A framework for modeling and representing temporal discourse structure

In this paper, we present a framework for modeling temporal reference in Chinese discourse. It generates the temporal relations (i.e. "before", "after" and "same-as") among the sentences by applying inference rules on a set of pre-specified linguistic features, including discourse features, such as theme, rhetorical relation as well as temporal features, such as tense/aspect markers and connectives. They are designed to determine time orientation or estimate the extent of time shift. A variety of experiments are conducted to compare the different impacts of these features on temporal reference determination and show that theme is the most significant one. The overall performance is promising and achieves 91.33% of accuracy on news articles. Meanwhile, we obtain the insight that the default forward movement of temporal reference, which is employed in previous researches, is not applicable to all genres of texts. In addition, we develop a graphic representation, namely temporal discourse map, which helps locating temporal references in a discourse.