From dictionary to corpus to self-organizing dictionary: learning valency associations in the face of variation and change

ing over specific lexically-governed particles and prepositions and specific predicate selectional preferences, but including some `derived' / `alternant' semi-productive, and therefore only semipredictable, bounded dependency constructions, such as particle or dative movement, there are at least 163 valency frames associated with verbal predicates in (current) English (Briscoe, 2000). In this paper, I will review the work that my colleagues and I have done to learn (semi-)automatically this very large number of associations between individual verbal predicates and valency frames. Access to a comprehensive and accurate valency lexicon is critical for the development of robust and accurate parsing technology capable of recovering predicate-argument relations (and thus logical forms) from free text or transcribed speech. Without this information it is possible to `chunk’ input into phrases but not to distinguish arguments from adjuncts or resolve most phrasal attachment ambiguities. Furthermore, for statistical parsers it is not enough to know the associations of predicates to valency frames, it is also critical to know the relative frequency of such associations given a specific predicate. Such information is a core component of that required to `lexicalize’ a probabilistic parser, and it is now well–established that lexicalization is essential for accurate disambiguation (e.g. Collins, 1997, Carroll et al, 1998). While state-of-the-art wide-coverage grammars of English, capable of recovering predicateargument structure and expressed as a unification-based phrase structure grammar, have on the order of 1000 rules, it is clear that the number of associations between valency frames and predicates needed in a lexicon for such a grammar will be much higher.