to2000 contained in files ldquoreut2-000sgmrdquo and ldquoreut2-001sgmrdquoThe rest of the files of Reuters-21578 are still involved in thetraining phase
The performances of GCAS GCAS-L and CL-CIS areevaluated with precision recall and F-measure [21] thetraditional measures that have been widely applied by mostinformation retrieval systems to analyze and evaluate theirperformance The F-measure is a harmonic combination ofthe precision and recall values used in information retrievalAs shown in Tables 6 7 and 8 the experimental results statethat (1) GCAS scores from 0681 to 0732 on precision andfrom 0581 to 0706 on recall (2) GCAS-L scores from 0719to 0793 on precision and from 0609 to 0734 on recall (3)CL-CIS scores from 0776 to 0897 on precision and from0661 to 0805 on recall Figure 5 explicitly states that (1)
Table 6 Experimental results of GCAS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0681 0593 0634reut2-001sgm (newid 1001ndash2000) 0706 0581 0637Both (reut2) (newid 1ndash2000) 0698 0602 0646BBC-News2000 0732 0706 0719
Table 7 Experimental results of GCAS-L
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0719 0621 0666reut2-001sgm (newid 1001ndash2000) 0755 0609 0674Both (reut2) (newid 1ndash2000) 0742 0616 0673BBC-News2000 0793 0734 0762
the general causality analysis system with learning (GCAS-L) performances better than the general causality analysissystem (GCAS) on all evaluation measures (2) the causalityanalysis system strengthened with causal and category learn-ing (CL-CIS) exceeds GCAS-L in the meantime
8 The Scientific World Journal
Table 8 Experimental results of CL-CIS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0776 0685 0728reut2-001sgm (newid 1001ndash2000) 0815 0661 0730Both (reut2) (newid 1ndash2000) 0809 0683 0741BBC-News2000 0897 0805 0849
5 Concluding Remarks
In recent years detection and clarification of cause-effectrelationships among texts events or objects has been elevatedto a prominent research topic of natural and social sciencesover the human knowledge development history
This paper demonstrates a novel comprehensive causalityextraction system (CL-CIS) integrated with the means ofcategory-learning CL-CIS considers cause-effect relations inboth explicit and implicit forms and especially practices therelation between category and causality in computation
CL-CIS is inspired with cognitive philosophy in categoryand causality In causality extraction and induction tasks CL-CIS implements a simple recurrent neural network (SRNN)to simulate human associative memory which has the abilityto associate different types of inputs when processing infor-mation
In elaborately designed experimental tasks CL-CIS hasbeen examined in full with two text collections Reuters-21578and BBC-News2000 The experimental results have testifiedthe capability and performance of CL-CIS in constructionof cause-effect relations and also confirmed the expectationthat the means of causal and category learning will improvethe precision and coverage of causality induction in a notablemanner
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work is financially supported by the National NaturalScience Foundation of China (Grant no 61003126)
References
[1] httpenwikipediaorgwikiCausality[2] J Pearl Causality Models Reasoning and Inference Cambridge
University Press New York NY USA 2000[3] D-S Chang and K-S Choi ldquoIncremental cue phrase learning
and bootstrapping method for causality extraction using cuephrase and word pair probabilitiesrdquo Information Processing andManagement vol 42 no 3 pp 662ndash678 2006
[4] S Kim R H Bracewell and K M Wallace ldquoA frameworkfor automatic causality extraction using semantic similarityrdquoin Proceedings of the ASME International Design EngineeringTechnical Conferences and Computers and Information in Engi-neering Conference (IDETCCIE rsquo07) pp 831ndash840 Las VegasNev USA September 2007
[5] T Inui K Inui andYMatsumoto ldquoAcquiring causal knowledgefrom text using the connective markersrdquo Transactions of theInformation Processing Society of Japan vol 45 no 3 pp 919ndash932 2004
[6] D Marcu and A Echihabi ldquoAn unsupervised approach torecognizing discourse relationsrdquo in Proceedings of the 40thAnnual Meeting of the Association for Computational LinguisticsConference (ACL rsquo02) pp 368ndash375 University of PennsylvaniaPhiladelphia Pa USA 2002
[7] C Pechsiri and A Kawtraku ldquoMining causality from texts forquestion answering systemrdquo IEICE Transactions on Informationand Systems vol 90 no 10 pp 1523ndash1533 2007
[8] R Girju ldquoAutomatic detection of causal relations for questionansweringrdquo in Proceedings of the 41st Annual Meeting of theAssociation for Computational Linguistics 2003
[9] K Torisawa ldquoAutomatic extraction of commonsense inferencerules from corporardquo in Proceedings of the 9th Annual Meeting ofthe Association for Natural Language Processing 2003
[10] Y Guo N Hua and Z Shao ldquoCognitive causality detectionwith associative memory in textual eventsrdquo in Proceedings ofthe IEEE International Symposium on Information Engineeringand Electronic Commerce (IEEC rsquo09) pp 140ndash144 TernopilUkraine May 2009
[11] P Thwaites J Q Smith and E Riccomagno ldquoCausal analysiswith Chain Event GraphsrdquoArtificial Intelligence vol 174 no 12-13 pp 889ndash909 2010
[12] M R Waldmann and Y Hagmayer ldquoCategories and causalitythe neglected directionrdquo Cognitive Psychology vol 53 no 1 pp27ndash58 2006
[13] G L MurphyTheBig Book of Concepts MIT Press CambridgeMass USA 2002
[14] B Rehder ldquoA causal-model theory of conceptual representationand categorizationrdquo Journal of Experimental Psychology Learn-ing Memory and Cognition vol 29 no 6 pp 1141ndash1159 2003
[15] B Rehder ldquoCategorization as casual reasoningrdquo Cognitive Sci-ence vol 27 no 5 pp 709ndash748 2003
[16] B Rehder and R Hastie ldquoCausal knowledge and categoriesthe effects of causal beliefs on categorization induction andsimilarityrdquo Journal of Experimental Psychology General vol 130no 3 pp 323ndash360 2001
[17] B Rehder and R Hastie ldquoCategory coherence and category-based property inductionrdquo Cognition vol 91 no 2 pp 113ndash1532004
[18] Y Lien and PWCheng ldquoDistinguishing genuine from spuriouscauses a coherence hypothesisrdquo Cognitive Psychology vol 40pp 87ndash137 2000
[19] G A Miller R Beckwith C Fellbaum D Gross and K JMiller ldquoIntroduction to WordNet an on-line lexical databaserdquoInternational Journal of Lexicography vol 3 no 4 pp 235ndash3121990
[20] A W Koornneef and J J A Van Berkum ldquoOn the use of verb-based implicit causality in sentence comprehension evidencefrom self-paced reading and eye trackingrdquo Journal of Memoryand Language vol 54 no 4 pp 445ndash465 2006
[21] C D Manning and H Schutze Foundations of StatisticalNatural Language Processing TheMIT Press Cambridge MassUSA 4th edition 2001
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
4 The Scientific World Journal
Table 2 Types of implicit causality sentences
Types Subtypes Exemplar sentences
Compoundsentences
Cause-effect sentencesconnected with ldquoandrdquo
Cause-effect(1)This was the first time I made an international call and my heart was beating fast(2) The filter is much more efficient than the primary filter and it removes allremaining solid particles from the fuel
Effect-cause(3) The crops had failed and there had not been any rain for months(4) Aluminum is used as engineering material for planes and spaceships and it islight and tough
Cause-effect sentenceswithout connectives
Effect-cause(5) My heart sank Some of them did not weigh more than 135 pounds and theothers looked too young to be in trouble(6) but the Voyagerrsquos captain and three crewmen had stayed in board They hadhoped the storm would die out and they could save the shipCause-effect(7) The red distress flare shot up from the sinking ship the Voyager Everymanaboard our Coast Guard cutter knew that time had run out
Relative clauses SV SVO SVC SVOC SVOOSVA SVOA
(8) To make an atom we have to use uranium in which the atoms are available forfission(9) We know that a cat whose eyes can take in more rays of light than our eyes cansee clearly in the night
If clauses
(10) This system of subsidies must be maintained if the farmer will sufferconsiderable losses if it is abolished(11) If the water will rise above this level we must warn everybody in theneighborhood
That clauses (12) The act was even the bolder that he stood utterly alone(13) The crying was all the more racking that she was not hysterical but hopeless
SVO-SVOC(14) Her falling ill spoiled everything(15) Timidity and shamefacedness caused her to stand back and looked indifferentlyaway
other miscellaneous connectives correspondingly hold 1813 6 6 and 7 of all the connectives in collections
