Acquisition and application of contextual role knowledge for coreference resolution

• US 20090326919A1
• Filed: 11/18/2004
• Published: 12/31/2009
• Est. Priority Date: 11/18/2003
• Status: Active Grant

• Current Assignee: University of Utah Research Foundation (State of Utah)
• Original Assignee: University of Utah Research Foundation (State of Utah)
• Inventors: Bean, David L.
• Granted Patent: US 7,813,916 B2
• Time in Patent Office: Days
• US Class Current: 704/9
• CPC Class Codes: G06F 40/289 Phrasal analysis, e.g. finite state techniques or chunking

Abstract

Coreference resolution is the process of identifying when two noun phrases (NP) refer to the same entity. Two main contributions to computational coreference resolution are made. First, this work contributes a new method for recognizing when an NP is anaphoric. Second, traditional approaches to coreference resolution typically select the most appropriate antecedent by recognizing word similarity, proximity, and agreement in number, gender, and semantic class. This work contributes a new source of evidence that focuses on the roles that an anaphor and antecedent play in particular events or relationships. I show that using contextual role knowledge as part of the coreference resolution process increases the number of anaphors that can be resolved, and I demonstrate an unsupervised method for acquiring contextual role knowledge that does not require an annotated training corpus. A probabilistic model based on the Dempster-Shafer model of evidence is used to incorporate contextual role knowledge with traditional evidence sources.

Claims

1. A method for associating anaphors with antecedents in a written work, the method comprising:

processing a training corpus containing textual documents that are topically related to the written work, said processing producing interpretive information useful to categorize noun phrases of the training corpus as independent or potentially anaphoric;

identifying noun phrases within the written work;

using the interpretive information, filtering those identified noun phrases to exclude noun phrases that can be identified to be independent in nature;

identifying a set of potentially anaphoric noun phrases occurring in the written work;

following said identifying a set of potentially anaphoric noun phrases, recognizing cases of unambiguous coreferences in the set of potentially anaphoric noun phrases, said recognizing associating a noun phrase with an antecedent for each case;

following said recognizing, identifying coreference combinations for unrecognized noun phrases from the set of potentially anaphoric noun phrases, each coreference combination including an unrecognized noun phrase and a potential antecedent;

applying a plurality of general knowledge sources and contextual role knowledge sources to the coreference combinations, wherein the contextual role knowledge sources include events and a manner of participation in the events to identify relatedness for the coreference combination at a thematic role level, said applying producing evidentiary values for the coreference combinations;

applying a factor to each of the produced evidentiary values to favor more credible knowledge sources;

for each unrecognized noun phrase, applying a probabilistic model to the produced evidentiary values associated with the noun phrase; and

for each application of the probabilistic model to an unrecognized noun phrase, selecting either an antecedent for that unrecognized noun phrase, if the coreference of that antecedent has a corresponding evidentiary value above a selected threshold value, or no antecedent otherwise.

2. A method according to claim 1, further comprising the step of filtering definite noun phrases that are restrictively modified from the same sentence.

3. A method according to claim 1, wherein said applying a plurality of knowledge sources applies at least two knowledge sources selected from the group of a gender agreement knowledge source, a number agreement knowledge source, a semantic agreement knowledge source, a scoping knowledge source, a syntactic role knowledge source, and a recency knowledge source.

4. A method according to claim 1, wherein the probabilistic model applied to the produced evidentiary values is a Dempster-Shafer Model.

5. A method according to claim 1, further comprising the step of collapsing an anaphoric chain into a single candidate.

6. A method according to claim 1, wherein said recognizing cases of unambiguous coreferences applies a set of syntactic heuristics.

7. A method according to claim 1, wherein said recognizing cases of unambiguous coreferences recognizes repeated proper nouns.

8. A method according to claim 1, further comprising the step of recognizing and filtering pleonastic noun phrases from the identified noun phrases.

20. The method as recited in claim 1, wherein events and a manner of participation in the events to identify relatedness for the coreference combinations include one or more of:

synonymy that identifies synonymous roles between contextual roles;

sequence that identifies sequentially related contextual roles;

set membership that includes a contextual roles that are performed by a noun phrase;

a thematic role associated with the event;

a combination of a predicate and an argument relation that identifies roles of possession, property, or condition.

9. A method for associating anaphors with antecedents in a written work, comprising the steps of

processing a training corpus containing textual documents that are topically related to the written work, said processing producing interpretive information useful to categorize noun phrases of the training corpus as independent or potentially anaphoric, said processing producing a definite—

only list of definite noun phrases, said processing optionally producing 81 extractions of noun phrases from the first sentences of texts comprising the training corpus, said processing further optionally producing existential head patterns, wherein said interpretive information includes the definite-only list of definite noun phrases and optionally the produced 81 extractions or the existential head patterns;

identifying noun phrases within the written work;

using the interpretive information, filtering those identified noun phrases to exclude noun phrases that can be identified to be independent in nature, said filtering using the produced definite-only list to exclude definite noun phrases;

identifying a set of potentially anaphoric noun phrases occurring in the written work;

following said identifying a set of potentially anaphoric noun phrases, recognizing cases of unambiguous coreferences in the set of potentially anaphoric noun phrases, said recognizing associating a noun phrase with an antecedent for each case;

following said recognizing, identifying coreference combinations for unrecognized noun phrases from the set of potentially anaphoric noun phrases, each coreference combination including an unrecognized noun phrase and a potential antecedent;

applying a plurality of general knowledge sources and contextual role knowledge sources to the coreference combinations, said applying producing evidentiary values for the coreference combinations, wherein the contextual role knowledge sources include events and a manner of participation in the events that identifies relatedness for the coreference combinations at a thematic role level;

applying a factor to each of the produced evidentiary values to favor more credible knowledge sources;

for each unrecognized noun phrase, applying a Dempster-Shafer model to the produced evidentiary values associated with the noun phrase; and

for each application of the model to an unrecognized noun phrase, selecting either an antecedent for that unrecognized noun phrase, if the coreference of that antecedent has a corresponding evidentiary value above a selected threshold value, or no antecedent otherwise.

