An Automatic Approach For Conceptual System Construction In Traffic Law Domain

In 2016,the Supreme People's Court of China for the first time proposed to build a”smart court” based on the forefront of the development of The Times.Under such circumstances,the construction of legal knowledge graph is an effective auxiliary means for the construction of ”smart court”,which has important value.The premise of constructing the domain knowledge graph is to construct its conceptual system,including the domain concept,the hypernymy between the domain concepts,and the concept attribute relationship.The traditional method of conceptual system construction is to use domain experts to construct manually,which needs a lot of manpower cost.In addition,the focus of this thesis is the domain concept,the hypernymy between concepts and the relationship of concept attributes,which belong to the category of conceptual system.Therefore,the automatic construction of concept system has become an important topic to accelerate the construction of domain knowledge graph.There are few researches on the existing methods of domain concept extraction and domain concept relationship detection,and a large number of annotated data sets are needed,which limit the development of automatic concept system construction methods.Based on the above background,this thesis studies the automatic construction method of concept system for the field of road traffic law,which can automatically construct the concept system of the field without manually annotating the data set.The main research contents are as follows:(1)A domain concept extraction method based on shallow syntactic information is proposed.In this method,candidate phrase sets are extracted from text corpus by using N-grams model through candidate phrase evaluation module,and then the candidate phrase sets are evaluated according to Point-Wise Mutual Information,Point-Wise KL Divergence and Term Frequency-Inverse Document Frequency(TF-IDF).Then,through the candidate domain concept evaluation module,combined with the encyclopedia knowledge base,remote training is used to evaluate the candidate domain concept set.Then,through POS evaluation module,the concept set of candidate domain is further evaluated according to the POS tagging sequence of the concept of candidate domain in the text.(2)A domain concept relationship detection method based on pre-training language model is proposed.First,find the corresponding text sequence according to the input concept.The input sequence is then encoded using an advanced pretraining language model.Then we use the mask sequence constructed by the position information of the concept reference in the text sequence to filter the output sequence of the pretraining language model.Then the attention mechanism and pooling operation are used to obtain the contextual embedding of the concept.Finally,this thesis uses a dual affine operation to score the concept pairs,and uses the Log Sum Exp function to aggregate the original scores to predict the input relational categories.(3)The domain corpus and domain concept relationship detection data set used in the text are constructed,and the effectiveness of the proposed method is verified through the overall experimental results,ablation experiments and error analysis.The significance of this work is that it provides an effective solution for the task of automatic construction of concept system.This scheme firstly extracts domain concept sets based on shallow syntactic information,and then identifies hypernymy and concept attribute relation based on domain concept relation detection based on pre-training language model.