Deceptive languageweb.stanford.edu/class/...handout-11-02-deception.pdf · Overview On deception...

Deceptive language

Chris Potts

Linguist 287 / CS 424P: Extracting Social Meaning andSentiment, Fall 2010

Nov 2

Overview On deception Cues to deception Newman et al. Hancock et al. Enos et al. Looking for new data

Overview

1 On deception

2 Cues to deception and truthfulness (drawing on the largemeta-survey of DePaulo et al. (2003)).

3 Three studies:• Newman, Pennebaker, Berry, Richards: ‘Lying Words:

Predicting Deception From Linguistic Styles’(Newman et al. 2003).

• Hancock, Toma, Ellison, Lying in online data profiles(Toma et al. 2007, 2008; Toma and Hancock 2010).

• Enos, Shriberg, Graciarena, Hirschberg, Stolcke, TheColumbia-SRI-Colorado Deception Corpus

(Enos et al. 2007).

4 Hunting for publicly-available data for deception research.


On deception

DePaulo et al. (2003:74):

We define deception as a deliberate attempt to misleadothers. Falsehoods communicated by people who aremistaken or self-deceived are not lies, but literal truthsdesigned to mislead are lies.


PerjurySolan and Tiersma (2005:212–213) summarize the legal definition:

Perjury consists of lying under oath: having sworn to tell thetruth, the witness speaks falsely. It is a serious crime, sincefalse testimony may cause the innocent to go to prison or allowthe guilty to go free.

It is not normally a crime to lie. To commit perjury, a personmust first have taken an oath to testify truthfully. Federal lawalso requires that the person “willfully and contrary to suchoath states or subscribes any material matter which he doesnot believe to be true.” [. . . ] If the speaker did not know thatthe actual and asserted state of affairs were different, shewould have made a mere mistake.

Not only must the accused make a false statement, but itmust be material. If the false statement relates to a minormatter or something that is unlikely to influence a trial or otherofficial proceeding, it does not constitute perjury, even thoughwe might still call the statement a lie.


Lying and bullshitting

Fania Pascal, from an anecdote in Rhees (1984), cited in Frankfurt1988:

I had my tonsils out and was in the Evelyn Nursing Homefeeling sorry for myself. Wittgenstein called. I croaked: “Ifeel just like a dog chat has been run over.” He wasdisgusted: “You don’t know what a dog that has been runover feels like.”

Frankfurt (1988:125):

“Her statement is grounded neither in a belief that it istrue nor, as a lie must be, in a belief that it is not true. It isjust this lack of connection to a concern with truth — thisindifference to how things really are — that I regard as ofthe essence of bullshit.


Relevance

Bronston v. United States, decided by the Supreme Court in1973. Bronston had filed for bankruptcy. Under oath:

Q: Do you have any bank accounts in Swiss banks,Mr. Bronston?

A: No, sir.

Q: Have you ever?

A: The company had an account there for about six months, inZurich.

The truth: Bronston also had a Swiss bank account in the past.

The outcome: He was convicted of perjury, but the decision wasreversed by the Supreme Count (9-0), on the grounds that (i) heuttered no literal falsehood; and (ii) it was the lawyer’s job topursue the whole truth.


Relevance

Bronston v. United States, decided by the Supreme Court in1973. Bronston had filed for bankruptcy. Under oath:Q: Do you have any bank accounts in Swiss banks,

Mr. Bronston?

A: No, sir.

Q: Have you ever?

A: The company had an account there for about six months, inZurich.

The truth: Bronston also had a Swiss bank account in the past.

The outcome: He was convicted of perjury, but the decision wasreversed by the Supreme Count (9-0), on the grounds that (i) heuttered no literal falsehood; and (ii) it was the lawyer’s job topursue the whole truth.


Informativity 1

Kyle and Ellen would like to see a movie. Kyle has $20 in hispocket. Ellen has $8.

Context 1Tickets cost $8 each.

Context 2Tickets cost $10 each.

Kyle: “I have $8.”

Speakers more likely to say that Kyle was untruthful in context 2than in context 1, though what he said is literally true in both cases.


Informativity 2

• Context 1: B knows Barbara’s full number, but he doesn’t wantthem to be able to call her.

• Context 2: B knows Barbara’s full number, but A and B bothknow that they can’t call numbers that begin with “413”.

A: What is Barbara’s phone number?

B: Hmm. It begins with 413.


Kinds of deceptionWhere the issue of whether p is relevant:

Type Speaker’s beliefs Speaker’s public commitment

Lie ¬p pBullshit (p or ¬p) p

Comments:• The actual facts of the matter seem not to matter much:

• If the speaker believes ¬p and claims p, and p turns out to betrue, we still call the speaker a liar.

• If the speaker believes ¬p and claims ¬p, but p turns out to betrue, we say the speaker was honest (but misinformed,perhaps irresponsible).

• Bluffing is a kind of lying without social stigma.

• Withholding information can be lying if the speaker falselycommits to ignorance.


Lying is common

• DePaulo et al. (2003)76: “Lying is a fact of everyday life.Studies in which people kept daily diaries of all of their liessuggest that people tell an average of one or two lies a day.[. . . ] People lie most frequently about their feelings, theirpreferences, and their attitudes and opinions. Less often, theylie about their actions, plans, and whereabouts. Lies aboutachievements and failures are also commonplace. [. . . ]Interspersed among these unremarkable lies, in much smallernumbers, are lies that people regard as serious.”

• Toma et al. (2007): “Deception was indeed frequentlyobserved: approximately nine out of ten (81%) of theparticipants lied on at least one of the assessed variables.”

• (There are probably many lies in the speed-dating corpus!)


Liars on lyingFrom DePaulo et al. (2003:76):

• “More than half the time, liars said that they based their lies onexperiences from their own lives, altering critical details.”

• “In the literature on cues to deception, as in everyday life, liesabout personal feelings, facts, and attitudes are the mostcommonplace.”

• “The results suggest that people regard their everyday lies aslittle lies of little consequence or regret. They do not spendmuch time planning them or worrying about the possibility ofgetting caught. Still, everyday lies do leave a smudge.Although people reported feeling only low levels of distressabout their lies, they did feel a bit more uncomfortable whiletelling their lies, and directly afterwards, than they had felt justbefore lying. Also, people described the social interactions inwhich lies were told as more superficial and less pleasantthan the interactions in which no lies were told.”


Machines vs. humans• Ekman and O’Sullivan (1991): Studies involving college

students as well as trained police officers generally showchance-level performance on liar detection.

• Enos et al. (2007): “With respect to accuracy at labelingGLOBAL LIES and TRUTHS in the CSC Corpus, the topic ofthe present work, human judges performed even worse: onaverage 47.8% versus a chance baseline of 63.6%.”

• Newman et al. (2003):

Although liars have some control over the content oftheir stories, their underlying state of mind may leak outthrough the style of language used to tell the story. Thedata presented here provide some insight into the lin-guistic manifestations of this state of mind.

Specifically, deceptive communication was character-ized by the use of fewer first-person singular pronouns(e.g., I, me, my), fewer third-person pronouns (e.g., he,she, they), more negative emotion words (e.g., hate, anger,enemy), fewer exclusive words (e.g., but, except, without),and more motion verbs (e.g., walk, move, go). Four ofthese linguistic categories—first-person pronouns, neg-ative emotion words, exclusive words, and motionverbs—were consistent with our specific predictions.However, the generalizability of these categories varieddepending on the topic. In this discussion, we first exam-ine the meaning of each linguistic predictor and thenaddress the generalization of these predictors.

ELEMENTS OF THE LINGUISTIC PROFILE

First, in the present studies, liars used first-person pro-nouns at a lower rate than truth-tellers. The lower rate ofself-references is consistent with previous literature (butsee our discussion below of DePaulo et al., 2003) and isthought to reflect an attempt by liars to “dissociate”themselves from the lie (Dulaney, 1982; Knapp et al.,1974; Mehrabian, 1971; for a review, see Knapp &Comadena, 1979; Vrij, 2000, Chap. 4). Self-referencesindicate that individuals are being “honest” with them-selves (Campbell & Pennebaker, in press; Davis & Brock,1975; Duval & Wicklund, 1972; Feldman Barrett et al., inpress; Shapiro, 1989). Because deceptive stories do notreflect one’s true attitudes or experiences, liars may wishto “stand back” (Knapp et al., 1974, p. 26) by investingless of themselves in their words.