313 Implicit Causality in Texts Causal effect relations aregeneral connections in the objective world and the causalitysentences exist in languages in a pervasive manner Theexplicit causality sentences are those conducted by backwardand forward causality connectives listed in Section 312while the implicit causality sentences are those connectedwith other connectives even without any one Our researchworks have testified that there are five types of implicitcausality sentences shown in Table 2
In the practical English texts there exist a few inter-personal verbs like ldquopraiserdquo and ldquoapologizerdquo thus supplyinginformation about whose behavior or state is the more likelyimmediate cause of the event at hand Because it is conveyedimplicitly as part of themeaning of the verb this probabilisticcue is usually referred to as implicit causality
Such exemplar implicit causality verbs [20] adopted inthis paper are listed in Table 3 together with their biasindicating the probabilities as causal cues for example 100is the causal baseline the higher the bias value is the morelikely is the cause
32 System Description Our causality induction system withassistance of category learning named CL-CIS is composedwith the following functional modules (shown in Figure 3)category learning classify exemplars into categories categoryand causal mapping causal learning building causal-effectrelations and the final testing module CL-CIS also includesthree reference libraries (databases) causality in verbscausality in discourse connectives and implicit causalityfor query and assistance Table 4 compares the compositiondifference among general causality analysis system (GCAS)general causality analysis system with learning (GCAS-L)and CL-CIS As the construction of three libraries is elabo-rated in Section 31 Methodology this section concentrateson the six functional modules
321 Category Learning (CL) The category learning modulebuilds up different distinct and exhaustive categories accord-ing to semantic contents (eg noun phrases verb phrases)of sentences in our text collections This module is a basicpreprocessing step and the target is to construct a collectionof distinct and exhaustive categories with the text learningmethods
The Scientific World Journal 5
Table 3 Exemplar implicit causality verbs
NP1 verbs Bias NP2 verbs BiasAmazed 119 Admire 200Annoy 119 Adore 186Apologize 100 Appreciate 200Be in the way 108 Comfort 191Beg 117 Compliment on something 196Bore 105 Congratulate 195Call 119 Criticize 200Confess 103 Envy 200Disappoint 103 Fear 200Disturb 114 Fire 196Fascinate 100 Hate 196Hurt 113 Hold in contempt 200Inspire 123 Hold responsible 191Intimidate 123 Loathe 186Irritate 122 Love 186Lie to 122 Praise 196Mislead 122 Press charges against 191Swindle 113 Punish 195Win 119 Respect 195Worry 119 Thank 182
Table 4 System composition comparison of GCAS GCAS-L andCL-CIS
System composition GCAS GCAS-L CL-CISModules
Category learning radic
Classify exemplars intocategories radic
Category and causal mapping radic
Causal learning radic radic
Building causal-effectrelations radic radic radic
Testing causal-effectrelations radic radic radic
LibrariesCausality in verbs radic radic radic
Causality in discourseconnectives Optional radic radic
Implicit causality Optional radic
322 Classify Exemplars into Categories (CEC) In the clas-sify exemplars into categories module each sentence istreated as an individual independent event and parsed intophrases The classification is based on the comparison ofa sentence with features of each category so as to set upthe category background knowledge for each sentence Forexample the concept bank can be clarified as a financeorganization or a river body boundary with correspondingcategory background knowledge of its sentence and contexts
Category learning
Classify exemplars into
categories
Category and causal
mapping
Causal learning
Building causal-effect
relations
Testing causal-effect
relations
Causality in verbs
Causality in discourse
connectivesImplicit causality
Figure 3 System framework of CL-CIS
Inputlayer (IL)
Recurrentlayer (RL)
Outputlayer (OL)
Contextlayer (CL)
Time delay
OU1 OU2
RU1RU2
WOR
IU1 IU2
WRI W
RC
CU1 CU2
middot middot middot
middot middot middot
middot middot middot middot middot middot
OU|O|
RU|R|
IU|I| CU|C|
Figure 4 Architecture of SRNN
323 Category and Causal Mapping (CCM) This moduleconstructs the category-causal-effect mapping relations forexample a virus from a predefined category causes specificdisease-related symptoms such as a swelling of the spleen(splenomegaly) The construction mechanism to a greatextent is based on collection storage and indexing ofmassive cases
324 Causal Learning (CauL) As analyzed before a sim-ple connectionist one-layer network that may be used tounderstand the task of when to activate category knowledgewhen learning about novel effects This module adopts asimple recurrent neural network (SRNN) to simulate theconnectionist model
Symbols in Figure 4 are defined in Table 5 the first orderweight matrices WRI and WOR fully connect the units ofthe input layer (IL) the recurrent layer (RL) and the outputlayer (OL) respectively as in the feed forward multilayerperceptron (MLP) The current activities of recurrent unitsRU(119905) are fed back through time delay connections to thecontext layer which is presented as CU(119905+1) = RU(119905)
Therefore each unit in recurrent layer is fed by activitiesof all recurrent units from previous time step through
6 The Scientific World Journal
Table 5 Definition of SRNN Symbols
Symbols DefinitionIU A unit of input layerRU A unit of recurrent layerCU A unit of context layerOU A unit of output layer|119868| The number of units in IL|119877| The number of units in RL|119862| The number of units in CL|119874| The number of units in OLWRI The weight vector from IL to RLWRC The weight vector from CL to RLWOR The weight vector from RL to OL
recurrent weight matrix WRC The context layer which iscomposed of activities of recurrent units from previous timestep can be viewed as an extension of input layer to therecurrent layer Above working procedure represents thememory of the network via holding contextual informationfrom previous time steps
The weight matrices 119882RI 119882RC and 119882OR are presentedas follows
119882RI= [(119908
RI1 )119879 (119908
RI2 )119879 (119908
RI|119877|)119879]
=
[[[[[
[
119908ri11 119908
ri12 sdot sdot sdot 119908
ri1|119877|
119908ri21 119908ri22 sdot sdot sdot 119908
ri1|119877|
d
119908ri|119868|1 119908
ri|119868|2 sdot sdot sdot 119908
ri|119868||119877|
]]]]]
]
119882RC= [(119908
RC1 )119879 (119908
RC2 )119879 (119908
RC|119877| )119879]
=
[[[[
[
119908rc11 119908rc12 sdot sdot sdot 119908
rc1|119877|
119908rc21 119908rc22 sdot sdot sdot 119908
rc1|119877|
d
119908rc|119862|1 119908
rc|119862|2 sdot sdot sdot 119908
rc|119862||119877|
]]]]
]
119882OR= [(119908
OR1 )119879 (119908
OR2 )119879 (119908
OR|119874| )119879]
=
[[[[
[
119908or11 119908or12 sdot sdot sdot 119908
or1|119874|
119908or21 119908or22 sdot sdot sdot 119908
or1|119874|
d
119908or|119877|1 119908
or|119877|2 sdot sdot sdot 119908
or|119877||119874|
]]]]
]
(1)
In the above formulations (119908RI119896 )119879 is the transpose of 119908RI119896
for the instance of119882RI where 119908RI119896 is a row vector and (119908RI119896 )119879
is the column vector of the same elements The vector 119908RI119896 =(119908ri1119896 119908
ri2119896 119908
ri|119868|119896) represents the weights from all the input
layer units to the recurrent (hidden) layer unit RU119896The sameconclusion applies with119882RC and119882OR
Given an input pattern in time t IU(t) = (IU(119905)1 IU(119905)2
IU(119905)|119868|) and recurrent activities RU(t) = (RU(119905)1 RU
(119905)2
RU(119905)|119877|) for the 119894th recurrent unit the net input RU1015840(119905)
119894and
output activity RU(119905)119894
are calculated as follows
RU1015840(119905)119894 = IU(t)sdot (119908
RI119894 )119879+ RU(tminus1) sdot (119908RC119894 )
119879
=
|119868|
sum119895=1
IU(119905)119895 119908ri119895119894 +
|119877|
sum119895=1
RU(119905minus1)119895 119908rc119895119894
RU(119905)119894 = 119891 (RU1015840(119905)
119894 )
(2)
For the kth output unit its net inputOU1015840(119905)119896
and output activityOU(119905)119896
are calculated as (3) Consider the following
OU1015840(119905)119896= RU(t) sdot (119908OR119896 )
119879=
|119877|
sum119895=1
RU(119905)119895 119908or119895119896
OU(119905)119896= 119891 (OU1015840(119905)
119896)
(3)
Here the activation function 119891 applies the logistic sigmoidfunction (4) in this paper Consider the following
119891 (119909) =1
1 + 119890minus119909=119890119909
1 + 119890119909 (4)
