Abstract

Answering questions that ask about temporal information involves several forms of inference. First, relations between events and temporal expressions need to be inferred, either in the question or in the answer. Second, semantic inference between events, entities and their definitions needs to be performed. Sometimes, semantic inference employs aspectual information to relate events expressed by different lexical units. Third, temporal inference is preconditioned by relations between events, or between events, their arguments and their time anchors. In this paper we present a Question Answering (QA) methodology for handling temporal inference by combining all these forms of inference. The Problem The AQUAINT 1 program is a multi-project effort aimed at creating capabilities of processing complex questions and finding their answers from heterogeneous collections of texts. An important component of this effort deals with the recognition and processing of temporal information for Question Answering (QA). When asking a question that refers directly or indirectly to a temporal expression, the answer is expected to validate the temporal constraints. To achieve such functionality, QA systems need to (1) incorporate relations between temporal expressions and events or entities mentioned in the question; and (2) to rely on temporal inference for justifying the answer. Answer justification was presented in (Harabagiu et al. 2001) as a multifeedback methodology, but this mechanism did not include any temporal constraints. Whenever the answer to a question needs to be justified, if temporal expressions are involved, the justification must contain some form of temporal inference. For example, the expected answer type of question Q 1 is a DATE: Q 1: “When did Iraq invade Kuwait?” The expected answer type is an argument of the event E1=“invade ” which has two more arguments: “Iraq ” and

Citations