a theory of reading_ from eye fixations to comprehension - copy.pdf

Carnegie Mellon UniversityResearch Showcase @ CMU

Department of Psychology Dietrich College of Humanities and Social Sciences

January 1980

A theory of reading: From eye fixations tocomprehensionMarcel Adam JustCarnegie Mellon University, [email protected]

Patricia A. CarpenterCarnegie Mellon University

Follow this and additional works at: http://repository.cmu.edu/psychologyPart of the Artificial Intelligence and Robotics Commons, Cognition and Perception Commons,

Cognitive Neuroscience Commons, Cognitive Psychology Commons, Computational NeuroscienceCommons, Developmental Neuroscience Commons, Discourse and Text Linguistics Commons,First and Second Language Acquisition Commons, and the Semantics and Pragmatics Commons

This Article is brought to you for free and open access by the Dietrich College of Humanities and Social Sciences at Research Showcase @ CMU. It hasbeen accepted for inclusion in Department of Psychology by an authorized administrator of Research Showcase @ CMU. For more information, pleasecontact [email protected].

Published InPsychological Review, 329- 354.

Psychological ReviewV O L U M E 87 N U M B E R 4 J U L Y 1980

A Theory of Reading: From Eye Fixations to ComprehensionMarcel Adam Just and Patricia A. Carpenter

Carnegie-Mellon University

This article presents a model of reading comprehension that accounts for theallocation of eye fixations of college students reading scientific passages. Themodel deals with processing at the level of words, clauses, and text units.Readers make longer pauses at points where processing loads are greater.Greater loads occur while readers are accessing infrequent words, integratinginformation from important clauses, and making inferences at the ends of sen-tences. The model accounts forthe gaze duration on each word of text as a func-tion of the involvement of the various levels of processing. The model is em-bedded in a theoretical framework capable of accommodating the flexibilityof reading.

Although readers go through many of thesame processes as listeners, there is onestriking difference between reading andlistening comprehensiona reader cancontrol the rate of input. Unlike a listener, areader can skip over portions of the text, re-read sections, or pause on a particular word.A reader can take in information at a pacethat matches the internal comprehensionprocesses. By examining where a readerpauses, it is possible to learn about thecomprehension processes themselves. Us-ing this approach, a process model of read-ing comprehension is developed that ac-counts for the gaze durations of collegestudents reading scientific passages.

The research was supported in part by Grant G-79-0119 from the National Institute of Education andGrant MH-29617 from the National Institute of MentalHealth.

We thank Allen Newell and Robert Thibadeau fortheir very helpful discussions.

The order of authorship was decided by the toss of acoin.

Requests for reprints should be sent to Marcel AdamJust, Department of Psychology, Carnegie-MellonUniversity, Pittsburgh, Pennsylvania 15213.

The following display presents an excerptfrom the data to illustrate some characteris-tics of eye fixations that motivate the model.This display presents a protocol of a col-lege student reading the first two sentencesof a passage about the properties of fly-wheels. The reader averages about 200words per minute on the scientific texts. Inthis study, the reader was told to read a para-graph with understanding and then recall itscontent. Consecutive fixations on the sameword have been aggregated into units calledgazes. The gazes within each sentence havebeen sequentially numbered above thefixated word with the gaze durations (inmsec) indicated below the sequence number.

One important aspect of the protocol isthat almost every content word is fixated atleast once. There is a common misconcep-tion that readers do not fixate every word,but only some small proportion of the text,perhaps one out of every two or threewords. However, the data to be presentedin this article (and most of our other datacollected in reading experiments) show thatduring ordinary reading, almost all content

Copyright 1980 by the American Psychological Association, Inc. 0033-295X/80/8704-0329$00.75

329

330 MARCEL ADAM JUST AND PATRICIA A. CARPENTER

Eye fixations of a college student reading a scientific passage. Gazes within each sentence are sequentiallynumbered above the fixated words with the durations (in msec) indicated below the sequence number.

1 2 3 4 5 6 7 8 9 1 21566 267 400 83 267 617 767 450 450 400 616

Flywheels are one of the oldest mechanical devices known to man. Every internal-

3 5 4 6 7 8 9 1 0 1 1 1 2 1 3517 684 250 317 617 1116 367 467 483 450 383

combustion engine contains a small flywheel that converts the jerky motion of the pistons into the14 15 16 17 18 19 20 21

284 383 317 283 533 50 366 566smooth flow of energy that powers the drive shaft.

words are fixated. This applies not only toscientific text but also to narratives writtenfor adult readers. The current data are notnovel in this regard. The eye fixation studiesfrom the first part of the century point tothe same conclusion (Buswell, 1937, chap.4; Dearborn, 1906, chap. 4; Judd & Buswell,1922, chap. 2). When readers are given a textthat is appropriate for their age level, theyaverage 1.2 words per fixation. The wordsthat are not always fixated tend to be shortfunction words, such as the, of, and a. Thenumber of words per fixation is even lowerif the text is especially difficult or if thereader is poorly educated. Of course, thisis not the case when adults are given simpletexts, such as children's stories; under suchcircumstances, these same studies show anincrease to an average of two words perfixation. Similarly, readers skip more wordsif they are speed-reading or skimming(Taylor, 1962). These old results and thecurrent results are consistent with the reportof McConkie and Rayner (1975; Rayner,1978) that readers generally cannot deter-mine the meaning of a word that is in periph-eral vision. These results have importantimplications for the present model; sincemost words of a text are fixated, we cantry to account for the total duration of com-prehension in terms of the gaze durationon each word.

The protocol also shows that the gazeduration varies considerably from word toword. There is a misconception that in-dividual fixations are all about 250 msec induration. But this is not true; there is alarge variation in the duration of individualfixations as well as the total gaze durationon individual words. As the preceding dis-play shows, some gaze durations are verylong, such as the gaze on the word Fly-

wheels. The model proposes that gaze dura-tions reflect the time to execute comprehen-sion processes. In this case the longerfixations are attributed to longer processingcaused by the word's infrequency and itsthematic importance. Also, the fixations atthe end of each sentence tend to be long. Forexample, this reader had gaze durations of450 and 566 msec on each of the last wordsof the first two sentences. The sentence-terminal pauses will be shown to reflect anintegrative process that is evoked at theends of sentences.

The link between eye fixation data and thetheory rests on two assumptions. The first,called the immediacy assumption, is that areader tries to interpret each content wordof a text as it is encountered, even at theexpense of making guesses that sometimesturn out to be wrong. Interpretation refersto processing at several levels such as en-coding the word, choosing one meaning ofit, assigning it to its referent, and deter-mining its status in the sentence and in thediscourse. The immediacy assumptionposits that the interpretations at all levelsof processing are not deferred; they occuras soon as possible, a qualification that willbe clarified later.

The second assumption, the eye-mindassumption, is that the eye remains fixatedon a word as long as the word is beingprocessed. So the time it takes to processa newly fixated word is directly indicatedby the gaze duration. Of course, compre-hending that word often involves the use ofinformation from preceding parts of the text,without any backward fixations. So the con-cepts corresponding to two different wordsmay be compared to each other, for exam-ple, whereas only the more recently en-

THEORY OF READING 331

countered word is fixated. The eye-mindassumption can be contrasted with an alter-native view that data acquired from severalsuccessive eye fixations are internally buf-fered before being semantically processed(Bouma & deVoogd, 1974). This alternativeview was proposed to explain a reading taskin which the phrases of a text were suc-cessively presented in the same location.However, the situation was unusual in twoways. First, there were no eye movementsinvolved, so the normal reading processesmay not have been used. Second, and moretelling, readers could not perform a simplecomprehension test after seeing the text thisway. By contrast, several studies of morenatural situations support the eye-mind as-sumption that readers pause on words thatrequire more processing (Just & Carpenter,1978; Carpenter & Daneman, Note 1). Theeye-mind assumption posits that there is noappreciable lag between what is beingfixated and what is being processed. Thisassumption has also been explored in spa-tial problem-solving tasks and has beensupported in that domain as well as in read-ing (Just & Carpenter, 1976). The im-mediacy and eye-mind assumptions areused to interpret gaze duration data in thedevelopment of the reading model.

The article has four major sections. Thefirst briefly describes a theoretical frame-work for the processes and structures inreading. The second section describes thereading task and eye fixation results ac-counted for by the model. The third sec-tion presents the model itself, with subsec-tions describing each component process ofthe model. The fourth section discussessome implications of the theory for languagecomprehension and relates this theory ofreading to other approaches.

