Recognition Of Temporal Relation In Chinese Texts

Recognizing temporal relations in texts is fundamental for deep language understanding and helpful for many applications such as question answering, information extraction, and document summarization. Temporal relations are the relations between the events, or an event and a temporal expression in a text. At present, temporal relation recognition, especially in Chinese, is not satisfying. This article mainly studies the tasks of temporal relation recognition in TempEval-2. And the corresponding corpus is used in the experiments. We analyze the efficiency of many language features used in temporal relation recognition, and the methods of feature extraction. We propose the temporal relation recognition method based on Maximum Entropy Model, and use rules to recognize temporal relations between events and document creation time. We also explore the recognition of temporal relations between main events in adjacent sentences.The main work of this paper is as follows:1. Phrase structures, and syntactic elements have been annotated according to the guidelines of Harbin Institute of Technology;2. Temporal relation and their recognition tasks have been introduced, and the difficulties have been analyzed in detail;3. Recognition of temporal relations in the same sentence has been researched. The task includes two sub-tasks:extracting the pairs of temporal entities and classifying. The first sub-task is focused on identifying whether a pair of temporal entity is the true relation that conveyed in the text. We use syntactic relation between events or event and temporal expressions to extract the temporal entity pairs. And we use some linguistic features and ME model to classify them;4. Temporal relations between events and the document creation time have been recognized based on rules. According to the appearance of a reference time, related signal word, events will be categorized into four cases: an event with the reference time, which is the precise time expression; an event with the reference time, which is a fuzzy time expression; an event without the reference time, but with relevant signal word; an event without the reference time and with relevant signal word. We set up various identification rules for these cases to identify the relations;5. Also, temporal relation recognition between main events in adjacent sentences has been explored. Firstly, we analyze the above relations’ influences on it. Then, we use ME to identify the relations. Finally, we use rules to improve the classification results.