10. A method according to claim 9, further comprising the step of filtering definite noun phrases that are restrictively modified from the same sentence.

11. A method according to claim 9, wherein said applying a plurality of knowledge sources applies at least two knowledge sources selected from the group of a gender agreement knowledge source, a number agreement knowledge source, a semantic agreement knowledge source, a scoping knowledge source, a syntactic role knowledge source, and a recency knowledge source.

12. A method according to claim 9, further comprising the step of collapsing an anaphoric chain into a single candidate.

13. A method for associating anaphors with antecedents in a written work, comprising the steps of:

processing a training corpus containing textual documents that are topically related to the written work, said processing producing interpretive information useful to categorize noun phrases of the training corpus as independent or potentially anaphoric;

processing the training corpus to extract a set of thematic caseframes;

identifying noun phrases within the written work;

using the interpretive information, filtering those identified noun phrases to exclude noun phrases that can be identified to be independent in nature;

identifying a set of potentially anaphoric noun phrases occurring in the written work;

following said identifying a set of potentially anaphoric noun phrases, recognizing cases of unambiguous coreferences in the set of potentially anaphoric noun phrases, said recognizing associating a noun phrase with an antecedent for each case;

following said recognizing, identifying coreference combinations for unrecognized noun phrases from the set of potentially anaphoric noun phrases, each coreference combination including an unrecognized noun phrase and a potential antecedent;

applying a plurality of general knowledge sources and a contextual role knowledge source to the coreference combinations, said applying producing evidentiary values for the coreference combinations, wherein the contextual knowledge sources apply extracted thematic caseframes that identify relatedness between the noun phrase and potential antecedent in each coreference combination at a thematic role level;

applying a factor to each of the produced evidentiary values to favor more credible knowledge sources;

for each unrecognized noun phrase, applying a probabilistic model to the produced evidentiary values associated with the noun phrase; and

for each application of the probabilistic model to an unrecognized noun phrase, selecting either an antecedent for that unrecognized noun phrase, if the coreference of that antecedent has a corresponding evidentiary value above a selected threshold value, or no antecedent otherwise.

14. A method according to claim 13, further comprising the step of constructing at least one of a caseframe lexical expectations knowledge base, a caseframe semantic expectations knowledge base, or a caseframe network knowledge base for caseframe co-occurrence relationships;

and wherein the constructed knowledge base is one of the plurality of knowledge sources applied to coreference combinations.

15. A method according to claim 13, further comprising the step of constructing a caseframe lexical expectations knowledge base, a caseframe semantic expectations knowledge base, and a caseframe network knowledge base for caseframe co-occurrence relationships;

and wherein said applying a plurality of knowledge sources applies the caseframe lexical expectations knowledge base, caseframe semantic expectations knowledge base, and caseframe network knowledge base.

16. A method according to claim 13, further comprising the steps of:

constructing a caseframe lexical expectations knowledge base;

following said recognizing cases of unambiguous coreferences, identifying thematic caseframes for noun phrases of unambiguous coreferences; and

expanding the caseframe lexical expectation knowledge base to register identified thematic caseframes for the noun phrases of unambiguous coreferences.

17. A method according to claim 13, further comprising the step of discarding thematic caseframes that occur at a low frequency.

18. A set of computer readable media containing computer instructions for operating a anaphor-antecedent associator, the set of computer readable media comprising at least one medium upon which is stored the computer instructions executable by a computing system to achieve the functions of:

(i) processing a training corpus containing textual documents that are topically related to the written work, said processing producing interpretive information useful to categorize noun phrases of the training corpus as independent or potentially anaphoric;

(ii) identifying noun phrases within the written work;

(iii) using the interpretive information, filtering those identified noun phrases to exclude noun phrases that can be identified to be independent in nature;

(iv) identifying a set of potentially anaphoric noun phrases occurring in the written work;

(v) following said identifying a set of potentially anaphoric noun phrases, recognizing cases of unambiguous coreferences in the set of potentially anaphoric noun phrases, said recognizing associating a noun phrase with an antecedent for each case;

(vi) following said recognizing, identifying coreference combinations for unrecognized noun phrases from the set of potentially anaphoric noun phrases, each coreference combination including an unrecognized noun phrase and a potential antecedent;

(vii) applying knowledge sources including a plurality of general knowledge sources and contextual role knowledge sources to the coreference combinations, wherein the contextual role knowledge sources include events and a manner of participation in the events to identify relatedness for the coreference combination at a thematic role level, said applying producing evidentiary values for the coreference combinations;

(viii) applying a factor to each of the produced evidentiary values to favor more credible knowledge sources;

(ix) for each unrecognized noun phrase, applying a probabilistic model to the produced evidentiary values associated with the noun phrase; and

(x) for each application of the probabilistic model to an unrecognized noun phrase, selecting either an antecedent for that unrecognized noun phrase, if the coreference of that antecedent has a corresponding evidentiary value above a selected threshold value, or no antecedent otherwise.

19. (canceled)