Deceptive Speech Frank Enos • April 19, 2006 Defining Deception Deliberate choice to mislead a target without prior notification (Ekman‘’01) Often to gain some advantage Excludes: Self-deception Theater, etc. Falsehoods due to ignorance/error Pathological behaviors Why study deception? Law enforcement / Jurisprudence Intelligence / Military / Security Business Politics Mental health practitioners Social situations Is it ever good to lie? Why study deception? What makes speech “believable”? Recognizing deception means recognizing intention. How do people spot a liar? How does this relate to other subjective phenomena in speech? E.g. emotion, charisma Problems in studying deception? Most people are terrible at detecting deception — ~50% accuracy (Ekman & O’sullivan 1991, Aamodt 2006, etc.) People use subjective judgments — emotion, etc. Recognizing emotion is hard People Are Terrible At This Group #Studies #Subjects Accuracy % Criminals 1 52 65.40 Secret service 1 34 64.12 Psychologists 4 508 61.56 Judges 2 194 59.01 Cops 8 511 55.16 Federal officers 4 341 54.54 122 8,876 54.20 Detectives 5 341 51.16 Parole officers 1 32 40.42 Students Problems in studying deception? Hard to get good data Real world (example) Laboratory Ethical issues Privacy Subject rights Claims of success But also ethical imperatives: Need for reliable methods Debunking faulty methods False confessions 20th Century Lie Detection Polygraph http://antipolygraph.org The Polygraph and Lie Detection (N.A.P. 2003) Voice Stress Analysis Microtremors 8-12Hz Universal Lie response http://www.love-detector.com/ http://news-info.wustl.edu/news/page/normal/669.html Reid Behavioral Analysis Interview Interrogation An Example… Frank Tells Some Lies Frank Tells Some Lies Maria: I’m buying tickets to Händel’s Messiah for me and my friends — would you like to join us? Frank: When is it? Maria: December 19th. Frank: Uh… the 19th… Maria: My two friends from school are coming, and Robin… Frank: I’d love to! How to Lie (Ekman‘’01) Concealment Falsification Misdirecting Telling the truth falsely Half-concealment Incorrect inference dodge. Frank Tells Some Lies Maria: I’m buying tickets to Handel’s Messiah for me and my friends — would you like to join us? Frank: When is it? Maria: December 19th. Frank: Uh… the 19th… Maria: My two friends from school are coming, and Robin… Frank: I’d love to! • Concealment • Falsification • Misdirecting • Telling the truth falsely • Half-concealment • Incorrect inference dodge. Reasons To Lie (Frank‘’92 ) Self-preservation Self-presentation *Gain Altruistic (social) lies How Not To Lie (Ekman‘’01) Leakage Part of the truth comes out Liar shows inconsistent emotion Liar says something inconsistent with the lie Deception clues Indications that the speaker is deceiving Again, can be emotion Inconsistent story How Not To Lie (Ekman‘’01) Bad lines Lying well is hard Fabrication means keeping story straight Concealment means remembering what is omitted All this creates cognitive load harder to hide emotion Detection apprehension (fear) Target is hard to fool Target is suspicious Stakes are high Serious rewards and/or punishments are at stake Punishment for being caught is great How Not To Lie (Ekman‘’01) Deception guilt Stakes for the target are high Deceit is unauthorized Liar is not practiced at lying Liar and target are acquainted Target can’t be faulted as mean or gullible Deception is unexpected by target Duping delight Target poses particular challenge Lie is a particular challenge Others can appreciate liar’s performance Features of Deception Cognitive Coherence, fluency Interpersonal Discourse features: DA, turn-taking, etc. Emotion Describing Emotion Primary emotions Acceptance, anger, anticipation, disgust, joy, fear, sadness, surprise One approach: continuous dim. model (Cowie/Lang) Activation – evaluation space Add control/agency Primary E’s differ on at least 2 dimensions of this scale (Pereira) Problems With Emotion and Deception Relevant emotions may not differ much on these scales Othello error People are afraid of the police People are angry when wrongly accused People think pizza is funny Brokow hazard Failure to account for individual differences Bulk of extant deception research… Not focused on verifying 20th century techniques Done by psychologists Considers primarily facial and physical cues “Speech is hard” Little focus on automatic detection of deception Modeling Deception in Speech Lexical Prosodic/Acoustic Discourse Deception in Speech (Depaulo ’03) Positive Correlates Interrupted/repeated