Linguistic Knowledge-Enhanced Graph Neural Network For Event Relation Extraction

Event relation extraction is an important issue in the field of information extraction,which aims to extract unstructured text into structured event knowledge and is very important for the downstream tasks such as question answering system and event logic graph construction.Traditional event relation extraction is mostly based on neural network to extract text features,but the supervised method based on neural network must be driven by a large amout of annotated data.However,in the real world,data is easy to obtain,but annotated data is rare,and the cost of manual annotated data is often huge,especially fields such as biology and medicine.In addition,it is difficult for existing deep learning methods to accurately capture the structural event information due to the complex expression patterns among the event information in the text.Therefore,it is also a great challenge to extract rich event relation features from sentences with complex semantic interactions.From the perspective of solving the above problems,this paper puts forward two kinds of event relation extraction models.The main contributions of this thesis including the following three aspects:1、Aiming at the exisiting deep learning methods are unable to accurately capture structured event information due to the lack of large-scale and high-quality annotated data and the complex semantic interaction between the event information in the text in the event relation extraction task,this paper proposes a linguistic knowledge-enhanced self-supervised graph convolutional network(GCN)for event relation extraction.The model introduces linguistic knowledge such as morphology and syntax,and uses GCN to model the syntactic dependency tree of sentences to extract the rich relation representations,conducts adaptive clustering based on the relation representation features and generates pseudo-labels as the self-supervision information,refines event pair representations via self-supervised training schema.Finally,the experimental results on two datasets to verify the effectiveness of the model.2、Aiming at the problem of the complex expression of the relationship between events in sentences lead to the imprecise parsing,and in order to fully extract the feature of event pair relation,this paper proposes multimodal bilinear pooling fusion graph neural network for event relation extraction.The model uses the probability matrix of all dependency arcs from a dependency parser to build a syntax-based graph convolutional network,and the graph attention network is used to extract the semantic features of sentences.Then the model applys multimodal bilinear pooling to fuse the features obtained from the two modules,and the enhanced features after fusion are used for relation classification.Finally,experiments on two datasets show that the performance of this model is significantly better than of the baseline model.3、Based on the event relation extraction model proposed in this paper,an event relation extraction system is designed and implemented.The system can automatically annotate input sentences and label corresponding relation types,which realizes the transformation from algorithm to application,greatly readuces the workload and has high accuracy.In future research,the system will help to realize the automation of event relation extraction,so as to further improve the application efficiency and accuracy of natural language processing technology in various fields.