The (Non)Utility of Predicate-Argument Frequencies for Pronoun Interpretation

The (Non)Utility of Predicate-Argument Frequencies for Pronoun Interpretation Andrew Kehler∗ UC San Diego akehler@ucsd.edu Douglas Appelt SRI International appelt@ai.sri.com Lara Taylor∗ UC San Diego lmtaylor@ucsd.edu Aleksandr Simma† UC San Diego asimma@ucsd.edu Abstract State-of-the-art pronoun interpretation sys- tems rely predominantly on morphosyntac- tic contextual features. While the use of deep knowledge and inference to improve these models would appear technically in- feasible, previous work has suggested that predicate-argument statistics mined from naturally-occurring data could provide a useful approximation to such knowledge. We test this idea in several system configu- rations, and conclude from our results and subsequent error analysis that such statis- tics offer little or no predictive information above that provided by morphosyntax. 1 Introduction The last several years has seen a number of works that use weight-based systems (trained either man- ually or via supervised learning) for pronoun in- terpretation, in addition to others that have ad- dressed the broader task of entity-level coreference (see Mitkov (2002) for a useful survey). These sys- tems typically rely on a variety of morphosyntactic factors that have been posited in the literature to affect the interpretation of pronouns in naturally- occurring discourse, including gender and number agreement, the distance between the pronoun and antecedent, the grammatical positions of the pro- noun and antecedent, and the linguistic form of the antecedent, among others. A common refrain is that the performance of systems that rely on such fea- tures is plateauing, and that further progress will re- quire the use of world knowledge and inference (ibid., Ch. 9, inter alia). World knowledge, after all, would seem to play a role in determining that the referent of it in example (1) is the entity denoted by his industry rather than Glendening’s initiative or the edge. ∗Department of Linguistics. †Department of Computer Science and Engineering. (1) He worries that Glendening’s initiative could push his industry over the edge, forcing it to shift operations elsewhere. Of course, no well-suited knowledge base and accom- panying inference procedure exists that can deliver such a capability robustly in an open domain. In lieu of this capability, previous authors have suggested that what can be viewed as a more su- perficial form of semantic information – predicate- argument statistics mined from naturally-occurring data – could be used to capture certain selectional regularities. For instance, such statistics might re- veal that forcing industry is a more likely verb- object combination in naturally-occurring data than forcing initiative or forcing edge. Assuming that such statistics imply that industries are more likely to be forced in the real world than are initiatives or edges, this information could be taken to establish a preference for his industry as the antecedent of it in (1). While there will always be cases that require ar- bitrarily deep knowledge for their interpretation, the empirical question of how far one can go by relying on this sort of selectional information remains. Our point of departure is the work of Lappin and Leass (1994, henceforth L&L) and Dagan et al. (1995). …

Content truncated to 500 words