We present first results using paraphrase as well astextual entailment data to test the language universalconstraint posited by Wu’s (1995, 1997) InversionTransduction Grammar (ITG) hypothesis. Inmachine translation and alignment, the ITG Hypothesisprovides a strong inductive bias, and has beenshown empirically across numerous language pairsand corpora to yield both efficiency and accuracygains for various language acquisition tasks. Monolingualparaphrase and textual entailment recognitiondatasets, however, potentially facilitate closertests of certain aspects of the hypothesis than bilingualparallel corpora, which simultaneously exhibitmany irrelevant dimensions of cross-lingual variation.We investigate this using simple generic BracketingITGs containing no language-specific linguisticknowledge. Experimental results on the MSRParaphrase Corpus show that, even in the absenceof any thesaurus to accommodate lexical variationbetween the paraphrases, an uninterpolated averageprecision of at least 76% is obtainable fromthe Bracketing ITG’s structure matching bias alone.This is consistent with experimental results on thePascal Recognising Textual Entailment ChallengeCorpus, which show surpisingly strong results for anumber of the task subsets.
展开▼