325 Building Causal-Effect Relations (BCER) This modulebuilds up the ldquocause-effectrdquo pairs and connections with theinputs of SRNN and corresponding outputs The causalitydetection results could include four possible types (1) singlecause-effect pairs in which any two pairs are independentfrom each other (2) cause-effect chains which are formedwithmore than one cause-effect pairs connected together (aneffect is a cause of another effect) (3) one-cause-multiple-effect pairs (4) multiple-cause-one-effect pairs All of thecausality connections are archived in a database Both CauLandBCERmodules exploit three libraries (databases) causal-ity in verbs causality in discourse connectives and implicitcausality for reference
326 Testing Causal-Effect Relations (TCER) This moduletests and verifies a new processed causal-effect relation withour existing and expanding collection ofmassive causal-effectrelations The concrete technology includes comparison oftriples ⟨CategoryCauseEffect⟩
4 Experiments and Results
Our experiments test general causality analysis system(GCAS) general causality analysis system with learning(GCAS-L) and CL-CIS together in order to examine andreveal the assertion that category information is a necessarysupplement for causality extraction and has impact on subse-quent learning-based cause-effect identification
In our experiments we have used two text collections (1)Reuters-21578 currently the most widely used test collectionfor text processing research (2) BBC-News2000 collected bya self-developed Web Crawler from httpwwwbbccouk
The Scientific World Journal 7
00
01
02
03
04
05
06
07
08
09
10
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
0681 0706 069807320719
0755 074207930776
0815 0809
0897
(a) Precision
00
01
02
03
04
05
06
07
08
09
10
0593 0581 0602
0706
0621 0609 0616
0734
0685 0661 0683
0805
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
(b) Recall
00
01
02
03
04
05
06
07
08
09
10
0634 0637 0646 0719
0666 0674 0673 0762
0728 0730 0741
0849
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
GCASGCAS-L
CL-CIS
(c) F-measure (balanced)
Figure 5 Evaluation results of GCAS GCAS-L and CL-CIS
BBC-News2000 collected 2000 news articles which havebeen pruned off unnecessary Web page elements such ashtml tags images URLs and so forth BBC-News2000 is notcategorized and is treated as a whole hybrid
30 judges are involved to manually search cause-effectpairs and construct standard causality references (SCR) SCR-Reuters for Reuters-21578 and SCR-BBC for BBC-News2000Due to limited human resources in current stage SCR-Reuters only covers 2000 documents newid ranges from 1 to2000 contained in files ldquoreut2-000sgmrdquo and ldquoreut2-001sgmrdquoThe rest of the files of Reuters-21578 are still involved in thetraining phase
The performances of GCAS GCAS-L and CL-CIS areevaluated with precision recall and F-measure [21] thetraditional measures that have been widely applied by mostinformation retrieval systems to analyze and evaluate theirperformance The F-measure is a harmonic combination ofthe precision and recall values used in information retrievalAs shown in Tables 6 7 and 8 the experimental results statethat (1) GCAS scores from 0681 to 0732 on precision andfrom 0581 to 0706 on recall (2) GCAS-L scores from 0719to 0793 on precision and from 0609 to 0734 on recall (3)CL-CIS scores from 0776 to 0897 on precision and from0661 to 0805 on recall Figure 5 explicitly states that (1)
Table 6 Experimental results of GCAS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0681 0593 0634reut2-001sgm (newid 1001ndash2000) 0706 0581 0637Both (reut2) (newid 1ndash2000) 0698 0602 0646BBC-News2000 0732 0706 0719
Table 7 Experimental results of GCAS-L
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0719 0621 0666reut2-001sgm (newid 1001ndash2000) 0755 0609 0674Both (reut2) (newid 1ndash2000) 0742 0616 0673BBC-News2000 0793 0734 0762
the general causality analysis system with learning (GCAS-L) performances better than the general causality analysissystem (GCAS) on all evaluation measures (2) the causalityanalysis system strengthened with causal and category learn-ing (CL-CIS) exceeds GCAS-L in the meantime
8 The Scientific World Journal
Table 8 Experimental results of CL-CIS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0776 0685 0728reut2-001sgm (newid 1001ndash2000) 0815 0661 0730Both (reut2) (newid 1ndash2000) 0809 0683 0741BBC-News2000 0897 0805 0849
5 Concluding Remarks
In recent years detection and clarification of cause-effectrelationships among texts events or objects has been elevatedto a prominent research topic of natural and social sciencesover the human knowledge development history
This paper demonstrates a novel comprehensive causalityextraction system (CL-CIS) integrated with the means ofcategory-learning CL-CIS considers cause-effect relations inboth explicit and implicit forms and especially practices therelation between category and causality in computation
CL-CIS is inspired with cognitive philosophy in categoryand causality In causality extraction and induction tasks CL-CIS implements a simple recurrent neural network (SRNN)to simulate human associative memory which has the abilityto associate different types of inputs when processing infor-mation
In elaborately designed experimental tasks CL-CIS hasbeen examined in full with two text collections Reuters-21578and BBC-News2000 The experimental results have testifiedthe capability and performance of CL-CIS in constructionof cause-effect relations and also confirmed the expectationthat the means of causal and category learning will improvethe precision and coverage of causality induction in a notablemanner
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work is financially supported by the National NaturalScience Foundation of China (Grant no 61003126)
References
[1] httpenwikipediaorgwikiCausality[2] J Pearl Causality Models Reasoning and Inference Cambridge
University Press New York NY USA 2000[3] D-S Chang and K-S Choi ldquoIncremental cue phrase learning
and bootstrapping method for causality extraction using cuephrase and word pair probabilitiesrdquo Information Processing andManagement vol 42 no 3 pp 662ndash678 2006
[4] S Kim R H Bracewell and K M Wallace ldquoA frameworkfor automatic causality extraction using semantic similarityrdquoin Proceedings of the ASME International Design EngineeringTechnical Conferences and Computers and Information in Engi-neering Conference (IDETCCIE rsquo07) pp 831ndash840 Las VegasNev USA September 2007
[5] T Inui K Inui andYMatsumoto ldquoAcquiring causal knowledgefrom text using the connective markersrdquo Transactions of theInformation Processing Society of Japan vol 45 no 3 pp 919ndash932 2004
[6] D Marcu and A Echihabi ldquoAn unsupervised approach torecognizing discourse relationsrdquo in Proceedings of the 40thAnnual Meeting of the Association for Computational LinguisticsConference (ACL rsquo02) pp 368ndash375 University of PennsylvaniaPhiladelphia Pa USA 2002
[7] C Pechsiri and A Kawtraku ldquoMining causality from texts forquestion answering systemrdquo IEICE Transactions on Informationand Systems vol 90 no 10 pp 1523ndash1533 2007
[8] R Girju ldquoAutomatic detection of causal relations for questionansweringrdquo in Proceedings of the 41st Annual Meeting of theAssociation for Computational Linguistics 2003
[9] K Torisawa ldquoAutomatic extraction of commonsense inferencerules from corporardquo in Proceedings of the 9th Annual Meeting ofthe Association for Natural Language Processing 2003
[10] Y Guo N Hua and Z Shao ldquoCognitive causality detectionwith associative memory in textual eventsrdquo in Proceedings ofthe IEEE International Symposium on Information Engineeringand Electronic Commerce (IEEC rsquo09) pp 140ndash144 TernopilUkraine May 2009
[11] P Thwaites J Q Smith and E Riccomagno ldquoCausal analysiswith Chain Event GraphsrdquoArtificial Intelligence vol 174 no 12-13 pp 889ndash909 2010
[12] M R Waldmann and Y Hagmayer ldquoCategories and causalitythe neglected directionrdquo Cognitive Psychology vol 53 no 1 pp27ndash58 2006
[13] G L MurphyTheBig Book of Concepts MIT Press CambridgeMass USA 2002
[14] B Rehder ldquoA causal-model theory of conceptual representationand categorizationrdquo Journal of Experimental Psychology Learn-ing Memory and Cognition vol 29 no 6 pp 1141ndash1159 2003
[15] B Rehder ldquoCategorization as casual reasoningrdquo Cognitive Sci-ence vol 27 no 5 pp 709ndash748 2003
[16] B Rehder and R Hastie ldquoCausal knowledge and categoriesthe effects of causal beliefs on categorization induction andsimilarityrdquo Journal of Experimental Psychology General vol 130no 3 pp 323ndash360 2001
[17] B Rehder and R Hastie ldquoCategory coherence and category-based property inductionrdquo Cognition vol 91 no 2 pp 113ndash1532004
[18] Y Lien and PWCheng ldquoDistinguishing genuine from spuriouscauses a coherence hypothesisrdquo Cognitive Psychology vol 40pp 87ndash137 2000