words References to “external” events Verbal/vocal uncertainty Vocal tension F0 Deception in Speech (Depaulo ’03) Negative Correlates Subject stays on topic Admitted uncertainties Verbal/vocal immediacy Admitted lack of memory Spontaneous corrections Problems, revisited Differences due to: Gender Social Status Language Culture Personality Columbia/SRI/Colorado Corpus With Julia Hirschberg, Stefan Benus, and colleagues from SRI/ICSI and U. C. Boulder Goals Examine feasibility of automatic deception detection using speech Discover or verify acoustic/prosodic, lexical, and discourse correlates of deception Model a “non-guilt” scenario Create a “clean” corpus Columbia/SRI/Colorado Corpus Inflated-performance scenario Motivation: financial gain and self-presentation 32 Subjects: 16 women, 16 men Native speakers of Standard American English Subjects told study seeks to identify people who match profile based on “25 Top Entrepreneurs” Columbia/SRI/Colorado Corpus Subjects take test in six categories: Interactive, music, survival, food, NYC geography, civics Questions manipulated 2 too high; 2 too low; 2 match Subjects told study also seeks people who can convince interviewer they match profile Self-presentation + reward Subjects undergo recorded interview in booth Indicate veracity of factual content of each utterance using pedals CSC Corpus: Data 15.2 hrs. of interviews; 7 hrs subject speech Lexically transcribed & automatically aligned lexical/discourse features Lie conditions: Global Lie / Local Lie Segmentations (LT/LL): slash units (5709/3782), phrases (11,612/7108), turns (2230/1573) Acoustic features (± recognizer output) Columbia University– SRI/ICSI – University of Colorado Deception Corpus: An Example Segment SEGMENT TYPE Breath Group LABEL LIE Obtained from subject pedal presses. um i was visiting a friend in venezuela and we went camping ACOUSTIC FEATURES max_corrected_pitch mean_corrected_pitch Produced using ASR output and other acoustic analyses pitch_change_1st_word pitch_change_last_word normalized_mean_energy unintelligible_words 5.7 5.3 -6.7 -11.5 0.2 0.0 LEXICAL FEATURES has_filled_pause positive_emotion_word uses_past_tense PREDICTION negative_emotion_word contains_pronoun_i verbs_in_gerund YES YES NO LIE NO YES YES Produced automatically using lexical transcription. CSC Corpus: Results Classification (Ripper rule induction, randomized 5-fold cv) Slash Units / Local Lies — Baseline 60.2% Lexical & acoustic: 62.8 %; + subject dependent: 66.4% Phrases / Local Lies — Baseline 59.9% Lexical & acoustic 61.1%; + subject dependent: 67.1% Other findings Positive emotion words deception (LIWC) Pleasantness deception (DAL) Filled pauses truth Some pitch correlation — varies with subject Example JRIP rules: (cueLieToCueTruths >= 2) and (TOPIC = topic_newyork) and (numSUwithFPtoNumSU <= 0) and (wu_ENERGY_NO_UV_STY_MAX__EG_ZNORMENERGY_NO_UV_STY_MIN__EG_ZNORM-D <= 5.846) => PEDAL=L (231.0/61.0) (cueLieToCueTruths >= 2) and (numSUwithFPtoNumSU <= 1) and (wu_ENERGY_NO_UV_STY_MAX__EG_ZNORMENERGY_NO_UV_STY_MIN__EG_ZNORM-D <= 5.68314) and (wu_ENERGY_NO_UV_RAW_MAX-ENERGY_NO_UV_RAW_MIN-D >= 8.41605) and (wu_F0_SLOPES_NOHD__LAST >= -2.004) => PEDAL=L (284.0/117.0) (cueLieToCueTruths >= 2) and (wu_F0_RAW_MAX >= 5.706379) and (wu_DUR_PHONE_SPNN_AV <= 1.0661) => PEDAL=L (262.0/115.0) CSC Corpus: A Perception Study With Julia Hirschberg, Stefan Benus, Robin Cautin and colleagues from SRI/ICSI 32 Judges Each judge rated 2 interviews Judge Labels: Local Lie using Praat Global Lie on paper Takes pre- and post-test questionnaires Personality Inventory Judge receives ‘training’ on one subject. By Judge 58.2% Acc. By Interviewee Personality Measure: NEO-FFI Costa & McCrae (1992) Five-factor model Openness to Experience Conscientiousness Extraversion Agreeability Neuroticism Widely used in psychology literature Neuroticism, Openness & Agreeableness correlate with judge performance WRT Global lies. These factors also provide strongly predictive models for accuracy at global lies. Other Perception Findings No effect for training Judges’ post-test confidence did not correlate with pre-test confidence Judges who claimed experience had significantly higher pre-test confidence But not higher accuracy! Many subjects used disfluencies as cues to D. In this corpus, disfluencies correlate with TRUTH! (Benus et al. ‘06) Our Future Work Individual differences Wizards of deception Predicting Global Lies Local lies as ‘hotspots’ New paradigm Shorter Addition of personality test for speakers Addition of cognitive load