Second, liars also used negative emotion words at ahigher rate than truth-tellers. Liars may feel guilty, either

because of their lie or because of the topic they are lyingabout (e.g., Vrij, 2000). Because of this tension and guilt,liars may express more negative emotion. In support ofthis, Knapp et al. (1974) found that liars made disparag-ing remarks about their communication partner at amuch higher rate than truth-tellers. In an early meta-analysis, Zuckerman, DePaulo, and Rosenthal (1981)identified negative statements as a significant marker ofdeception. The “negative emotion” category in LIWCcontains a subcategory of “anxiety” words, and it is possi-ble that anxiety words are more predictive than overallnegative emotion. However, in the present studies, anxi-ety words were one of the categories omitted due to lowrate of use.

Third, liars used fewer “exclusive” words than truth-tellers, suggesting lower cognitive complexity. A personwho uses words such as but, except, and without is makinga distinction between what is in a given category andwhat is not within a category. Telling a false story is ahighly cognitively complicated task. Adding informationabout what did not happen may require cognitiveresources that the typical liar does not possess. Fourth,liars used more “motion” verbs than truth-tellers, alsosuggesting lower cognitive complexity. Because liars’ sto-ries are by definition fabricated, some of their cognitiveresources are taken up by the effort of creating a believ-able story. Motion verbs (e.g., walk, go, carry) provide sim-ple, concrete descriptions and are more readily accessi-ble than words that focus on evaluations and judgments(e.g., think, believe).

In addition, liars in the present studies unexpectedlyused third-person pronouns at a lower rate than truth-tellers. This is inconsistent with previous literature: Liarstypically use more other-references than truth-tellers(e.g., Knapp et al., 1974). It is possible that this reflectsthe subject matter—abortion attitudes—in the majorityof the present studies. Talking about abortion necessar-ily involves talking about women, but this can be doneusing pronouns (she, her), more specific nouns (a woman,my sister), or even proper names. This word use mayreflect differences in the underlying psychologybetween liars and truth-tellers such that people lyingabout their attitudes added concrete details by referringto specific people instead of using the generic she.Although this interpretation is admittedly post hoc, it isconsistent with our finding that liars tend to be concreterather than abstract (see also Knapp et al., 1974).

Previous investigations of the linguistic differencesbetween liars and truth-tellers have yielded mixedresults, largely due to substantive differences in the wayslinguistic categories have been defined and assessed. In arecent exhaustive meta-analysis, DePaulo et al. (2003)reviewed the combined evidence for 158 different cuesto deception. Overall, liars appear to be less forthcom-

672 PERSONALITY AND SOCIAL PSYCHOLOGY BULLETIN

TABLE 5: Comparison of Human Judges’ Ratings With LIWC’s Pre-diction Equations in Three Abortion Studies

Predicted

Deceptive Truthful

LIWC equationsActual

Deceptive (n = 200) 68% (135) 32% (65)Truthful (n = 200) 34% (69) 66% (131)

Human judgesActual

Deceptive (n = 200) 30% (59) 71% (141)Truthful (n = 200) 27% (53) 74% (147)

NOTE: LIWC = Linguistic Inquiry and Word Count. N = 400 communi-cations. The overall hit rate was 67% for LIWC and 52% for judges;these were significantly different, z = 6.25, p < .001. LIWC performedsignificantly better than chance, z = 6.80, p < .001, but the judges didnot, z = .80, ns. See the text for an analysis of error rates.

at STANFORD UNIV on October 25, 2010psp.sagepub.comDownloaded from


Cues to deception

• Cues to deception (both linguistic and other) are generallyassumed to be weak. There are no absolute “tells”.

• The individual cues that have been performed are allindicative of other emotional states as well.

• We might, though, hold out hope that a cluster of featuresreliably detects just deception.


DePaulo et al. (2003) surveyAre liars less forthcoming?

Appendix A

Definitions of Cues to Deception

Cue Definition

Are Liars Less Forthcoming Than Truth Tellers?

001 Response length Length or duration of the speaker’s message002 Talking time Proportion of the total time of the interaction that the speaker spends

talking or seems talkative003 Length of interaction Total duration of the interaction between the speaker and the other person004 Details Degree to which the message includes details such as descriptions of

people, places, actions, objects, events, and the timing of events;degree to which the message seemed complete, concrete, striking, orrich in details

005 Sensory information (RM) Speakers describe sensory attributes such as sounds and colors006 Cognitive complexity Use of longer sentences (as indexed by mean length of the sentences),

more syntactically complex sentences (those with more subordinateclauses, prepositional phrases, etc.), or sentences that includes morewords that precede the verb (mean preverb length); use of the wordsbut or yet; use of descriptions of people that are differentiating anddispositional

007 Unique words Type–token ratio; total number of different or unique words008 Blocks access to information Attempts by the communicator to block access to information, including,

for example, refusals to discuss certain topics or the use of unnecessaryconnectors (then, next, etc.) to pass over information (The volunteeringof information beyond the specific information that was requested wasalso included, after being reversed.)

009 Response latency Time between the end of a question and the beginning of the speaker’sanswer

010 Rate of speaking Number of words or syllables per unit of time011 Presses lips (AU 23, 24) Lips are pressed together

Do Liars Tell Less Compelling Tales Than Truth Tellers?

012 Plausibility Degree to which the message seems plausible, likely, or believable013 Logical structure (CBCA) “Consistency and coherence of statements; collection of different and

independent details that form a coherent account of a sequence ofevents” (Zaparniuk, Yuille, & Taylor, 1995, p. 344)

014 Discrepant, ambivalent Speakers’ communications seem internally inconsistent or discrepant;information from different sources (e.g., face vs. voice) seemscontradictory; speaker seems to be ambivalent

015 Involved, expressive (overall) Speaker seems involved, expressive, interested016 Verbal and vocal involvement Speakers describe personal experiences, or they describe events in a

personal and revealing way; speakers seems vocally expressive andinvolved

017 Facial expressiveness Speaker’s face appears animated or expressive018 Illustrators Hand movements that accompany speech and illustrate it019 Verbal immediacy Linguistic variations called verbal nonimmediacy devices, described by

Wiener and Mehrabian (1968) as indicative of speakers’ efforts todistance themselves from their listener, the content of theircommunications, or the act of conveying those communications.Wiener and Mehrabian (1968) described 19 categories andsubcategories, such as spatial nonimmediacy (e.g., “There’s Johnny” ismore nonimmediate than “Here’s Johnny”), temporal nonimmediacy(the present tense is more immediate than other tenses), and passivity(the passive voice is more nonimmediate than the active voice).

020 Verbal immediacy, temporal A subcategory of verbal immediacy in which speakers use the presenttense instead of past or future tenses

021 Generalizing terms Generalizing terms (sometimes called levelers) such as everyone, no one,all, none, and every; statements implying that unspecified others agreewith the speaker

022 Self-references Speakers’ references to themselves or their experiences, usually indexedby the use of personal pronouns such as I, me, mine, and myself

023 Mutual and group references Speakers’ references to themselves and others, usually indexed by the useof second-person pronouns such as we, us, and ours

(Appendixes continue)

113CUES TO DECEPTION


DePaulo et al. (2003) surveyAre liars less compelling?

Appendix A

Definitions of Cues to Deception

Cue Definition

Are Liars Less Forthcoming Than Truth Tellers?

001 Response length Length or duration of the speaker’s message002 Talking time Proportion of the total time of the interaction that the speaker spends

talking or seems talkative003 Length of interaction Total duration of the interaction between the speaker and the other person004 Details Degree to which the message includes details such as descriptions of

people, places, actions, objects, events, and the timing of events;degree to which the message seemed complete, concrete, striking, orrich in details

005 Sensory information (RM) Speakers describe sensory attributes such as sounds and colors006 Cognitive complexity Use of longer sentences (as indexed by mean length of the sentences),

more syntactically complex sentences (those with more subordinateclauses, prepositional phrases, etc.), or sentences that includes morewords that precede the verb (mean preverb length); use of the wordsbut or yet; use of descriptions of people that are differentiating anddispositional

007 Unique words Type–token ratio; total number of different or unique words008 Blocks access to information Attempts by the communicator to block access to information, including,

for example, refusals to discuss certain topics or the use of unnecessaryconnectors (then, next, etc.) to pass over information (The volunteeringof information beyond the specific information that was requested wasalso included, after being reversed.)