Theoretical FrameworkReading can be construed as the coor-

dinated execution of a number of process-ing stages such as word encoding, lexicalaccess, assigning semantic roles, and relat-ing the information in a given sentence toprevious sentences and previous knowl-edge. Some of the major stages of theproposed model are depicted schematicallyin Figure 1. The diagram depicts both pro-cesses and structures. The stages of readingin the left-hand column are shown in theirusual sequence of execution. The long-termmemory on the right-hand side is the store-house of knowledge, including the pro-cedural knowledge used in executing the

M

~-lGet Next Input:

Move Eyes

^Extract PhysicalFeatures

i

Encode Word andAccess Lexicon

Assign CaseRoles

Integrate withRepresentation

of Previous Textt

K\>

-

/I

WORKING MEMORY

physical featureswordsmeanings

clausestext unitsdomain of discourse

variable- binding memory

LONG TERM

Productions thatrepresent

orthographyphonology

semanticspragmaticsdiscourse stucturescheme of

domainepisodic knowledge

^ J

Figure 1. A schematic diagram of the major processes and structures in reading comprehension. (Solidlines denote data-flow paths, and dashed lines indicate canonical flow of control.)


stages on the left. The working memory inthe middle mediates the long-term memoryand the comprehension processes. Al-though it is easy to informally agree on thegeneral involvement of these processes inreading, it is more difficult to specify thecharacteristics of the processes, their inter-relations, and their effects on reading per-formance.

The nature of comprehension processesdepends on a larger issue, namely the ar-chitecture of the processing system in whichthey are embedded. Although the humanarchitecture is very far from being known,production systems have been suggested asa possible framework because they haveseveral properties that might plausibly beshared by the human system. Detailed dis-cussions of production systems as models ofthe human architecture are presented else-where (Anderson, 1976; Newell, 1973,1980). The following three major propertiesare of particular relevance here.

1. Structural and procedural knowledgeis stored in the form of condition-actionrules, such that a given stimulus conditionproduces a given action. The productions"fire" one after the other (serially), and itis this serial processing that consumes timein comprehension and other forms ofthought. In addition to the serial produc-tions, there are also fast, automatic pro-ductions that produce spreading activationamong associated concepts (Anderson,1976; Collins & Loftus, 1975). These auto-matic productions operate in parallel to theserial productions and in parallel to eachother (Newell, 1980). These productions arefast and automatic because they operateonly on constants; that is, they directly as-sociate an action with a particular condition(such as activating the concept dog on de-tecting cat). By contrast, serial productionsare slow because they operate on variablesas well as constants; they associate anaction with a class of conditions. A serialproduction can fire only after the particularcondition instance is bound to the variablespecified in the production. It may be thebinding of variables that consumes time andcapacity (Newell, 1980). This architecturalfeature of two kinds of productions permitsserial comprehension processes to operatein the foreground, whereas in the back-

ground, automatic productions activaterelevant semantic and episodic knowledge.

2. Productions operate on the symbols ina limited-capacity working memory. Thesymbols are the activated concepts that arethe inputs and outputs of productions. Itemsare inserted into working memory as a resultof being encoded from the text or beinginserted by a production. Retrieval fromlong-term memory occurs when a produc-tion fires and activates a concept, causingit to be inserted into working memory.Long-term memory is a collection of pro-ductions that are the repositories of bothprocedural and declarative knowledge. Inthe case of reading, this knowledge includesorthography, phonology, syntax, and seman-tics of the language, as well as schemasfor particular topics and discourse types(Schank & Abelson, 1977). A new knowl-edge structure is acquired in long-termmemory if a new production is createdto encode that structure (Newell, 1980).This occurs if the structure participatesin a large number of processing episodes.

One important property of workingmemory is that its capacity is limited, sothat information is sometimes lost. Oneway in which capacity can be exceeded(causing forgetting) is that the level of ac-tivation of an item may decay to some sub-threshold level through disuse over time(Collins & Loftus, 1975; Hitch, 1978; Reit-man, 1974). A second forgetting mech-anism allows for processes and structuresto displace each other, within some limits(Case, 1978). Heavy processing require-ments in a given task may decrease theamount of information that can be main-tained, perhaps by generating too manycompeting structures or by actively inhibit-ing the maintenance of preceding informa-tion. There is recent evidence to suggestthat working memory capacity (as opposedto passive memory span) is strongly cor-related with individual differences in read-ing comprehension performance, pre-sumably because readers with greater ca-pacity can integrate more elements of thetext at a given time (Daneman & Carpenter,in press).

3. Production systems have a mechanismfor adaptive sequencing of processes. Theitems in working memory at a given time


enable a given production to fire and insertnew items, which in turn enable anotherproduction, and so on. In this way, the inter-mediate results of the comprehension pro-cess that are placed in working memorycan influence or sequence subsequent pro-cessing. There is no need for a super-ordinate controlling program to sequencethe mental actions.

The self-sequencing nature of produc-tions is compatible with the model depictedin Figure 1. The composition of each stageis simply a collection of productions thatshare a common higher level goal. The pro-ductions within a stage have similar enablingconditions and produce actions that serveas conditions for other productions in thesame stage. The productions within a stageneed not be bound to each other in anyother way. Thus the ordering of stages witha production system is accomplished notby direct control transfer mechanisms butan indirect self-sequencing accomplishedby one production helping to create the con-ditions that enable the "next" productionto fire.

This architecture permits stages to be exe-cuted not only in canonical orders but alsoin noncanonical orders. There are occa-sions when some stages of reading seem tobe partially or entirely skipped; some stagesseem to be executed out of sequence, andsome "later" stages sometimes seem to beable to influence "earlier" stages (Levy, inpress). Stages can be executed earlier thannormal if their enabling conditions existearlier than normal. For example, if a con-text strongly primes a case role, then thecase assignment could precede the lexicalaccess of a word. Having read John poundedthe nail with a , a reader can assignthe last word to the instrumental case on thebasis of cues provided by the words poundand nail, before encoding hammer. This or-ganization can permit "context effects" incomprehension, where a strong precedingcontext shortens reading time on a givenword or clause. This might occur if a pro-cessing stage that is normally intermediatebetween two others is partially or entirelyeliminated. It could be eliminated if the pre-ceding stage plus the context providedsufficient enabling conditions for the laterstage. Analogously, a misleading context

could lengthen comprehension time by pro-viding elements that enable conflictingprocesses.

The production system organization canalso explain how "later" stages can in-fluence "earlier" stages, so that higher levelschemas can affect word encoding, forexample. If the productions of the normallylater stage are enabled earlier than usual,then their outputs can serve as inputs tothe normally earlier stage. The ordering ofstages does not have to be entirely reversedto obtain this top-down influence. It may besufficient for just a portion of the produc-tions of the "later" stage to fire in order toinfluence the "earlier" stage.

In this view of processing stages, severalstages can be executed cotemporaneously inthe sense that firings of productions of twoor more stages may be interleaved. Conse-quently, data and control can be transferredback and forth among different stages,somewhat similarly to computer programsorganized into coroutines. Coroutines aretwo or more subprograms that have equalstatus (i.e., there is no master-slave rela-tionship). When one coroutine obtains con-trol, it executes until it detects a conditionindicating it should relinquish control, andthen another coroutine executes, and so on.One interesting difference between co-routines and the production system modelis that coroutines generally transfer databetween each other only along specifiedpaths, used especially for this purpose. Bycontrast, productions "transfer" data byplacing it in the working memory, so thatall processes have access to it. In this sense,the working memory serves as a messagecenter, and communication among stages isby means of the items in working memory.This is distinct from one stage feeding itsoutput directly to another stage.

