Sentiment analysis in the lexicon and in thediscourse

Josef Ruppenhofer

July 1, 2009

Overview

I What do you mean, Sentiment?

I Sentiment denotation v. invited inference

I Attribution

I Variety

I Context dependence

I (Non-)Compositionality

I Automatic processing of sentiment expressions

The core cases

I Expressions of judgment/evaluation/emotionI Wow, this is my 4th Olympus camera.I Staley declared it to be one hell of a collection.

Why do sentiment analysis?

I Fact-based question answeringI Q: When is the first day of spring in 2007?I Q: Does the us have a tax treaty with cuba?

I Opinion question answeringI Q: What is the international reaction to the reelection of

Robert Mugabe as President of Zimbabwe?

I Product Review MiningI Is that restaurant/book/movie/laptop any good?

A word on terminology

I There are quite a few terms in use in the area of sentimentanalysis

I Subjectivity analysisI Opinion analysisI Sentiment analysis

I Often they are used interchangeably but in some cases theyare meant to evoke a particular theoretical background orhistory

I Generally, subjectivity analysis takes the broadest view of theobject of study

Less central cases

I BeliefI Most voters believe that he’s not going to raise their taxes

I Suasion/ArguingI One more word out of you and I’ll have you investigated.I We can enhance our sales, if we add speech recognition.

Inspirations of Subjectivity analysis

I literary theoryI Subjective sentences (Unspeakable Sentences,Ann Banfield,

1982)(1) ”What are you doing in here?” (2) Suddenly she [Zoe] wasfurious with him [Joe]. (3) ”Spying, of course.” (4) ”Well ofall dumb things! (5) I thought you ran away.” (6) Joe Bunchwas awful.

I Subjective sentences take the point of view of particularcharacter.

I Subjective sentences have subtle but identifiable grammaticalproperties.

I A few drops of rain were now falling.I He said he would go home now/then.

Inspirations of Subjectivity analysis

I linguistic theoryI Private states (Quirk et al. 1985)I The private subtype consists of verbs expressing intellectual

states such as belief and intellectual acts such as discovery.These states and acts are ‘private’ in the sense that they arenot observable: a person may be observed to assert that Godexists, but not to believe that God exists.

Inspirations of Subjectivity analysis

Some kinds of linguistic choices that depend on the situatedexperience of a particular person are ignored by subjectivityanalysis.

I Spatial, temporal, and person deixisI Turn left v. Turn westI yesterday v. October 5

I Fictive construals of motionI Trees flew past usI The road winds through the mountains.

Information Targeted By Sentiment Analysis

I Source (whose psychological/emotional state; aka OpinionHolder)

I Target (who or what it’s about; aka Topic)

I Polarity [aka Valence] (positive or negative; pro or contra; ...)

I Attitude (Sentiment; Arguing; ...)

I Intensity (loathe > hate > dislike > disprefer > not like verymuch)

Sentiment viewed linguistically

I Some interesting features of sentiment expressionI denotation v. invited inferenceI attribution problemI varietyI frequent dependence on context

I subjective or notI polarity

I non-compositionality

Opinons v.polar facts

I Some statements invite emotional appraisal but do notexplicitly denote appraisal.

I While such polar facts may in a particular context seem tohave an obvious value, their evaluation may be very differentin another one.

XP not Vista

I Pros: Price/Value. XP OS NOT VISTA! Screen good even inbright daylignt. Easy to access USB, lightweight.

I Cons: machine has the older XP not vista.

Attribution

Attribution

I I like ice-cream.

I Bill says that Sue hates Harry.

I In Yugoslavia’s dying days, Slobodan Milosevic slanderedMesic as Ustashe and blocked his presidency.

I She told her she is going to marry that idiot.

Attribution: types of sentiment expressions

I References to private statesI His alarm grewI He was boiling with anger

I References to speech or writing events expressing privatestates:

I UCC leaders roundly condemned the Iranian PresidentI ”He’s a quack”, He said.

I Expressive subjective elementsI He would be quite a catchI He’s a quack.

Attribution

Attribution and the level of analysis

Research commonly distinguishes between document,sentence/clause, and expression-level sentiment analysis. Inparticular tasks, attribution is (often) ignored.

I product review mining, e.g. on Amazon

I movie reviews

I news paper editorials

I blogs

Variety

Variety

I The list of linguistic expressions conveying sentiment isextremely long

I darlingI gemI sweetheartI marvelI dimwitI blockheadI idiotI . . .

Variety

I Sentiment can be expressed by many different linguistic means

I your husband is a darlingI your dear husbandI your husband, that darling,I I like your husbandI your darling of a husband

Context

Context dependence: Presence of subjectivity

I Word sensesI Neutral: Perhaps, ironically, it has something to do with the

idea that, historically, the liberal arts are connected to thecultivation of gentlemen and gentlewomen.

