Learning Methods for Combining Linguistic Indicators to Classify Verbs

Learning Methods for Combining Linguistic Indicators to Classify Verbs Eric V. Siegel Department of Computer Science Columbia University New York, NY 10027 evs@cs.columbia.edu Abstract Fourteen linguistically-motivated numeri- cal indicators are evaluated for their abil- ity to categorize verbs as either states or events. The values for each indicator are computed automatically across a corpus of text. To improve classification perfor- mance, machine learning techniques are employed to combine multiple indicators. Three machine learning methods are com- pared for this task: decision tree induction, a genetic algorithm, and log-linear regres- sion. 1 Introduction The ability to distinguish states, e.g., "Mark seems happy," from events, e.g., "Rende ran down the street," is a necessary prerequisite for interpreting certain adverbial adjuncts, as well as identifying temporal constraints between sentences in a dis- course (Moens and Steedman, 1988; Doff, 1992; Kla- vans, 1994). Furthermore, stativity is the first of three fundamental temporal distinctions that com- pose the aspectual class of a clause. Aspectual clas- sification is a necessary component for a system that analyzes temporal constraints, or performs lex- ical choice and tense selection in machine transla- tion (Moens and Steedman, 1988; Passonneau, 1988; Doff, 1992; Klavans, 1994). Researchers have used empirical analysis of cor- pora to develop linguistically-based numerical indi- cators that aid in aspectual classification (Klavans and Chodorow, 1992; Siegel and McKeown, 1996). Specifically, this technique takes advantage of lin- guistic constraints that pertain to aspect, e.g., only clauses that describe an event can appear in the pro- gressive. Therefore, a verb that appears more fre- quently in the progressive is more likely to describe an event. In this paper, we evaluate fourteen quantitative linguistic indicators for their ability to classify verbs according to stativity. The values of these indicators are computed automatically across a corpus of text. Classification performance is then measured over an unrestricted set of verbs. Our analysis reveals a pre- dictive value for several indicators that have not tra- ditionally been linked to stativity in the linguistics literature. Then, in order to improve classification performance, we apply machine learning methods to combine multiple indicators. Three machine learn- ing techniques are compared for this task: decision tree induction, a genetic algorithm, and log-linear regression. In the following sections, we further detail and motivate the distinction between states and events. Next, we describe our approach, detailing the set of linguistic indicators, the corpus and tools used, and the machine learning methods. Finally, we present experimental results and discuss conclusions and fu- ture work. 2 Stative and Event Verbs Stativity must be identified to detect temporal con- straints between clauses attached with when. For ex- ample, in interpreting, "She had good strength when objectively tested, ,1 the have-state began before or at the beginning of the test-event, and ended after or at the end of the test-event. However, in inter- preting, "Phototherapy was discontinued when the bilirubin came down to 13," the discontinue-event began at the end of the come-event. As another ex- ample, the simple present reading of an event, …

Content truncated to 500 words