ResearchTexts

This section describes the texts that wereused in the reading research because theirproperties, both local and global, have alarge influence on the processing. The globalorganization of a narrative text has beenshown to influence how a reader recalls


Definition Setting Come Contiqueneg \ Def in i t ion /Selling Couae C

Figure 2. A schematic diagram of the major text-grammatical categories of information in the scientificparagraphs.

the text (Kintsch & van Dijk, 1978; Mandler& Johnson, 1977; Meyer, 1975; Rumelhart,1977b; Thorndyke, 1977). The experimentreported next shows that the organizationhas at least part of its effect when thetext is being read. Scientific texts wereselected from Newsweek and Time becausetheir content and style is typical of whatstudents read to learn about technicaltopics. The passages discussed a variety oftopics that were generally unfamiliar to thereaders in the study. When readers wereasked to rate their familiarity with the topicof each passage on a 5-point scale, the modalrating was at the "entirely unfamiliar" endof the scale. There were 15 passages,averaging 132 words each. Although thetexts are moderately well written, they areon the borderline between ' 'fairly difficult''and "difficult" on Flesch's (1951) reada-bility scale, with 17 words per sentenceand 1.6 syllables per word. The following isan example of one of the passages:Flywheels are one of the oldest mechanical devicesknown to man. Every internal-combustion engine con-tains a small flywheel that converts the jerky motionof the pistons into the smooth flow of energy thatpowers the drive shaft. The greater the mass of a fly-wheel and the faster it spins, the more energy canbe stored in it. But its maximum spinning speed islimited by the strength of the material it is made from.If it spins too fast for its mass, any flywheel willfly apart. One type of flywheel consists of round sand-wiches of fiberglass and rubber providing the maxi-mum possible storage of energy when the wheel isconfined in a small space as in an automobile. Anothertype, the "superflywheel," consists of a series of rim-less spokes. This flywheel stores the maximum energywhen space is unlimited.

The content of the passages was analyzedby segmenting the text into idea units andcategorizing these units by means of a sim-ple text grammar. First, all of the 15 pas-sages were segmented into text units calledsectors, producing 274 sectors. The averagesector length was seven words. Each sectorwas judged to be a single meaningful piece ofinformation, whether it consisted of a word,phrase, clause, or sentence. The general cri-teria for segmentation into sectors weresimilar to those used by Meyer and Mc-Conkie (1973), who related such text units torecall performance.

A simplified grammar was developed tocategorize the sectors of the texts. Thegrammar (shown schematically in Figure 2)classifies the text units into a structure thatis quasi-hierarchical. This abbreviatedgrammar captures most of the regularitiesin our short passages (see Vesonder, 1979,for a more complete grammar for longerscientific passages). The initial sentencesgenerally introduced a topica scientificdevelopment or event. The beginnings ofthe passage sometimes gave details of thetime, place, and people involved with thediscovery. Familiar concepts were simplynamed, whereas unusual concepts were ac-companied by an explicit definition. Themain topic itself was developed throughspecific examples or through subtopics thatwere then expanded with further descrip-tions, explanations, and concrete examples.Consequences, usually toward the end ofthe passage, stated the importance of theevent for other applications. Table 1 showshow each text unit or sector in the "Fly-wheel" passage was classified according tothese categories. Each of the 274 sectorswas assigned to one of the five levels of thegrammar by one of the authors. The levelsof the grammar were further confirmed bya pretest involving 16 subjects who ratedthe importance of each sector in its pas-sage on a 7-point scale. The mean im-portance ratings differed reliably amongthe five levels F(4, 270) = 40.04, p


Table 1A Classification of the "Flywheel" Passage Into Text-Grammatical Categories

Category Sector

Topic Flywheels are one of the oldest mechanical devicesTopic known to manExpansion Every internal-combustion engine contains a small flywheelExpansion that converts the jerky motion of the pistons into the smooth flow of energyExpansion that powers the drive shaftCause The greater the mass of a flywheel and the faster it spins,Consequence the more energy can be stored in it.Subtopic But its maximum spinning speed is limited by the strength of the materialSubtopic it is made from.Expansion If it spins too fast for its mass,Expansion any flywheel will fly apart.Definition One type of flywheel consists of round sandwiches of fiberglas and rubberExpansion providing the maximum possible storage of energyExpansion when the wheel is confined in a small spaceDetail as in an automobile.Definition Another type, the "superflywheel," consists of a series of rimless spokes.Expansion This flywheel stores the maximum energyDetail when space is unlimited.

reading will be demonstrated with the eyefixation data. The next section presents thedata collection and analysis procedures,followed by the model and results.

Method and Data AnalysisThe readers were 14 undergraduates who read 2

practice texts followed by the 15 scientific texts inrandom order. Although the readers were asked torecall each passage immediately after reading it,they also were told to read naturally without memoriz-ing. They were also asked not to reread the passageor parts of it. The texts were presented on a televisionmonitor using uppercase and lowercase letters anda conventional paragraph layout. To initiate the read-ing of a passage, the reader had to look at a fixationpoint (located where the first word of the paragraphwould later appear) and press a "ready" button. Ifthe reader's point of regard (as measured by the eyetracker) was within 1 of the fixation point, then exactly500 msec later the passage appeared in its entirety onthe screen. The passage appeared instantaneously (i.e.,within one video frame) and remained there until thereader signaled that he had finished reading by push-ing a response button.

The reader's pupil and corneal reflections weremonitored relatively unobtrusively by a televisioncamera that was 75 cm away. The monitoring sys-tem, manufactured by Applied Science Laboratories,computed the reader's point of regard (as opposedto eye or head position) every 16.7 msec. The accuracyof the tracker was verified before and after each pas-sage was read by having the reader look at a fixationpoint and determining whether the obtained point ofregard was within 1 of that point. This procedure

indicated that accuracy was maintained during thereading of 195 of the 210 experimental passages in theentire experiment; the data from the 15 inaccuratetrials were discarded.

Data reduction procedures converted the 60 ob-servations per sec into fixations and then into gazes oneach word. While the data were being acquired, anew "fixation" was scored as having occurred if thepoint of regard changed by more than 1 (the size of athree-letter syllable). The durations of blinks that werepreceded and followed by fixations on the same loca-tion were attributed to the reading time on that loca-tion. Another program aggregated consecutive fixa-tions on the same word into gazes and computed theduration of gaze on each of the 1,936 words in the 15passages. Fixations on interword spaces were at-tributed to the word on the right because the perceptualspan is centered to the right of the point of regard,at least for readers of left-to-right languages (McConkie& Rayner, 1976; Schiepers, 1980). The durations ofsaccades, blinks that occurred between words, regres-sions, and rereading were not included in the dataanalysis. Because of the instructions not to reread,these categories account for relatively little of thetotal reading time, approximately 12% in all. The meanduration of gaze on each word was computed byaveraging over readers; these 1,936 mean gaze dura-tions constitute the main dependent measure of interest.

The model presents a number of factors that in-fluence various reading processes; some factors havetheir effect on individual words and some on largerunits, such as clauses. The data were fit to the modelwith a multiple linear regression in which the in-dependent variables were the factors postulated to af-fect reading time and the dependent variable was themean gaze duration on each word. Since the modelalso applies at the level of clauses and phrases, a


second regression analysis was done at the phrase/clause level. The independent variables for the latteranalysis were the factors postulated to affect read-ing time at the clause level, and the independentvariable was the mean gaze duration on each of the274 sectors described previously.

The psychological interpretation of the independentvariables in the two regression analyses will bedescribed in detail in the sections that follow. Theequation for the analysis of the gaze duration on in-dividual words was

GW, = am Xim + fi,

where GW( is the gaze duration on a word /, am is theregression weight in msec for independent variableXm, Xlm are the independent variables that code thefollowing seven properties of word i:

(a) length, (b) the logarithm of its normative fre-quency, (c) whether the word occurs at the beginningof a line of text, (d) whether it is a novel word to thereader, (e) its case grammatical role (one of 11 pos-sibilities), (f) whether it is the last word in a sentence,(g) whether it is the last word in a paragraph.

The equation for the analysis of the gaze durationon individual sectors was

GSj = b0 + 5> Zln + 6,where GSj is the gaze duration on sector j, and bn isthe regression weight in msec for independent variableZn. The ZJn are the independent variables that code thefollowing eight properties of sector j:

(a) its text grammatical level, multiplied by the num-ber of content words; (b) length; (c) the sum of thelogarithms of the frequencies of its component words;(d) the number of line-initial words it contains; (e) thenumber of novel words it contains; (f) the sum of thecase role regression weights of its component words;(g) whether it is the last sector in a sentence; (h)whether it is the last sector in a paragraph.

ResultsThe mean gaze duration on each word

(239 msec) indicated reading rates that aretypical for texts of this difficulty. If the 239msec per word is incremented by 12% toallow for saccades, blinks, and occasionalrereading, the reading rate is 225 words permin. The standard deviation of the 239-msec gaze mean was 168 msec, indicatingconsiderable variability in gaze durationfrom word to word. The results of the re-gression analyses are shown in Table 2.The table is divided into three sections, cor-responding to the three major processingstages postulated by the model, encodingand lexical access, case role assignment,and interclause integration. The regressionweights shown in Table 2 for the word-by-word analysis (above the double line) are de-

rived from a regression equation involving17 independent variables (11 of which arethe case role indicator variables). The stand-ard error of estimate of this model was 88msec, and the ^?2 value was .72. The resultsof the interclause integration stage make useof both the word-by-word analysis and thesector-by-sector analysis. (The latter analy-sis will be explained in more detail in thesection on interclause integration). Sincethe gaze durations on successive words andphrases are time-series data, it is interestingto note that there was no reliable positiveserial correlation among the residuals in theword-by-word regression or the sector-by-sector regression.

