Two-way Bridge Between Language and Logic

About the project:

• Research Outline
• Participants
• Publications

Research Outline

Question-answering systems have tended to operate at one of two ends of a spectrum: the language end, or the knowledge/reasoning end. Open-domain full-text systems have broad coverage, but they tend to lose precision by not taking the linguistic context of source information into account, and they tend to lose recall because they are too sensitive to the linguistic form of source information. Knowledge/reasoning systems can provide higher precision by selecting answers that give the right information in the right context, and higher recall by operating on canonical representations of the source information and by composing answers that require proofs from information gathered from multiple sources. But KR systems tend to be narrow in coverage, only providing answers in domains that have been carefully axiomatized.

We built a crosscutting linguistic technology that provides a two-way bridge that links these effective but quite different approaches. This enables a user to ask questions in natural language and have broad coverage from a large text database while also leveraging the deep capabilities of a first-class multi-source reasoning engine. Building on our high-performance parsing and generation engine and broad-coverage English grammar, we provide:

• Layered Mapping from Language to Logic: We automatically transform questions, source material and answer candidates from free-text English into logical formulas that have the same meaning, but can be interpreted by a deductive reasoning system. The mapping consists of multiple layers of linguistic description:
  • functional structures (f-structures) that capture the predicate argument structure found in the text itself. The mapping from strings to f-structures is done using a broad-coverage LFG grammar run in the XLE linguistics environment. The grammar can be used with both the XLE parser and generator, thus allowing for a two-way mapping between strings and f-structures.
  • semantic structures that canonicalize away from the presenting linguistic form. The mapping from f-structures to linguistic semantic structures is done using the XLE rewriting (transfer) component, which allows for packed rewriting of representations.
  • Abstract KR structures which can be used by a reasoning system. The mapping from linguistic semantic structures to the more abstract KR-compatible structures is done using the XLE rewriting (transfer) component, which allows for packed rewriting of representations.

  This identifies answers with higher precision and better recall, especially for questions of some complexity.

• Mapping from Logic to Language: We automatically generate language strings from AKR logical formulas provided by the question-answering system. This is used to create well-structured English for answers and explanations that are provided in a formal language by the deductive system. This is done by mapping the AKR logical formulas directly into skeletal f-structures, bypassing the level of linguistic semantic representation. These underspecified f-structures are then run through the XLE generator using a slightly modified version of the parsing grammar with a combination of generation-specific robustness techniques.

• Entailment and Contradiction Detection: We have developed a matching component that compares the AKR representations of question texts with proposed answer texts. This supports higher precision for ranking answers for open-domain questions where there may not be coverage in the deductive system, and it reduces ambiguity in the transformation of question and answer texts to logic.

For descriptions of the components of this project, including lists of relevant publications, see:

• Parsing, generation, rewriting, and matching (XLE linguistics environment) and the broad-coverage LFG grammar
• Reasoning about Document Collections project (RDC)

This project is a component of Asker, our ambiguity-enabled, scalable knowledge repository, which allows for natural-language queries over massive document collections.

Participants

• Natural Language Theory and Technologies group (NLTT) in the Intelligent Systems Lab (ISL)
• Interns: Olya Gurevich, Anubha Kothari, Xiaofei Lu, Jean-Philippe Marcotte, Rowan Nairn, Mike Orme, Kiyoko Uchiyama

Selected Publications

• Processing Paraphrases and Phrasal Implicatives in the Bridge Question-Answering System. 2008. K. B. Pichotta. Undergraduate honors thesis, Stanford University. [pdf]
• Computing Relative Polarity for Textual Inference. 2006. Rowan Nairn, Cleo Condoravdi and Lauri Karttunen. In the Proceedings of ICoS-5 (Inference in Computational Semantics). April 20-21, 2006. Buxton, UK. [pdf]
• Semantics via F-structure Rewriting 2006. Dick Crouch and Tracy Holloway King. In the Proceedings of LFG06 Conference. CSLI On-line Publications. [pdf preprint]
• Deverbal Nouns in Knowledge Representation. 2006 Olga Gurevich, Richard Crouch, Tracy Holloway King and Valeria de Paiva. In Proceedings of FLAIRS'06. [pdf]
• Local Textual Inference: Can it be Defined or Circumscribed? 2005. Annie Zaenen, Lauri Karttunen, and Richard Crouch. In Proceedings of the ACL. [pdf]
• Packed rewriting for mapping semantics to KR. 2005. Richard Crouch. In Proceedings of the International Workshop on Computational Semantics. Tilburg, The Netherlands. [pdf]
• A Basic Logic for Textual Inference. 2005. Danny Bobrow, Cleo Condoravdi, Richard Crouch, Ronald Kaplan, Lauri Karttunen, Tracy Holloway King, Valeria de Paiva, and Annie Zaenen. In Proceedings of the AAAI Workshop on Inference fro Textual Question Answering, Pittsburgh, PA. [pdf]
• Unifying lexical resources. 2005. Dick Crouch and Tracy Holloway King. In Proceedings of the Verb Workshop. Saarbruecken, Germany. [pdf]

Papers concerning related components

• XLE natural language parsing, generation, and rewrite (transfer) system (publications at bottom of page)
• Parallel Grammar Project and the English grammar publications
• Reasoning about document collections KR mapping (RDC) publications