009 Response latency Time between the end of a question and the beginning of the speaker’sanswer

010 Rate of speaking Number of words or syllables per unit of time011 Presses lips (AU 23, 24) Lips are pressed together

Do Liars Tell Less Compelling Tales Than Truth Tellers?

012 Plausibility Degree to which the message seems plausible, likely, or believable013 Logical structure (CBCA) “Consistency and coherence of statements; collection of different and

independent details that form a coherent account of a sequence ofevents” (Zaparniuk, Yuille, & Taylor, 1995, p. 344)

014 Discrepant, ambivalent Speakers’ communications seem internally inconsistent or discrepant;information from different sources (e.g., face vs. voice) seemscontradictory; speaker seems to be ambivalent

015 Involved, expressive (overall) Speaker seems involved, expressive, interested016 Verbal and vocal involvement Speakers describe personal experiences, or they describe events in a

personal and revealing way; speakers seems vocally expressive andinvolved

017 Facial expressiveness Speaker’s face appears animated or expressive018 Illustrators Hand movements that accompany speech and illustrate it019 Verbal immediacy Linguistic variations called verbal nonimmediacy devices, described by

Wiener and Mehrabian (1968) as indicative of speakers’ efforts todistance themselves from their listener, the content of theircommunications, or the act of conveying those communications.Wiener and Mehrabian (1968) described 19 categories andsubcategories, such as spatial nonimmediacy (e.g., “There’s Johnny” ismore nonimmediate than “Here’s Johnny”), temporal nonimmediacy(the present tense is more immediate than other tenses), and passivity(the passive voice is more nonimmediate than the active voice).

020 Verbal immediacy, temporal A subcategory of verbal immediacy in which speakers use the presenttense instead of past or future tenses

021 Generalizing terms Generalizing terms (sometimes called levelers) such as everyone, no one,all, none, and every; statements implying that unspecified others agreewith the speaker

022 Self-references Speakers’ references to themselves or their experiences, usually indexedby the use of personal pronouns such as I, me, mine, and myself

023 Mutual and group references Speakers’ references to themselves and others, usually indexed by the useof second-person pronouns such as we, us, and ours

(Appendixes continue)



DePaulo et al. (2003) surveyAre liars less compelling?

Appendix A (continued)

Cue Definition

Do Liars Tell Less Compelling Tales Than Truth Tellers? (continued)

024 Other references Speakers’ references to others or their experiences, usually indexed by theuse of third-person pronouns such as he, she, they, or them

025 Verbal and vocal immediacy(impressions)

Speakers respond in ways that seem direct, relevant, clear, and personalrather than indirect, distancing, evasive, irrelevant, unclear, orimpersonal

026 Nonverbal immediacy Speakers are nonimmediate when they maintain a greater distance fromthe other person, lean away, face away, or gaze away, or when theirbody movements appear to be nonimmediate.

027 Eye contact Speaker looks toward the other person’s eyes, uses direct gaze028 Gaze aversion Speakers look away or avert their gaze029 Eye shifts Eye movements or shifts in the direction of focus of the speaker’s eyes030 Tentative constructions Verbal hedges such as “may,” “might,” “could,” “I think,” “I guess,” and

“it seems to me” (Absolute verbs, which include all forms of the verbto be, were included after being reversed.)

031 Verbal and vocal uncertainty(impressions)

Speakers seem uncertain, insecure, or not very dominant, assertive, oremphatic; speakers seem to have difficulty answering the question

032 Amplitude, loudness Intensity, amplitude, or loudness of the voice033 Chin raise (AU 17) Chin is raised; chin and lower lip are pushed up034 Shrugs Up and down movement of the shoulders; or, the palms of the hand are

open and the hands are moving up and down035 Non-ah speech disturbances Speech disturbances other than “ums,” “ers,” and “ahs,” as described by

Kasl and Mahl (1965); categories include grammatical errors,stuttering, false starts, incomplete sentences, slips of the tongue, andincoherent sounds

036 Word and phrase repetitions Subcategory of non-ah speech disturbances in which words or phrases arerepeated with no intervening pauses or speech errors

037 Silent pauses Unfilled pauses; periods of silence038 Filled pauses Pauses filled with utterances such as “ah,” “um,” “er,” “uh,” and

“hmmm”039 Mixed pauses Silent and filled pauses (undifferentiated)040 Mixed disturbances (ah plus

non-ah)Non-ah speech disturbances and filled pauses (undifferentiated)

041 Ritualized speech Vague terms and cliches such as “you know,” “well,” “really,” and “Imean”

042 Miscellaneous dysfluencies Miscellaneous speech disturbances; speech seems dysfluent043 Body animation, activity Movements of the head, arms, legs, feet, and/or postural shifts or leans044 Postural shifts Postural adjustments, trunk movements, or repositionings of the body045 Head movements

(undifferentiated)Head movements (undifferentiated)

046 Hand movements(undifferentiated)

Hand movements or gestures (undifferentiated)

047 Arm movements Movements of the arms048 Foot or leg movements Movements of the legs and/or feet

Are Liars Less Positive and Pleasant Than Truth Tellers?

049 Friendly, pleasant (overall) Speaker seems friendly, pleasant, likable (Impressions of negative affectwere also included after being reversed.)

050 Cooperative Speaker seems cooperative, helpful, positive, and secure051 Attractive Speaker seems physically attractive052 Negative statements and

complaintsDegree to which the message seems negative or includes negativecomments or complaints (Measures of positive comments wereincluded after being reversed.)

053 Vocal pleasantness Voice seems pleasant (e.g., positive, friendly, likable)054 Facial pleasantness Speaker’s face appears pleasant; speakers show more positive facial

expressions (such as smiles) than negative expressions (such as frownsor sneers)

055 Head nods Affirmative head nods; vertical head movements056 Brow lowering (AU 4) Eyebrows are lowered057 Sneers (AU 9, 10) Upper lip is raised058 Smiling (undifferentiated) Smiling as perceived by the coders, who were given no specific definition

or were given a definition not involving specific AUs (e.g., “corners ofthe mouth are pulled up”); laughing is sometimes included too

059 Lip corner pull (AU 12) Corners of the lips are pulled up and back

114 DEPAULO ET AL.


DePaulo et al. (2003) surveyAre liars less positive?


Cue Definition

Do Liars Tell Less Compelling Tales Than Truth Tellers? (continued)

024 Other references Speakers’ references to others or their experiences, usually indexed by theuse of third-person pronouns such as he, she, they, or them

025 Verbal and vocal immediacy(impressions)

Speakers respond in ways that seem direct, relevant, clear, and personalrather than indirect, distancing, evasive, irrelevant, unclear, orimpersonal

026 Nonverbal immediacy Speakers are nonimmediate when they maintain a greater distance fromthe other person, lean away, face away, or gaze away, or when theirbody movements appear to be nonimmediate.

027 Eye contact Speaker looks toward the other person’s eyes, uses direct gaze028 Gaze aversion Speakers look away or avert their gaze029 Eye shifts Eye movements or shifts in the direction of focus of the speaker’s eyes030 Tentative constructions Verbal hedges such as “may,” “might,” “could,” “I think,” “I guess,” and

“it seems to me” (Absolute verbs, which include all forms of the verbto be, were included after being reversed.)

031 Verbal and vocal uncertainty(impressions)

Speakers seem uncertain, insecure, or not very dominant, assertive, oremphatic; speakers seem to have difficulty answering the question

032 Amplitude, loudness Intensity, amplitude, or loudness of the voice033 Chin raise (AU 17) Chin is raised; chin and lower lip are pushed up034 Shrugs Up and down movement of the shoulders; or, the palms of the hand are

open and the hands are moving up and down035 Non-ah speech disturbances Speech disturbances other than “ums,” “ers,” and “ahs,” as described by

Kasl and Mahl (1965); categories include grammatical errors,stuttering, false starts, incomplete sentences, slips of the tongue, andincoherent sounds

036 Word and phrase repetitions Subcategory of non-ah speech disturbances in which words or phrases arerepeated with no intervening pauses or speech errors

037 Silent pauses Unfilled pauses; periods of silence038 Filled pauses Pauses filled with utterances such as “ah,” “um,” “er,” “uh,” and

“hmmm”039 Mixed pauses Silent and filled pauses (undifferentiated)040 Mixed disturbances (ah plus

non-ah)Non-ah speech disturbances and filled pauses (undifferentiated)

041 Ritualized speech Vague terms and cliches such as “you know,” “well,” “really,” and “Imean”

042 Miscellaneous dysfluencies Miscellaneous speech disturbances; speech seems dysfluent043 Body animation, activity Movements of the head, arms, legs, feet, and/or postural shifts or leans044 Postural shifts Postural adjustments, trunk movements, or repositionings of the body045 Head movements

(undifferentiated)Head movements (undifferentiated)

046 Hand movements(undifferentiated)

Hand movements or gestures (undifferentiated)

047 Arm movements Movements of the arms048 Foot or leg movements Movements of the legs and/or feet

Are Liars Less Positive and Pleasant Than Truth Tellers?

