Verb Sense and Subcategorization: Using Joint Inference to Improve Performance on Complementary Task

Verb Sense and Subcategorization: Using Joint Inference to Improve Performance on Complementary Tasks Galen Andrew, Trond Grenager, and Christopher Manning Computer Science Department Stanford University Stanford, CA 94305-9040 {pupochik, grenager, manning}@cs.stanford.edu Abstract We propose a general model for joint inference in corre- lated natural language processing tasks when fully anno- tated training data is not available, and apply this model to the dual tasks of word sense disambiguation and verb subcategorization frame determination. The model uses the EM algorithm to simultaneously complete partially annotated training sets and learn a generative probabilis- tic model over multiple annotations. When applied to the word sense and verb subcategorization frame determina- tion tasks, the model learns sharp joint probability dis- tributions which correspond to linguistic intuitions about the correlations of the variables. Use of the joint model leads to error reductions over competitive independent models on these tasks. 1 Introduction Natural language processing research has tradition- ally been divided into a number of separate tasks, each of which is believed to be an important sub- task of the larger language comprehension or gener- ation problem. These tasks are usually addressed separately, with systems designed to solve a sin- gle problem. However, many of these tasks are not truly independent; if solutions to one were known they would facilitate finding solutions to the others. For some sets of these problems, one would like to be able to do joint inference, where information of one kind can influence decisions about information of another kind and vice versa. For instance, in- formation about named entities can usefully inform the decisions of a part-of-speech tagger, but equally, part-of-speech information can help a named entity recognizer. If one had a large corpus annotated with all the information types of interest, one could es- timate a joint distribution over all of the variables simply by counting. However, it is more often the case that one lacks any jointly annotated corpus, or at least one that is sufficiently large, given that the joint distribution is necessarily sparser than the marginal distributions. It would therefore be useful to be able to build a model for this joint inference task using only partially supervised data. In this System Name Accuracy kunlp 57.6 jhu-english-JHU-fi nal 56.6 SMUls 56.3 LIA-Sinequa-Lexsample 53.5 manning-cs224n 52.3 Table 1: Performance of the top 5 Senseval-2 word sense disambiguation systems when considering accuracy only on the 29 verbs. Systems not guessing on all instances have been omitted. paper we examine these problems in the context of joint inference over verb senses and their subcate- gorization frames (SCFs). 1.1 Verb Sense and Subcategorization Of the syntactic categories tested in the Senseval word sense disambiguation (WSD) competitions, verbs have proven empirically to be the most dif- ficult. In Senseval-1, Kilgarriff and Rosenzweig (2000) found a 10-point difference between the best systems’ performance on verbs compared with other parts-of-speech. In Senseval-2, Yarowsky and Florian (2002) also found that while accuracies of around 73% were possible for adjectives and nouns, even the most competitive …

Content truncated to 500 words