Research And Implementation Of Entity Recognition And Relationship Extraction Method Based On Deep Learning

In recent years,named entity recognition(NER)and entity relationship extraction(RE)have become research hotspots in the field of natural language processing(NLP),and they are also the key tasks of knowledge graph(KG).Named entity recognition is used to identify the corresponding entities from free text.Relationship extraction aims to extract the relationship between two entities from text corpus.With the deepening of the research on Chinese named entity recognition and Chinese entity relationship extraction tasks,there are many problems in these two tasks:(1)The sequence modeling layer network bidirectional LSTM commonly used in NER task has complex structure,it will reduce the speed and make the model lack of parallelism.(2)BERT(Bidirectional Encoder Representation from Transformers)model solves the problem of "one word polysemy" that can not be solved by static word vectors such as Word2 vec,but its static mask mechanism makes the semantic representation generated by pre-training only be character level and lack features of word level.(3)At the character level,NER task fusion of word meaning and other information to rich semantic information is one of the focuses of research in recent years,but how to effectively fuse word vectors and reduce the impact of word segmentation errors while using word information is also a major difficulty.(4)Simple attention mechanism has limited performance improvement on RE tasks.How to make better use of attention mechanism is also a research direction.Therefore,this paper mainly studies Chinese entity recognition and Chinese relationship extraction in general domain,which are used to improve the efficiency of NER task and RE task respectively.The main research work of this paper is as follows:(1)In the Chinese named entity recognition task,an entity recognition model integrating dynamic mask pre-training and dilated convolution network is proposed.Firstly,in order to enrich the syntactic and semantic information,Ro BERTa model based on dynamic mask is introduced in the pre-training stage.The training mechanism of dynamic mask can better represent the semantic information of word meaning in the text.In addition,the character vector is enhanced by the word vector that can be formed by the word in the context,and the word segmentation information is also used.Secondly,in order to improve the efficiency of feature extraction,the dilated convolution network is introduced in the sequence modeling layer to improve the complex structure of the traditional LSTM network,which affects the speed,and follow the design idea of hybird dilated convolution to avoid the generation of grid effect.Experiments show that the designed model improves the recognition performance of MSRA and people’s daily data sets compared with the baseline model.(2)In the task of Chinese relation extraction,a Chinese relation extraction model combined with word level and sentence level double-layer attention mechanism is proposed.Firstly,considering that the entity words carry relevant information in the relationship extraction task,in order to make the word embedding layer contain entity information,the model introduces the location information of the two entities in this layer to enrich the semantic representation.Secondly,in order to obtain the more important sentence features in the text and the more important word features in the sentence,the model introduces the word level and sentence level double-layer attention mechanism for feature extraction,so that the final generated vector can better represent the text information.Experiments show that the designed model is better than the baseline model on the specified data set.