049 Friendly, pleasant (overall) Speaker seems friendly, pleasant, likable (Impressions of negative affectwere also included after being reversed.)

050 Cooperative Speaker seems cooperative, helpful, positive, and secure051 Attractive Speaker seems physically attractive052 Negative statements and

complaintsDegree to which the message seems negative or includes negativecomments or complaints (Measures of positive comments wereincluded after being reversed.)

053 Vocal pleasantness Voice seems pleasant (e.g., positive, friendly, likable)054 Facial pleasantness Speaker’s face appears pleasant; speakers show more positive facial

expressions (such as smiles) than negative expressions (such as frownsor sneers)

055 Head nods Affirmative head nods; vertical head movements056 Brow lowering (AU 4) Eyebrows are lowered057 Sneers (AU 9, 10) Upper lip is raised058 Smiling (undifferentiated) Smiling as perceived by the coders, who were given no specific definition

or were given a definition not involving specific AUs (e.g., “corners ofthe mouth are pulled up”); laughing is sometimes included too

059 Lip corner pull (AU 12) Corners of the lips are pulled up and back

114 DEPAULO ET AL.


DePaulo et al. (2003) surveyAre liars more tense?


Cue Definition

Are Liars More Tense Than Truth Tellers?

060 Eye muscles (AU 6), notduring positive emotions

Movement of the orbicularis oculi, or muscles around the eye, duringemotions that are not positive

061 Nervous, tense (overall) Speaker seems nervous, tense; speaker makes body movements that seemnervous

062 Vocal tension Voice sounds tense, not relaxed; or, vocal stress as assessed by thePsychological Stress Evaluator, which measures vocal micro-tremors,or by the Mark II voice analyzer

063 Frequency, pitch Voice pitch sounds high; or, fundamental frequency of the voice064 Relaxed posture Posture seems comfortable, relaxed; speaker is leaning forward or

sideways065 Pupil dilation Pupil size, usually measured by a pupillometer066 Blinking (AU 45) Eyes open and close quickly067 Object fidgeting Speakers are touching or manipulating objects068 Self-fidgeting Speakers are touching, rubbing, or scratching their body or face069 Facial fidgeting Speakers are touching or rubbing their faces or playing with their hair070 Fidgeting (undifferentiated) Object fidgeting and/or self-fidgeting and/or facial fidgeting

(undifferentiated)

Do Lies Include Fewer Ordinary Imperfections and Unusual Contents Than Do Truths?

071 Unstructured production(CBCA)

“Narratives are presented in an unstructured fashion, free from anunderlying pattern or structure.” (Zaparniuk et al., 1995, p. 344)

072 Spontaneous corrections(CBCA)

“Spontaneous correction of one’s statements” (Zaparniuk et al., 1995, p.344)

073 Admitted lack of memory,unspecified (CBCA)

Admission of lack of memory

074 Self-doubt (CBCA) “Raising doubts about one’s own testimony; raising objections to theaccuracy of recalled information” (Zaparniuk et al., 1995, p. 344)

075 Self-deprecation (CBCA) “Inclusion of unfavorable, self-incriminating details” (Zaparniuk et al.,1995, p. 344)

076 Contextual embedding(CBCA)

“Statements that place the event within its spatial and temporal context”(Zaparniuk et al., 1995, p. 344)

077 Verbal and nonverbalinteractions (CBCA)

“Verbatim reproduction of dialogue” and “descriptions of interrelatedactions and reactions” (Zaparniuk et al., 1995, p. 344)

078 Unexpected complications(CBCA)

“The reporting of either an unforseen interruption or difficulty, orspontaneous termination of the event” (Zaparniuk et al., 1995, p. 344)

079 Unusual details (CBCA) “Inclusion of detail that is not unrealistic, but has a low probability ofoccurrence” (Zaparniuk et al., 1995, p. 344)

080 Superfluous details (CBCA) “Vivid and concrete descriptions of superfluous details” (Zaparniuk et al.,1995, p. 344)

081 Related external associations(CBCA)

“Reference to events or relationships that are external to the event ofimmediate focus” (Zaparniuk et al., 1995, p. 344)

082 Another’s mental state(CBCA)

“Statements inferring the cognitive and emotional state of others involvedin the event” (Zaparniuk et al., 1995, p. 344)

083 Subjective mental state(CBCA)

“Accounts of the witness’s own cognitive and emotional state at the timeof the event” (Zaparniuk et al., 1995, p. 344)

Cues Listed in Appendix Ba

084 Number of segments Perceived number of behavioral units085 Idiosyncratic information

(RM)Speakers mention idiosyncratic information

086 Facial shielding Speakers appear to be shielding their face087 Realism (RM) The story is realistic and makes sense088 Intensity of facial expression Speaker’s facial expression appears to be intense; rated intensity of AUs089 Face changes Changes in facial expressions; onset, offset, and apex phases; face seems

mobile090 Indifferent, unconcerned Speaker seems indifferent, unconcerned091 Seems planned, not

spontaneousMessage seems planned or rehearsed

092 Cognitively busy Speaker seems to be making mental calculations093 Serious Speaker seems serious, formal094 Pitch variety Variation in fundamental frequency095 Pitch changes Frequency of changes in the pitch of the voice096 Rate change Rate of speaking in the second half of the message minus rate of

speaking in the first half(Appendixes continue)



Negation and negativity

HypothesisLiars will use more negations and negative words than truth-tellers.




RationaleNewman et al. (2003:666): “liars may feel guilty either about lyingor about the topic they are discussing [. . . ] Diary studies of small“everyday” lies suggest that people feel discomfort and guilt whilelying and immediately afterward [. . . ] If this state of mind isreflected in patterns of language use, then deceptivecommunications should be characterized by more words reflectingnegative emotion.”




ConcernsDePaulo et al. (2003:75): “Liars’ feelings about lying are notnecessarily negative ones. Ekman (1985/1992) suggested thatliars sometimes experience ‘duping delight,’ which could includeexcitement about the challenge of lying or pride in succeeding atthe lie. [. . . ] The duping delight hypothesis has not yet beentested.”

DePaulo et al. (2003:81): “Yet those who tell the truth about theirtransgressions or failings may feel even greater guilt and shamethan those whose shortcomings remain hidden by their lies.”




Increased use of negative words could signal negativity.

● ● ● ●

●

●

●

●● ●

NEG good (20,447 tokens)

1 2 3 4 5 6 7 8 9 10

0.03

0.1

0.16● ●

●●

●

●● ● ●

●

depress(ed/ing) (18,498 tokens)

1 2 3 4 5 6 7 8 9 10

0.080.110.13

●

●

●

●

●

●

●

●● ●

bad (368,273 tokens)

1 2 3 4 5 6 7 8 9 10

0.04

0.12

0.21

●

●

●

●

●

●

●● ● ●

terrible (55,492 tokens)

1 2 3 4 5 6 7 8 9 10

0.03

0.16

0.28

Pr(

c|w

)

Rating




Increased use of negative words could signal sympathy.

hugs

rock

s

teeh

ee

unde

rsta

nd

just

wow

NEG good (478 tokens)

0.12

0.22

0.31

hugs

rock

s

teeh

ee

unde

rsta

nd

just

wow

depress(ed/ing) (910 tokens)

0.08

0.21

0.34

hugs

rock

s

teeh

ee

unde

rsta

nd

just

wow

bad (3,085 tokens)

0.15

0.19

0.23

hugs

rock

s

teeh

ee

unde

rsta

nd

just

wow

terrible (388 tokens)

0.15

0.21

0.28

Pr(

c|w

)




Increased use of negative words could signal sympathy.Similar to the finding of Ranganath et al. (2009) that flirtatiouspeople use more negative words to express sympathy.


