Whats New Discourse Dialogue CMSC 35900 1 October
What’s New? Discourse & Dialogue CMSC 35900 -1 October 28, 2004
Agenda • Attention and Information • Given/New dichotomy • Implications & Applications – Given/New-based paraphrase – Speech recognition & synthesis • Stress and accent – Gestural synthesis
Discourse • So far – Analytic models (G&S, M&T) – Discourse structure recognition, segmentation • Now, generation and synthesis – Sentence/paragraph surface realization • Grammatical forms of discourse entities – pron, def NP • Sentence ordering of information – subj/not • Acoustic form of entities- accented/not; accent type
Attention and Information • Perspective: Focus of Attention – Coherence, Reference • Perspective: Information Flow – Goal of discourse: • Communication of information – Speaker to hearer
Given/New Dichotomy • Each “information unit” contains – “New” information • the “News” • New to hearer • New to discourse – “Given/Old” information • What is being talked about • Known to hearer • Already evoked in discourse
Given/New Effects • Influences structure of utterance – Word order – Form of referring expression – Prosodic prominence • Guides interpretation by hearer
Given/New & Word Order • Default word order (English, declarative) – Left-to-right increase in “New-ness” – Subject -> Given • Discourse-old - present in context – Predicate -> New
Given/New & Referring Expressions • Hierarchy of salience – Tied to Given/New status • Given+Salient -> Pronoun • Given, less salient -> Definite NP • New -> Indefinite NP
Given/New & Prosody • Prosody – Pitch, Loudness, Duration, … – Tone group = Information Unit • Given + New • Unstressed -> Given, salient • Stressed -> New, less salient
Application of Information Flow • Paraphrase (Mc. Keown 1983) • Natural language is ambiguous – Semantic - word senses - e. g. bank – Syntactic - structural • E. g. prepositional phrase attachment – Reference… • Paraphrase makes explicit system interpretation – Especially modification
Given/New Perspective • Word order affected by role in sentence – What speaker thinks hearer knows or not – Wh-items are “new”, rest “given”, assume true • Question: 3 parts – 2: Lack of knowledge: wh-item with no subclause – 3: Angle: Direct/Indirect modifiers of wh-items – 1: Given info: Everything else
Example • Q: Which active users advised by Tom Wirth work on projects in area 3? • P: Assuming that there are projects in area 3, which active users work on those projects? Look for users advised by Wirth. New: lack info Work on Active users New: Angle Advised by TW projects In area 3 Given
Syntax & Information Structure • Link parse tree to Given/new info • Root = Main verb = Inorder traversal – Left subtree= Subject = Preorder – Right subtree = Object = Preorder • Traversal order + Part information+Transform – > Linearization
Paraphrase by Given/New • Advantages: – Corrective response: e. g. if given info isn’t – More flexible/portable that template-based paraphrase
Applications of Info Structure • Speech recognition and synthesis – Prosody • Pitch, loudness, length – New - more likely stressed; Old: often unstressed – “Tunes” for given/new
Understanding Acoustic Realization • Motivation – Synthetic speech – Experimental evidence • Key components – Prosody – Syntax • Contextually “appropriate” speech synthesis
Speech Synthesis • Generally INTELLIGIBLE – But not NATURAL – Requires high attention to listen to • “Default” sentence intonation – May be misleading – Speaking of BILL, • A) JOHN thought he would WIN, but he DIDN’T • B) JOHN thought he would WIN, but HE didn’t
Accent Assignment: Analysis • Accent: – Increased loudness, duration, pitch movement • Basic view: – “available”/Given: no accent; New(er): accent • Attend to new information • Questions: – Does accent continue to decrease with repetition? – How does discourse “structure” affect accent?