[19] G A Miller R Beckwith C Fellbaum D Gross and K JMiller ldquoIntroduction to WordNet an on-line lexical databaserdquoInternational Journal of Lexicography vol 3 no 4 pp 235ndash3121990
[20] A W Koornneef and J J A Van Berkum ldquoOn the use of verb-based implicit causality in sentence comprehension evidencefrom self-paced reading and eye trackingrdquo Journal of Memoryand Language vol 54 no 4 pp 445ndash465 2006
[21] C D Manning and H Schutze Foundations of StatisticalNatural Language Processing TheMIT Press Cambridge MassUSA 4th edition 2001
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World Journal 5
Table 3 Exemplar implicit causality verbs
NP1 verbs Bias NP2 verbs BiasAmazed 119 Admire 200Annoy 119 Adore 186Apologize 100 Appreciate 200Be in the way 108 Comfort 191Beg 117 Compliment on something 196Bore 105 Congratulate 195Call 119 Criticize 200Confess 103 Envy 200Disappoint 103 Fear 200Disturb 114 Fire 196Fascinate 100 Hate 196Hurt 113 Hold in contempt 200Inspire 123 Hold responsible 191Intimidate 123 Loathe 186Irritate 122 Love 186Lie to 122 Praise 196Mislead 122 Press charges against 191Swindle 113 Punish 195Win 119 Respect 195Worry 119 Thank 182
Table 4 System composition comparison of GCAS GCAS-L andCL-CIS
System composition GCAS GCAS-L CL-CISModules
Category learning radic
Classify exemplars intocategories radic
Category and causal mapping radic
Causal learning radic radic
Building causal-effectrelations radic radic radic
Testing causal-effectrelations radic radic radic
LibrariesCausality in verbs radic radic radic
Causality in discourseconnectives Optional radic radic
Implicit causality Optional radic
322 Classify Exemplars into Categories (CEC) In the clas-sify exemplars into categories module each sentence istreated as an individual independent event and parsed intophrases The classification is based on the comparison ofa sentence with features of each category so as to set upthe category background knowledge for each sentence Forexample the concept bank can be clarified as a financeorganization or a river body boundary with correspondingcategory background knowledge of its sentence and contexts
Category learning
Classify exemplars into
categories
Category and causal
mapping
Causal learning
Building causal-effect
relations
Testing causal-effect
relations
Causality in verbs
Causality in discourse
connectivesImplicit causality
Figure 3 System framework of CL-CIS
Inputlayer (IL)
Recurrentlayer (RL)
Outputlayer (OL)
Contextlayer (CL)
Time delay
OU1 OU2
RU1RU2
WOR
IU1 IU2
WRI W
RC
CU1 CU2
middot middot middot
middot middot middot
middot middot middot middot middot middot
OU|O|
RU|R|
IU|I| CU|C|
Figure 4 Architecture of SRNN
323 Category and Causal Mapping (CCM) This moduleconstructs the category-causal-effect mapping relations forexample a virus from a predefined category causes specificdisease-related symptoms such as a swelling of the spleen(splenomegaly) The construction mechanism to a greatextent is based on collection storage and indexing ofmassive cases
324 Causal Learning (CauL) As analyzed before a sim-ple connectionist one-layer network that may be used tounderstand the task of when to activate category knowledgewhen learning about novel effects This module adopts asimple recurrent neural network (SRNN) to simulate theconnectionist model
Symbols in Figure 4 are defined in Table 5 the first orderweight matrices WRI and WOR fully connect the units ofthe input layer (IL) the recurrent layer (RL) and the outputlayer (OL) respectively as in the feed forward multilayerperceptron (MLP) The current activities of recurrent unitsRU(119905) are fed back through time delay connections to thecontext layer which is presented as CU(119905+1) = RU(119905)
Therefore each unit in recurrent layer is fed by activitiesof all recurrent units from previous time step through
6 The Scientific World Journal
Table 5 Definition of SRNN Symbols
Symbols DefinitionIU A unit of input layerRU A unit of recurrent layerCU A unit of context layerOU A unit of output layer|119868| The number of units in IL|119877| The number of units in RL|119862| The number of units in CL|119874| The number of units in OLWRI The weight vector from IL to RLWRC The weight vector from CL to RLWOR The weight vector from RL to OL
recurrent weight matrix WRC The context layer which iscomposed of activities of recurrent units from previous timestep can be viewed as an extension of input layer to therecurrent layer Above working procedure represents thememory of the network via holding contextual informationfrom previous time steps
The weight matrices 119882RI 119882RC and 119882OR are presentedas follows
119882RI= [(119908
RI1 )119879 (119908
RI2 )119879 (119908
RI|119877|)119879]
=
[[[[[
[
119908ri11 119908
ri12 sdot sdot sdot 119908
ri1|119877|
119908ri21 119908ri22 sdot sdot sdot 119908
ri1|119877|
d
119908ri|119868|1 119908
ri|119868|2 sdot sdot sdot 119908
ri|119868||119877|
]]]]]
]
119882RC= [(119908
RC1 )119879 (119908
RC2 )119879 (119908
RC|119877| )119879]
=
[[[[
[
119908rc11 119908rc12 sdot sdot sdot 119908
rc1|119877|
119908rc21 119908rc22 sdot sdot sdot 119908
rc1|119877|
d
119908rc|119862|1 119908
rc|119862|2 sdot sdot sdot 119908
rc|119862||119877|
]]]]
]
119882OR= [(119908
OR1 )119879 (119908
OR2 )119879 (119908
OR|119874| )119879]
=
[[[[
[
119908or11 119908or12 sdot sdot sdot 119908
or1|119874|
119908or21 119908or22 sdot sdot sdot 119908
or1|119874|
d
119908or|119877|1 119908
or|119877|2 sdot sdot sdot 119908
or|119877||119874|
]]]]
]
(1)
In the above formulations (119908RI119896 )119879 is the transpose of 119908RI119896
for the instance of119882RI where 119908RI119896 is a row vector and (119908RI119896 )119879
is the column vector of the same elements The vector 119908RI119896 =(119908ri1119896 119908
ri2119896 119908
ri|119868|119896) represents the weights from all the input
layer units to the recurrent (hidden) layer unit RU119896The sameconclusion applies with119882RC and119882OR
Given an input pattern in time t IU(t) = (IU(119905)1 IU(119905)2
IU(119905)|119868|) and recurrent activities RU(t) = (RU(119905)1 RU
(119905)2
RU(119905)|119877|) for the 119894th recurrent unit the net input RU1015840(119905)
119894and
output activity RU(119905)119894
are calculated as follows
RU1015840(119905)119894 = IU(t)sdot (119908
RI119894 )119879+ RU(tminus1) sdot (119908RC119894 )
119879
=
|119868|
sum119895=1
IU(119905)119895 119908ri119895119894 +
|119877|
sum119895=1
RU(119905minus1)119895 119908rc119895119894
RU(119905)119894 = 119891 (RU1015840(119905)
119894 )
(2)
For the kth output unit its net inputOU1015840(119905)119896
and output activityOU(119905)119896
are calculated as (3) Consider the following
OU1015840(119905)119896= RU(t) sdot (119908OR119896 )
119879=
|119877|
sum119895=1
RU(119905)119895 119908or119895119896
OU(119905)119896= 119891 (OU1015840(119905)
119896)
(3)
Here the activation function 119891 applies the logistic sigmoidfunction (4) in this paper Consider the following
119891 (119909) =1
1 + 119890minus119909=119890119909
1 + 119890119909 (4)
325 Building Causal-Effect Relations (BCER) This modulebuilds up the ldquocause-effectrdquo pairs and connections with theinputs of SRNN and corresponding outputs The causalitydetection results could include four possible types (1) singlecause-effect pairs in which any two pairs are independentfrom each other (2) cause-effect chains which are formedwithmore than one cause-effect pairs connected together (aneffect is a cause of another effect) (3) one-cause-multiple-effect pairs (4) multiple-cause-one-effect pairs All of thecausality connections are archived in a database Both CauLandBCERmodules exploit three libraries (databases) causal-ity in verbs causality in discourse connectives and implicitcausality for reference
326 Testing Causal-Effect Relations (TCER) This moduletests and verifies a new processed causal-effect relation withour existing and expanding collection ofmassive causal-effectrelations The concrete technology includes comparison oftriples ⟨CategoryCauseEffect⟩
4 Experiments and Results
Our experiments test general causality analysis system(GCAS) general causality analysis system with learning(GCAS-L) and CL-CIS together in order to examine andreveal the assertion that category information is a necessarysupplement for causality extraction and has impact on subse-quent learning-based cause-effect identification
In our experiments we have used two text collections (1)Reuters-21578 currently the most widely used test collectionfor text processing research (2) BBC-News2000 collected bya self-developed Web Crawler from httpwwwbbccouk
The Scientific World Journal 7
00
01
02
03
04
05
06
07
08
09
10
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
0681 0706 069807320719
0755 074207930776
0815 0809
0897
(a) Precision
00
01
02
03
04
05
06
07
08
09
10
0593 0581 0602
0706
0621 0609 0616
0734
0685 0661 0683
0805
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
