1 / 17

Lecture 21

Lecture 21. Algorithms for Reference Resolution. Issues. Which features should we make use of? How should we order them? I.e. which override which? What should appear in our discourse model? I.e., what types of information do we need to keep track of? How to evaluate?. Lappin & Leass ‘94.

jconners
Télécharger la présentation

Lecture 21

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Lecture 21 Algorithms for Reference Resolution CS 4705

  2. Issues • Which features should we make use of? • How should we order them? I.e. which override which? • What should appear in our discourse model? I.e., what types of information do we need to keep track of? • How to evaluate?

  3. Lappin & Leass ‘94 • Weights candidate antecedents by recency and syntactic preference (86% accuracy) • Two major functions to perform: • Update the discourse model when an NP that evokes a new entity is found in the text, computing the salience of this entity for future anaphora resolution • Find most likely referent for current anaphor by considering possible antecedents and their salience values • Partial example for 3P, non-reflexives

  4. Salience Factors and Weights (Trained)

  5. Implicit ordering of arguments: subj/exist pred/obj/indobj-oblique/dem.advPP On the sofa, the cat was eating bonbons. sofa: 100+40+80=220 cat: 100+80+50+80=310 bonbons: 100+50+50+80=280 • Update: • Weights accumulate over time • Cut in half after each sentence processed • Salience values for subsequent referents accumulate for equivalence class of correferential items (exceptions, e.g. multiple references in same sentence)

  6. The bonbons were clearly very tasty. sofa: 140/2=70 cat: 310/2=155 bonbons: 280/2 +(100+80+50+80)=450 • Additional salience weights for grammatical role parallelism (35) and cataphora (-175) calculated when pronoun to be resolved • Additional constraints on gender/number agrmt/syntax They were a gift from an unknown admirer. sofa: 70/2=35 cat: 155/2=77.5 bonbons: 450/2=225 (+35)

  7. Reference Resolution • Collect potential referents (up to four sentences back): {sofa,cat,bonbons} • Remove those that don’t agree in number/gender with pronoun {bonbons} • Remove those that don’t pass intra-sentential syntactic coreference constraints The cat washed it. (itcat) • Add applicable values for role parallelism (+35) or cataphora (-175) to current salience value for each potential antecedent • Select referent with highest salience; if tie, select closest referent in string

  8. Solution They were a gift from an unknown admirer. (they=bonbons) bonbons: 225/2=112.5 (+35) Why? Only potential antecedent to agree in gender/number Also has highest salience (147.5) Or… It ate them all. (it=?) sofa: 70/2=35 cat: 155/2=77.5 bonbons: 225/2=112.5

  9. Centering (Grosz et al ‘95;Brennan et al ‘87) • Adds concept of discourse center • Un: an utterance • Backward-looking center Cb(Un): current focus after Un interpreted • Forward-looking centers Cf(Un): ordered list of potential focii referred to in Un • Cb(Un+1) is highest ranked member of Cf(Un) • Cp(Un): preferred center of Cf(Un) • Other Cf ordered subj<exist. Prednom<obj<indobj-oblique<dem. advPP (Brennan et al)

  10. Transitions from Un to Un+1

  11. Rules • If any element of Cf(Un) is pronominalized in Un+1, then Cb(Un+1) must be • Preference: Continue > Retain > Smooth-Shift > Rough-Shift • Algorithm • Generate Cb-Cf combinations for all possible reference assignments • Filter by constraints (syntactic coreference, selectional restrictions,…) • Rank by preference among transition orderings

  12. Example U1:George saw a mouse on the floor. U2:He showed it to Harry. U3:He ate it. • Grammatical role hierarchy • Cf(U1): {George,mouse,floor} • Cp(U1): George • Cb(U1): undefined • Assume it=mouse • Cf(U2): {George, mouse,Harry} • Cp(U2): George • Cb(U2): George • Continue (Cp(U2)=Cb(U2); Cb(U1) undefined

  13. Assume it=floor • Cf(U2): {George,floor,Harry} • Cp(U2): George • Cb(U2): George • Continue (Cp(U2)=Cb(U2); Cb(U1) undefined) • What did George show Harry? • How to break a tie? Ordering of previous Cf list? U3: He ate it. Assume he=George, it=mouse • Cf(U3): {George,mouse} • Cp(U3): George • Cb(U3): George • Continue (Cp(U3)=Cb(U3); Cb(U3)=Cb(U2))

  14. Assume he=Harry, it=mouse • Cf(U3): {Harry,mouse} • Cp(U3): Harry • Cb(U3): Harry • Continue (Cp(U3)=Cb(U3); Cb(U3)Cb(U2)) • So, who ate the mouse?

  15. Centering Theory vs. Lappin & Leass • Centering sometimes prefers an antecedent Lappin and Leass (or Hobbs) would consider to have low salience • Always prefers a single pronominalization strategy: prescriptive, assumes discourse coherent • Constraints too simple: grammatical role, recency, repeated mention • Assumes correct syntactic information available as input

  16. Hobbs ‘78: Syntax-Based Reference Resolution • Search for antecedent in parse tree of current sentence, then prior sentences in order of recency • For current S, search for NP nodes to the left of a path p from the pronoun up to the first NP or S node (X) above it in L2R, breadth-first • Propose as pronoun’s antecedent any NP you find as long as it has an NP or S node between itself and X • If X is highest node in sentence, search prior sentences, L2R breadth-first, for candidate NPs • O.w., continue searching current tree by going to next S or NP above X before going to prior sentences

  17. Evaluation • Centering only now being specified enough to be tested automatically on real data • Specifying the Parameters of Centering Theory: A Corpus-Based Evaluation using Text from Application-Oriented Domains (Poesio et al., ACL 2000) • Walker ‘89 manual comparison of Centering vs. Hobbs ‘78 • Only 281 examples from 3 genres • Assumed correct features given as input to each • Centering 77.6% vs. Hobbs 81.8% • Lappin and Leass’ 86% accuracy on test set from computer training manuals

More Related