Discourse Markers Discourse Dialogue CS 35900 November 25

Discourse Markers Discourse & Dialogue CS 35900 November 25, 2006

Agenda • Definition • Motivation • Understanding Discourse Markers in Human-Human Dialogue • Using Discourse Markers in HCI – Structuring instructions – Improving recognition – Recognizing conversational and dialogue acts

Discourse Markers • “Cue”/”Clue” phrases – E. g. “By the way”, “so”, “now”, “anyway” • Signal discourse structure – E. g. asides, new topics, returns • Signal relations between segments – Cohesion, continuation, contrast, etc • Appropriate use clarifies dialogue – Misuse causes confusion

Lost in the Mail (Application) • U: Read old messages. • S: Skipping to first old message. Message 1 from Shaw Wen Yang, Subject: Question • U: Read it to me. • S: Call me about the tool, shaw x 22630 • U: Send a reply. • S: Preparing message to shaw. Do you want to include the current message? • U: Yes • S: Begin recording after the tone……Say cancel, send or review. • U: Send • S: Message sent. • U: <silence>…. <touch tones>. . .

Why Discourse Markers? • Prevalent: – Begin 65% human-human task utterances • 44% excluding acknowledgments • Related to discourse meaning/structure • How do linguistic features of DMs interact with discourse properties to create function? – What is the scope? – At what level is relationship marked? – How do meanings contribute to function?

Discourse Relations • Markers apply on different planes of talk – exchange, action, idea, participation framework, information status • E. g. ‘oh’: primary: info state; 2 nd: action, participation • Semantics and syntax contribute to role of DM – E. g. “so” -> result – Mark or disambiguate relations • Don’t create relations

Levels of Contribution • Wide scope -> less communicative force – Narrow scope -> more communicative force • More markers -> decrease individual impact – Clearer discourse structure -> fewer DMs

Indexical Relations • Utterance situated with respect to – Speaker, hearer, time, place • Discourse markers related to context – Proximal (near) vs distal (far) relations – Linkage to preceding/following utterance • preceding - “oh” - proximal; “well”: link pre & post – Speaker - proximal =“oh”; hearer- distal

Discourse Markers & Coherence • Establish coherence by – Locate utterance on “plane of talk” • Exchange, action, idea, participation frame, info status, etc – Turn-taking, task step, semantics, attitude, etc – Index local context reference for interpretation • Establish direction of relation to surrounding utts

A Solution • U: Read old messages. • S: Skipping to first old message. Message 1 from Shaw Wen Yang, Subject: Question • U: Read it to me. • S: Call me about the tool, shaw x 22630 • U: Send a reply. • S: Preparing message to shaw. Do you want to include the current message? • U: Yes • S: Begin recording after the tone……Say cancel, send or review. • U: Send • S: Message sent. What now? • U: Next message.

Generating Discourse Markers • Integrate intentional & informational – Grosz & Sidner; Mann & Thompson – Identify nuclear DSP and contributing relations • Pairwise relations of rhetorical types (intent) – E. g. concession: nucleus • Relations based on task/domain (inform) – E. g. step: prev-result – Interpret cues wrt discourse structure/relations • “since”: contributor: nucleus; ”because”: nucleus: contributor • No duplication of cues within embedded relations – Duplicate in sequence

Characterizing Discourse Markers • Problem: Ambiguity – Discourse use vs sentential use • E. g. “now”: Topic initiation vs temporal meaning – Overall: 1/3 ambiguous; coord conj: ½ • Disambiguation: – Prosody: 84% (non-conj: 93%) • DM: own intermediate phrase; or first, no accent • Sentence: no separation, H* or complex accent – Text: 89% • DM: preceding punctuation – POS weaker cue

Improving Recognition • Discourse markers as special case of POS tagging – POS = part of speech • E. g. noun, verb, conjunction, etc – Discourse marker POS: • • Acknowledgment: “okay”; “uh-huh” Interjection DM: “oh”, ”well” Conjunction DM: “and”, ”but” Adverb DM: “now”, ”then”

Recognizing Markers • Build joint model of word+POS recognition – Expand ASR model • Build decision trees to identify equivalences – Handle sparseness • Build binary classification trees – Successive merging with least information loss • Apply to POS and word+POS pairs – Cluster unambiguously • Joint modeling improves POS tagging – Additional discourse features further improve • Boundary tones, repairs, silence

Discourse Markers & Structure • Discourse markers correlated with conversational moves – E. g. “so”- summarize; “well” - dissent • Discourse markers NOT correlated with subsequent speech acts – Correlated with PRIOR talk • Previous turn initiates adjacency pair -> no DM • Previous turn concludes adjacency pair -> DM • Clear expectation: no DM; unclear -> DM

Discourse Markers • Short cue word, phrases that signal – relation of utterance to its context – locate utterance in the “plane” of talk • Important role in disambiguating meanings – Signal shift in topic – Signal changes in footing – Key in loosely organized contexts