Understanding time is essential to understanding events in the world. Knowing what has happened, what is happening, and what may happen in the future is critical for reasoning about those events. It is thus an important natural language processing (NLP) task to understand time.This thesis advances the study of time by developing new insights into some aspects of the problem of reasoning about time in text, new algorithmic and machine learning approaches, and new datasets that would support continuing work on these problems by the research community. We also discuss a few research directions suggested by this work that could further improve our understanding of time in natural language text. The thesis specifically addresses three key aspects of the temporal reasoning problem: time expression understanding, temporal relation extraction, and temporal common sense acquisition. Time expressions (e.g., yesterday or last month) often provide absolute time anchors for events. Temporal relations (e.g., event A is before or after event B) provide relative order information between events, which is complementary to time expressions. Temporal common sense (e.g., duration and frequency) is another important component in temporal reasoning, but is usually absent in a single piece of text because people do not say things that are obvious. The bulk of this thesis is devoted to the important problem of identifying temporal relations between events, a problem that has been studied a lot by the research community. The work in the thesis introduces new machine learning methods and a novel conceptual view of the problem that together result in an improvement of more than 20% over the previous state-of-the-art.
PDF
download
878987797
028-85220240
