|
|
COMPUTING LINGUISTICALLY-BASED TEXTUAL INFERENCES
Danny Bobrow, Cleo Condoravdi, Lauri Karttunen, Tracy Holloway King, Valeria de Paiva, and Annie Zaenen PARC Natural Language Group Friday, January 18, 3:30pm, MJH Rm 126 In this talk we give an overview and a demo of PARC's Bridge system. The particular task that we focus on is entailment and contradiction detection (ECD), a more refined variant of the PASCAL RTE (Recognizing Textual Entailment) challenge. Given a passage of text and a query, does the query sentence follow from the text in the passage, is it contradicted by it, or neither? Here are examples of all three cases: Passage: Oswald assassinated Kennedy. Query: Did Kennedy die? Response: YES Passage: Bill forgot to shave this morning. Query: Did Bill shave this morning? Response: NO Passage: There is a cat in the yard. Query: Is there a black cat in the yard? Response: UNKNOWN The entailment and contradiction detection algorithm operates on the level of Abstract Knowledge Structure (AKR) without the need of disambiguation. An AKR representation, derived from the syntactic and semantic analyses of a sentence, is a flat set of facts that involves concepts, roles, and contexts. Texts are parsed to produce packed f-structures and these are rewritten and canonicalized, without unpacking, into AKR. Canonicalization is determined both by the structure of the representations and the lexical items involved. The system includes knowledge about words and their relations between them that are encoded in resources such as WordNet and VerbNet. It also includes knowledge about lexically or constructionally triggered presuppositions and entailments. The ECD process first aligns context and concept terms and then computes specificity relations between the aligned concept terms. Some special case reasoners support identification of named objects, comparison of specificity of WordNet synsets, and compatibility of cardinality restrictions. All the query facts that are entailed by the corresponding passage facts get removed. If no query facts remain, the system responds YES. A conflict in the instantiation claims of two aligned terms marks a contradiction. In this case the system responds NO. If some query facts remain at the end, the response is UNKNOWN. The linguistic phenomena we illustrate in this presentation include lexical entailments (kill => die), relations between lexical predicates or phrasal constructions and their embedded complements (forget that A => A, forget to A => not A, take the trouble to S => S, waste an opportunity to S => not S), and inferring temporal relations from temporal modifiers. 
|