Research On Technologies Of Relation Extraction Based On Frame Semantics

Relation extraction is the key task of information extraction,and its purpose is to identify the relation between entities and entities from given plain text,and transform it into a human-machine readable structured form for storage.It not only plays an important role in the construction and expansion of knowledge base,but also has important application value in intelligent search,automatic question answering,knowledge reasoning and other tasks.Traditional relation extraction methods rely on feature engineering and relation extraction patterns,which can not meet the needs of dynamic massive text relation extraction.In recent years,with the successful application of deep learning technology in different tasks of natural language processing,deep learning has become the mainstream methods of relation extraction.Relation extraction based on deep learning can automatically learn the semantic features and relation patterns for relation expression from large-scale text,which has significantly surpassed the traditional model in performance,but there are still some problems to be solved in semantic representation,distant supervision noise,data imbalance.Frame semantic knowledge base,from the perspective of human cognition,takes frame as the research object,and describes entities,the relations between entities and the events that entities participate in through the semantic features such as lexical units,frame elements,frame relations.In relation extraction model based on deep learning,frame semantic features are introduced to realize entity semantic scene representation,which can enhance the semantic representation of entity context and improve the performance of relation extraction model.Therefore,this paper focuses on the challenge of relation extraction,combines with the frame semantic knowledge,and studies from four aspects: frame identification and the extraction of frame semantic features,the semantic representation of relation extraction,the noise of distantly supervised relation extraction,and the data imbalance of relation extraction.The main research contents and contributions of this paper are as follows:(1)Frame recognition and frame semantic features extraction.Frame recognition is a basic work of all the researches in this paper,which is directly related to the accuracy of semantic scene for entity.Therefore,this paper proposes a Chinese frame recognition model based on DNN.In this model,the context of target words is represented by dependency relation and word vector of dependency words;the context features of target words are learned automatically by neural network.In order to get more attention to important features,a frame recognition model based on convolutional neural network is proposed,which exploits twolevel attention mechanism.The model introduces attention mechanism in input layer and pooling layer respectively,trying to guide the model to pay more attention to the word and n-gram features closely related to the target word in the learning process;moreover,the model adopts multi-dimensional convolution kernel,which can capture features of different granularity,and the effect is significantly better than all baseline models.On the basis of frame recognition,single frame semantic feature extraction algorithm and extended frame semantic feature extraction algorithm are proposed to provide technical support for related tasks of relation extraction based on frame semantics.(2)Relation extraction based on frame semantics and sequence features.Aiming at the problem of semantic representation of neural relation extraction model,a relation extraction model based on frame semantics and sequence representation is proposed.The single-frame semantic features fusion method and the extended-frame semantic features fusion method based on attention mechanism are studied,and a multi-level frame semantic features fusion model is proposed.The semantic features within and between frames is fully considered in the representation of entity semantic scene.On this basis,a method based on the combination of frame semantic features and sequence features is proposed.Then,Transformer neural network architecture based on self attention mechanism is used to model the long-distance dependence between entities,which is superior to all baseline models in both Chinese and English datasets.(3)Context-aware based on frame-semantics for distantly supervised relation extraction.Aiming at the noise problem of distantly supervised relation extraction,a rule-based instance selection method and a distantly supervised relation extraction model based on frame-semantic context aware are proposed.Firstly,the rule-based method is used to select the effective instances for the bag;a hierarchical frame-semantic representation model is proposed,which encourages the model to pay more attention to the frame and semantic scenarios that are important to the target relation through the two-layer attention mechanism,so as to realize the distantly supervised relation extraction based on the frame semantics.Experimental results show that the proposed method can effectively reduce the noise problem of distantly supervised relation extraction.(4)Imbalanced-data relation extraction based on frame semantics and multi-instance learning.In order to solve the problem of imbalanced data in relation extraction,a relation extraction model based on frame semantics and multi-task learning is proposed.This model introduces the representation of frame-semantic scene for entity pair,which aims to learn a general entity context representation mode from few-sample relation instances,and alleviate the sample imbalance problem;moreover,increasing the representation of semantic scene in the model can enhance the distinction between positive and negative instances,and weaken the impact of negative instances on the model performance.Aiming at the problem of too many negative samples,the model adopts the multi-task learning method;the relation extraction task is divided into two sub tasks:relation recognition and relation classification.The parameter sharing mechanism is used to jointly optimize the objective function of the two tasks,which reduces the impact of negative samples on the performance of relation extraction.