The Automatic Acquisition of Frequencies of Verb Subcategorization Frames from Tagged Corpora

The Automatic Acquisition of Frequencies of Verb Subcategorization Frames from Tagged Corpora Akira Ushioda, David A. Evans, Ted Gibson, Alex Waibel Computational Linguistics Program Carnegie Mellon University Pittsburgh, PA 15P13-3890 aushioda @icl. cmu. edu Abstract We describe a mechanism for automatically acquiring verb subcategorization frames and their frequencies in a large corpus. A tagged corpus is first partially parsed to identify noun phrases and then a finear grammar is used to estimate the appropri- ate subcategorization frame for each verb token in the corpus. In an experiment involving the identification of six fixed subcategorization frames, our current system showed more than 80% accuracy. In addition, a new statistical approach substan- tially improves the accuracy of the frequency estimation. 1 Introduction When we construct a grammar, there is always a trade-off between the coverage of the grammar and the ambiguity of the grammar. If we hope to develop an efficient high- coverage parser for unrestricted texts, we must have some means of dealing with the combinatorial explosion of syntactic ambiguities. While a general probabilistic optimiza- tion technique such as the Inside-Outside algorithm ([Baker, 1979], [Lauri and Young, 1990], [Jelinek el ai., 1990], [Carroll and Charniak, 1992]) can be used to reduce ambi- guity by providing estimates on the applicability of the context-free rules in a grammar (for example), the algorithm does not take advantage of lexical information, including such information as verb subcategorization frame preferences. Discovering or acquiring lexically-sensitive linguistic structures from large corpora may offer an essential comple- mentary approach. Verb subcategorization (verb-subcat) frames represent one of the most important ele- ments of grammatical/lexical knowledge for efficient and reliable parsing. At this stage in the computational-linguistic exploration of corpora, dictionaries are still probably more re- liable than automatic acquisition systems as a source of subcategorization (subcat) frames for verbs. The Oxford Advanced Learners Dictionary (OALD) [Hornby, 1989], for exam- ple, uses 32 verb patterns to describe a usage of each verb for each meaning of the verb. However, dictionaries do not provide quantitative information such as how often each verb is used with each of the possible subcat frames. Since dictionaries are repositories, primarily, of what is possible, not what is most likely, they tend to contain information about rare usage [de Marken, 1992]. But without information about the frequencies of the subcat frames we find in dictionaries, we face the prospect of having to treat each frame as equiprobable in parsing. This can lead to serious inefficiency. We also know that the frequency of subcat frames can vary by domain; frames that are very rare in one domain can be quite common in another. If we could automatically determine the frequencies of subcat frames for domains, we would be able to tailor parsing with domain-specific 95 heuristics. Indeed, it would be desirable to have a subcat dictionary for each possible domain. This paper describes a mechanism for automatically acquiring subcat frames and their frequencies based on a tagged corpus. The method utilizes a tagged corpus because (i) we don't …

Content truncated to 500 words