Using Word-level Features to Better Predict Student Emotions during Spoken Tutoring Dialogues Mihai Rotaru Diane J. Litman DoD Group Meeting Presentation Introduction Why is important to detect/handle emotions? Emotion annotation Classification task Previous work 2 (Spoken) Tutoring dialogues Education Classroom setting Human (one-on-one) tutoring Computer tutoring (ITS – Intelligent Tutoring Systems) Addressing the learning gap between human and computer tutoring Dialogue-based ITS (Ex: Why2) Improve the language understanding module of ITS Incorporate affective reasoning Connection between learning and student emotional state Adding human-provided emotional scaffolding to a reading tutor increases student persistence (Aist et al., 2002) 3 Human-Computer Excerpt Tutor26: Let me ask my question in a more general way. What happens to the velocity of a body when there is no force acting on it? Student27: dammit (ASR: it is) Tutor28 : Could you please repeat that? Student29 : same (ASR: i same) Tutor30 : Right. And what happens to the HORIZONTAL velocity of a body when there is no HORIZONTAL force acting on it? Student31: zero (ASR: the zero) Tutor32 : Well... OK, let's consider how forces in the vertical direction affect the motion of a body in a horizontal direction in a different scenario <…omitted…> Student33: oh gravity you already said this (ASR: beats gravity you're exert less) Tutor34: Fine. Are there any other forces acting on the apple as it falls? Student35: no why are you doing this again (ASR: no y and to it yes) Tutor36: Correct. The only force acting on the apple as it falls to the earth is the earth's gravitational force. In which direction does gravitational force act? Student37: downward you computer (ASR: downward you computer) Affective reasoning Prerequisites Dialogue-based ITS : Why2 Interaction via speech : ITSPOKE (Intelligent Tutoring SPOKEn dialogue system) Affective reasoning Detect student emotions Handle student emotions 5 • Back-end is Why2-Atlas system (VanLehn et al., 2002) • Sphinx2 speech recognition and Cepstral text-to-speech 6 • Back-end is Why2-Atlas system (VanLehn et al., 2002) • Sphinx2 speech recognition and Cepstral text-to-speech 7 • Back-end is Why2-Atlas system (VanLehn et al., 2002) • Sphinx2 speech recognition and Cepstral text-to-speech 8 Student emotions Emotion annotation Perceived, intuitive expressions of emotion Relative to other turns in context and tutoring task 3 Main emotion classes Negative - e.g. uncertain, bored, irritated, confused, sad; (question turns) Positive - e.g. confident, enthusiastic Neutral - no strong expression of negative or positive emotion; (grounding turns) Corpora Human-Human (453 student turns from 10 dialogues) Human-Computer (333 student turns from 15 dialogues) 9 Annotation example Tutor: Uh let us talk of one car first. Student: ok. (EMOTION = NEUTRAL) Tutor: If there is a car, what is it that exerts force on the car such that it accelerates forward? Student: The engine. (EMOTION = POSITIVE) Tutor: Uh well engine is part of the car, so how can it exert force on itself? Student: um… (EMOTION = NEGATIVE) 10 Classification task 3 Levels of Annotation Granularity NPN - Negative, Positive, Neutral NnN - Negative, Non-Negative EnE - Emotional, Non-Emotional positives and neutrals are conflated as Non-Negative negatives and positives are conflated as Emotional neutrals are Non-Emotional useful for triggering system adaptation (HH corpus analysis) Agreed subset Predict the class of each student turn 11 Previous work - Features Human-Human 5 feature types 3 feature types amplitude, pitch, duration Acoustic-prosodic Lexical Other automatic Manual Identifiers Combinations Human-Computer Acoustic-prosodic amplitude, pitch, duration Lexical Other automatic Manual Identifiers Combinations Current turn Contextual Local – previous two turns Global – all turns so far 12 Previous work - Results HH EnE HC EnE Kappa 0.55 0.30 Baseline 51.71% 58.64% Accuracy 88.86% 72.91% Rel. improv. 