Chinese Temporal Information Extraction In Financial Field

Temporal information is an important component of an event. Some researches show that, the percent of text content expressed by temporal information is about 27%, only less than proper noun, which percent is about 31%. Thus, temporal information processing is a very important component in Natural Language Processing, and it plays a significant role in Information Extraction, Information Retrieval, Quest and Answer, Text Summary and Data Mining. Therefore, building an Information-Extraction-based temporal system is critical for temporal information extraction and temporal information analysis.This paper puts the emphasis on both the research about the theoretic frame of Chinese temporal system and the work to build an applied temporal processing platform that includes two main parts: Chinese temporal expression recognition and temporal relationship analysis.To recognize the Chinese temporal expression, we divide the task into two sub-tasks: the explicit temporal expression recognition and the implicit temporal expression recognition. For explicit temporal expression recognition, the "two steps" strategy is applied. Firstly, separate simple temporal expressions are recognized. Than by matching context-sensitive template rules, these simple temporal expressions are combined to form complicated temporal expressions. In open test, the F-Measure of complicated temporal expression recognition is 95.0% and it is primarily up to the requirement of latter tasks. For implicit temporal expression recognition, witch is also called as Chinese situation analysis. The Bayes Classification is used for Chinese verb classification. In open test, the F-Measure of Chinese verb classification is 86.66% and it reaches coequal performance of international researches on Chinese verb classification. Moreover, the Bayes verb classification overcomes the disadvantage of traditional methods that depend on dictionary resource and syntactic resource. Another work of this paper comes down to an important task in temporal relationship research: time-event mapping analysis. Different from traditional rule-based methods, we use a machine learning method, transformation-based error-driven learning algorithm to determine the time-event mapping relationship, which can automatically acquire the analytical rules. In open and close test, by using the transformation rule set, the error rate of time-event mapping analysis is 17.25% and 27.27% respectively. The experiment indicates that the error-driven learning algorithm is a good patch for based-rule method.