Complexity

HypothesisLiars speech will be less complex than truth-tellers speech:

1 Newman et al. (2003:667): Fewer exclusive words.(“individuals who use a higher number of “exclusive” words[. . . ] are generally healthier than those who do not”)

2 Newman et al. (2003:667): More simple-past tense verbs ofmotion: “When people are attempting to construct a falsestory, we argue that simple, concrete actions are easier tostring together than false evaluations.”

3 Shorter sentences, shorter turns.


Complexity





RationaleLying imposes an extra cognitive load.


Complexity





ConcernsThis might not distinguish lying from storytelling, or from anysituation in which the speaker feels rushed, anxious, etc.


PronounsHypothesisLiars will use fewer 1st-person pronouns and more 3rd-personpronouns.



RationaleNewman et al. (2003:666): “the use of the first-person singular is asubtle proclamation of one’s ownership of a statement.”



Concern: Other correlationsPennebaker and colleagues have shown that increased use offirst-person pronouns correlates with lots of other things. Forexample, Chung and Pennebaker (2007) report on patternssuggesting that 1st person pronoun use is negatively correlatedwith status and positively correlated with individualism.



Concern: Definitely not universalPronoun use is heavily conditioned by the nature of the pronominalparadigm, which can often include null pronouns. In addition, notall pronominal paradigms divide up in a way that easily feeds intothis hypothesis.


Pronouns in WALS Online

http://wals.info/

http://wals.info/


DisfluenciesHypothesisLiars will speak more slowly and exhibit more disfluencies. (Kasland Mahl (1965) remove ‘ah’ from this category on the groundsthat it signals something different.)



RationaleLiars are more uncertain, and will thus require more time to sortout what they want to say.



Concerns• DePaulo et al. (2003:94): “Results of the fluency indices

suggest that speech disturbances have little predictive poweras cues to deceit. [. . . ] Only one type of speech disturbance,the repetition of words and phrases, produced a statisticallyreliable effect”

• Levels of disfluency are highly speaker and context specific.To measure a change for a specific person, we need to knowwhat the baseline is.


Acoustic features

• DePaulo et al. (2003): According to Ekman (1985), “the cuesindicative of detection apprehension are fear cues. Theseinclude higher pitch, faster and louder speech, pauses,speech errors, and indirect speech. The greater the liars’detection apprehension, the more evident these fear cuesshould be.”

• Enos et al. (2007): “extreme values for energy — either highor low — correlate with deception.”


Domain-specificity

• Newman et al. (2003): The studies vary considerably in theirsubject matter, and this affects the models they fit.

• Larcker and Zakolyukina (2010): Deceptive CEOs inconference calls make fewer references to “shareholdersvalue and value creation”.

• Toma and Hancock (2010): “liars produced fewer, rather thanmore, negative emotion words. This could be due to the factthat people who lied more were more eager to make a goodimpression, and thus avoided sounding negative — which isusually a turnoff in dating situations. Future work is needed toclarify the nature of this indicator.”


Difficult-to-extract features: Your ideas

Drawn from the DePaulo et al. (2003) feature survey:

1 Involved, expressive (overall): “Speaker seems involved,expressive, interested”

2 Plausibility: “Degree to which the message seems plausible,likely, or believable”

3 Ritualized speech: “Vague terms and cliches such as ‘youknow,’ ‘well,’ ‘really,’ and ‘I mean’.”

4 Vocal tension: “Voice sounds tense, not relaxed; or, vocalstress as assessed by the Psychological Stress Evaluator,which measures vocal micro-tremors, or by the Mark II voiceanalyzer.”


Newman et al. (2003)’s experimentsAbortion attitudesWithin-subjects accurate/inaccurate reporting of attitudes onabortion, taped (Study 1), typed (S2), and handwritten (S3).

Feelings about friends: 3-min videotaped monologuesFriend 1 Friend 2 Friend 3 Friend 4

Truth Like Like Dislike DislikeStory Like Dislike Like Dislike

Mock crimeOne group was told to wait, while the other was told to “steal” adollar. Both participant-types were then accused of theft, after priorinstruction to deny it. They were then hooked-up to what they weretold was a lie-detector (it actually took various physiologicalmeasurements) and given a systematic 2-min interview in whichthey denied wrong-doing.


Predictive models

1 Data preparation: Transcribe data as necessary and removecontent words, low-frequency words, and disfluencies.

2 Categorization into 29 LIWC categories.

3 Model-building: begin from a null model and add predictorsstepwise, keeping only those that are significant.

4 For each study S, fit a model trained on the other four studiesand then use that model to make predictions about S.


ResultsThis model was built using just the predictors that were significantin at least two studies:

human judges. Recall that judges rated the perceivedtruthfulness of all communications dealing with abor-tion attitudes (Studies 1 through 3; N = 400 communica-tions) and the proportion of judges who believed eachcommunication to be truthful had been calculated. Weused these data to calculate a “hit rate” for our judgesand compared it to LIWC’s ability to correctly identifydeception. More specifically, a dichotomous classifica-tion was made for each communication. If the propor-tion of judges believing a particular communication tobe truthful was greater than 50%, it was defined as

“judged truthful.” The remaining communications wereconsidered “judged false.”

We then calculated the proportion of communica-tions that had been correctly identified as truthful ordeceptive by the judges and compared them with theLIWC classifications based on the general predictionequation (see above). As seen in Table 5, LIWC correctlyclassified 67% of the abortion communications and thejudges correctly classified 52% of the abortion commu-nications. These proportions were significantly different(z = 6.25, p < .0001). LIWC performed significantly betterthan chance (z = 6.80, p < .001) but the judges did not (z =.80, ns).

To correct for a potential positivity bias in judges’responses, we conducted a signal detection analysis. Wefirst converted the proportion of correct hits (identify-ing truthful communications as truthful) and the pro-portion of false positives (identifying deceptive commu-nications as truthful) to z scores. These values of .74 and.71 were converted to z scores of z = .64 and z = .55,respectively. We then calculated d ! by subtracting theproportion of false positives from the proportion of cor-rect hits (d ! = .09, ns).

In addition, the two detection strategies showed dif-ferent patterns of error. The judges were significantlymore likely to make “false positive” identifications than“false negative” identifications (71% vs. 27% of judges’errors, respectively; z = 14.67, p < .001). LIWC, in con-trast, was equally likely to make “false positive” identifica-tions and “false negative” identifications (49% vs. 51% ofLIWC errors, respectively; z = .50, ns).

Discussion

Successfully lying to another person usually involvesthe manipulation of language and the careful construc-tion of a story that will appear truthful. In addition to cre-ating a convincing story, liars also must present it in astyle that appears sincere (Friedman & Tucker, 1990).

Newman et al. / LINGUISTIC STYLE AND DECEPTION 671

TABLE 3: Predictors of Deception: Logistic Regression CoefficientsUsed in Prediction Equations

% Accuracy (Lie/ AdjustedLIWC Category Truth/Overall) R2

Predicting Study 1:Studies 2-5 combined 60/58/59* 4

Negative emotion –.268Senses .270Exclusive words .452Motion verbs –.310

Predicting Study 2:Studies 1 and 3-5 combined 61/57/59* 6

Negative emotion –.227Exclusive words .286Motion verbs –.358

Predicting Study 3:Studies 1-2 and 4-5 combined 66/69/67** 17

First-person pronouns .209Third-person pronouns .254Exclusive words .362Motion verbs –.213

Predicting Study 4:Studies 1-3 and 5 combined 52/54/53 1

First-person pronouns .240Articles –.264Negative emotion –.382Exclusive words .463

Predicting Study 5:Studies 1-4 combined 53/43/48 0


General prediction equation:All 5 studies combined 59/62/61** 8

First-person pronouns .260Third-person pronouns .250Negative emotion –.217Exclusive words .419Motion verbs –.259

NOTE: LIWC = Linguistic Inquiry and Word Count. For % accuracy,the three percentages listed for each equation are (a) % of liars identi-fied accurately, (b) % of truth-tellers identified accurately, and (c)overall accuracy. For overall accuracy rates, *p < .05 and **p < .001when compared to chance performance of 50%. Coefficients in thenegative direction mean that liars used the category at a higher rate. Allcoefficients are significant at p < .05 or better.