76.93% 34.50% Litman and Forbes, ACL 2004 13 How to improve? Use word-level features instead of turn-level features Extend the pitch features set Simplified word-level emotion model 14 Why word-level features? Emotion might not be expressed over the entire turn “This is great” Angry Happy 15 Why word-level features? (2) Can approximate pitch contour better at sub-turn levels. Especially for longer turns 350 300 250 200 150 100 50 This is great 16 Extended pitch features set Previous work Min, Max Avg, Stdev Extend with Start, End Regression coefficient and regression error Quadratic regression coefficient from Batliner et al. 2003 17 But wait… Features Student turn 321654615, asdakd, 342.234234 Asdhkas, a34334, 324,7657755 Machine learning Turn emotional class Turn-level Word-level Word 1 321654615, asdakd, 342.234234 Asdhkas, a34334, 324,7657755 … … Word n 321654615, asdakd, 342.234234 Asdhkas, a34334, 324,7657755 ? Turn emotional class 321654615, asdakd, 342.234234 Asdhkas, a34334, 324,7657755 Sönmez et al., 1998 18 Word-level emotion model Features Student turn 321654615, asdakd, 342.234234 Asdhkas, a34334, 324,7657755 Machine learning Turn emotional class Turn-level Word-level Word 1 321654615, asdakd, 342.234234 Asdhkas, a34334, 324,7657755 … … Word n 321654615, asdakd, 342.234234 Asdhkas, a34334, 324,7657755 Word-level emotion … Turn emotional class Word-level emotion 19 Word-level emotion model Training phase Each word labeled with turn class Extra features to identify the position of the word in the turn (distance in words from the beginning and end of the turn) Learn emotion model at the word level Test phase Predict each word class based on the learned model Use majority/weighted voting to label the turn based on its word classes Ties are broken randomly 20 Questions to answer Will word level feature work better than turn level features for emotion prediction? If yes, where does the advantage comes from? Yes Better prediction of longer turns Is there a feature set that offers robust performance? Yes. Combination of pitch and lexical features at word level. 21 Experiments EnE classification, agreed turns Two contrasting corpora Two contrasting learners (WEKA) IB1 – nearest neighbor classifier ADA – boosted decision trees 22 Feature sets Only pitch and lexical features 6 sets of features Turn level: Word level: Lex-Turn – only lexical Pitch-Turn – only pitch PitchLex-Turn – lexical and prosodic Lex-Word – only lexical + positional Pitch-Word – only pitch + positional PitchLex-Word – lexical and prosodic + positional Baseline: majority class 10 x 10 cross validation 23 Results – IB1 on HH Word-level features significantly outperform turn-level features Word-level better than turn-level on longer turns Best performers: Lex-Word, PitchLex-Word 100% 90% 85% 90% Lex-Turn 80% Lex-Word 75% 80% 70% Pitch-Turn 65% Pitch-Word 70% 60% 55% PitchLexTurn PitchLexWord 60% 50% Baseline Lex Pitch PitchLex Turn level (square-pattern bars), Word level (no-pattern bars) 50% single short medium long 24 Results – ADA on HH Turn-level performance increases a lot Word-level significantly better than turn-level on features sets with pitch Word-level better than turn-level on longer turns but the difference is smaller Best performers: Lex-Turn, Lex-Word, PitchLex-Word 100% 90% 85% 90% 80% Lex-Turn 75% Lex-Word 80% 70% Pitch-Turn 65% 70% Pitch-Word 60% 55% PitchLexTurn PitchLexWord 60% 50% Baseline Lex Pitch PitchLex Turn level (square-pattern bars), Word level (no-pattern bars) 50% single short medium long 25 Results – IB1 on HC Word-level features significantly outperform turn-level features Lexical information less helpful than on HH corpus Word-level better than turn-level on longer turns Best performers: Pitch-Word, PitchLex-Word 75% 90% 70% 80% Lex-Turn 65% Lex-Word 70% Pitch-Turn 60% 60% Pitch-Word 55% PitchLexTurn PitchLexWord 50% 50% Baseline Lex Pitch PitchLex Turn level (square-pattern bars), Word level (no-pattern bars) 40% 1 2 3 more3 26 Results – ADA on HC Difference not significant anymore IB1 better than ADA on word-level features ADA has bigger variance on this corpus Word-level better than turn-level on longer turns but the difference is smaller Best performers: Pitch-Turn, Pitch-Word, PitchLex-Turn, PitchLex-Word 75% 90% 70% 80% Lex-Turn 65% Lex-Word 70% Pitch-Turn 60% 60% Pitch-Word 55% PitchLexTurn PitchLexWord 50% 50% Baseline Lex Pitch PitchLex Turn level (square-pattern bars), Word level (no-pattern bars) 40% 1 2 3 more3 27 Discussion Lexical features at turn and word-level are similar Pitch features differ significantly Performance dependent on corpus and learner Word-level better than turn-level (4/6) PitchLex-Word a consistent best performer Our best accuracies comparable with previous work 28 Conclusions & Future work Word-level better than turn-level for emotion prediction Even under a very simple word-level emotion model Word-level better at predicting longer turns PitchLex-Word a consistent best performer Future work: More refined word-level emotion models HMMs Co-training Filter irrelevant words Use the prosodic information left out See if our conclusions generalize on detecting student uncertainty Experiment with other sub-turn units (breath groups) 29