Research On Multi-feature Perception Entity Relation Extraction Model Based On Attention Mechanism

With the comprehensive construction of a smart grid and the promotion of a globalization strategy,the electric power field has accumulated a large amount of unstructured Chinese and English text data that has great value to the development of the industry.How to automatically mine the required structured information from the massive unstructured power text data has become a key research direction for text data mining in the power industry,and has important research significance and application value for the construction of smart grid.As a key technology in natural language processing,relation extraction technology can identify semantic relations between entity pairs from unstructured text data and transform unstructured text into structured data for storage.The article aims to investigate the problems of existing relation extraction models in inter-entity semantic feature extraction,entity context key phrase information extraction,and Chinese semantic information characterization,design and implement three entity relation extraction frameworks,and experimentally validate the constructed models on general domain data.The main contributions of this paper are summarized as follows.(1)A multi-feature fusion entity relation extraction model is proposed for the problem of inadequate semantic feature extraction among entities.The model first introduces character-level word vectors on top of the traditional word vectors to enhance the semantic representation of word vectors.Secondly,the multi-window CNN network and segmented Bi LSTM attention network are constructed for capturing multiple local features on the shortest dependency path between entities and overall features in text sequences,respectively.Finally,multiple local features and overall features are fused for relation extraction.Experimental comparisons with the mainstream relational extraction model on the Semeval-2010 Task 8 dataset are conducted,and the experimental results show that the proposed model can effectively improve the performance of relation extraction.(2)A BERT gated multi-window attention model for entity relation extraction is proposed to address the problem that the relation extraction model cannot mine the key phrase information of entity context.The model first guides the original sentence through the constructed constraint information.Secondly,the key phrase extraction network is used to capture the entity context key phrase features and enhance the semantic information representation of the phrase itself by building a global gating mechanism.Thirdly,the key phrase features are filtered and globally sensed using a classification feature perception network.Finally,the relation extraction is combined with a predictive output network.The experimental results on the Semeval-2010 Task 8 dataset show that the method has further improved over the existing optimal method,with an F1-score of 90.25%.(3)A relation extraction model based on cross-attention and multi-feature perception is proposed to address the problem that the Chinese relation extraction model cannot accurately characterize the semantic information of sentences by using only character or word vectors.The model is based on character vectors and word vectors of sentences,and the character and word cross-attention mechanism are designed for fusing important word features in sentences.A graph convolutional network and a deep separable convolutional network are used to capture syntactic information and local semantic information of word features for relation extraction.The experimental results on the Chinese San Wen dataset show that the proposed model outperforms the mainstream neural network relation extraction methods with an F1-score of 71.57%.In summary,the three relation extraction mappings constructed in this study effectively solve the problems of existing relation extraction models,and provide technical support for the subsequent construction of relation extraction systems in the electric power field.