Inferring Selectional Preferences from Part-Of-Speech N-grams

Proceedings of the 13th Conference of the European Chapter of the Association for Computational Linguistics, pages 377–386, Avignon, France, April 23 - 27 2012. c⃝2012 Association for Computational Linguistics Inferring Selectional Preferences from Part-Of-Speech N-grams Hyeju Jang and Jack Mostow Project LISTEN (www.cs.cmu.edu/~listen), School of Computer Science Carnegie Mellon University Pittsburgh, PA 15213, USA hyejuj@cs.cmu.edu, mostow@cs.cmu.edu Abstract We present the PONG method to compute selectional preferences using part-of-speech (POS) N-grams. From a corpus labeled with grammatical dependencies, PONG learns the distribution of word relations for each POS N-gram. From the much larger but unlabeled Google N-grams corpus, PONG learns the distribution of POS N-grams for a given pair of words. We derive the probability that one word has a given grammatical relation to the other. PONG estimates this probability by combining both distributions, whether or not either word occurs in the labeled corpus. PONG achieves higher average precision on 16 relations than a state-of-the-art baseline in a pseudo-disambiguation task, but lower coverage and recall. 1 Introduction Selectional preferences specify plausible fillers for the arguments of a predicate, e.g., celebrate. Can you celebrate a birthday? Sure. Can you celebrate a pencil? Arguably yes: Today the Acme Pencil Factory celebrated its one-billionth pencil. However, such a contrived example is unnatural because unlike birthday, pencil lacks a strong association with celebrate. How can we compute the degree to which birthday or pencil is a plausible and typical object of celebrate? Formally, we are interested in computing the probability Pr(r | t, R), where (as Table 1 specifies), t is a target word such as celebrate, r is a word possibly related to it, such as birthday or pencil, and R is a possible relation between them, whether a semantic role such as the agent of an action, or a grammatical dependency such as the object of a verb. We call t the “target” because originally it referred to a vocabulary word targeted for instruction, and r its “relative.” Notation Description R a relation between words t a target word r, r' possible relatives of t g a word N-gram gi and gj ith and jth words of g p the POS N-gram of g Table 1: Notation used throughout this paper Previous work on selectional preferences has used them primarily for natural language analytic tasks such as word sense disambiguation (Resnik, 1997), dependency parsing (Zhou et al., 2011), and semantic role labeling (Gildea and Jurafsky, 2002). However, selectional preferences can also apply to natural language generation tasks such as sentence generation and question generation. For generation tasks, choosing the right word to express a specified argument of a relation requires knowing its connotations – that is, its selectional preferences. Therefore, it is useful to know selectional preferences for many different relations. Such knowledge could have many uses. In education, they could help teach word connotations. In machine learning they could help computers learn languages. In machine translation, they could help generate more natural wording. This paper introduces a method named PONG (for Part-Of-Speech …

Content truncated to 500 words