The Reading ModelThe next five subsections describe the

major stages shown in Figure 1: get next in-put, encoding and lexical access, case roleassignment, interclause integration, andsentence wrap-up. Each subsection de-scribes the processes in that stage togetherwith the factors that affect the durationof those processes, and hence the gazedurations.

Get Next InputThis is the first stage of a cycle that finds

information, encodes it, and processes it.When the perceptual and semantic stageshave done all of the requisite processing ona particular word, the eye is directed toland in a new place where it continues torest until the requisite processing is done,and so forth. The specification of what con-stitutes "all of the requisite processing"is contained in a list of conditions that mustbe satisfied before the reader terminates thegaze on the current word and fixates thenext one.These conditions include a speci-fication of the goals of normal reading. Forinstance, one condition may be that a mean-ing of the word be accessed and anothercondition may be that a case role be as-signed. These conditions can also reflectmore specific reading goals. A reader who istrying to memorize a text may have as acondition that the word or phrase be trans-ferred to long-term memory. By setting theconditions appropriately, the reader can


adjust his processes to the situation at hand.When the goal conditions for processing aword are satisfied, the resulting action isget next input.

The command to get next input usuallyresults in a saccade to the next part of thetext, one or two words forward. The processthat selects the placement of the next for-ward fixation does not have to be very com-plex or intelligent. The choice of where toplace the next forward fixation appears todepend primarily on the length of the nextword or two to the right of the currentfixation (McConkie & Rayner, 1975). Thelength information, which is encoded para-foveally, is then used to program a right-ward saccade. However, if only the rightmargin is visible in the parafovea, then theeye is directed to the first word of the next

line, producing a return sweep. In this casethe information in peripheral vision is notadequate for accurate targeting. The returnsweep is typically too short; the eye oftenlands on the second word of the new line fora brief amount of time (50 or 75 msec) andthen makes a corrective saccade leftward tothe first word of the line (Bayle, 1942). Onoccasion, a comprehension stage may re-quire a review of previously read text toreencode it or process it to deeper levels.In those cases, the get next input stageresults in a regressive saccade to the rele-vant portion of the text.

The duration of the get next input stage isshort, consisting of the time for a neural sig-nal to be transmitted to the eye muscles.In monkeys, this takes about 30 msec(Robinson, 1972). This duration must not be

Table 2Application of the Regression Model to the Gaze Duration on Each Word (Above Double Line)and to Each Sector (Below Double Line)

Processing stage Factor

Encoding and lexical access no. of syllableslog frequencybeginning of linenovel word

Case role assignment agent (86)instrument (110)direct or indirect object (174)adverb/manner (35)place or time (64)possessive (genitive) (39)verb (368)state/adjective (451)rhetorical word (15)determiner (243)connective (351)

Interclause integration last word in sentencelast word in paragraph

Regression weight(msec)

52**53**30**

802**51**53**25*29231633**44**70**26**9

71**157**

Integration time per content wordfrom regression analysis ofdata aggregated into sectors

topic (22)definition/cause/consequence (23)subtopic (48)expansion (68)detail (113)

72*94*78*73*60'

Note. Frequency of occurrence of case roles is in parentheses.* t = p < 0.05; ** ( = p


confused with the typical 150- to 200-mseclatency of a saccade to a visual stimulusthat has spatial or temporal uncertainty(Westheimer, 1954). These latencies includestimulus detection, interpretation, andselection of the next fixation target. In nor-mal reading, there is very little uncertaintyabout direction of the next saccade (it isalmost always rightward for forward fixa-tions, except for the return sweeps), nor isthere much uncertainty about distance. Onthe average, the saccade distance may besimply the mean center-to-center distancebetween words, a distance that does notvary much, relative to the physically pos-sible variation in eye movements. Thus it isreasonable to suppose that the preprogram-ming time is very short here, consistingusually of a "go" signal and the time ittakes that signal to be translated into a motormovement, about 30 msec (Robinson, 1972).The actual movements, the saccades, con-stitute about 5%-10% of the total readingtime. Recent analyses suggest that the sac-cade itself may destroy the visual per-sistence of the information from the preced-ing fixation so that it does not mask the in-put from the new fixation (Breitmeyer,1980). Consequently, it is reasonable to as-sume that stimulus encoding can commencesoon after the eye arrives at a new location.

Word Encoding and Lexical AccessThe reading process involves encoding a

word into an internal semantic format. It isassumed that prior to this encoding, thetransduction from the printed word to thevisual features has already taken place, andthat the features have been deposited intothe working memory. Perceptual encodingproductions use the visual features as condi-tions; their action is to activate the repre-sentation of the word. Once the representa-tion of the word has been sufficientlyactivated, its corresponding concept is ac-cessed and inserted into working memory.The concept serves as a pointer to a morecomplete representation of the meaning,which consists of a small semantic networkrealized as a set of productions. The majornodes of the network are the possible mean-ings of the word, the semantic and syntactic

properties of the meanings, and informa-tion about the contexts in which they usuallyoccur (see Rieger, 1979, for a related pro-posal).The word meanings are representedas abstract predicates, defined by their rela-tions to other predicates.

The productions that encode a wordgenerally trigger on orthographically basedsubword units such as syllables (Mewhort& Beal, 1977; Spoehr & Smith, 1973; Taft,1979). However, there are times when al-ternative codes, including orthographic,phonological, and whole-word codes, areused (Baron, 1977; Kleiman, 1975; LaBerge& Samuels, 1974). Since the syllablelikeencoding is believed to be the dominantmode, the data were analyzed in terms of thenumber of syllables in each word. Encodingtime increased by 52 msec for each syllable,as shown in Table 2.

The mechanism underlying lexical accessis the activation of a word's meaning repre-sentation by various sources. There arethree ways that a concept's level of ac-tivation can be temporarily increased aboveits base level. One activation mechanismis perceptual encoding; the encoded repre-sentation of a word can activate its meaning.A second source is the parallel productionsthat produce spreading activation throughthe semantic and episodic knowledge baseof the reader. The third source is activa-tion by the serial productions that do themajor computations in all of the stages ofprocessing. When a concept has been ac-tivated above some threshold by one ormore of these sources, a pointer to its mean-ing is inserted into working memory. Theactivation level gradually decays to a sub-threshold level unless some process reac-tivates it. If the word soon reoccurs in thetext while the concept is still activated,lexical access will be facilitated because theactivation level will still be close to thresh-old. When the activation level does de-crease, it decreases to an asymptote slightlyhigher than the old base level. In this way,the system can learn from both local andlong-term word repetitions. Frequentlyused words will have a high base level ofactivation, and consequently will requirerelatively less additional activation to re-trieve them. Thus, frequent words should


take less time to access than infrequentwords (Morton, 1969). Similarly, thevarious possible interpretations of eachword will have different base activationlevels, such that the more common inter-pretations have higher base activationlevels. For example, although the word doeshas at least two very different meanings, the"third-person-singular verb" interpretationwould have a higher base activation becauseit is more common than the ' 'female deer''interpretation (Carpenter & Daneman, Note1). The more common interpretation wouldthen be accessed faster, since less additionalactivation would be required to bring theactivation level to threshold. This model oflexical access can account for word fre-quency effects, priming effects, and repeti-tion effects in reading.

The gaze duration showed both frequencyand repetition effects. Frequency was ana-lyzed by relating gaze duration to thelogarithm of the normative frequency ofeach word, based on the Kucera and Francis(1967) norms. It was expected that gazeduration would decrease with the logarithmof the word's frequency; that is, smalldifferences among infrequent words wouldbe as important as much larger differencesamong frequent words (Mitchell & Green,1978). For algebraic convenience, thenormative frequencies were increased byone (to eliminate the problem of taking thelogarithm of zero), and the logarithm wascomputed and then subtracted from 4.85,the logarithm of the frequency of the mostfrequent English word. The analysis in-dicated a clear relation between this mea-sure of frequency and gaze duration. Asshown in Table 2, gaze duration increasedby 53 msec for each log unit of decreasein word frequency. A moderately frequentword like water (with a frequency of 442)was accessed 140 msec faster than a wordthat did not appear in the norms.

