| Named Entity Recognition(NER),as an important subtask of information extraction,has the task goal of identifying predefined categories of entities in structured or unstructured text,which can be specifically divided into two subtasks of determining entity boundaries and entity types,and accurate word separation results are the key to determining entity boundaries.There are no boundary markers in Chinese text that explicitly mark words like spaces in English text,so the early Chinese Named Entity Recognition(CNER)task needs to first split the target text,which is the main reason why Chinese NER is more challenging compared to English.Since wrong word splitting results bring error propagation,most traditional Chinese NER models use methods based on splitting text by characters.In this paper,we proceed on the basis of the word-based segmentation method and improve the information and coding and decoding methods for the characteristics of Chinese context,and the main work is as follows.(1)A model text representation layer incorporating lexical information is designed to solve the problem that the traditional CNER methods based on word separation cannot utilize the boundary and semantic information of words in the text.In this paper,the lexical information is added to each relevant character by referring to the FLAT model approach,and a self-dependent gating unit is added to the Transformer structure to enhance the character representation itself.In this way,it is realized to utilize lexical boundary and semantic information while retaining the own features of characters that are important for entity recognition to enhance the entity recognition effect.And considering that Chinese as an ideographic script also contains a certain degree of semantic information in its glyphs and pinyin,this paper proposes a CNER method based on multi-feature fusion by further incorporating Chinese glyph and pinyin features based on the combination of lexical information.Firstly,each Chinese character in the training data is converted into a Wubi code based on the principle of Wubi character splitting to obtain a sequence of Wubi codes corresponding to the input.The input is converted into a pinyin sequence in sentence units by the online pinyin conversion tool,and the pinyin of each Chinese character is split according to three parts: vowel,rhyme and tone.The Wubi coding sequence and the pinyin sequence are input into the text convolution layer for convolution on a character-by-character basis to obtain the glyph and pinyin features of each Chinese character.The meaningless glyph and pinyin features are filtered out by linear attention,and finally stitched with the combined lexical information character features as the input to the decoder.In this paper,the Chinese NER model achieves F1 scores of 72.44% and 96.16% on two publicly available Chinese flat entity datasets,Weibo and MSRA,respectively.(2)Considering that the traditional CNER methods are based on the sequence annotation task framework,which only focus on planar entities and ignore the possible nesting between entities.In this paper,we propose a Chinese nested NER model based on span encoding,which further encodes character vectors into span representation based on the CNER model with multi-feature fusion,and achieves the recognition of Chinese nested entities by converting the sequence annotation task into a span classification task.Firstly,all possible two-by-two combinations of each character in the input are enumerated as spans,and the multi-feature fusion vectors of the span constituent characters are computed to model the dependency between the head and tail of the span,and on top of that,the span internal information is added to further enhance the semantic representation of the span.The span representation is fed into the classifier to obtain the probability distribution of the span corresponding entity classes to achieve the recognition of the model for nested entities,and the experimental results on the Chinese People’s Daily and CMe EE datasets containing nested entities reach 94.03% and 67.20%,respectively,verifying the effectiveness of the span representation enhancement method by fusing multiple features and intra-span information. |