Accent Assignment: Results • “Topic” status & First/Later mention vs – De-/Accenting, form of referring expression • Results: – First, +Topic: Accented, Full NP • Later, +Topic: De-accented , probably pronoun • Later, +Topic, +Refinement: Accented (even Pron) – First, -Topic: Accented Full NP • Later, -Topic: Accented Full NP, Implicit • Later, -Topic, +past-topic/+contrast: Accented NP (mod)
“To. BI” Intonation Framework • To. BI: Tone and Break Indices • Describe English sentence intonation • Tones: – Two pitch levels: H(igh) and L(ow) • * - on stressed syllable, e. g. H*, L+H* – Types: Pitch accents, Phrase Tones (L-, L%) • Last accent in phrase = ‘nuclear’ accent – Units: Intermediate and Intonational Phrases
“To. BI” Intonation Framework • Break indices – Mark groupings in speech – 0 - most closely linked; 5 - most disjoint – 4 = Intermediate phrase boundary (-) • ~ comma – 5 = Intonational phrase boundary (%, $) • ~ period - sentence
To. BI Examples
Contrast Examples
Contrast Examples
Contrast Examples
Syntax & Information Status • Intonation units more flexible than standard syntactic constituents, e. g. subject, predicate • CCG - Combinatory Categorial Grammar – Allows multiple analyses (parses) to fit – Link syntactic, semantic, and pragmatic/prosodic function with each unit
Generating Appropriate Intonation • Basic “previous mention” strategy – Accent first mention of content words – De-accent closed class function words – De-accent content words already mentioned • Inadequate – Need contrastive stress TOO
Generating Appropriate Intonation • • Identify theme (topic: links to previous info) Identify rheme (contributes new information) Shared propositional content Assign appropriate basic intonation contour – rheme: H* L-L%; – theme: L+H* L-H% (at most)
Generating Appropriate Intonation • Identify focus element in theme/rheme – Word to get accent • Focus – First mention, and – Contrastive • What is contrastive? ?
Contrastive Items: Domain • For each entity x: – 0: find alternatives in discourse and KB – 1: RSET= x and alternatives, – PROPS= features of x – CSET= features of x to mark for contrast – 2: For each p in PROPS, r in RSET, • IF p is not property of r, add p to CSET. – 3: Focus p of x • E. g. She broke her left LEG, NOT her RIGHT leg.
Contrastive Items: Word. Net • Word. Net: Semantic KB – 4 parts of speech: N, V, Adj, Adv – Category/word: one or more synonym sets – Hierarchies linked by relations: e. g. IS-A • Content Word W is new if NOT: – In focus history or history’s equivalence class • Equiv. Class: reachable by N hypernym/synset links • Content Word W is contrastive if: – In history’s contrast list • Contrast: hyponyms of hypernyms of W
Examples • • • (84) Q: I know which AMPLIFIER produces clean BASS, but WHICH amplifier produces clean TREBLE? L+H* L(H%) H* LL$ A: The BRITISH amplifier produces clean TREBLE. H* L(L%) L+H* LH$ (85) Q: I know which AMPLIFIER produces MUDDY treble, but WHICH amplifier produces CLEAN treble? L+H* L(H%) H* LL$ A: The BRITISH amplifier produces CLEAN treble. H* L(L%) L+H* LH$
Summary • Assigns contextually based intonation – Uses given/new information status – Extended to fine-grained contrastive status • Identifies contrast based on – Knowledge base if available – Word. Net Lexical DB for greater generality
Conclusions • Theme/Rheme identification difficult • Contrast/Similarity measures for Word. Net – Still oversimplified • Evaluation: How do you tell if it’s right? – Many alternatives • Incorporate in larger discourse structure – Discourse segments, plans, ….
Examples • • The X 4 is a SOLID-state AMPLIFIER L+H*LH* H* L- L$ The X 5 is a TUBE amplifier. L+H*LH* L-L$ It COSTS EIGHT HUNDRED DOLLARS, H* H* L-H% IT costs NINE hundred dollars, L+H*LH* L-H%
- Slides: 35