TABLE 4: Effect Sizes and Reliability of Linguistic Predictors AcrossFive Studies

Study

1 2 3 4 5 Mean d Reliability

First person .31 .85 .75 –.02 –.24 .36 .43Third person .22 .30 .07 .24 –.21 .16 .28Negative emotion –.19 –.27 –.42 .40 –.33 –.15 .36Exclusive words .40 1.23 .91 .02 .30 .54 .55Motion verbs –.14 .09 –.31 –.40 –.29 –.20 .40

NOTE: Numbers for each study represent Cohen’s d comparing liarsand truth-tellers. Effect sizes in the negative direction mean that liarsused the linguistic category at a higher rate. Mean d is a weightedmean of these effect sizes. Reliability was calculated using Cronbach’salpha.


Table: Newman et al. (2003). A negative coefficient means that thevariable correlates with lying. A positive one means that the variablecorrelates with truth-telling.


Deception in online dating profiles

Toma et al. (2007, 2008); Toma and Hancock (2010):

• 80 participants (40 men, 40 women)

• When they came into the lab, they were given copies of theirown profiles and asked to assess their accuracy on a scale,1 (least accurate) to 5 (most accurate).

• Objective measures: participants’ height, weight, and agewere measured by the experimenters.

• Deception index: standardized mean deviance from the truthin three categories (with some tolerances built in to accountfor measurement error).


Degrees of deceptiveness

“Participants rated their self-descriptions as very accurate. On the1 (extremely inaccurate) to 5 (extremely accurate) scale used,self-descriptions were rated as 4.79 (SD = 0.41, min = 4.00, max =5.00), suggesting that daters considered them to be almost free ofdeceptions.”

Overall Males Females

Lied about height 48.10 55.30 41.50Lied about weight 59.70 60.50 59.00

Lied about age 18.70 24.30 13.20Lied in any category 81.30 87.20 75.60

Table: People lied a lot.


Deceptiveness by gender and category


Hypotheses

H1: Highly deceptive online dating profiles will have fewerself-references but more negations and negative emotionwords than less deceptive profiles.

H2: Highly deceptive profiles will have fewer exclusive words andincreased motion words, but a lower overall word count thanless deceptive profiles.

H3: Emotionally-related linguistic cues to deception shouldaccount for more variance in deception scores thancognitively-related linguistic cues in online dating profiles.


Predictive model

A linear model predicting the deception index. Negativecoefficients ≈ deception, positive coefficients ≈ truthfulness. Thecombined model accounts for 23% of the deception index variance.

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

M(."$ *#)+,-#*./01%023$ 6(*$ .234#/%#-$ %2$ ($ %#&%$ +0)#$ (3-$ /'3$%"/2'8"$9:;<C$N2/$#(."$*#)+,-#*./01%023E$9:;<$1/2-'.#-$(3$2'%1'%$03-0.(%038$%"#$62/-$+/#O'#3.5$+2/$#(."$.(%#82/5$H#C8CE$+0/*%,1#/*23$ 1/232'3*E$ 3#8(%023*E$ 3#8(%04#$ #D2%023$ 62/-*E$D2%023$ 62/-*$ (3-$ #&.)'*04#$ 62/-*IC$;2/-$ +/#O'#3.0#*$ (/#$#&1/#**#-$ (*$ ($ 1#/.#3%(8#$ 2+$ %"#$ %2%()$ 3'DK#/$ 2+$ 62/-*$.23%(03#-$03$#(."$+0)#$H*##$!(K)#$PIC$

!"#$%&#'()*'*+#,$##+-)'9:;<$.(1%'/#-$Q?CQRS$2+$%"#$62/-*$03$-(%#/*T$%#&%'()$*#)+,-#*./01%023*C$ U3$ (4#/(8#E$ %"#$ *#)+,-#*./01%023*$6#/#$ PFGCPG$62/-*$ )238$ H!"$ V$ PPQCFAIE$ 60%"$ 32$ 8#3-#/$ -0++#/#3.#*$ 03$)#38%"C$$

W(/%0.01(3%*$/(%#-$%"#0/$*#)+,-#*./01%023*$(*$4#/5$(..'/(%#C$U3$%"#$P$ H#&%/#D#)5$ 03(..'/(%#I$ %2$F$ H#&%/#D#)5$(..'/(%#I$ *.()#$'*#-E$*#)+,-#*./01%023*$6#/#$/(%#-$(*$AC?R$H!"$V$>CAPE$#$%$V$AC>>E$#&'$V$FC>>IE$*'88#*%038$%"(%$-(%#/*$.23*0-#/#-$%"#D$%2$K#$()D2*%$+/##$2+$-#.#1%023*C$$

!2$(**#**$6"#%"#/$1/2+0)#$)0#*$6#/#$.2//#)(%#-$60%"$)038'0*%0.$."(38#*$ 03$ %"#$ 21#3,#3-#-$ *#)+,-#*./01%023*E$ 6#$ K'0)%$*#1(/(%#$ /#8/#**023$D2-#)*$ +2/$ %"#$ #D2%023()$ (3-$ .2830%04#$03-0.(%2/*$2+$-#.#1%023E$(3-$(3$24#/())$D2-#)$.23%(03038$K2%"$%51#*$ 2+$ 03-0.(%2/*C$ X23#$ 2+$ %"#$ 03-0.(%2/*$ 6#/#$ .2//#)(%#-$60%"$ #(."$ 2%"#/E$ 03-0.(%038$ %"(%$ D')%0.2))03#(/0%5$6(*$ 32%$ ($1/2K)#DC$

!"#$%&"'%( )*+#,-$( .$+%( /!(

;2/-$.2'3%$$ $ PFGCPG$ PPQCFA$P*%,1#/*23$*038')(/$$1/232'3*$$

:E$D5E$D#E$D03#$ ?CRRS$ AC>?S$

X#8(%023*$ X2E$3#4#/E$32%$ PCAAS$ PCLGS$X#8(%04#$#D2%023*$ Y(%#E$"'/%E$'8)5$ PCPFS$ PC==S$M&.)'*04#$62/-*$ Z'%E$60%"2'%E$ =CRAS$ PCR?S$[2%023$62/-*$$ ;()JE$D24#E$82$ PCGRS$ PCFFS$

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

*0830+0.(3%$H*##$!(K)#$=IE$1/240-038$*'112/%$+2/$YPC$Y26#4#/E$%"#$ .2#++0.0#3%$ +2/$ 3#8(%04#$ #D2%023$ 62/-*$ 6(*$ 03$ %"#$2112*0%#$-0/#.%023$%"(3$1/#-0.%#-C$$

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

<=>?(6+7$8'5@( /7;3( ( ,(MD2%023()$.'#*$ $ $ $

:,1/232'3*$ ,>C=FA$ >C>=$X#8(%023*$ >C=QP$ >C>P$X#8C$#D2%023*$ ,>C=RG$ >C>>Q$

<2830%04#$.'#*$ $ $ $34$5$%&6+#7/86 $ $ $

;2/-$.2'3%$ ,>C==Q$ >C>G$M&.)'*04#$62/-*$ >C>>F$ >CR?$[2%023$62/-*$ >C>=A$ >CQA$

*89$18/+#7/86 $ $ $;2/-$.2'3%$ ,>C==Q$ >C>A$

U4#/())$D2-#)$ $ $ $;2/-$.2'3%$ ,>C=RP$ >C>>F$:,1/232'3*$ ,>C=?R$ >C>>Q$X#8(%023*$ >CL=P$ >C>>L$X#8C$#D2%023*$ ,>C=RL$ >C>>G$

0+1-$(A3(/7+%;+5;"B$;(5$85$&&"'%(6'$99"6"$%7&(9'5(-"%8C"&7"6("%;"6+7'5&('9(;$6$,7"'%3(

$!2$ *'DD(/07#E$ %"0*$ *%'-5$ 034#*%08(%#-$6"#%"#/$-#.#1%023$ 03$23)03#$-(%038$1/2+0)#*$0*$(**2.0(%#-$60%"$)038'0*%0.$."(38#*$03$%"#$%#&%'()$-#*./01%023C$;#$#&(D03#-$6"#%"#/$.'#*$/#)(%#-$%2$%"#$ #D2%023()$ (3-$ .2830%04#$ (*1#.%*$ 2+$ -#.#1%023$ 62')-$#D#/8#$03$%"#$'30O'#$D#-0($.23%#&%$2+$23)03#$-(%038$1/2+0)#*E$6"#/#$ HPI$ )0#*$ %#3-$ %2$ 2..'/$ (*$ 1(/%$ 2+$ %"#$ 24#/())$ *#)+,1/#*#3%(%023$ /(%"#/$ %"(3$ 03$ %"#$ %#&%'()$ *#)+,-#*./01%023^$ H=I$

7

human judges. Recall that judges rated the perceivedtruthfulness of all communications dealing with abor-tion attitudes (Studies 1 through 3; N = 400 communica-tions) and the proportion of judges who believed eachcommunication to be truthful had been calculated. Weused these data to calculate a “hit rate” for our judgesand compared it to LIWC’s ability to correctly identifydeception. More specifically, a dichotomous classifica-tion was made for each communication. If the propor-tion of judges believing a particular communication tobe truthful was greater than 50%, it was defined as

“judged truthful.” The remaining communications wereconsidered “judged false.”