(b) Recall
00
01
02
03
04
05
06
07
08
09
10
0634 0637 0646 0719
0666 0674 0673 0762
0728 0730 0741
0849
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
GCASGCAS-L
CL-CIS
(c) F-measure (balanced)
Figure 5 Evaluation results of GCAS GCAS-L and CL-CIS
BBC-News2000 collected 2000 news articles which havebeen pruned off unnecessary Web page elements such ashtml tags images URLs and so forth BBC-News2000 is notcategorized and is treated as a whole hybrid
30 judges are involved to manually search cause-effectpairs and construct standard causality references (SCR) SCR-Reuters for Reuters-21578 and SCR-BBC for BBC-News2000Due to limited human resources in current stage SCR-Reuters only covers 2000 documents newid ranges from 1 to2000 contained in files ldquoreut2-000sgmrdquo and ldquoreut2-001sgmrdquoThe rest of the files of Reuters-21578 are still involved in thetraining phase
The performances of GCAS GCAS-L and CL-CIS areevaluated with precision recall and F-measure [21] thetraditional measures that have been widely applied by mostinformation retrieval systems to analyze and evaluate theirperformance The F-measure is a harmonic combination ofthe precision and recall values used in information retrievalAs shown in Tables 6 7 and 8 the experimental results statethat (1) GCAS scores from 0681 to 0732 on precision andfrom 0581 to 0706 on recall (2) GCAS-L scores from 0719to 0793 on precision and from 0609 to 0734 on recall (3)CL-CIS scores from 0776 to 0897 on precision and from0661 to 0805 on recall Figure 5 explicitly states that (1)
Table 6 Experimental results of GCAS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0681 0593 0634reut2-001sgm (newid 1001ndash2000) 0706 0581 0637Both (reut2) (newid 1ndash2000) 0698 0602 0646BBC-News2000 0732 0706 0719
Table 7 Experimental results of GCAS-L
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0719 0621 0666reut2-001sgm (newid 1001ndash2000) 0755 0609 0674Both (reut2) (newid 1ndash2000) 0742 0616 0673BBC-News2000 0793 0734 0762
the general causality analysis system with learning (GCAS-L) performances better than the general causality analysissystem (GCAS) on all evaluation measures (2) the causalityanalysis system strengthened with causal and category learn-ing (CL-CIS) exceeds GCAS-L in the meantime
8 The Scientific World Journal
Table 8 Experimental results of CL-CIS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0776 0685 0728reut2-001sgm (newid 1001ndash2000) 0815 0661 0730Both (reut2) (newid 1ndash2000) 0809 0683 0741BBC-News2000 0897 0805 0849
5 Concluding Remarks
In recent years detection and clarification of cause-effectrelationships among texts events or objects has been elevatedto a prominent research topic of natural and social sciencesover the human knowledge development history
This paper demonstrates a novel comprehensive causalityextraction system (CL-CIS) integrated with the means ofcategory-learning CL-CIS considers cause-effect relations inboth explicit and implicit forms and especially practices therelation between category and causality in computation
CL-CIS is inspired with cognitive philosophy in categoryand causality In causality extraction and induction tasks CL-CIS implements a simple recurrent neural network (SRNN)to simulate human associative memory which has the abilityto associate different types of inputs when processing infor-mation
In elaborately designed experimental tasks CL-CIS hasbeen examined in full with two text collections Reuters-21578and BBC-News2000 The experimental results have testifiedthe capability and performance of CL-CIS in constructionof cause-effect relations and also confirmed the expectationthat the means of causal and category learning will improvethe precision and coverage of causality induction in a notablemanner
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work is financially supported by the National NaturalScience Foundation of China (Grant no 61003126)
References
[1] httpenwikipediaorgwikiCausality[2] J Pearl Causality Models Reasoning and Inference Cambridge
University Press New York NY USA 2000[3] D-S Chang and K-S Choi ldquoIncremental cue phrase learning
and bootstrapping method for causality extraction using cuephrase and word pair probabilitiesrdquo Information Processing andManagement vol 42 no 3 pp 662ndash678 2006
[4] S Kim R H Bracewell and K M Wallace ldquoA frameworkfor automatic causality extraction using semantic similarityrdquoin Proceedings of the ASME International Design EngineeringTechnical Conferences and Computers and Information in Engi-neering Conference (IDETCCIE rsquo07) pp 831ndash840 Las VegasNev USA September 2007
[5] T Inui K Inui andYMatsumoto ldquoAcquiring causal knowledgefrom text using the connective markersrdquo Transactions of theInformation Processing Society of Japan vol 45 no 3 pp 919ndash932 2004
[6] D Marcu and A Echihabi ldquoAn unsupervised approach torecognizing discourse relationsrdquo in Proceedings of the 40thAnnual Meeting of the Association for Computational LinguisticsConference (ACL rsquo02) pp 368ndash375 University of PennsylvaniaPhiladelphia Pa USA 2002
[7] C Pechsiri and A Kawtraku ldquoMining causality from texts forquestion answering systemrdquo IEICE Transactions on Informationand Systems vol 90 no 10 pp 1523ndash1533 2007
[8] R Girju ldquoAutomatic detection of causal relations for questionansweringrdquo in Proceedings of the 41st Annual Meeting of theAssociation for Computational Linguistics 2003
[9] K Torisawa ldquoAutomatic extraction of commonsense inferencerules from corporardquo in Proceedings of the 9th Annual Meeting ofthe Association for Natural Language Processing 2003
[10] Y Guo N Hua and Z Shao ldquoCognitive causality detectionwith associative memory in textual eventsrdquo in Proceedings ofthe IEEE International Symposium on Information Engineeringand Electronic Commerce (IEEC rsquo09) pp 140ndash144 TernopilUkraine May 2009
[11] P Thwaites J Q Smith and E Riccomagno ldquoCausal analysiswith Chain Event GraphsrdquoArtificial Intelligence vol 174 no 12-13 pp 889ndash909 2010
[12] M R Waldmann and Y Hagmayer ldquoCategories and causalitythe neglected directionrdquo Cognitive Psychology vol 53 no 1 pp27ndash58 2006
[13] G L MurphyTheBig Book of Concepts MIT Press CambridgeMass USA 2002
[14] B Rehder ldquoA causal-model theory of conceptual representationand categorizationrdquo Journal of Experimental Psychology Learn-ing Memory and Cognition vol 29 no 6 pp 1141ndash1159 2003
[15] B Rehder ldquoCategorization as casual reasoningrdquo Cognitive Sci-ence vol 27 no 5 pp 709ndash748 2003
[16] B Rehder and R Hastie ldquoCausal knowledge and categoriesthe effects of causal beliefs on categorization induction andsimilarityrdquo Journal of Experimental Psychology General vol 130no 3 pp 323ndash360 2001
[17] B Rehder and R Hastie ldquoCategory coherence and category-based property inductionrdquo Cognition vol 91 no 2 pp 113ndash1532004
[18] Y Lien and PWCheng ldquoDistinguishing genuine from spuriouscauses a coherence hypothesisrdquo Cognitive Psychology vol 40pp 87ndash137 2000
[19] G A Miller R Beckwith C Fellbaum D Gross and K JMiller ldquoIntroduction to WordNet an on-line lexical databaserdquoInternational Journal of Lexicography vol 3 no 4 pp 235ndash3121990
[20] A W Koornneef and J J A Van Berkum ldquoOn the use of verb-based implicit causality in sentence comprehension evidencefrom self-paced reading and eye trackingrdquo Journal of Memoryand Language vol 54 no 4 pp 445ndash465 2006
[21] C D Manning and H Schutze Foundations of StatisticalNatural Language Processing TheMIT Press Cambridge MassUSA 4th edition 2001
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
6 The Scientific World Journal
Table 5 Definition of SRNN Symbols
Symbols DefinitionIU A unit of input layerRU A unit of recurrent layerCU A unit of context layerOU A unit of output layer|119868| The number of units in IL|119877| The number of units in RL|119862| The number of units in CL|119874| The number of units in OLWRI The weight vector from IL to RLWRC The weight vector from CL to RLWOR The weight vector from RL to OL
recurrent weight matrix WRC The context layer which iscomposed of activities of recurrent units from previous timestep can be viewed as an extension of input layer to therecurrent layer Above working procedure represents thememory of the network via holding contextual informationfrom previous time steps
The weight matrices 119882RI 119882RC and 119882OR are presentedas follows
119882RI= [(119908
RI1 )119879 (119908
RI2 )119879 (119908
RI|119877|)119879]
=
[[[[[
[
119908ri11 119908
ri12 sdot sdot sdot 119908
ri1|119877|
119908ri21 119908ri22 sdot sdot sdot 119908
ri1|119877|
d
119908ri|119868|1 119908
ri|119868|2 sdot sdot sdot 119908
ri|119868||119877|
]]]]]
]
119882RC= [(119908
RC1 )119879 (119908
RC2 )119879 (119908