I Paradigmatic choicesI Ask Daddy if he can take you to school.I Ask that man if he can take you to school. (where “that

man” is the addressee’s father and the speaker’s husband)I Konnten Sie mir helfen?I Konntest Du mir helfen? (rude if addressed to an unfamiliar

adult in a formal context)

Context dependence: constructional influences

I NP be hell: That plane ride was hell.

I a hell of a NP: Are You Ready for a Hell of a Good Time?

I one hell of a NP: I am having one hell of a time trying toconsolidate and organize my music collection.

Context dependence: Presence of subjectivity

I Associations are not stable across timeI for word senses (E knave - G Knabe)I for paradigmatic choices

I estate tax v. death taxI French fries v. freedom fries

Context dependence: Presence of subjectivity

I Associations differ between speakers (ideology)I liberal

I The liberal elite have clearly fallen out of touch with theAmerican voter.

I I’m a Fighting Liberal, and Proud of What that Means

A latte-drinking, Volvo-driving, New York Times-reading,Hollywood-loving liberal

Context dependence: Polarity

I Some constructions mark presence of subjectivity but leavepolarity open

I the next N.N.A leading black conservative says Democratic presidentialfrontrunner Barack Obama is not the next John F.Kennedy. Instead, says Ken Blackwell, he’s the next GeorgeMcGovern and ”it’s time people learned the facts.”!

I Person be such a NI You’re such a good kisserI You, you’re such a big star to me.I Oh, Lisa, you’re such a girl!

Idiomaticity

Idiomaticity/non-compositionality

I a lot of subjective language consists of MWEsI but none of the components may be sentiment evoking in

isolationI be not a altar boy (”Your son is no altar boy,” the mother is

told.)I have had it with N (We’ve had it with the attacks on our

patriotism)I be the third rail of (Antidumping is the ”third rail” of U.S.

trade politics)I Go Down Like A Lead Balloon

Idiomaticity/non-compositionality

I and even if there are such components, they may not add upor be hard to integrate

I goody two-shoes (Everyone despises a Miss Goody Two Shoes,and Isidingo’s Thandi has to be the epitom of goody twoshoes.)

I love to death (I just love her to death.)I no love lost (There is no love lost between them)

Idiomaticity/non-compositionality

I subjective expressions can be embedded under, or modified by,other expressions with the effect that their polarity and/orintensity is modulated

I I have faith in the president.I I lack faith in the president.I I have hardly any faith in the president

I In sentiment analysis, people often describe the effect ofwords like lack on faith as valence shifting (Polanyi &Zaenen 2005).

I It would be nice to compose the sentiment values of linguisticexpressions (Moilanen & Pulman 2007) ...

Idiomaticity/ non-compositionality

... but despite many simple cases that is a difficult challenge

I Cheers to Timothy Whitfield for the wonderfully horridvisuals.

I I don’t like reggae. I love it

I He’s never met a dictator he didn’t like.

I Has he ever seen a tax he didn’t like?

I I doubt she’s ever come across a military spending bill shedidn’t like.

I Name me even just one tax increase she didn’t like!

I At least, she’s never held up a tax bill she didn’t like.

Sentimental inferences

I Often we can infer sentiments beyond those that are directlystated

I I am glad that Chelsea lost.I Negative towards Chelsea: I hate that team.I Neutral: I want Arsenal to become champions.I Positive: This way they don’t have so many games and can

concentrate on the championship.

I Interestingly, the simple inference (Negative towards Chelsea)is correct in the majority of instances.

Sentiment is in the grammar

I Some comparisons of equality must be interpretedmetaphorically/evaluatively

I Tom is as fat as Bill.I Tom is as fat as a cow.I *Tom is fat as BillI Tom is fat as a cow.

I There are idiomatic constructions dedicated to expressingsubjectivity:

I Since you’re going to be there a while, why not go andchit-chat? (why+VP)

I Customer to Waiter: ”What is that fly doing in my soup?”Waiter. ”Looks like the back stroke. (WXDY)

Using context to process subjective expressions

I for the purpose of lexical analysisI use distributional hypothesis to find expressions that

I are also subjective/objective (Wiebe 2000)I given subjective status, have same (prior) polarity as seed

words

I use contextual coherence to learn polarity during lexiconacquisition (Kanayama et al. 2006)

I use syntagmatic co-occurrence with seed terms as measure of

polarity (Turney 2002, Hatzivassiloglou & McKeown 1997)

I for analysis of particular usesI use contextual coherence to decide polarity of expressions with

flexible polarityI look for valence shifters that may influence polarity of

expressions on particular uses (Wilson 2008)

Non-contextual methods

I Analyzing dictionaries, esp. Wordnet (Gyamfi, Wiebe, &Mihalcea, R. 2009)

I Acquire subjective expressions (+polarity) via aligned-corpora(Banea, Mihalcea, & Wiebe, J. 2008)

Resources

I Jan Wiebe’s websiteI bibliographyI Opinionfinder, including dictionariesI MPQA corpus

I Esuli and Sebastiani’s SentiWordNet

I Bing Liu, who has a product mining angle