Automatic Acquisition of Subcategorization Frames from Tagged Text

Automatic Acquisition of Subcategorization Frames from Tagged Text Michael R. Brent and Robert C. Berwick MIT Artificial Intelligence Lab 545 Technology Square Cambridge, MA 02139 ABSTRACT This paper describes an implemented program that takes a tagged text corpus and generates a partial list of the subcategorization frames in wtfich each verb occurs. The completeness of the out- put list increases monotonically with the total occurrences of each verb in the training corpus. False positive rates are one to three percent. Five subeategorization frames are currently detected and we foresee no impediment to detecting many more. Ulti- mately, we expect to provide a large subcategorization dictio- nary to the NLP community and to train dictionaries for specific corpora. INTRODUCTION Accurate parsing requires knowing the subcategoriza- tion frames of verbs, as shown by (1). (1) a. I expected [NP the man who smoked NP] to eat ice-cream b. I doubted [NP the man who liked to eat ice-cream NP] Current high-coverage parsers tend to use either custom, hand-generated lists of subcategorization frames (e.g., [7]), or published, hand-generated lists like the Ox[ord Advanced Learner's Dictionary of Contemporary English, [9] (e.g., [5]). In either case, such lists are expensive to build and to main- tain in the face of evolving usage. In addition, they tend not to include rare usages or specialized vocabularies like financial or military jargon. Further, they are often incom- plete in arbitrary ways. For example, Webster's Ninth New Collegiate Dictionary lists the sense of strike meaning "to occur to", as in "it struck him that...", but it does not list that same sense of hit. (Our program discovered both.) To address these problems we have implemented a program that takes a tagged text corpus and generates a partial list of the subcategorization frames in which each verb occurs. The program uses only a small, finite-state grammar for a fragment of English. The completeness of the output list in- creases monotonically with the total number of occurrences of each verb in the training corpus. Automatically learning subcategorization frames (SFs) is impeded by a bootstrapping problem -- you can't parse without knowing SFs and you can't learn from examples without parsing them. For instance, the obvious approach to identifying verbs that take infinitival complements would be to look for a verb followed by an infinitive. Unfortu- nately, as shown by (1), finding such a case does not license any definite conclusions. Our system bootstraps by rec- ognizing those sentences that it can parse without already knowing the SFs -- mainly sentences involving pronouns or proper names rather than full noun-phrases in certain argu- ment positions. It simply ignores other sentences. The dis- tributional constraints on pronouns and full noun-phrases are almost identical, so lessons learned in the easy-to-parse cases apply to all cases. The remainder of this paper consists of a section de- scribing and quantifying our results, a section describing the methods used to obtain them, and a section discussing related work. RESULTS So far, we have concentrated on the five …

Content truncated to 500 words