RC|119877| )119879]
=
[[[[
[
119908rc11 119908rc12 sdot sdot sdot 119908
rc1|119877|
119908rc21 119908rc22 sdot sdot sdot 119908
rc1|119877|
d
119908rc|119862|1 119908
rc|119862|2 sdot sdot sdot 119908
rc|119862||119877|
]]]]
]
119882OR= [(119908
OR1 )119879 (119908
OR2 )119879 (119908
OR|119874| )119879]
=
[[[[
[
119908or11 119908or12 sdot sdot sdot 119908
or1|119874|
119908or21 119908or22 sdot sdot sdot 119908
or1|119874|
d
119908or|119877|1 119908
or|119877|2 sdot sdot sdot 119908
or|119877||119874|
]]]]
]
(1)
In the above formulations (119908RI119896 )119879 is the transpose of 119908RI119896
for the instance of119882RI where 119908RI119896 is a row vector and (119908RI119896 )119879
is the column vector of the same elements The vector 119908RI119896 =(119908ri1119896 119908
ri2119896 119908
ri|119868|119896) represents the weights from all the input
layer units to the recurrent (hidden) layer unit RU119896The sameconclusion applies with119882RC and119882OR
Given an input pattern in time t IU(t) = (IU(119905)1 IU(119905)2
IU(119905)|119868|) and recurrent activities RU(t) = (RU(119905)1 RU
(119905)2
RU(119905)|119877|) for the 119894th recurrent unit the net input RU1015840(119905)
119894and
output activity RU(119905)119894
are calculated as follows
RU1015840(119905)119894 = IU(t)sdot (119908
RI119894 )119879+ RU(tminus1) sdot (119908RC119894 )
119879
=
|119868|
sum119895=1
IU(119905)119895 119908ri119895119894 +
|119877|
sum119895=1
RU(119905minus1)119895 119908rc119895119894
RU(119905)119894 = 119891 (RU1015840(119905)
119894 )
(2)
For the kth output unit its net inputOU1015840(119905)119896
and output activityOU(119905)119896
are calculated as (3) Consider the following
OU1015840(119905)119896= RU(t) sdot (119908OR119896 )
119879=
|119877|
sum119895=1
RU(119905)119895 119908or119895119896
OU(119905)119896= 119891 (OU1015840(119905)
119896)
(3)
Here the activation function 119891 applies the logistic sigmoidfunction (4) in this paper Consider the following
119891 (119909) =1
1 + 119890minus119909=119890119909
1 + 119890119909 (4)
325 Building Causal-Effect Relations (BCER) This modulebuilds up the ldquocause-effectrdquo pairs and connections with theinputs of SRNN and corresponding outputs The causalitydetection results could include four possible types (1) singlecause-effect pairs in which any two pairs are independentfrom each other (2) cause-effect chains which are formedwithmore than one cause-effect pairs connected together (aneffect is a cause of another effect) (3) one-cause-multiple-effect pairs (4) multiple-cause-one-effect pairs All of thecausality connections are archived in a database Both CauLandBCERmodules exploit three libraries (databases) causal-ity in verbs causality in discourse connectives and implicitcausality for reference
326 Testing Causal-Effect Relations (TCER) This moduletests and verifies a new processed causal-effect relation withour existing and expanding collection ofmassive causal-effectrelations The concrete technology includes comparison oftriples ⟨CategoryCauseEffect⟩
4 Experiments and Results
Our experiments test general causality analysis system(GCAS) general causality analysis system with learning(GCAS-L) and CL-CIS together in order to examine andreveal the assertion that category information is a necessarysupplement for causality extraction and has impact on subse-quent learning-based cause-effect identification
In our experiments we have used two text collections (1)Reuters-21578 currently the most widely used test collectionfor text processing research (2) BBC-News2000 collected bya self-developed Web Crawler from httpwwwbbccouk
The Scientific World Journal 7
00
01
02
03
04
05
06
07
08
09
10
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
0681 0706 069807320719
0755 074207930776
0815 0809
0897
(a) Precision
00
01
02
03
04
05
06
07
08
09
10
0593 0581 0602
0706
0621 0609 0616
0734
0685 0661 0683
0805
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
(b) Recall
00
01
02
03
04
05
06
07
08
09
10
0634 0637 0646 0719
0666 0674 0673 0762
0728 0730 0741
0849
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
GCASGCAS-L
CL-CIS
(c) F-measure (balanced)
Figure 5 Evaluation results of GCAS GCAS-L and CL-CIS
BBC-News2000 collected 2000 news articles which havebeen pruned off unnecessary Web page elements such ashtml tags images URLs and so forth BBC-News2000 is notcategorized and is treated as a whole hybrid
30 judges are involved to manually search cause-effectpairs and construct standard causality references (SCR) SCR-Reuters for Reuters-21578 and SCR-BBC for BBC-News2000Due to limited human resources in current stage SCR-Reuters only covers 2000 documents newid ranges from 1 to2000 contained in files ldquoreut2-000sgmrdquo and ldquoreut2-001sgmrdquoThe rest of the files of Reuters-21578 are still involved in thetraining phase
The performances of GCAS GCAS-L and CL-CIS areevaluated with precision recall and F-measure [21] thetraditional measures that have been widely applied by mostinformation retrieval systems to analyze and evaluate theirperformance The F-measure is a harmonic combination ofthe precision and recall values used in information retrievalAs shown in Tables 6 7 and 8 the experimental results statethat (1) GCAS scores from 0681 to 0732 on precision andfrom 0581 to 0706 on recall (2) GCAS-L scores from 0719to 0793 on precision and from 0609 to 0734 on recall (3)CL-CIS scores from 0776 to 0897 on precision and from0661 to 0805 on recall Figure 5 explicitly states that (1)
Table 6 Experimental results of GCAS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0681 0593 0634reut2-001sgm (newid 1001ndash2000) 0706 0581 0637Both (reut2) (newid 1ndash2000) 0698 0602 0646BBC-News2000 0732 0706 0719
Table 7 Experimental results of GCAS-L
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0719 0621 0666reut2-001sgm (newid 1001ndash2000) 0755 0609 0674Both (reut2) (newid 1ndash2000) 0742 0616 0673BBC-News2000 0793 0734 0762
the general causality analysis system with learning (GCAS-L) performances better than the general causality analysissystem (GCAS) on all evaluation measures (2) the causalityanalysis system strengthened with causal and category learn-ing (CL-CIS) exceeds GCAS-L in the meantime
8 The Scientific World Journal
Table 8 Experimental results of CL-CIS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0776 0685 0728reut2-001sgm (newid 1001ndash2000) 0815 0661 0730Both (reut2) (newid 1ndash2000) 0809 0683 0741BBC-News2000 0897 0805 0849
5 Concluding Remarks
In recent years detection and clarification of cause-effectrelationships among texts events or objects has been elevatedto a prominent research topic of natural and social sciencesover the human knowledge development history
This paper demonstrates a novel comprehensive causalityextraction system (CL-CIS) integrated with the means ofcategory-learning CL-CIS considers cause-effect relations inboth explicit and implicit forms and especially practices therelation between category and causality in computation
CL-CIS is inspired with cognitive philosophy in categoryand causality In causality extraction and induction tasks CL-CIS implements a simple recurrent neural network (SRNN)to simulate human associative memory which has the abilityto associate different types of inputs when processing infor-mation
In elaborately designed experimental tasks CL-CIS hasbeen examined in full with two text collections Reuters-21578and BBC-News2000 The experimental results have testifiedthe capability and performance of CL-CIS in constructionof cause-effect relations and also confirmed the expectationthat the means of causal and category learning will improvethe precision and coverage of causality induction in a notablemanner
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work is financially supported by the National NaturalScience Foundation of China (Grant no 61003126)
References
[1] httpenwikipediaorgwikiCausality[2] J Pearl Causality Models Reasoning and Inference Cambridge
University Press New York NY USA 2000[3] D-S Chang and K-S Choi ldquoIncremental cue phrase learning
and bootstrapping method for causality extraction using cuephrase and word pair probabilitiesrdquo Information Processing andManagement vol 42 no 3 pp 662ndash678 2006
[4] S Kim R H Bracewell and K M Wallace ldquoA frameworkfor automatic causality extraction using semantic similarityrdquoin Proceedings of the ASME International Design EngineeringTechnical Conferences and Computers and Information in Engi-neering Conference (IDETCCIE rsquo07) pp 831ndash840 Las VegasNev USA September 2007
[5] T Inui K Inui andYMatsumoto ldquoAcquiring causal knowledgefrom text using the connective markersrdquo Transactions of theInformation Processing Society of Japan vol 45 no 3 pp 919ndash932 2004
