| Named Entity Recognition(NER)aims to extract entities such as person,place,institution,or other proper nouns from unstructured text.As one of the basic tasks in the field of natural language processing,it plays an important role in various natural language processing tasks such as dialogue question answering,machine translation,information retrieval,and knowledge graph,and plays a key role in many landing applications.In recent years,the technology of named entity recognition is becoming mature,but it still faces many challenges,especially in the direction of Chinese named entity recognition,the problems of blurred word boundaries and loss of word semantics have not been well solved.At the moment,the mainstream named entity recognition method is based on the hybrid model of characters and words,and the method of lexical enhancement is used based on the character-level model to improve the performance of named entity recognition.Compared with character-level models,such methods can increase the recognition of word boundaries in text and the semantic information of vocabulary.Compared with the word-level model,it can reduce the problem of poor performance caused by word segmentation errors,which can be said to be a combination of the advantages of both.But this method still has a series of problems: a single vocabulary enhancement method may cause bias problems,make characters and words too much attention,and lead to the loss of global semantic information;The introduction of lexical information by dictionary matching is generally limited by the quality of the dictionary,and it is likely to introduce wrong words or invalid words;The word hybrid model will inevitably lead to an increase in model complexity,with a larger amount of parameters or calculation,and performance improvement will also lead to a decrease in efficiency.In response to the above problems,this thesis proposes a multi-metadata embedding method integrating Chinese character structure information,a multi-task learning method introducing word prediction,and a lexical enhancement method based on the inter-attention mechanism,which are used to solve the problem of multi-metadata embedding and multifeature fusion in Chinese named entity recognition.The main research work and innovation points of this thesis include:1.A multi-metadata embedding method integrating the structural information of Chinese characters.The mainstream methods introduce lexical information based on the character-level model to solve the problem of blurred word boundaries and semantic loss in Chinese named entity recognition.However,it may introduce additional noise words,and there is also the problem of strong correlation between characters and words.This thesis proposes a dual-stream neural network model based on Transformer,which uses Chinese character structure information to constrain the high attention problem of characters and words in the cross-attention network,and reasonably applies the Chinese character structure information to the Chinese named entity recognition task.By providing richer structural and semantic information in Chinese characters,the shortcomings of the hybrid model of characters and words are optimized.Outperforms the current state-of-the-art performance on four Chinese NER benchmark datasets,and improves by up to 3% on the Weibo dataset.2.Introduce a multi-task learning method for word prediction.In the current mainstream Chinese named entity recognition tasks,word hybrid models are used to solve the problems of blurred Chinese word boundaries and lack of lexical information.However,this method will bring more wrong information in the case of poor dictionary quality,resulting in errors in entity recognition.In this thesis,a multi-task learning method is adopted.On the baseline model FLAT,an additional vocabulary prediction task is designed by rationally using the masked part of the output.Essentially,a binary classification task is used to predict the effectiveness of these vocabulary words.More prior information can guide the model to learn the correct word information,thereby further improving the effect of named entity recognition.At the same time,a method of quadratic input is proposed to share parameters,which reduces the complexity of the model and shows superior performance on three public datasets.3.A lexical enhancement method based on an inter-attention mechanism.Taking FLAT as an example,when lexical information is introduced,words are spliced into Chinese characters using two position encodings,which makes the input sequence too long and the computational complexity of the model is too high.This thesis designs an inter-attention mechanism based on the TENER model,which is used to construct a non-flat lattice structure that integrates lexical information,which can greatly shorten the input sequence,reduce the amount of calculation,and make the performance reach the level of FLAT.Compared with FLAT,this method reduces unnecessary attention calculations between “words-characters” and"words-words”,reduces memory usage,and can use higher batches to reduce training time.Great results are obtained on four open datasets. |
