Discriminative Learning of Selectional Preference from Unlabeled Text

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 59–68, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Discriminative Learning of Selectional Preference from Unlabeled Text Shane Bergsma Department of Computing Science University of Alberta Edmonton, Alberta Canada, T6G 2E8 bergsma@cs.ualberta.ca Dekang Lin Google, Inc. 1600 Amphitheatre Parkway Mountain View California, 94301 lindek@google.com Randy Goebel Department of Computing Science University of Alberta Edmonton, Alberta Canada, T6G 2E8 goebel@cs.ualberta.ca Abstract We present a discriminative method for learn- ing selectional preferences from unlabeled text. Positive examples are taken from ob- served predicate-argument pairs, while nega- tives are constructed from unobserved combi- nations. We train a Support Vector Machine classifier to distinguish the positive from the negative instances. We show how to parti- tion the examples for efficient training with 57 thousand features and 6.5 million training instances. The model outperforms other re- cent approaches, achieving excellent correla- tion with human plausibility judgments. Com- pared to Mutual Information, it identifies 66% more verb-object pairs in unseen text, and re- solves 37% more pronouns correctly in a pro- noun resolution experiment. 1 Introduction Selectional preferences (SPs) tell us which argu- ments are plausible for a particular predicate. For example, Table 2 (Section 4.4) lists plausible and implausible direct objects (arguments) for particu- lar verbs (predicates). SPs can help resolve syntac- tic, word sense, and reference ambiguity (Clark and Weir, 2002), and so gathering them has received a lot of attention in the NLP community. One way to determine SPs is from co-occurrences of predicates and arguments in text. Unfortunately, no matter how much text we use, many acceptable pairs will be missing. Bikel (2004) found that only 1.49% of the bilexical dependencies considered by Collins’ parser during decoding were observed dur- ing training. In our parsed corpus (Section 4.1), for example, we find eat with nachos, burritos, and tacos, but not with the equally tasty quesadillas, chimichangas, or tostadas. Rather than solely re- lying on co-occurrence counts, we would like to use them to generalize to unseen pairs. In particular, we would like to exploit a number of arbitrary and potentially overlapping properties of predicates and arguments when we assign SPs. We do this by representing these properties as fea- tures in a linear classifier, and training the weights using discriminative learning. Positive examples are taken from observed predicate-argument pairs, while pseudo-negatives are constructed from unob- served combinations. We train a Support Vector Ma- chine (SVM) classifier to distinguish the positives from the negatives. We refer to our model’s scores as Discriminative Selectional Preference (DSP). By creating training vectors automatically, DSP enjoys all the advantages of supervised learning, but with- out the need for manual annotation of examples. We evaluate DSP on the task of assigning verb- object selectional preference. We encode a noun’s textual distribution as feature information. The learned feature weights are linguistically interesting, yielding high-quality similar-word lists as latent in- formation. Despite its representational power, DSP scales to real-world data sizes: examples are parti- tioned by predicate, …

Content truncated to 500 words