Spoken Dialog with Humans and Machines Context and Prosody in the Interpretation of Cue Phrases in Dialogue Julia Hirschberg Columbia University and KTH 11/22/07 In collaboration with Agustín Gravano, Stefan Benus, Héctor Chávez, Shira Mitchell, and Lauren Wilcox With thanks to Gregory Ward and Elisa Sneed German 2 Managing Conversation How do speakers indicate conversational structure in human/human dialogue? How do they communicate varying levels of attention, agreement, acknowledgment? What role does lexical choice play in these communicative acts? Phonetic realization? Prosodic variation? Prior context? Can human/human behavior be modeled in Spoken Dialogue Systems? 3 Cue Phrases/Discourse Markers/Cue Words/ Discourse Particles/Clue Words Linguistic expressions that can be employed to convey information about the discourse structure, or to make a semantic (literal?) contribution. Examples: now, well, so, alright, and, okay, first, on the other hand, by the way, for example, … 4 Some Examples that’s pretty much okay Speaker 1: between the yellow mermaid and the whale Speaker 2: okay Speaker 1: and it is okay we gonna be placing the blue moon 5 A Problem for Spoken Dialogue systems How do speakers produce and hearers interpret such potentially ambiguous terms? How important is acoustic/prosodic information? Phonetic variation? Discourse context? 6 Research Goals Learn which features best characterize the different functions of single affirmative cue words. Determine how these can be identified automatically. Important in Spoken Dialogue Systems: Understand user input. Produce output appropriately. 7 Overview Previous research The Columbia Games Corpus Collection paradigm Annotations Perception Study of Okays Experimental design Analysis and results Machine Learning Experiments on Okay Future work: Entrainment and Cue Phrases 8 Previous Work General studies Schriffin ’82, ‘87; Reichman ’85; Grosz & Sidner ‘86 Cues to cue phrase disambiguation Hirschberg & Litman ’87, ’93; Hockey ’93; Litman ’94 Cues to Dialogue Act identification Jurafsky et al ’98; Rosset & Lamel ’04 Contextual cues to the production of backchannels Ward & Tsukahara ’00; Sanjanhar & Ward ’06 9 The Columbia Games Corpus Collection 12 spontaneous task-oriented dyadic conversations in Standard American English (9h 8m speech) 2 subjects playing a series of computer games, no eye contact (45m 39s mean session time) 2 sessions per subject, w/different partners Several types of games, designed to vary the way discourse entities became old, or ‘given’ in the discourse to study variation in intonational realization of information status 10 Cards Game #1 Player 1 (Describer) Player 2 (Searcher) • Short monologues • Vary frequency and order of occurrence of objects on the cards. 11 Cards Game #2 Player 1 (Describer) Player 2 (Searcher) • Dialogue • Vary frequency and order of occurrence of objects on the cards across speakers. 12 Objects Game Follower must place the target object where it appears on the Describer’s screen solely via the description provided (4h 19m) Describer: Follower: 13 The Columbia Games Corpus Recording and Logging Recorded on separate channels in soundproof booth, digitized and downsampled to 16k All user and system behaviors logged 14 The Columbia Games Corpus Annotation Orthographic transcription and alignment (~73k words). Laughs, coughs, breaths, smacks, throat-clearings. Self-repairs. Intonation, using ToBI conventions. Function (10 categories) of affirmative cue words (alright, mm-hm, okay, right, uh-huh, yeah, yes, …). Question form and function. Turn-taking behaviors. 15 Perception Study Selection of Materials Acknowledgment Agreement Speaker 1: yeah um there's like there's/ some space there's Speaker 2: okay I think I got it Backchannel okay Cue beginning discourse segment Speaker 1: but it's gonna be below the onion Speaker 2: okay Speaker 1: okay alright I'll try it okay Speaker 2: okay the owl is blinking 18 Perception Study Experiment Design 54 instances of ‘okay’ (18 for each function). 2 tokens for each ‘okay’: Isolated condition: Only the word ‘okay’. Contextualized condition: 2 full speaker turns: The turn containing the target ‘okay’; and The previous turn by the other speaker. speakers okay contextualized ‘okay’ 19 Perception Study Experiment Design 1/3 each: 3 labelers agreed, 2…, none Two conditions: Part 1: 54 isolated tokens Part 2: 54 contextualized tokens Subjects asked to classify each token of ‘okay’ as: Acknowledgment / Agreement, or Backchannel, or Cue beginning discourse segment. 20 Perception Study Definitions Given to the Subjects Acknowledge/Agreement: The function of okay that indicates “I believe what you said” and/or “I agree with what you say”. Backchannel: The function of okay in response to another speaker's utterance that indicates only “I’m still here” or “I hear you and please continue”. Cue beginning discourse segment The function of okay that marks a new segment of a discourse or a new topic. This use of okay could be replaced by now. 21 Perception Study Subjects and Procedure Subjects: 20 paid subjects (10 female, 10 male). Ages between 20 and 60. Native speakers of English. No hearing problems. GUI on a laboratory workstation with headphones. 22 Results: Inter-Subject Agreement Kappa measure of agreement with respect to chance (Fleiss ’71) Isolated Condition Contextualized Condition Overall .120 .294 Ack / Agree vs. Other .089 .227 Backchannel vs. Other .118 .164 Cue beginning vs. Other .157 .497 23 Results:Cues to Interpretation Phonetic transcription of okay: Isolated Condition Strong correlation for realization of initial vowel Backchannel Ack/Agree, Cue Beginning Contextualized Condition No strong correlations found for phonetic variants. 