At one extreme of the frequency dimen-sion are words that a reader has never en-countered before. In scientific passages, thenovel words tend to be technical terms. Toread these words, a reader cannot depend oncontacting some prior perceptual and se-mantic representation; neither exists. Thereader must construct some perceptual

representation (perhaps phonological as wellas orthographic), associate this with thesemantic and syntactic properties of theconcept that can be inferred from the pas-sage, and then possibly construct a lexicalentry. These processes seem to take a greatdeal more time than ordinary encoding andaccess processes. Two judges identifiedseven words in the texts (that had zero fre-quency) as probably entirely novel to thereaders. Novelty was coded as an indicatorvariable, and it was found that these wordstook an additional 802 msec on average toprocess, as shown in Table 2. However,there was considerable variability amongthe words; their gaze durations rangedfrom 913 msec (for staphyllococci) to 2,431msec (for thermoluminescence).

Once a word has been encoded and ac-cessed once, it should be easier to accessit when it occurs again. Other researchhas suggested that frequency and repetitionhave their primary effect on lexical accessrather than encoding (Dixon & Rothkopf,1979; Glanzer & Ehrenreich, 1979; Scar-borough, Cortese, & Scarborough, 1977),although the possibility of some small ef-fects on the encoding process does exist.According to the model, repetition effectsshould occur in reading because the firsttime a word meaning is accessed, it shouldtemporarily achieve a higher activation levelsimilar to the level of a more frequentword. This mechanism particularly predictsrepetition effects for infrequent words,whose activation levels are low to startwith, but not for the highly frequent wordsthat occur in natural text. Generally, repeti-tion effects are larger for low-frequencywords (Scarborough et al., 1977). "Lowfrequency" in the Scarborough study wasdefined as less than 28 occurrences per mil-lion, the boundary of 28 emerging from amedian split of the frequencies of theirstimuli. So the analysis of repetition ef-fects was limited to words with frequenciesof 25 occurrences per million or less. Therewere 346 such instances in the text; 251were initial occurrences and 95 were repeti-tions. The repetitions were words with thesame morphological stem, disregarding af-fixes. An analysis of covariance on this sub-set of the data examined the effects of repe-


titions covarying out the number of syl-lables. The adjusted mean gaze durationswere 49 msec longer on the initial appear-ance of these words than on the subsequentappearances, f(343) = 2.21, p < .03. Mostof this effect (43 msec) was obtained on thesecond appearance of a word. These resultsindicate that once an infrequent word ap-pears in a text, processing time on that wordis decreased on subsequent appearances.

Lexical access is complicated by the factthat some words have more than one mean-ing, so the appropriate interpretation mustbe selected, or at least guessed at. When apolysemous word is accessed, the wordrepresentation that is retrieved is a pointerto a semantic network that includes the mul-tiple representations. The interpretationthat is selected is the one with the highestactivation level, and several factors can af-fect the activation. First, some interpreta-tions start off with a higher activation level;for instance, the "third-person-singular"interpretation of does has a higher baseactivation level than the "deer" interpreta-tion. Second, the automatic productionsthat produce spreading activation can con-tribute selectively to the activation level ofone particular interpretation. The spreadingactivation can emanate from the precedingsemantic and syntactic context, from thereader's knowledge of the domain, and fromknowledge of the discourse style. Third, theoutput of other stages operating on the sameword may activate a particular interpreta-tion. For example, although hammer can beinterpreted as a noun or a verb, a sentencecontext that suggests an instrument to thecase role assignment stage (e.g., John hit thenail with a ) may help activate thenoun interpretation. Fourth, when a wordwith many highly related meanings occurs inan impoverished context, there may be nosingle interpretation with higher activationthan the others, and the superordinate con-cept may be the selected interpretation ofthe word. This probably occurs for wordsthat have many closely related interpreta-tions, such as get and take.

The selection of only one interpretationof each word, posited by the immediacy as-sumption, provides a measure of cognitiveeconomy. Selecting just one interpretation

allows the activation of the unselected inter-pretations to decay, preventing them fromactivating their associates. Thus, the con-textual effects would remain focused in theappropriate semantic domain. This permitsa limited-capacity working memory to copewith the information flow in a spreading ac-tivation environment that may activatemany interpretations and associations forany lexical item. This method of processingalso avoids the combinatorial explosionthat results from entertaining more thanone interpretation for several succes-sive words.

This aspect of the model is consistent withsome recent results on lexical access that in-dicate that although multiple meanings of aword are initially activated, only onemeaning remains activated after a few hun-dred milliseconds. In one experiment, thesubjects simultaneously listened to asentence and pronounced a visually pre-sented word. When an ambiguous word(rose) was presented auditorally in asyntactic context (e.g., They all rose), thespeed of pronouncing a simultaneous visualprobe related to either meaning (stood orflower) was faster than in a control condi-tion (Tanenhaus, Leiman, & Seidenberg,1979). In another experiment, the subjectslistened to a sentence and performed a lexi-cal decision task on visually presentedstimuli. When an ambiguous word (bug)was presented in a semantic context (Johnsaw several spiders, roaches, and bugs), thespeed of a simultaneous lexical decisionrelated to either meaning (insect or spy) wasfaster than a control (Swinney, 1979). Inboth studies, the facilitation of the inap-propriate meaning was obtained only withina few hundred milliseconds of the occur-rence of the ambiguous word. If the probewas delayed longer, the inappropriate inter-pretation was no faster than the control.These results suggest that both meanings areavailable when an ambiguous word is beingaccessed, but the inappropriate meaning islost from working memory after a short time.

As the interpretation of the text is con-structed, a corresponding representation ofthe extensive meaningthe things beingtalked aboutis also being built. If thereferents of the words in a passage cannot


be determined, the text will be more dif-ficult to understand. One example of thisproblem is highlighted in a passage fromBransford and Johnson (1973) concerning aprocedure that involved arranging "thingsinto groups. Of course, one may be suf-ficient depending on how much thereis. . . ." (p. 400). Subjects who were notgiven the title washing clothes though thestory was incomprehensible. The referentialrepresentation helps the reader disam-biguate referents, infer relations, and inte-grate the text.

The immediacy assumption posits that anattempt to relate each content word to itsreferent occurs as soon as possible. Some-times this can be done when the word is firstfixated, but sometimes more information isrequired. For example, although the se-mantic interpretation of a relative adjectivelike large can be computed immediately,the extensive meaning depends on the wordit modifies (e.g., large insect vs. large build-ing). The referent of the entire noun phrasecan be computed only after both wordsare processed. The immediacy assumptiondoes not state that the relating is done im-mediately on each content word, but ratherthat it occurs as soon as possible. This is animportant distinction that will be made againin the discussion on integrative processes.

Assigning Case RolesComprehension involves determining

the relations among words, the relationsamong clauses, and the relations amongwhole units of text. This section describesthe first of these processes, that of deter-mining the relations among the words in aclause (or in Schank's, 1972, terms, deter-mining the dependencies among the con-cepts). These relations can be categorizedinto semantic cases, such as agent, recipi-ent, location, time, manner, instrument,action, or state (Chafe, 1970; Fillmore,1968). The case role assignment processusually takes as input a representation of thefixated word, including information aboutits possible case roles and syntactic proper-ties. For example, hammers tend to be instru-ments rather than locations or recipients,and information about a word's usual case

role can be an important contributor to theassignment process. But this normative in-formation generally is not sufficient to as-sign its case role in a particular clause. Con-sequently, the assignment process relies onheuristics that use the word meaning to-gether with information about the priorsemantic and syntactic context, as well aslanguage-based inferences. The output ofthe process is a representation of the word'ssemantic role with respect to the otherconstituents in its clause.

Just as certain meanings suggest par-ticular case roles, so, too, can the contextprime a particular case role. Consider thesentence John was interrogated by the

The semantic and syntactic cues sug-gest that the missing word will be an agent,such as detective. The strength of the con-text becomes evident if the primed casedoes not occur, for example, John wasinterrogated by the window. The priorsemantic context can precede the affectedcase assignment by more than a few words.In the sentences The lawyer wanted to knowwhere in the room John had been interro-gated and Mary told him that John wasinterrogated by the window, the thematicfocus of the first sentence on a locationalters the interpretation of by and facilitatesa locative case role assignment for window.

