High-Fidelity Lexical Axiom Construction from Verb Glosses

Proceedings of the Fifth Joint Conference on Lexical and Computational Semantics (*SEM 2016), pages 34–44, Berlin, Germany, August 11-12, 2016. High-Fidelity Lexical Axiom Construction from Verb Glosses Gene Kim University of Rochester Department of Computer Science gkim21@cs.rochester.edu Lenhart Schubert University of Rochester Department of Computer Science schubert@cs.rochester.edu Abstract This paper presents a rule-based approach to constructing lexical axioms from Word- Net verb entries in an expressive semantic representation, Episodic Logic (EL). EL differs from other representations in be- ing syntactically close to natural language and covering phenomena such as general- ized quantification, modification, and in- tensionality while still allowing highly ef- fective inference. The presented approach uses a novel preprocessing technique to improve parsing precision of coordina- tors and incorporates frames, hand-tagged word senses, and examples from WordNet to achieve highly consistent semantic in- terpretations. EL allows the full content of glosses to be incorporated into the for- mal lexical axioms, without sacrificing in- terpretive accuracy, or verb-to-verb infer- ence accuracy on a standard test set. Evaluation of semantic parser perfor- mance is based on EL-match, introduced here as a generalization of the smatch met- ric for semantic structure accuracy. On gloss parses, the approach achieves an EL- match F1 score of 0.83, and a whole- axiom F1 score of 0.45; verb entailment identification based on extracted axioms is competitive with the state-of-the-art. 1 Introduction Words encapsulate a great deal of knowledge, and in conjunction with language syntax, allow human beings to construct sentences that convey novel ideas to one another. Any system intended for broad natural language understanding will need to be able to perform inferences on the words that are the building blocks of language. For this reason, Gloss – slam2.v : “strike violently” Axiom – ((x slam2.v y) ** e) → ((x (violently.adv (strike.v y))) ** e) Figure 1: Example of rule extraction from ma- chine readable dictionaries for WordNet entry of slam2.v. there have been many attempts to transduce in- formal lexical knowledge from machine readable dictionaries into a formally structured form (Cal- zolari, 1984; Chodorow et al., 1985; Harabagiu et al., 1999; Moldovan and Rus, 2001; Hobbs, 2008; Allen et al., 2013). Consider an example of the types of knowl- edge these approaches seek to extract in Figure 1. WordNet defines slam2.v, i.e., sense 2 of the verb slam, as “strike violently”. This gloss states an im- plication that if “x slams y” characterizes an event e, then “x strikes y violently” also characterizes event e. All language phenomena must be able to be represented and reasoned about for such ax- ioms to be useful in a language understanding sys- tem. This is where previous approaches share a common shortcoming: the logical representations that the lexical knowledge is mapped into are in- sufficient for representing many common natural language devices or for performing inference. The contributions of this paper are the follow- ing: • We demonstrate limitations in previous ap- proaches to extracting lexical knowledge from machine readable dictionaries, partic- ularly in their choices of …

Content truncated to 500 words