[6] D Marcu and A Echihabi ldquoAn unsupervised approach torecognizing discourse relationsrdquo in Proceedings of the 40thAnnual Meeting of the Association for Computational LinguisticsConference (ACL rsquo02) pp 368ndash375 University of PennsylvaniaPhiladelphia Pa USA 2002
[7] C Pechsiri and A Kawtraku ldquoMining causality from texts forquestion answering systemrdquo IEICE Transactions on Informationand Systems vol 90 no 10 pp 1523ndash1533 2007
[8] R Girju ldquoAutomatic detection of causal relations for questionansweringrdquo in Proceedings of the 41st Annual Meeting of theAssociation for Computational Linguistics 2003
[9] K Torisawa ldquoAutomatic extraction of commonsense inferencerules from corporardquo in Proceedings of the 9th Annual Meeting ofthe Association for Natural Language Processing 2003
[10] Y Guo N Hua and Z Shao ldquoCognitive causality detectionwith associative memory in textual eventsrdquo in Proceedings ofthe IEEE International Symposium on Information Engineeringand Electronic Commerce (IEEC rsquo09) pp 140ndash144 TernopilUkraine May 2009
[11] P Thwaites J Q Smith and E Riccomagno ldquoCausal analysiswith Chain Event GraphsrdquoArtificial Intelligence vol 174 no 12-13 pp 889ndash909 2010
[12] M R Waldmann and Y Hagmayer ldquoCategories and causalitythe neglected directionrdquo Cognitive Psychology vol 53 no 1 pp27ndash58 2006
[13] G L MurphyTheBig Book of Concepts MIT Press CambridgeMass USA 2002
[14] B Rehder ldquoA causal-model theory of conceptual representationand categorizationrdquo Journal of Experimental Psychology Learn-ing Memory and Cognition vol 29 no 6 pp 1141ndash1159 2003
[15] B Rehder ldquoCategorization as casual reasoningrdquo Cognitive Sci-ence vol 27 no 5 pp 709ndash748 2003
[16] B Rehder and R Hastie ldquoCausal knowledge and categoriesthe effects of causal beliefs on categorization induction andsimilarityrdquo Journal of Experimental Psychology General vol 130no 3 pp 323ndash360 2001
[17] B Rehder and R Hastie ldquoCategory coherence and category-based property inductionrdquo Cognition vol 91 no 2 pp 113ndash1532004
[18] Y Lien and PWCheng ldquoDistinguishing genuine from spuriouscauses a coherence hypothesisrdquo Cognitive Psychology vol 40pp 87ndash137 2000
[19] G A Miller R Beckwith C Fellbaum D Gross and K JMiller ldquoIntroduction to WordNet an on-line lexical databaserdquoInternational Journal of Lexicography vol 3 no 4 pp 235ndash3121990
[20] A W Koornneef and J J A Van Berkum ldquoOn the use of verb-based implicit causality in sentence comprehension evidencefrom self-paced reading and eye trackingrdquo Journal of Memoryand Language vol 54 no 4 pp 445ndash465 2006
[21] C D Manning and H Schutze Foundations of StatisticalNatural Language Processing TheMIT Press Cambridge MassUSA 4th edition 2001
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
The Scientific World Journal 7
00
01
02
03
04
05
06
07
08
09
10
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
0681 0706 069807320719
0755 074207930776
0815 0809
0897
(a) Precision
00
01
02
03
04
05
06
07
08
09
10
0593 0581 0602
0706
0621 0609 0616
0734
0685 0661 0683
0805
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
(b) Recall
00
01
02
03
04
05
06
07
08
09
10
0634 0637 0646 0719
0666 0674 0673 0762
0728 0730 0741
0849
Reut2-000sgm Reut2-001sgm Both (reut2) BBC-News2000
GCASGCAS-L
CL-CIS
(c) F-measure (balanced)
Figure 5 Evaluation results of GCAS GCAS-L and CL-CIS
BBC-News2000 collected 2000 news articles which havebeen pruned off unnecessary Web page elements such ashtml tags images URLs and so forth BBC-News2000 is notcategorized and is treated as a whole hybrid
30 judges are involved to manually search cause-effectpairs and construct standard causality references (SCR) SCR-Reuters for Reuters-21578 and SCR-BBC for BBC-News2000Due to limited human resources in current stage SCR-Reuters only covers 2000 documents newid ranges from 1 to2000 contained in files ldquoreut2-000sgmrdquo and ldquoreut2-001sgmrdquoThe rest of the files of Reuters-21578 are still involved in thetraining phase
The performances of GCAS GCAS-L and CL-CIS areevaluated with precision recall and F-measure [21] thetraditional measures that have been widely applied by mostinformation retrieval systems to analyze and evaluate theirperformance The F-measure is a harmonic combination ofthe precision and recall values used in information retrievalAs shown in Tables 6 7 and 8 the experimental results statethat (1) GCAS scores from 0681 to 0732 on precision andfrom 0581 to 0706 on recall (2) GCAS-L scores from 0719to 0793 on precision and from 0609 to 0734 on recall (3)CL-CIS scores from 0776 to 0897 on precision and from0661 to 0805 on recall Figure 5 explicitly states that (1)
Table 6 Experimental results of GCAS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0681 0593 0634reut2-001sgm (newid 1001ndash2000) 0706 0581 0637Both (reut2) (newid 1ndash2000) 0698 0602 0646BBC-News2000 0732 0706 0719
Table 7 Experimental results of GCAS-L
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0719 0621 0666reut2-001sgm (newid 1001ndash2000) 0755 0609 0674Both (reut2) (newid 1ndash2000) 0742 0616 0673BBC-News2000 0793 0734 0762
the general causality analysis system with learning (GCAS-L) performances better than the general causality analysissystem (GCAS) on all evaluation measures (2) the causalityanalysis system strengthened with causal and category learn-ing (CL-CIS) exceeds GCAS-L in the meantime
8 The Scientific World Journal
Table 8 Experimental results of CL-CIS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0776 0685 0728reut2-001sgm (newid 1001ndash2000) 0815 0661 0730Both (reut2) (newid 1ndash2000) 0809 0683 0741BBC-News2000 0897 0805 0849
5 Concluding Remarks
In recent years detection and clarification of cause-effectrelationships among texts events or objects has been elevatedto a prominent research topic of natural and social sciencesover the human knowledge development history
This paper demonstrates a novel comprehensive causalityextraction system (CL-CIS) integrated with the means ofcategory-learning CL-CIS considers cause-effect relations inboth explicit and implicit forms and especially practices therelation between category and causality in computation
CL-CIS is inspired with cognitive philosophy in categoryand causality In causality extraction and induction tasks CL-CIS implements a simple recurrent neural network (SRNN)to simulate human associative memory which has the abilityto associate different types of inputs when processing infor-mation
In elaborately designed experimental tasks CL-CIS hasbeen examined in full with two text collections Reuters-21578and BBC-News2000 The experimental results have testifiedthe capability and performance of CL-CIS in constructionof cause-effect relations and also confirmed the expectationthat the means of causal and category learning will improvethe precision and coverage of causality induction in a notablemanner
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work is financially supported by the National NaturalScience Foundation of China (Grant no 61003126)
References
[1] httpenwikipediaorgwikiCausality[2] J Pearl Causality Models Reasoning and Inference Cambridge
University Press New York NY USA 2000[3] D-S Chang and K-S Choi ldquoIncremental cue phrase learning
and bootstrapping method for causality extraction using cuephrase and word pair probabilitiesrdquo Information Processing andManagement vol 42 no 3 pp 662ndash678 2006
[4] S Kim R H Bracewell and K M Wallace ldquoA frameworkfor automatic causality extraction using semantic similarityrdquoin Proceedings of the ASME International Design EngineeringTechnical Conferences and Computers and Information in Engi-neering Conference (IDETCCIE rsquo07) pp 831ndash840 Las VegasNev USA September 2007
[5] T Inui K Inui andYMatsumoto ldquoAcquiring causal knowledgefrom text using the connective markersrdquo Transactions of theInformation Processing Society of Japan vol 45 no 3 pp 919ndash932 2004
[6] D Marcu and A Echihabi ldquoAn unsupervised approach torecognizing discourse relationsrdquo in Proceedings of the 40thAnnual Meeting of the Association for Computational LinguisticsConference (ACL rsquo02) pp 368ndash375 University of PennsylvaniaPhiladelphia Pa USA 2002
[7] C Pechsiri and A Kawtraku ldquoMining causality from texts forquestion answering systemrdquo IEICE Transactions on Informationand Systems vol 90 no 10 pp 1523ndash1533 2007
[8] R Girju ldquoAutomatic detection of causal relations for questionansweringrdquo in Proceedings of the 41st Annual Meeting of theAssociation for Computational Linguistics 2003
[9] K Torisawa ldquoAutomatic extraction of commonsense inferencerules from corporardquo in Proceedings of the 9th Annual Meeting ofthe Association for Natural Language Processing 2003