The specific heuristics that are used incase role assignment have received someattention (see Clark & Clark, 1977, for someexamples). Many proposals contain thesuggestion that readers use the verb as apivotal source of information to establishthe necessary and possible case roles andthen fit the noun phrases into those slots(Schank, 1972). But the immediacy assump-tion posits that the case role assignmentfor an item preceding the verb is not post-poned in anticipation of the verb. Similar tothe lexical access stage, the case assign-ment stage makes a best guess about aword's case when the word is fixated, ratherthan making the decision contingent on sub-sequent words. So, the model would not ac-cord any special status to verbs. Anothersuggested heuristic (that children appearto use) is to assign a sequence consist-ing of animate noun-verb-noun to thecase roles of agent-action-object (Bever,


1970). Like all heuristics, this one some-times fails, so young children sometimesmisinterpret passive sentences (Eraser,Bellugi, & Brown, 1963). This heuristic maybe employed by adults, but in a modifiedversion that conforms to the immediacy as-sumption. Rather than waiting for the threemajor constituents before assigning caseroles, the reader should assign an animatenoun to the agent role as soon as it is en-countered, in the absence of contraryprior context.

The immediate assignment of a case roleimplies that readers will sometimes makeerrors and have to revise previous deci-sions. For example, an adult who assigns therole of agent to an animate noun and thenencounters a passive verb will have to revisethe agent assignment. (Presumably, youngchildren do not make this revision.) The im-mediacy of the case assignment process isevident in the reading of sentences suchas Mary loves Jonathan. . . . The im-mediacy assumption suggests that a readerwould assign to Jonathan the role ofrecipient; this would in turn result in an in-correct assignment if the sentence con-tinued Mary loves Jonathan apples.

Because case roles are assigned withinclauses, the assignment process must in-clude a segmentation procedure to deter-mine clause boundaries within sentences.Sentences can sometimes be segmented intoclauses on the basis of explicit markers,such as a subordinating conjunction (e.g.,because, when). More often, the reader can-not tell with certainty where one clause endsand another starts until beyond the clauseboundary (or potential boundary). A generalstrategy for dealing with such cases hasbeen suggested, namely to assign a word tothe clause being processed, if possible(Frazier & Fodor, 1978). For example, theword soil in the sentence When farmers areplowing the soil . . . can continue theinitial clause (When farmers are plowing thesoil, it is most fertile) or start a new one(When farmers are plowing the soil ismost fertile). The suggested strategy is tocontinue the initial clause until contrary in-formation is encountered. Interestingly, thestrategy discussed by Frazier and Fodor(1978) presupposes the immediacy assump-tion; the segmentation decision arises be-

cause case roles are assigned as soon asthe words are encountered.

There is no direct mapping between par-ticular case roles and the duration of theassignment process. For example, there isno a priori reason to expect that assignmentof instruments takes more or less time thanlocations. The time for a particular assign-ment might depend more on the context andproperties of the word than on the particularcase role being assigned. Detailed specifica-tion of the process is not within the scope ofthis article; it probably requires a large-scalesimulation model to examine the complexinteractions of different levels of process-ing. Nevertheless, we examined whether,all things being equal, different case roleassignments tend to take different amountsof time.

The analysis included the usual case rolesjust noted (Fillmore, 1968), as well as othercategories such as determiners and adjec-tives that are not cases but still play a part inthe parsing and assignment process. Eachword was classified into 1 of 11 categories:verb, agent, instrument, indirect or directobject, location or time, adverb, adjective orstate, connective (preposition or conjunc-tion), possessive, determiner, and rhetoricalword (such as well). Some cases werepooled (such as location and time) becausethey were relatively infrequent in thetext and because they have some concep-tual similarity. The case roles were codedas indicator variables and were all enteredinto the regression with the intercept forcedto zero.

The results of the case role assignmentanalysis, shown in Table 2, indicate thatthere are some variations among the cases.As expected, verbs did not take particularlylong (33 msec), and in fact, although thetime was significantly different from 0, itwas not greater than the agent or instrumentcases (51 msec and 53 msec, respectively).Four cases had parameters that were notsignificantly different from 0, connectives (9msec), adverb/manner (29 msec), place ortime (23 msec), and possessives (16 msec).These parameters could reflect someproperties of particular word classes, inaddition to parsing and case role assign-ment processes. For example, if a connec-tive (e.g., and or but) simply takes less


time to access than other words, the ad-vantage should appear in this parameter.However, the parameters are not due solelyto length or frequency, since these variablesmake a separate contribution to the regres-sion equation. Although this analysis doesnot examine any of the contextual ef-fects thought to be of some importance inthe case assignment process, it does indicateroughly the relative amount of time spentassigning various categories of words totheir case roles in a clause. Later theorieswill have to account for the precise pat-tern of case assignment durations in termsof specific operations that use prior con-text and word meanings to assign thevarious cases.

Interclause IntegrationClauses and sentences must be related to

each other by the reader to capture the co-herence in the text. As each new clauseor sentence is encountered, it must be in-tegrated with the previous information ac-quired from the text or with the knowledgeretrieved from the reader's long-term mem-ory. Integrating the new sentence with theold information consists of representing therelations between the new and the oldstructures.

Several search strategies may be used tolocate old information that is related to thenew information. One strategy is to check ifthe new information is related to the otherinformation that is already in working mem-ory either because it has been repeatedlyreferred to or because it is recent (Car-penter & Just, 1977a; Kintsch & van Dijk,1978). Using this strategy implies that ad-jacency between clauses and sentences willcause a search for a possible relation. Forinstance, the adjacent sentences Mary hurtherself 'and John laughed seem related (Johnmust be a cad) even though there is no ex-plicit mention of the relation. This strategyalso entails trying to relate new informa-tion to a topic that is active in workingmemory. This is a good strategy, sinceinformation in a passage should be relatedto the topic.

A second strategy is to search for specificconnections based on cues in the new sen-tence itself. Sentences often contain old in-

formation as well as new. Sometimes the oldinformation is explicitly marked (as in cleftconstructions and relative clauses), butoften it is simply some argument repeatedfrom the prior text. The reader can use thisold information to search his or her long-term text representation and referentialrepresentation for potential points of at-tachment between the new information andthe old (Haviland & Clark, 1974). Thissecond strategy may take more time thanthe first. In fact, it takes longer to reada sentence that refers to information intro-duced several sentences earlier than onethat refers to recently introduced informa-tion (Carpenter & Just, 1977a).

There are two main points at which inte-gration can occur. First, as each ensuingword of the text is encountered, there isan attempt to relate it to previous informa-tion (Just & Carpenter, 1978). Second, arunning representation of the clause ismaintained, with an updating as each wordof the clause is read. This running clauserepresentation consists of the configurationof clause elements arranged according totheir case relations. This second type ofintegration involves an attempt to relatethe running clause representation to pre-vious information at each update. Integra-tion occurs whenever a linking relation canbe computed. Consider the sentence Al-though he spoke softly, yesterday's speakercould hear the little boy's question. Thepoint of this example is not so much that theinitial integration of he and speaker is incor-rect, but that the integration is attemptedat the earliest opportunity. This model im-plies that integration time may be dis-tributed over fixations on different parts ofa clause. Moreover, the duration of theprocess may depend on the number of con-cepts in the clause; as these increase, thenumber of potential points of contact be-tween the new clause and previous informa-tion will increase. There is also evidencefor integration triggered by the end of thesentence; this process is discussed next inmore detail.

Integration results in the creation of a newstructure. The symbol representing thatstructure is a pointer to the integrated con-cepts, and this superordinate symbol is thenavailable for further processing. In this


way, integration can chunk the incomingtext and allows a limited working memoryto deal with large segments of prose. Themacrorules proposed by Kintsch and vanDijk (1978) can be construed as productionsthat integrate.

Integration can also lead to forgetting inworking memory. As each new chunk isformed, there is a possibility that it will dis-place some previous information fromworking memory. Particularly vulnerableare items that are only marginally activated,usually because they were processed muchearlier and have not recently participatedin a production. For instance, the repre-sentation of a clause will decay if it wasprocessed early in a text and was not re-lated to subsequent information. This mech-anism can also clear working memory of"lower level" representations that are nolonger necessary. For example, the ver-batim representation of a previously readsentence may be displaced by the processesthat integrated the sentence with other in-formation (Jarvella, 1971). By contrast, thesemantic elements that participate in an in-tegration production obtain an increasedactivation level. This increases the proba-bility that they will become a permanentpart of long-term memory.

