Research On Entity Relation Extraction Method Based On Deep Learning

With the rapid development of Internet technology and the advancement of storage technology,text data is increasing day by day.How to extract valuable information from it and make it serve society has always attracted the attention of researchers.The task of entity relation extraction is the core task of information extraction technology in the field of natural language processing.Its main purpose is to identify entities from natural language texts while extracting entity-relation triples containing the relation between entity pairs.Valuable information can be extracted from irregular data,and it can also provide basic technical support for downstream tasks.The current entity relation extraction methods based on deep learning have achieved good results in closed-domain scenarios,but most of them cannot solve the problem of entity overlap in natural text very well.Since the parameters of the model are obtained from the training set in a closed domain,the model in this scenario can only extract predefined entity relations and cannot detect unknown entity relations and continue learning.Therefore,starting from the task of entity relation extraction for closed domains,this thesis first studies the entity relation extraction method in the entity overlapping scenario.Then studies the entity relation extraction method in the open domain scenario with unknown entity relation types.Finally,Introduce continuous learning into the entity relation extraction task,and research the continuous relation extraction method.The main research contents are as follows:(1)A method for entity relation extraction based on sequence annotation is proposed.By dividing the entity-relation extraction task into two sub-tasks: entity extraction and subject-object alignment,the constructed relational matrix is used to solve the problem of relational overlap existing in complex scenes.First,obtain the vectorized representation of the sentence through BERT.Then use the entity extraction component to extract all the entities in the sentence.Finally construct a relation matrix to align the subject and object,thereby extracting the triples corresponding to the text.The method proposed in this thesis is evaluated on the international public relation extraction benchmark dataset.The experimental results show that the method proposed in this thesis shows better entity relation extraction performance than the benchmark model.(2)A new unknown relation detection method based on the local outlier factor algorithm is proposed.By introducing the local outlier factor algorithm to process the extracted deep-level text features,it is possible to extract known relations and detect unknown entity relations at the same time.First,the model encodes the input sentences and annotated entities into relation representation vectors.Then,the semantic features of entity relations are learned using the cross-entropy function.Finally,the local outlier factor algorithm is used to detect whether the input instance belongs to an unknown class in the testing phase.The method proposed in this thesis is evaluated on the public relation extraction benchmark dataset.The experimental results show that the performance of the method proposed in this thesis is better than the existing methods,and the robustness of the method is further demonstrated through auxiliary experiments.(3)A persistent relation extraction method based on relative entropy is proposed.By iteratively replaying the representative samples in the memory bank and using the relative entropy to constrain them,the consistency of the relational embedding in the feature space is ensured,thereby alleviating the catastrophic forgetting phenomenon encountered by the model’s continuous learning of new relations,so that the model has continuous learning ability.First,supervised learning is used for each new task,so that the model can effectively learn the relation representation of the instance.Then,after the current task training,the representative samples of each type of relation are stored in the memory bank.Finally,when learning a new task While iteratively replaying the samples in the memory bank,the relative entropy of the samples is used to make the model retain the previously learned knowledge as effectively as possible.Experimental results on public datasets show that the proposed method outperforms the baseline models.In addition,this thesis further demonstrates the effectiveness of the proposed method by adding traditional topology loss for fusion experiments.Experiments show that the method proposed in this thesis has achieved better performance in the corresponding research direction,and by solving the core problems in the entity relation extraction task in complex scenes,it provides different methods for entity relation extraction researchers to a certain extent research ideas.