We then calculated the proportion of communica-tions that had been correctly identified as truthful ordeceptive by the judges and compared them with theLIWC classifications based on the general predictionequation (see above). As seen in Table 5, LIWC correctlyclassified 67% of the abortion communications and thejudges correctly classified 52% of the abortion commu-nications. These proportions were significantly different(z = 6.25, p < .0001). LIWC performed significantly betterthan chance (z = 6.80, p < .001) but the judges did not (z =.80, ns).

To correct for a potential positivity bias in judges’responses, we conducted a signal detection analysis. Wefirst converted the proportion of correct hits (identify-ing truthful communications as truthful) and the pro-portion of false positives (identifying deceptive commu-nications as truthful) to z scores. These values of .74 and.71 were converted to z scores of z = .64 and z = .55,respectively. We then calculated d ! by subtracting theproportion of false positives from the proportion of cor-rect hits (d ! = .09, ns).

In addition, the two detection strategies showed dif-ferent patterns of error. The judges were significantlymore likely to make “false positive” identifications than“false negative” identifications (71% vs. 27% of judges’errors, respectively; z = 14.67, p < .001). LIWC, in con-trast, was equally likely to make “false positive” identifica-tions and “false negative” identifications (49% vs. 51% ofLIWC errors, respectively; z = .50, ns).

Discussion

Successfully lying to another person usually involvesthe manipulation of language and the careful construc-tion of a story that will appear truthful. In addition to cre-ating a convincing story, liars also must present it in astyle that appears sincere (Friedman & Tucker, 1990).

Newman et al. / LINGUISTIC STYLE AND DECEPTION 671

TABLE 3: Predictors of Deception: Logistic Regression CoefficientsUsed in Prediction Equations

% Accuracy (Lie/ AdjustedLIWC Category Truth/Overall) R2

Predicting Study 1:Studies 2-5 combined 60/58/59* 4

Negative emotion –.268Senses .270Exclusive words .452Motion verbs –.310

Predicting Study 2:Studies 1 and 3-5 combined 61/57/59* 6

Negative emotion –.227Exclusive words .286Motion verbs –.358

Predicting Study 3:Studies 1-2 and 4-5 combined 66/69/67** 17


Predicting Study 4:Studies 1-3 and 5 combined 52/54/53 1

First-person pronouns .240Articles –.264Negative emotion –.382Exclusive words .463

Predicting Study 5:Studies 1-4 combined 53/43/48 0


General prediction equation:All 5 studies combined 59/62/61** 8

First-person pronouns .260Third-person pronouns .250Negative emotion –.217Exclusive words .419Motion verbs –.259

NOTE: LIWC = Linguistic Inquiry and Word Count. For % accuracy,the three percentages listed for each equation are (a) % of liars identi-fied accurately, (b) % of truth-tellers identified accurately, and (c)overall accuracy. For overall accuracy rates, *p < .05 and **p < .001when compared to chance performance of 50%. Coefficients in thenegative direction mean that liars used the category at a higher rate. Allcoefficients are significant at p < .05 or better.

TABLE 4: Effect Sizes and Reliability of Linguistic Predictors AcrossFive Studies

Study

1 2 3 4 5 Mean d Reliability

First person .31 .85 .75 –.02 –.24 .36 .43Third person .22 .30 .07 .24 –.21 .16 .28Negative emotion –.19 –.27 –.42 .40 –.33 –.15 .36Exclusive words .40 1.23 .91 .02 .30 .54 .55Motion verbs –.14 .09 –.31 –.40 –.29 –.20 .40

NOTE: Numbers for each study represent Cohen’s d comparing liarsand truth-tellers. Effect sizes in the negative direction mean that liarsused the linguistic category at a higher rate. Mean d is a weightedmean of these effect sizes. Reliability was calculated using Cronbach’salpha.


Table: Newman et al. (2003).


Enos et al. (2007): CSC Deception Corpus

1 Subjects answered six questions. These were manipulated sothey underperformed relative to a target on two questions,overperformed on two, and matched the target on two.

2 Subjects were then shown a comparison of their responsesrelative to the target and told that actual goal of the study wasto find people who could successfully convince a naiveinterview that they matched the target profile.

3 During these interviews (25-50 min), subjects used a footpedal to indicate whether their current statement was true orfalse.

4 The corpus includes high-quality audio and was transcribed.


Data selection

Extracting probable hot spots1 “Include segments that are responses to questions that

directly ask the subject for his or her score on a particularsection.”

2 “Include segments that respond to immediate follow-upquestions requesting a justification of the claimed score, whensuch a question is posed by the interviewer.”

3 “Omit everything else.”

Subsets• Critical: 465 sentence-like units (SUs) based on rule 1.

Feature selection yields 22 features.

• Critical-Plus: 675 SUs based on rules 1 and 2. Featureselection yields 56 features.


ResultsDecision-tree classifier

From the corpus of 9068 SUs, we thus produced two setsof CRITICAL SEGMENTS: one set of 465 based only on Rule1 (termed Critical) and one set of 675 based on Rules 1 and2 (termed Critical-Plus). Feature selection was employed toreduce the feature set to 22 features for the Critical set and 56for the Critical-Plus set.4.2. Coping with skewed class distributionsIt is well known that classification algorithms — particularlythose using decision trees, such as c4.5 [14] — can be neg-atively affected by datasets in which the class distribution isskewed (c.f [16, 17, 18]). In simple terms, this results in a biason the part of the induced decision tree toward the majority classdue to the ‘over-prevalence’ [16] of majority class examples.For CRITICAL SEGMENTS, the CSC Corpus is such a

dataset. The present sets of CRITICAL SEGMENTS contain a ma-jority of LIE examples: (67.5% for Critical, 62% for Critical-Plus). Because initial classification results on the natural classdistribution were poor but exceeded chance, we hypothesizedthat adjusting the class imbalance might allow the learner toinduce more effective rules. We follow a commonly used ap-proach to adjust the imbalance.In this approach, termed under-sampling1 [17], examples

from the majority class are eliminated in order to create a bal-anced distribution. For the Critical-Plus dataset, combinedtraining/test sets of 508 examples2 were used. Under-sampledtraining/test sets were created as follows: for each of 10 train-ing/test sets, randomly select 50 examples (25 TRUTH, 25LIE) for the test set; from the remaining examples, randomlyselect 458 (229 TRUTH, 229 LIE) for the test set. An anal-ogous approach was used with the Critical dataset, producingsets of 272 training and 30 test examples.For each dataset, the above procedure was repeated 10

times with different random seeds to account for the exclusionof some data; results reported here thus reflect average perfor-mance on 100 individual training/test sets for each dataset.

5. Results and DiscussionIn Table 1 we report classification results for the two datasets,both for the original samples (using 10-fold cross-validation)and for the under-sampled datasets, using 100 random trials asdescribed in Section 4.2. Both raw accuracy and improvementrelative to chance are reported. Given the difference in base-lines, the relative scores represent the best basis for comparisonsince these scores are normalized with respect to the baselinechance accuracy, which varies among the configurations of thedata. Performance on the original samples is poor but exceedschance: 5.8% relative to chance for the Critical-Plus dataset,1.6% for the Critical dataset. Results for the under-sampleddatasets show 22.2% relative improvement for the Critical-Plus set and 23.8% relative improvement for the Critical set.This lends support to our hypothesis with respect to the skewof the distribution: in cases where the over-prevalence of oneclass interferes with c4.5’s modeling, resampling can render thelearner more capable of producing useful rules[16, 18].There are no previous results for classification of GLOBAL

LIES and TRUTHS on the corpus to provide a standard for com-parison. Some context is provided, however, by the perfor-mance of humans at the analogous task of labeling GLOBAL

1Under-sampling is generally preferable to over-sampling; see [17]for details.2The total number of examples available after subtracting the 167

‘excess’ LIE examples is 508.