The main types of interclause relationsin the scientific passages correspond tothe text-grammatical categories describedpreviously, such as definitions, causes, con-sequences, examples, and so forth. Textroles that are usually more important to thetext and to the reader's goals, such astopics or definitions, are integrated dif-ferently than less important units, such asdetails. The more central units will initiatemore retrievals of relevant previous knowl-edge of the domain (schematic knowledge)and retrievals of information acquired fromthe text but no longer resident in the work-ing memory. In addition, more relationswill be computed between the semanticallycentral propositions and previous informa-tion because centrality inherently entailsrelations with many other units. By con-trast, details are often less important to thereader's goals and to the text. Moreover,when a detail is to be integrated, the processis simpler because details are often con-

crete instantiations of an immediatelypreceding statement (at least in thesescientific texts), so they can be quickly ap-pended to information still present in theworking memory. Thus, higher level unitswill take more time to integrate becausetheir integration is usually essential tothe reader's goals, and because integrationof higher units involves more relations to becomputed and more retrievals to be made.

The nature of the link relating two struc-tures may be explicitly denoted either in thetext (with connectives like because, there-fore, and for instance) or it may have to beinferred on the basis of schematic knowl-edge of the domain. For example, the causalrelation between the sentences Cynthia felloff the rocking horse and She cried bittertears is inferred from the reader's knowl-edge about the temporal and causal relationbetween falling and hurting oneself (Char-niak, Note 2).

The model predicts that the gaze durationon a sector depends on its text-grammaticalrole and on the number of concepts it con-tains. Because integration can occur atmany points in a sector, the gaze durationassociated with integration cannot belocalized to a particular word. Thus, to dothe clause level of analysis, the gaze dura-tions on the individual words of a sectorwere cumulated, producing a total of 274sector gaze durations as the dependentvariable. The independent variables werethe aggregates of the word-level variables,except for case roles. The independentvariable that coded the case-role effect fora sector was the sum of the case-coefficients(obtained from the word-by-word regres-sion analysis) for each of the words in thesector. A new independent variable codedthe text-grammatical role of a sector and itsnumber of content words; it was the inter-action of the indicator variables that repre-sented the five text-grammatical levels andthe number of content words in the sector,with content words defined as in Hock-ett (1958).

The results indicate that the integrationtime for a given sector depends on its text-grammatical role. The portion of Table 2below the double line shows the integrationtime per content word for each type of sec-


tor. Generally, more important or centralsectors take longer to integrate. The modeldescribes this effect in terms of the inte-grative processes initiated by the seman-tics of the different types of information andtheir relevance to the reader's goals. Ananalysis of covariance examined the effectof text roles covarying out the number ofsyllables. The adjusted mean gaze durationsdiffered reliably, F(4, 268) = 8.82,p < .01;paired comparisons indicated that detailstook significantly less time than all otherroles, and expansions took significantly lesstime than topics and definitions/causes/consequences (all ps < .01). These resultsquantitatively and qualitatively replicatethose reported previously for a slightly dif-ferent paradigm (Carpenter & Just, inpress). The previously obtained coefficientsfor the five text-grammatical categorieswere 65, 106, 81, 76, and 47 msec per con-tent word, respectively, corresponding tothe newly obtained 72, 94, 78, 73, and 60.The model accounts very well for the sector-level data. The R2 value was .94, and thestandard error of estimate was 234. Themean gaze duration on a sector was 1,690msec, with a standard deviation of 902msec,and the mean sector length was 4.9 words.1

One cost of immediate interpretation,case role assignment, and integration is thatsome decisions will prove to be incorrect.There must be mechanisms to detect andrecover from such errors. The detection ofa misinterpretation often occurs whennew information to be integrated is incon-sistent with previous information. Thus,misinterpretation detection may be con-strued as inconsistency detection. Forexample, the sentence There were tears inher brown dress causes errors initially be-cause the most frequent interpretation oftears is not the appropriate one here, and theinitial interpretation is incompatible withdress. The eye fixations of subjects readingsuch garden path sentences clearly indicatethat readers do detect inconsistencies,typically at the point at which the incon-sistency is first evident (Carpenter & Dane-man, Note 1). At that point, they use a num-ber of error-recovery heuristics that enablethem to reinterpret the text. They do notstart reinterpreting the sentence from

its beginning. The heuristics point them tothe locus of the probable error. Readersstart the backtracking with the word thatfirst reveals the inconsistency, in this case,dress. If that word cannot be reinterpreted,they make regressions to the site of otherwords that were initially difficult to inter-pret, such as ambiguous words on which abest guess about word meaning had to bemade. The ability to return directly to thelocus of the misinterpretation and to recoverfrom an error makes the immediacy strategyfeasible.

Sentence Wrap-UpA special computational episode occurs

when a reader reaches the end of a sentence.This episode, called sentence wrap-up, isnot a stage of processing defined by its func-tion, but rather by virtue of being executedwhen the reader reaches the end of a sen-tence. The processes that occur duringsentence wrap-up involve a search for ref-erents that have not been assigned, the con-struction of interclause relations (with theaid of inferences, if necessary), and an at-tempt to handle any inconsistencies thatcould not be resolved within the sentence.

The ends of sentences have two importantproperties that make them especially goodplaces for integration. First, within-sen-tence ambiguities are usually clarified by theend of the sentence. For example, if a sen-tence introduces a new object or personwhose identity cannot be inferred from thepreceding context, some cue to theiridentity is generally given by the end of thesentence. For that reason, if readers can-

1 It might be argued that the variables coding the

text-grammatical roles ought to be independent of thenumber of content words. One might argue that adefinition, for example, takes a fixed amount of timeto integrate, regardless of the number of content wordsit contains. Although the model predicts a length-sensitive duration, the analysis can also be done withfive simple indicator variables to encode the fivelevels of the grammar. This analysis produced a fitthat was almost as good (R* = .93). The weights(assuming a zero intercept) were 250, 341, 257, 214,and 118 msec for the five categories, from topics todetails. Although this alternative is not ruled out bythe data, we will continue to retain the view thatintegration time depends on the number of contentwords involved.


not immediately determine the referent of aparticular word, then they can expect to betold the referent or given enough informa-tion to infer it by the end of the sentence.Indeed, readers do use the ends of sentencesto process inconsistencies that they cannotresolve within the sentence (Carpenter &Daneman, Note 1). The second property isthat the end of a sentence unambiguouslysignals the end of one thought and the be-ginning of a new one. It can be contrastedwith weaker cues that signal within-sen-tence clause boundaries such as commas,relative pronouns, and conjunctions thatcan signal other things besides the end ofa clause. Since ends of sentences are unam-biguous, they have the same role acrosssentences, and they may be processed moreuniformly than the cues to within-sentenceclause boundaries.

There is ample empirical support for theintegrative processing at the ends of sen-tences. Previous eye fixation studies showthat when a lexically based inference mustbe made to relate a new sentence to someprevious portion of the text, there is astrong tendency to pause at the lexical itemin question and at the end of sentence thatcontains it (Just & Carpenter, 1978). Read-res were given paragraphs containing pairs ofrelated sentences; the first noun in thesecond sentence was the agent or instru-ment of the verb in the first sentence:(la) It was dark and stormy the night the millionaire

was murdered.(Ib) The killer left no clues for the police to trace,In another condition, the integrating in-ference was less direct:(2a) It was dark and stormy the night the millionaire

died.(2b) The killer left no clues for the police to trace.It took about 500 msec longer to processSentence 2b than Ib, presumably due to themore difficult inference linking killer todie. There were two main places in whichthe readers paused for those 500 msec, in-dicating the points at which the inferencewas being computed. One point was on theword killer, and the other was on the end ofthe sentence containing killer. Another eyefixation study showed that integration link-ing a pronoun to its antecedent can occur

either when the pronoun is first encounteredor at the end of the sentence containing thepronoun (Carpenter & Just, 1977b).

Reading-time studies also have shownthat there is extra processing at the end ofa sentence. When subjects self-pace theword-by-word or phrase-by-phrase pre-sentation of a text, they tend to pauselonger at the word or phrase that terminatesa sentence (Aaronson & Scarborough, 1976;Mitchell & Green, 1978). The pause hasbeen attributed to contextual integrationprocesses, similar to the proposed inter-clause integration process here. Yet anothersource of evidence for sentence wrap-upprocesses is that verbatim memory for re-cently comprehended text declines after asentence boundary (Jarvella, 1971; Perfetti& Lesgold, 1977). The model attributes thedecline to the interference between sen-tence wrap-up processes and the main-tenance of verbatim information in workingmemory. Finally, another reason to expectsentence wrap-up processes is that we haveobserved pauses at sentence terminations inan eye fixation study similar to the one re-ported here (Carpenter & Just, in press).However, the current study provides strongerevidence because the text was presentedall at once.

