Question answering based on temporal inference
Sanda M. Harabagiu, Cosmin Adrian Bejan · 2005
Answering questions that ask about temporal information in-volves 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 infer-ence between events, entities and their definitions needs to be performed. Sometimes, semantic inference employs aspec-tual information to relate events expressed by different lexi-cal units. Third, temporal inference is preconditioned by rela-tions between events, or between events, their arguments and their time anchors. In this paper we present a Question An-swering (QA) methodology for handling temporal inference by combining all these forms of inference. The Problem The AQUAINT1 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 an-swer is expected to validate the temporal constraints. To achieve such functionality, QA systems need to (1) incor-porate relations between temporal expressions and events or entities mentioned in the question; and (2) to rely on tem-poral inference for justifying the answer. Answer justifica-tion was presented in (Harabagiu et al. 2001) as a multi-feedback 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 infer-ence. For example, the expected answer type of question Q 1 is a DATE: