Research On Frame Identification Based On Syntactic And Semantic Role Information

Frame Identification is to find an active frame for the target word in a given sentence.It is a subtask of frame semantic analysis.Because of the polysemy of the target word,it is a challenge for frame identification to distinguish the meaning of the target word according to the context and determine the activated frame.Integrating the syntactic and semantic role information of Prop Bank with the target word plays an important role in analyzing the frame of the target word.Therefore,the focus of this paper is on investigating how to integrate syntactic and Prop Bank semantic role information with the representation of the target word.The primary contributions and achievements of this work are as follows:(1)Aiming at the fusion of the target word with its syntactic and semantic role information in the framework identification task,this paper presents a graph-based model for frame identification model that combines target words with syntactic and semantic roles.The model learns syntactic and semantic role information of target words through the graph convolutional network.In order to solve the problem that a single model cannot capture the difference of syntactic and semantic role information of multiple parts of speech,a syntactic or semantic role training model is used in turn,and the difference of syntactic and semantic role information of different parts of speech is captured by multi-model voting.The effectiveness of the approach was demonstrated by a 1.76% improvement over the baseline model in the CFN dataset and a 0.93% improvement over the current best model in the Frame Net1.7 dataset.(2)Aiming at the differences of syntactic and semantic role information between different part of speech of target words in frame recognition tasks,which cannot be effectively represented by a single model,this paper proposes a frame recognition method based on Multiple POS-aware Mixture of Experts network.In this method,multiple experts are designed in a Mixture of expert network to learn different syntactic and semantic role information,so as to capture the information differences caused by different parts of speech,so as to achieve fine-grained information fusion.According to the experimental results on Frame Net1.5 dataset,the proposed method achieves 1.38% higher of accuracy than that of the current best model.On the Frame Net1.7 dataset,the proposed method achieves similar performance to the voting method in accuracy while reducing the number of model parameters required,demonstrating the effectiveness of the model.(3)Aiming at the problem of frame selection in knowledge base construction,this paper propose a frame selection module by using a graph-based frame identification model that integrates target lexical syntactic and semantic roles,which provides an auxiliary tool for knowledge base construction.This module has better representation ability than the previous module,and achieves a 4.5 times improvement in reasoning speed,which proves the effectiveness of the method.The main contributions of this paper include: 1)A graph-based model for frame identification model that combines target words with syntactic and semantic roles is proposed to solve the fusion problem of syntactic and semantic role information and target words;2)A frame recognition method based on Multiple POS-aware Mixture of Experts network is proposed,which integrates syntactic and semantic role information between target words of different parts of speech,and alleviates the problem that single model cannot effectively represent different syntactic and semantic role information;3)Based on the above framework recognition model,a framework selection module is developed and integrated into the Chinese Frame Net Knowledge Base construction system,providing auxiliary tools for man-machine collaboration framework creation,lexical extension and data annotation.