Constructing a Temporal Relation Identification System of Chinese based on Dependency Structure Analysis

"Temporal information (Time)" has been a subject of study in many disciplines particularly in philosophy, physics, and is an important dimension of natural language processing. The temporal information includes temporal expressions, event and temporal relations. There are many researches dealing with the temporal expressions and event expressions. However, researches on temporal relation identification and the construction of temporal relation annotated corpus are still limited. There is a well-known temporal information annotated guideline for English, TimeML (Pustejovsky, 2006 [59]) . However, there is no such a research that focuses on this in Chinese. Our research is the first work of the temporal relation identification between verbs in Chinese texts. In this research, we propose a temporal information annotation guideline for Chinese and a machine learning-based temporal relation identification method. Following the observation of our investigation, the distribution of events and temporal expressions is un-balance. The temporal information processing includes two independent tasks: anchoring the temporal expressions on a timeline and ordering the events to temporal order. Our research focuses on ordering the events, which is to identify the temporal relations between events. Because identifying the nominal event is difficult, we limit the events to the verbs in articles. The proposed annotation guideline is based on the TimeML language. We newly introduce dependency structure information to limit target temporal relations. The proposed method reduces the manual 2 Doctoral Dissertation, Department of Information Process, Graduate School of Information Science, Nara Institute of Science and Technology, NAIST-IS-DD0561040, February 7, 2008. vi efforts in constructing the annotated corpus. To annotate temporal relations of all combinations of events requires n(n-1)/2 manual judges. Our proposed method requires at most 3n manual judges. While the dependency structure based attributes reduce manual annotation costs, the limited relations preserve the majority of the temporal relations. We use a syntactic parsed corpus Penn Chinese treebank as the original data for annotating a basic annotated corpus. For using the dependency structure in temporal relation identification, we first construct a dependency analyzer for Chinese and combine it into the temporal relation annotating system. The accuracy of the dependency analyzer is 88% for word dependency analysis and this is better than existed Chinese dependency analyzer. The process of temporal relation identification includes following steps: to analyze the dependency structure, to analyze the temporal relation attributes of events and to extend the relation using the inference rule. We define events as those expressed by verbs and define the temporal relation types of event pairs which include the adjacent event pairs, the headmodifier event pairs and the sibling event pairs. These relations include most meaningful information, and we extend these relations using the inference rules to acquire long distance relations. We train a machine learner with our temporal relation annotated corpus to construct the temporal relation identifying system. Support Vector Machine is used as the machine learner in this system. We survey the coverage of our system with a small corpus. The accuracies of the annotating experiments are 68%~71% for annotating the temporal relation attributes. The result shows that our proposed system covers about 53% of temporal relations of all possible event pairs.