RI: Paraphrasing using Lexico-Semantic Resources

International Computer Science Institute, Berkeley CA

Investigators

Abstract

This one-year pilot proof-of-concenpt project generates paraphrases of semantically labeled input sentences using the semantics and syntax encoded in the publicly available FrameNet corpus and the lexical semantic information from the publicly available WordNet corpus. Arbitrary text is first processed using an existing frame-semantic parser (such as Shalmaneser), and then passed to the paraphrasing algorithm. The algorithm generates a large number of paraphrases with a wide range of syntactic and semantic distances from the input. For example, given the input ``I like eating cheese'', the system outputs the syntactically distant ``Eating cheese is liked by me'', the semantically distant ``I fear sipping juice'', and thousands of other sentences. The large variety of generated paraphrases makes the algorithm ideal for a range of statistical machine learning problems such as language modeling and machine translation, as well as other semantics-dependent tasks such as query and language generation. The system is tested both against human judgments of paraphrase quality and by measuring the efficacy of using paraphrases in a language modeling task. Paraphrases are generated from the Nuclear Threat Initiative corpus, which has been hand-marked with ground-truth frame-semantic annotations. This allows an analysis of: A. the quality of the paraphrases vs. human judgments, B. the relative accuracy and quantity vs. other paraphrasing methods, and C. the relative accuracy and quantity of using automatic frame parsing vs. gold-standard annotations.