Research On Chinese Hedge Scope Detection Based On Phrase

Hedge cues are defined as “words which make things vague”.The information dominated by hedges is called hedge information.Hedge information detection aims to distinguish the factual information and uncertain information,so as to improve the authenticity and reliability of information extraction.Researches on hedge information detection is of great significance for information extraction.Hedge information detection contains two subtasks: hedge identification and hedge scope detection.Scope detection method based on word regards words as candidate classification units,which will produce a large number of negatives because the sentence is too long.Meanwhile,due to the similarity of structural information and contextual information between adjacent positives and negatives,it is difficult for classifier to identify the boundary of scope precisely.Therefore,this paper proposes the Chinese hedge scope detection method based on phrase which regards phrases as candidate classification units.Hedge information is a phrase or clause which is dominated by hedge cue,and semantically it refers to the whole semantic scope governed by hedges.It shows that structural information and semantic information is useful for scope detection.This paper carries out the study of Chinese hedge scope detection on Chinese biomedical hedge information corpus from three aspects: structural information representation,semantic information representation,and the fusion of structural information and semantic information.Main contents are as follows:Research on Chinese hedge scope detection based on composite kernel model.Regarding phrases as candidate classification units,this paper could capture shallow flat information and syntactic structure information based on polynomial kernel function and tree kernel function respectively.Moreover,this paper explores the representation method of syntactic structure information.This paper uses the composite kernel function to fuse the polynomial kernel function and the tree kernel function to build Chinese hedge scope detection system,and the sentence level F1 value(F1_sen)of scope detection system obtains 70.19%,which is 4.06%,20.27% higher than individual kernels.The results show that the method based on composite kernel model is effective and could capture shallow flat information and structural information.Research on Chinese hedge scope detection based on neural network model.This paper uses neural network model to capture deep semantic information and explores the semantic information representation of Chinese hedge scope.Experimental results indicate that the performance of hedge scope detection system obtains 67.84% F1_sen.It shows that the method for scope detection based on neural network model could effectively capture deep semantic information of Chinese hedge scope,and improve the performance of scope detection.Research on Chinese hedge scope detection based on composite kernel and neural network model.The method for hedge scope detection based on composite kernel and neural network model could capture flat and structural information,and semantic information respectively.In order to leverage three kinds of information,this paper explores the fusion method of results for scope detection based on weight adjustment,the fusion method of results for scope detection based on XGBoost,and constructs Chinese hedge scope detection system based on composite kernel and neural network model.The performance of scope detection system achieves 71.68% F1_sen,which is 1.49%,3.84% higher than individual system based on composite model,neural network model.It shows that the fusion method could effectively capture flat information,structural information and semantic information of hedge scope,moreover,the integration of three kinds of information could further improve the performance of Chinese hedge scope detection system.