220 likes | 361 Views
Reference Resolution And Cognitive Grammar. Susanne Salmon-Alt Laurent Romary Loria - Nancy, France ICCS-01 San Sebastian, May 2001. General context. Computational objectives Designing man-machine dialogue systems with graphical user feedback and gestural designation
E N D
Reference Resolution AndCognitive Grammar Susanne Salmon-AltLaurent Romary Loria - Nancy, France ICCS-01 San Sebastian, May 2001
General context • Computational objectives • Designing man-machine dialogue systems with graphical user feedback and gestural designation • Allowing the user to express himself spontaneously • Linguistic objectives • Deriving a model that widely covers the range of possible referring expressions and their use in context • Narrowing the discrepancy between computational models and linguistic descriptions • Can a cognitive model be a means to achieve this?
Reference Resolution hearer puts a circle on the screen... hearer puts a line on the screen... Interpretation Evolution of the context ? Take a big circle. Okay, and now put a small lineon the left. Don’t stick itto the circle. Referring expressions Context model “World”: perception and gestures
Reference Resolution Reference Resolution Take a big circle. Okay, and now put a small lineon the left. Don’t stick itto the circle. {C1} L1 C2 C1 {L2} C3 L3 L2 {C1} {L2} Referring expression Context model Referents • Associate referring expressions to identifying representations for contextual objects
The general background • Observations: reference in task-oriented dialogues • A wide variety of referring expressions: indefinites, definites, demonstratives, pronouns • Both anaphoric and deictic uses • Modeling work is generally reduced to pronoun resolution • E.g. Centering (Grosz et al., 1995), Mitkov (1998) • Specificity of anaphoric expressions? • DRT / S-DRT: no essential difference between pronouns and definites (linking as the main mechanism) • How to integrate demonstratives (+gestures)?
Reference = Linking ? • Current strategies are basically co-referential • if indefinite : introduce a new discourse referent • if anaphoric : filter the context model on semantic constraints and choose a suitable referent • Problems • empirically inappropriate for definites (Poesio &Vieira, 1998) • need of some additional mechanisms for...
Difficulties with Linking visual information Delete the triangle. Bridging Take a triangle. Color the base in blue • one-anaphora • Thegreen blocksupports thebig pyramid, • but not thered/smallone. ordinals and other-expressions Take two lines. Move one line to the left. Delete the other line. • Need of cognitive structures rather than (or in complement to) of discourse variables
Cognitive Grammar (1) • Theoretical foundations (Langacker 1986, 1991) • language not self-contained, but part of cognitive processing • speaker’s knowledge : inventory of symbolic units (phonological and semantic pole) • semantic structures characterized relative to presupposed « cognitive domains » (concepts, perceptual experiences, knowledge systems)
Cognitive Grammar (2) NOUN roof • Semantic structure of nouns knife Abstract schema: Delimitation of a region in some domain Instantiation: Profiling a sub-structure of one (ore more) presupposed domains
Cognitive Grammar (3) tr lm horizontal [ ON THE LEFT OF] • Meaning (not truth-conditional): assembly and profiling of semantic units [ THE LINE ON THE LEFT OF THE CIRCLE ] [ LINE ] [ CIRCLE ] Atemporal relation Assembly and profiling
Suitability • Interesting properties : interpretation of NPs • not linking, but profiling within a given domain => encompasses all kind of anaphoric expressions • conceptual domains not primarily linguistic constructs => reference to percepual entities, gestures • meaning = imposing a profile on a domain => prediction of preferred referential access : The green block supports the big pyramid, but not the red/smallone. • Problem : formalisation
The Model — Overview Assembly of abstract schemas Search for a suitable conceptual domain Profiling of a regionof the domain abstract schema for determiners restructured domain context model (domains) complex schema for noun phrases abstract schema for nouns selected domain Calculus of an underspecified domain Unification with a domain of the context model Focusing an itemof a partition of the domain
The Context Model (1) @T1 @F1 @C1 @L1 Type = CIRCLE Card = 1 Type =LINE Card = 1 Type = TRIANGLE Card = 1 Properties = {(size: big)} Type = FIGURE Card = 2 Properties = {(size: small)} Diff-Crit = Type CIRCLELINE @C1@L1 Basic units : domains
The Context Model (2) @F1 @C1 @L1 Type = CIRCLE Card = 1 Type =LINE Card = 1 Type = FIGURE Card = 2 Properties = {(size: small)} Diff-Crit = Type • Grouping • Triggers • co-ordination • prepositions • argument structure • perceptual criteria • Result • partitioned domain • common type • differentiation criterion • focus structure THE LINE ON THE LEFT OF THE CIRCLE CIRCLELINE @C1 @L1
Underspecified Domains (1) Type = N N N N Determiner semantics: grounding (how to locate the thing within the given domain) Determiner semantics: the item is located within a domain of elements of type N abstract schema for determiners complex schema for noun phrases Indefinite NPs « a N » (a line) abstract schema for nouns Noun semantics: delineate an item of a partitioned domain Noun semantics: delineate an item of type N within a domain
Underspecified Domains (2) Type > N N ¬N ¬ N Determiner semantics: grounding (how to locate the thing within the given domain) Determiner semantics: the item is located within a domain of elements of a super-type of N abstract schema for determiners complex schema for noun phrases Definite NPs « the N » (the line) abstract schema for nouns Noun semantics: delineate an item of a partitioned domain Noun semantics: delineate an item of type N within a domain
Underspecified Domains (3) Type = ? Determiner semantics: grounding (how to locate the thing within the given domain) Determiner semantics: Ø (the item has to be located from its focal position focus) abstract schema for determiners complex schema for noun phrases Pronouns “it” abstract schema for nouns Noun semantics: delineate an item of a partitioned domain Pronoun semantics: delineate an item of a partitioned domain
Restructuration (1) context model (domains) underspecified domain selected domain restructured domain Type = N Type = N N N N N N N Indefinites « a N » Focusing one item of the partition Profiling of a region in the domain
Restructuration (2) Profiling of a region in the domain context model (domains) underspecified domain selected domain restructured domain Type > N Type > N N ¬ N ¬N N ¬N ¬N Definites “the N” Focusing item N of the partition
Restructuration (3) Profiling of a region in the domain context model (domains) underspecified domain selected domain restructured domain Type= ? Type = ? Pronouns « it » No change
Application - Example L1 C1 C1 L2 C2 C2 L3 C3 C3 C4 L4 C4 ... ... a big circle a small line @L @C @ ? @C1 @C @ ? @C&L @ ? @C&L @L @L1 Type = FIGURE CD = Type CD = Position Type = LINE Prop = SMALL Type = ? Type = CIRCLE Type = LINE Type = FIGURE CD = Type Type = LINE Prop = SMALL Type = CIRCLE Prop = BIG Type = CIRCLE Prop = BIG Type = CIRCLE Prop = BIG Type = LINE Prop = SMALL ... L1 L2 L3 L4 CIRCLE ¬CIR. C1 L1 it the circle Take a big circle. Okay, and now put a small line on the left. Don’t stick itto the circle.
Discussion • Single mechanism for different kinds of reference • not linking, but extraction • bridging, one-anaphora, other-expressions • integrated treatment of demonstratives (cf. full paper) • takes into account visual information (required for dialogues) • Formalization of Cognitive Grammar • Implementation into a real dialogue platform • Partial validation on a corpus of human dialogues • other-expressions • Being even for formal • Expressing the constraints in the framework of S-DRT