The results indicate that readers did pauselonger on the last word in a sentence. AsTable 2 shows, the duration of the sentencewrap-up period is 71 msec.

It is possible that wrap-up episodes couldoccur at the ends of text units smaller orlarger than a sentence. For example, thedata of Aaronson and Scarborough (1976)suggest that there are sometimes wrap-upprocesses at the ends of clauses. It is alsopossible that wrap-up could occur undersome circumstances at the ends of para-graphs. The decision of when and if to do awrap-up may be controlled by the desireddepth of processing. For example, skim-ming may require wrap-up only at paragraphterminations, whereas understanding a legalcontract may require wrap-up at clauseboundaries. In fact, the clause-boundary ef-fects obtained by Aaronson and Scar-borough are sensitive to the subjects' read-ing goals. The current analysis indicatedthat the final word in the paragraph might


also be a wrap-up point; it received an ad-ditional 157 msec of fixation. However,since readers also pressed a button to in-dicate that they had finished reading thepassage, this parameter might be influencedby their motor response.

Finally, the model included one other fac-tor that involves a physical property ofreading, namely the return sweep of the eyesfrom the right-hand side of one line of textto the left-hand side of the next line. Returnsweeps are often inaccurate, landing to theright of the first word in a line. The inac-curacy is often corrected by a leftwardsaccade to the first word. As a result ofthis error and recovery, the first word on aline eventually receives an increased gazeduration, relative to a line-medial word. Al-most all readers we have studied displaythe undershoot, but there are considerableindividual differences in whether they com-pensate for it by making an extra leftwardfixation to the first word. In fact, some re-searchers have associated these correctiveleftward movements with poor readers(Bayle, 1942). To test for increased gaze

durations on line-initial words, an indicatorvariable coded whether a word was the firstone on aline. As Table 2 shows, these wordsreceived an additional 30 msec of fixation.

Fit of the ModelTo see how well the model accounts

for the data, one can informally comparehow closely the estimated gaze durationsmatch the observed gaze durations. The dis-play that follows shows the estimated (initalics) and observed (in msec) gaze dura-tions for two sentences from the "Fly-wheel" passage. The estimated durationscan be computed by an appropriate com-bination of the weights given in Table 2.These estimates take into account theprocesses of encoding, lexical access, case-role assignment, sentence wrap-up, and thebeginning of the line effect; they do notinclude integration time for text roles, sincethere is no way to distribute this time on aword-by-word basis. In spite of this, thematch is satisfactory, and as mentionedearlier, the standard error of estimate was88 msec overall.

Observed mean gaze durations (msec) on each word of a text sample and estimates (italicized, from theword-by-word regression analysis.

169 215 165 295 290 73 196 504 29 482 0 328 431 51165 236 75 409 304 75 249 438 75 413 80 338 349 78

. . . One type of flywheel consists of round sandwiches of fiberglas and rubber providing the

369 326 308 22 272 253 128 199 69 336 32 41 267 197 70 164 195354 318 297 75 378 138 77 239 128 326 87 102 206 209 112 87 127

maximum possible storage of energy when the wheel is confined in a small space as in an

340 323 182 72 626 276 46 21 346 60 467 519465 334 236 77 513 304 75 102 289 75 361 319

automobile. Another type, the "superflywheel," consists of a series of rimless spokes . . .

Table 3 presents an analogous compari-son from the sector-by-sector analysis; thisincludes integration time. Again, the esti-mates from the model match the observeddata quite well. The standard error ofestimate was 234 msec overall.

Another way to evaluate the goodness offit is to compare the regression results tothose of another model that lacks most ofthe theoretically interesting independentvariables and contains only the variablethat codes the number of syllables. For theword-by-word analysis, this rudimentary

model produces an R2 of .46, compared to.72 for the complete model. For the sector-by-sector analysis, the rudimentary modelaccounts for a large portion of the variancebetween the gaze durations on sectors (R2 =.87). This is not surprising, since there isconsiderable variation in their lengths. Thecomplete sector-by-sector model accountsfor 94% of the variance, or 54% of thevariance unaccounted for by the re-duced model.

The regression equations were also fit to


Table 3Observed and Estimated Gaze Durations (msec) on Each Sector of the ' 'Flywheel'' Passage,According to the Sector-By-Sector Regression Analysis of the Group Data

Sector Observed Estimated

Flywheels are one of the oldest mechanical devices 1,921 1,999known to man. 478 680Every internal-combustion engine contains a small flywheel 2,316 2,398that converts the jerky motion of the pistons into the smooth flow of energy 2,477 2,807that powers the drive shaft. 1,056 1,264The greater the mass of a flywheel and the faster it spins, 2,143 2,304the more energy can be stored in it. 1,270 1,536But its maximum spinning speed is limited by the strength of the material 2,440 2,553it is made from. 615 780If it spins too fast for its mass, 1,414 1,502any flywheel will fly apart. 1,200 1,304One type of flywheel consists of round sandwiches of fiberglas and rubber 2,746 3,064providing the maximum possible storage of energy 1,799 1,870when the wheel is confined in a small space 1,522 1,448as in an automobile. 769 718Another type, the "superflywheel," consists of a series of rimless spokes. 2,938 2,830This flywheel stores the maximum energy 1,416 1,596when space is unlimited. 1,289 1,252

the gaze durations of each of the 14 readersindividually. The subjects varied in theirreading skill, with self-reported ScholasticAptitude Test scores ranging from 410 to660, which were correlated with their read-ing speeds in the experiment, ranging from186 words per min. to 377 words per min.r(12) = .54, p < .05. The mean R2 of the14 readers was .36 on the word-by-wordanalysis and .75 on the sector-by-sectoranalysis. This indicates substantial noise ineach reader's word-by-word data. Some ofthe regression weights of the readers in-dicated considerable individual differenceswith respect to certain processes. Forexample, 4 of the 14 readers spent no extratime on the last word of a sentence. Anotherparameter of great variability among readerswas the extra time spent on novel words,which ranged from 94 msec to 1,490 msec.

Although the sector-by-sector regressionanalysis uses an independent variable (thesum of the case role coefficients) that isestimated from the same data, this pro-cedure does not do violence to the results.To estimate the effect of this procedure, the14 subjects were divided randomly into twosubgroups, and the case-role coefficientswere obtained for each subgroup in a word-by-word analysis. Then these coefficientswere aggregated and used as independent

variables in a sector-by-sector analysis,such that one subgroup's coefficients wereused in the analysis of the other subgroup'ssector gaze durations. The results indicatedno difference of any importance betweenthe two subanalyses, and generally con-firmed that using the case role coefficientsfrom the word analysis in the sector analysiswas an acceptable procedure.

Some of the variables that were reliablein the word-by-word analysis were notreliable in the sector analysis. For example,sectors that included a line-initial word didnot have reliably longer durations, and sec-tors that included the end of a sentencetook 57 msec longer, but the reliability ofthe effect was marginal (p < .08). The sumof the logarithms of the frequencies of thewords in a sector did not reliably affectgaze duration on the sector. These differ-ences between the two levels of analysisindicate that some effects that are wordspecific are not reliable or large enough tobe detected when the data are aggregatedover groups of words. Nevertheless, someof these effects can be detected at the sec-tor level if the appropriate analysis is done.For example, the reason that the frequencyeffect was not reliable is that the aggrega-tion of the logarithms smooths over thedifferences between infrequent words and


frequent words. A regression analysis ofthe sector data shows a reliable word fre-quency effect if the independent variableencodes the number of infrequent words (ar-bitrarily defined as less than 25 in Kucera& Francis, 1967) occurring for the firsttime. This latter analysis indicates 82-msecextra spent for each infrequent word, andhas an R2 of .94. (Carpenter & Just, in press,reported a 51-msec effect for this variable).

Recall PerformanceThe recall of a given part of a text should

depend in part on what happens to the infor-mation as it is read. A clause that is thoroughlyintegrated with the representation of the textshould tend to be stored in long-term mem-ory, and therefore should be recalled better.There are two factors that determine howwell a clause will be integrated. First, thosesectors on which more integration time hasbeen spent, like topics and definitions,should be recalled better. As predicted theintegration parameter for a text role (i.e.,the five weights at the bottom of Table 2)reliably affected the probability that a sec-tor would be recalled, ?(271) = 2.01, p

a theory of reading_ from eye fixations to comprehension - copy.pdf

Documents

model of reading comprehension

theory of reading

cognitive psychology

robotics commons

perception commons

text linguistics commons

cognitive neuroscience

language acquisition