[10] Y Guo N Hua and Z Shao ldquoCognitive causality detectionwith associative memory in textual eventsrdquo in Proceedings ofthe IEEE International Symposium on Information Engineeringand Electronic Commerce (IEEC rsquo09) pp 140ndash144 TernopilUkraine May 2009
[11] P Thwaites J Q Smith and E Riccomagno ldquoCausal analysiswith Chain Event GraphsrdquoArtificial Intelligence vol 174 no 12-13 pp 889ndash909 2010
[12] M R Waldmann and Y Hagmayer ldquoCategories and causalitythe neglected directionrdquo Cognitive Psychology vol 53 no 1 pp27ndash58 2006
[13] G L MurphyTheBig Book of Concepts MIT Press CambridgeMass USA 2002
[14] B Rehder ldquoA causal-model theory of conceptual representationand categorizationrdquo Journal of Experimental Psychology Learn-ing Memory and Cognition vol 29 no 6 pp 1141ndash1159 2003
[15] B Rehder ldquoCategorization as casual reasoningrdquo Cognitive Sci-ence vol 27 no 5 pp 709ndash748 2003
[16] B Rehder and R Hastie ldquoCausal knowledge and categoriesthe effects of causal beliefs on categorization induction andsimilarityrdquo Journal of Experimental Psychology General vol 130no 3 pp 323ndash360 2001
[17] B Rehder and R Hastie ldquoCategory coherence and category-based property inductionrdquo Cognition vol 91 no 2 pp 113ndash1532004
[18] Y Lien and PWCheng ldquoDistinguishing genuine from spuriouscauses a coherence hypothesisrdquo Cognitive Psychology vol 40pp 87ndash137 2000
[19] G A Miller R Beckwith C Fellbaum D Gross and K JMiller ldquoIntroduction to WordNet an on-line lexical databaserdquoInternational Journal of Lexicography vol 3 no 4 pp 235ndash3121990
[20] A W Koornneef and J J A Van Berkum ldquoOn the use of verb-based implicit causality in sentence comprehension evidencefrom self-paced reading and eye trackingrdquo Journal of Memoryand Language vol 54 no 4 pp 445ndash465 2006
[21] C D Manning and H Schutze Foundations of StatisticalNatural Language Processing TheMIT Press Cambridge MassUSA 4th edition 2001
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
8 The Scientific World Journal
Table 8 Experimental results of CL-CIS
Experiment files Precision Recall 119865-measurereut2-000sgm (newid 1ndash1000) 0776 0685 0728reut2-001sgm (newid 1001ndash2000) 0815 0661 0730Both (reut2) (newid 1ndash2000) 0809 0683 0741BBC-News2000 0897 0805 0849
5 Concluding Remarks
In recent years detection and clarification of cause-effectrelationships among texts events or objects has been elevatedto a prominent research topic of natural and social sciencesover the human knowledge development history
This paper demonstrates a novel comprehensive causalityextraction system (CL-CIS) integrated with the means ofcategory-learning CL-CIS considers cause-effect relations inboth explicit and implicit forms and especially practices therelation between category and causality in computation
CL-CIS is inspired with cognitive philosophy in categoryand causality In causality extraction and induction tasks CL-CIS implements a simple recurrent neural network (SRNN)to simulate human associative memory which has the abilityto associate different types of inputs when processing infor-mation
In elaborately designed experimental tasks CL-CIS hasbeen examined in full with two text collections Reuters-21578and BBC-News2000 The experimental results have testifiedthe capability and performance of CL-CIS in constructionof cause-effect relations and also confirmed the expectationthat the means of causal and category learning will improvethe precision and coverage of causality induction in a notablemanner
Conflict of Interests
The authors declare that there is no conflict of interestsregarding the publication of this paper
Acknowledgment
This work is financially supported by the National NaturalScience Foundation of China (Grant no 61003126)
References
[1] httpenwikipediaorgwikiCausality[2] J Pearl Causality Models Reasoning and Inference Cambridge
University Press New York NY USA 2000[3] D-S Chang and K-S Choi ldquoIncremental cue phrase learning
and bootstrapping method for causality extraction using cuephrase and word pair probabilitiesrdquo Information Processing andManagement vol 42 no 3 pp 662ndash678 2006
[4] S Kim R H Bracewell and K M Wallace ldquoA frameworkfor automatic causality extraction using semantic similarityrdquoin Proceedings of the ASME International Design EngineeringTechnical Conferences and Computers and Information in Engi-neering Conference (IDETCCIE rsquo07) pp 831ndash840 Las VegasNev USA September 2007
[5] T Inui K Inui andYMatsumoto ldquoAcquiring causal knowledgefrom text using the connective markersrdquo Transactions of theInformation Processing Society of Japan vol 45 no 3 pp 919ndash932 2004
[6] D Marcu and A Echihabi ldquoAn unsupervised approach torecognizing discourse relationsrdquo in Proceedings of the 40thAnnual Meeting of the Association for Computational LinguisticsConference (ACL rsquo02) pp 368ndash375 University of PennsylvaniaPhiladelphia Pa USA 2002
[7] C Pechsiri and A Kawtraku ldquoMining causality from texts forquestion answering systemrdquo IEICE Transactions on Informationand Systems vol 90 no 10 pp 1523ndash1533 2007
[8] R Girju ldquoAutomatic detection of causal relations for questionansweringrdquo in Proceedings of the 41st Annual Meeting of theAssociation for Computational Linguistics 2003
[9] K Torisawa ldquoAutomatic extraction of commonsense inferencerules from corporardquo in Proceedings of the 9th Annual Meeting ofthe Association for Natural Language Processing 2003
[10] Y Guo N Hua and Z Shao ldquoCognitive causality detectionwith associative memory in textual eventsrdquo in Proceedings ofthe IEEE International Symposium on Information Engineeringand Electronic Commerce (IEEC rsquo09) pp 140ndash144 TernopilUkraine May 2009
[11] P Thwaites J Q Smith and E Riccomagno ldquoCausal analysiswith Chain Event GraphsrdquoArtificial Intelligence vol 174 no 12-13 pp 889ndash909 2010
[12] M R Waldmann and Y Hagmayer ldquoCategories and causalitythe neglected directionrdquo Cognitive Psychology vol 53 no 1 pp27ndash58 2006
[13] G L MurphyTheBig Book of Concepts MIT Press CambridgeMass USA 2002
[14] B Rehder ldquoA causal-model theory of conceptual representationand categorizationrdquo Journal of Experimental Psychology Learn-ing Memory and Cognition vol 29 no 6 pp 1141ndash1159 2003
[15] B Rehder ldquoCategorization as casual reasoningrdquo Cognitive Sci-ence vol 27 no 5 pp 709ndash748 2003
[16] B Rehder and R Hastie ldquoCausal knowledge and categoriesthe effects of causal beliefs on categorization induction andsimilarityrdquo Journal of Experimental Psychology General vol 130no 3 pp 323ndash360 2001
[17] B Rehder and R Hastie ldquoCategory coherence and category-based property inductionrdquo Cognition vol 91 no 2 pp 113ndash1532004
[18] Y Lien and PWCheng ldquoDistinguishing genuine from spuriouscauses a coherence hypothesisrdquo Cognitive Psychology vol 40pp 87ndash137 2000
[19] G A Miller R Beckwith C Fellbaum D Gross and K JMiller ldquoIntroduction to WordNet an on-line lexical databaserdquoInternational Journal of Lexicography vol 3 no 4 pp 235ndash3121990
[20] A W Koornneef and J J A Van Berkum ldquoOn the use of verb-based implicit causality in sentence comprehension evidencefrom self-paced reading and eye trackingrdquo Journal of Memoryand Language vol 54 no 4 pp 445ndash465 2006
[21] C D Manning and H Schutze Foundations of StatisticalNatural Language Processing TheMIT Press Cambridge MassUSA 4th edition 2001
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Submit your manuscripts athttpwwwhindawicom
Computer Games Technology
International Journal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Distributed Sensor Networks
International Journal of
Advances in
FuzzySystems
Hindawi Publishing Corporationhttpwwwhindawicom
Volume 2014
International Journal of
ReconfigurableComputing
Hindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Applied Computational Intelligence and Soft Computing
thinspAdvancesthinspinthinsp
Artificial Intelligence
HindawithinspPublishingthinspCorporationhttpwwwhindawicom Volumethinsp2014
Advances inSoftware EngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Electrical and Computer Engineering
Journal of
Journal of
Computer Networks and Communications
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporation
httpwwwhindawicom Volume 2014
Advances in
Multimedia
International Journal of
Biomedical Imaging
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
ArtificialNeural Systems
Advances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
RoboticsJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Computational Intelligence and Neuroscience
Industrial EngineeringJournal of
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014
The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014
Human-ComputerInteraction
Advances in
Computer EngineeringAdvances in
Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014