24 Results: Cues to Interpretation Isolated Condition Contextualized Condition Shorter /k/ Shorter latency between turns Shorter pause before okay Higher final pitch slope Longer 2nd syllable Lower intensity Higher final pitch slope More words by S2 before okay Fewer words by S1 after okay Lower final pitch slope Lower overall pitch slope Lower final pitch slope Longer latency between turns More words by S1 after okay Ack / Agree Backchannel Cue beginning S1 = Utterer of the target ‘okay’. S2 = The other speaker. 25 Results: Cues to Interpretation Phrase-final intonation (ToBI) (Both isolated and contextualized conditions.) H-H% Backchannel H-L% L-H% Ack/Agree, Backchannel L-L% Ack/Agree, Cue beginning 26 Perception Study: Conclusions Agreement: Availability of context improves inter-subject agreement. Cue beginnings easier to disambiguate than the other two functions. Cues to interpretation: Contextual features override word features Exception: Final pitch slope of okay in both conditions. 27 Machine Learning Experiments: Okay Can we identify the different functions of okay in our larger corpus reliably? What features perform best? How do these compare to those that predict human judgments? 28 Method ML Algorithm JRip: Weka’s implementation of the propositional rule learner Ripper (Cohen ’95). We also tried J4.8, Weka’s implementation of the decision tree learner C4.5 (Quinlan ’93, ’96), with similar results. 10-fold cross validation in all experiments. 29 Units of Analysis IPU (Inter-pausal unit) Maximal sequence of words delimited by pause > 50ms. Conversational Turn Maximal sequence of IPUs by the same speaker, with no contribution from the other speaker. 30 Experimental features Text-based features (from transcriptions) Word ident, POS tags (auto); position of word in IPU / turn IPU, turn length in words; prev turn same spkr? Timing features (from time alignment) Word / IPU / turn duration; amount of spkr overlap Time to word beg/end in IPU, turn Acoustic features {min, mean, max, stdev} x {pitch, intensity} Slope of pitch, stylized pitch, and intensity, over the whole word, and over its last 100, 200, 300ms. Acoustic features from last IPU of prior speaker’s turn. 31 Results: Classification of individual words Classification of each individual word into its most common functions. alright Ack/Agree, Cue Begin, Other mm-hm Ack/Agree, Backchannel okay Ack/Agree, Backchannel, Cue Begin, Ack+CueBegin, Ack+CueEnd, Other right Ack/Agree, Check, Literal Modifier yeah Ack/Agree, Backchannel 32 Results: Classification of ‘okay’ Feature Set Error Rate Majority Label F-Measure Ack / BackCue Ack/Agree + Ack/Agree + Cue End Agree channel Begin Cue Begin 1137 121 548 68 232 Text-based 31.7 .76 .16 .77 .09 .33 Acoustic 40.2 .69 .24 .64 .03 .25 Text-based + Timing 25.6 .79 .31 .82 .18 .67 Full set 25.5 .80 .46 .83 .21 .66 Baseline (1) 48.3 .68 .00 .00 .00 .00 Human labelers (2) 14.0 .89 .78 .94 .56 .73 (1) Majority class baseline: ACK/AGREE. (2) Calculated wrt each labeler’s agreement with the majority labels. 34 Conclusions: ML Experiments Context and timing features Like perception in context results: timing Pause after okay, not before # of succeeding words Acoustic features impoverished No phonetic features No pitch slope But ToBI labels (where available) didn’t help 35 Future Work Experiments with full ToBI labeling Other features Lexical, Acoustic-Prosodic, and Discourse Entrainment and Dis-Entrainment Positive correlations for affirmative cue words Affirmative cue word entrainment and game scores Affirmative cue word entrainment and overlaps and interruptions in turn-taking 36 Tack! Other Work Benus et al, 2007 “The prosody of backchannels in American English”, ICPhS 2007, Saarbrücken, Germany, August 2007. Gravano et al, 2007 “Classification of discourse functions of affirmative words in spoken dialogue”, Interspeech 2007, Antwerp, Belgium, August 2007. 38 Importance for Spoken Dialogue Systems Convey ambiguous terms with the intended meaning Interpret the user’s input correctly 39 Experiment Design Goal: Study the relation between the downstepped contour and Information status Syntactic position Discourse position Spontaneous speech Both monologue and dialogue 40 Experiment Design Three computer games. Two players, each on a different computer. They collaborate to perform a common task. Totally unrestricted speech. 41 Objects Game Player 1 (Describer) Player 2 (Searcher) • Dialogue • Vary target and surrounding objects (subject and object position). 42 Games Session Repeat 3 times: Cards Game #1 Cards Game #2 Short break (optional) Repeat 3 times: Objects Game Each subject participated in 2 sessions. 12 sessions 43 Subjects Postings: Columbia’s webpage for temporary job adds. Craig’s list http://www.craigslist.org Category: Gigs Event gigs Problem: People are unreliable ~50% did not show up, or cancelled with short notice. 44 Subjects Possible solutions: Give precise instructions to e-mail ALL required info: Name, native speaker?, hearing impairments?, etc. Ask for a phone number. Call them and explain why it is so important for us that they show up (or cancel with adecuate notice). Increase the pay after each session. Example: $5, $10, $15 instead of $10, $10, $10. 45 Recording Sound-proof booth 2 subjects + 1 or 2 confederates. Head-mounted mics. Digital Audio Tape (DAT): one channel per speaker. Wav files One mono file per speaker. Sample rate: 48000 Downsampled to 16000 (but kept original files!) ~20 hours of speech 2.8 GB (16k) 46 Logs Log everything the subjects do to a text file. Example: 17:03:55:234 17:04:04:868 17:04:31:837 17:04:38:426 17:05:03:873 ... BEGIN_EXECUTION NEXT_TURN RESULTS 97 points awarded. NEXT_TURN RESULTS 92 points awarded. Later, this may be used (e.g.) to divide each session into smaller tasks or conversations. 47