Learning Subcategorization

A method to identify the subcategorized constituents of a verb (its complements) automatically in a sentence is useful in various areas of Natural Language Processing (e.g. automatic acquisition of subcategorization lexicons, parsing, acquisition of verb semantics, information retrieval). I will describe a method for subcategorization identification that uses memorybased learning. Train and test material is extracted from the Penn Treebank II Wall Street Journal Corpus. It makes use of the part-of-speech tags of the words in the sentence and the chunks: base (i.e. non-recursive) noun phrases, adjectival and adverbial phrases. In the training material, each word or chunk has an associated class that describes its relation to the verb; whether it is directly dependent on the verb (the attachment), and if so, what its syntactic category and grammatical function are (e.g. predicative NP, temporal adjunct PP). All the training examples are stored in memory. Then the memory-based learner classifies new test instances by extrapolating from the class of the most similar training example. In doing so, it achieves a generalization accuracy of 88.3% on the word/chunk level, which corresponds to 91% of correctly identified subcategorization frames. LEARNING SUBCATEGORIZATION