Table 1: Accuracy Classifying Global Lies and Truths

RelativeDataset Improvement Accuracy Baseline

Critical-Plus 5.8% 65.6 62.0

Critical 1.6% 68.6 67.5

Critical-Plus / Under-sampled 22.2% 61.1 50.0

Critical / Under-sampled 23.8% 61.9 50.0

LIES with respect to each section of the interview: 32 humanlisteners scored on average 47.8% versus a chance baseline of63.6%[4].An interesting aspect of these results is that performance

is slightly better for the Critical dataset than for the Critical-Plus dataset, despite the smaller size of the Critical set (272training examples in each trial, versus 414). We suspect thatthis difference is due to the increased cognitive and emotionalstakes of the questions involved: The Critical dataset containsonly subject segments that respond directly to the interviewer’smost salient questions (e.g., ‘What was your score on sectionX?’); the Critical-Plus dataset includes additional segmentsthat contextualize that question but do not respond directly toit. It is possible that the latter differ enough with respect toemotional and cognitive load to produce a less effective learnerwhen included with the smaller Critical set.5.1. Importance of critical segmentsThe findings we report here are particularly relevant to the gen-eral deception detection task since our CRITICAL SEGMENTSare those that point directly to the topic of most interest withregard to the interview: the test scores claimed by the subjects.Earlier studies have attempted the separate task of classifyingall segments in the corpus with respect to LOCAL LIES with rel-ative accuracy gains of 7–10% above chance. However, the pri-mary task embodied by the paradigm (and attempted by humanlisteners with little success in an earlier perception study [4]) isto determine the veracity of the subjects’ claims with regard totheir scores. Thus, while performance achieved here is modest,it is significant since this performance is obtained specificallyon the segments whose veracity is of greatest interest, those thatreflect the GLOBAL LIE category.We have also shown that a more powerful classifier can

be trained using resampling techniques that compensate for thecorpus’s skewed class distributions. The substantially improvedperformance indicates that the learner is better able to infermore useful rules when the present data are distributed evenly— and more importantly that such rules exist.5.2. Relevant featuresBecause the bagging/boosting approach used here in 100 trialsper dataset produced a large number of c4.5 decision trees, itis impractical to give an exhaustive description of the featuresemployed in the models. We can, however, make some generalobservations about features that applied to a large number ofcases in the induced trees.Many of the rules induced from the current dataset paint

a very plausible picture of the correlates of deception and onethat is consistent with previous literature. First, lexical cues that

Comments• ‘Under-sampled’: average over 100 random balanced

selections.

• The baselines reflect the class distributions.

• Performance is better for the Critical class. “We suspect thatthis difference is due to the increased cognitive and emotionalstakes of the questions involved.”


FeaturesGeneralizations based on inspection of the models:

Truthfulness• Positive emotion words.

• Direct denials of lying.

• Filled pauses.

• Self-repairs.

Deception• Assertive terms’ (‘yes’, ‘no’).

• Qualifiers (‘absolutely’ or ‘really’).

• “extreme values for energy — either high or low”

“A difference between our two datasets is that the presence of pasttense verbs appears to correlate with deception in the Critical-Plusdataset, while it is not employed in the Critical set.”


Looking for new data

All the deception studies I’ve seen involve private collections ofdata. The reasons for this are clear, but it’s an obstacle toresearch. Might we be able to develop deception corpora frompublicly available data?


Perjurers

I have textual data for all of the following:

1 Michael Brown (U.S. Senate; Katrina response)

2 Roland Burris (Illinois House; contacts with Blagojevich)

3 Roger Clemens (U.S. House; steroid use)

4 Bill Clinton (deposition; Paula Jones)

5 Mark Fuhrman (court transcript; use of racial epithets)

6 Scooter Libby (Grand Jury testimony; Plame affair)

7 Bernie Madoff (Courtroom transcript; Ponzi scheme)

8 Ernie Sosa (U.S. House; steroid use)

9 Tobacco execs (Waxman hearings; tobacco addiction)


Politfact’s Truth-o-meter

http://politifact.com/

http://politifact.com/


Politfact’s Truth-o-meter

http://politifact.com/wisconsin/statements/2010/oct/21/

scott-walker/scott-walker-says-scientists-agree-adult-stem-cell/

http://politifact.com/wisconsin/statements/2010/oct/21/scott-walker/scott-walker-says-scientists-agree-adult-stem-cell/

http://politifact.com/wisconsin/statements/2010/oct/21/scott-walker/scott-walker-says-scientists-agree-adult-stem-cell/


Politfact data

Joe Lieberman: In the U.S. Senate, Barack Obama “has notreached across party lines to get anything significant done.” [False]

Category Texts

True 334Mostly True 246

Half-True 298Barely True 214

False 309Pants on Fire 121

Total 1,522

21,391 quote tokens; 3,782 types; mean quote length: 14

References I

Chung, Cindy and James W. Pennebaker. 2007. The psychological function of functionwords. In Klaus Fiedler, ed., Social Communication, 343–359. New York: PsychologyPress.

DePaulo, Bella M.; James J. Lindsay; Brian E. Malone; Laura Muhlenbruck; Kelly Charlston;and Harris Cooper. 2003. Cues to deception. Psychological Bulletin 129(1):74–118.

Ekman, Paul. 1985. Telling Lies: Clues to Deceit in the Marketplace, Politics, and Marriage.New York: Norton.

Ekman, Paul and Maureen O’Sullivan. 1991. Who can catch a liar? American Psychologist46(9):913–920.

Enos, Frank; Elizabeth Shriberg; Martin Graciarena; Julia Hirschberg; and Andreas Stolcke.2007. Detecting deception using critical segments. In Proceedings Interspeech,1621-1624, 1621–1624. ACL.

Frankfurt, Harry G. 1988. The Importance of What we Care About. Cambridge UniversityPress.

Kasl, Stanislav V. and George F. Mahl. 1965. The relationship of disturbances andhesitations in spontaneous speech to anxiety. Journal of Personality and SocialPsychology 1:425–433.

Larcker, David F. and Anastasia A. Zakolyukina. 2010. Detecting deceptive discussions inconference calls. Ms., Stanford University.

Newman, Matthew L.; James W. Pennebaker; Diane S. Berry; and Jane M. Richard. 2003.Lying words: Predicting deception from linguistic styles. Journal of Language and SocialPsychology 29(5):665–675.

References II

Ranganath, Rajesh; Daniel Jurafsky; and Daniel McFarland. 2009. It’s not you, it’s me:Detecting flirting and its misperception in speed-dates. In Proceedings of the 2009Conference on Empirical Methods in Natural Language Processing, 334–342. ACL.

Rhees, Rush, ed. 1984. Recollections of Wittgenstein. Oxford University Press.Solan, Lawrence M. and Peter M. Tiersma. 2005. Speaking of Crime: The Language of

Criminal Justice. Chicago, IL: University of Chicago Press.Toma, Catalina L and Jeffrey T. Hancock. 2010. Reading between the lines: Linguistic cues

to deception in online dating profiles. In Proceedings of the 2010 ACM Conference onComputer Supported Cooperative Work, 5–8. ACM.

Toma, Catalina L; Jeffrey T. Hancock; and Nicole B. Ellison. 2007. The truth about lying inonline dating profiles. In Proceedings of Computer–Human Interaction 2007, 449–452.ACM.

Toma, Catalina L; Jeffrey T. Hancock; and Nicole B. Ellison. 2008. Separating fact fromfiction: An examination of deceptive self-presentation in online dating profiles.Personality and Social Psychology Bulletin 1023–1036.

Deceptive languageweb.stanford.edu/class/...handout-11-02-deception.pdf · Overview On deception...

Documents

Transcript of Deceptive languageweb.stanford.edu/class/...handout-11-02-deception.pdf · Overview On deception...