Research On Multi-feature Fusion Entity Relationship Extraction Model Based On Deep Learning

The rapidly developing Internet is generating a large amount of valuable information all the time,most of which exists in unstructured form.How to quickly identify and extract these high-value information has become a hot research field that has attracted much attention.Relation extraction technology can automatically and accurately extract the semantic relationships between entities in text and convert them into structured triplet data,which provides a series of knowledge services for downstream tasks such as knowledge graph and information retrieval,and has great research significance in the development of NLP related technologies.This thesis focuses on the problems of relying on domain knowledge database features,lacking further research on the optimal combination of network and features,insufficient feature extraction contained in the data itself,focusing on the English field and relatively few Chinese,and so on in previous research work.Two entity relationship extraction models have been improved and implemented,and experimental verification has been carried out on a common dataset.The specific work contents are as follows:(1)A relation extraction model based on multi feature fusion combining entity context is proposed to address the limitations of feature extraction using different neural networks and the insufficient utilization of entity related information in sentences.In this model,multi-window CNN structure and maximum pooling are used to fully dig multi-granularity local features in sentences,Bi LSTM network is used to combine attention mechanism to extract sentence global features based on words,and word entity attention network is used to capture relevant background word features in entity context,and then weighted fusion of the three features is used for relation extraction.The features were optimized and combined on the public data set Sem Eval-2010 Task 8 to verify the feasibility and effectiveness of the model,and compared with the classical mainstream model.The results show that the model performs better than other comparison models without introducing external features.(2)A general relation extraction model in Chinese and English that enhances entity semantics is proposed to address the problem of inadequate contextual semantic information mining and to meet the requirements of model building with different language characteristics.The model uses the BERT pre-training model as the word embedding layer,which can better understand the semantic information of Chinese and English sentences.Then Bi GRU network is used to further extract features and learn deep semantic information.The entity labeling mechanism was used to locate the entity to improve the sensitivity of the model to the entity,and the semantic representation of the entity was obtained through average pooling.Combined with the attention mechanism,the weighted sentence features that contribute significantly to the classification of relation were obtained.Finally,entity representation,sentence feature and sentence level feature represented by encoder head node are splice and fused for relation extraction.Experiments on Chinese open source data sets and English open source data sets show that the model has a certain universality in processing Chinese and English data,and has better performance than the classical mainstream model.