Research On Entity Relation Extraction For Biomedical Texts

At present,a large number of biomedical texts are stored in various knowledge bases,and the number is still growing exponentially.These documents are the crystallization of biomedical technology.Mining valuable knowledge hidden in biomedical texts through information extraction technology is an important research topic.Entity relation extraction is a key sub-task of information extraction.It is usually formulated as classifying texts containing candidate biomedical entity pairs into the predefined relation types.The research on entity relation extraction for biomedical texts aims to building a deep learning model that can automatically extract structured biomedical entity relation triples from unstructured biomedical texts at the moment.Different from the general field,biomedical texts are generally longer and more complex.Besides,there are many noise words in the biomedical texts,such as some special medical symbols.These bring great challenges to biomedical entity relation extraction.Therefore,this thesis first proposes a self-supervised graph attention network in the sentence-level biomedical relation extraction task.The method incorporates self-supervision within the standard graph attention mechanism.Graph attention network can automatically learn the relation importance between words.The model pays more attention to learning useful information,and thus reduces the influence of noisy words.With the self-supervision of dependency-based parse trees,the graph attention network not only can improve its capacity of learning syntactic information but also can alleviate its lack of interpretability.The experimental results on several public datasets suggest that the model performs well in sentence-level biomedical relation extraction task.In real-world biomedical texts,a large number of relations between entities are expressed over multiple sentences.The goal of document-level relation extraction is to extract different relations from multiple entity pairs in the document.In general,different relations in a document are often interrelated.Thus this thesis proposes an entity-level relation graph-based method to document-level relation extraction.Firstly,we utilize an encoder module to capture the context information of entities,and a fully connected graph is formed with all entities as nodes and relations as edges.Then we leverage the convolutional neural network and selfattention mechanism to update the representations of relational edges,which focus on local and global relational dependencies respectively.Experimental results demonstrate that the method can effectively improve the performance of document-level relation extraction.Now most biomedical relation extraction models identify relations according to text and entity information,but ignore the important role of triggers.The triggers appearing in texts can generally directly represent the relation types of the entity pairs,which can obviously guide the model to predict the relations correctly.Therefore,we construct an annotation system to obtain the dataset firstly.Then we propose a deep learning method to extract relations and triggers jointly.We apply the pre-trained language model and bi-directional long-short term memory to extract global semantic features and sequence features,respectively.Then these shared features are used to extract relations and identify triggers.The experimental results show that the joint relation and trigger extraction can improve the performance of both relation extraction and trigger recognition.