Research On Fine-grained Entity Typing Method Based On Neural Networks

In the existing large-scale knowledge base,the lack of entity type information is particularly serious.However,entity type is the most basic and essential attribute of the entity.It is crucial for many natural language processing tasks.Therefore,the entity classification problem is very important and urgent to be solved.The entity classification problem aims to identify the semantic type of an entity in a particular piece of text.In recent years,the application of neural networks to entity classification tasks is a trend.However,many neural network-based classification methods usually only extract features separately from the entity and its text,ignoring their intrinsic links and rich background information of the entity.In order to utilize the rich semantic information of the knowledge base text as much as possible,we introduce the contextual entities and propose a fine-grained entity classification model based on neural network structure to classify the target entity described by a piece of text and obtain the fine-grained types of the target entity.Our model is an encoder-decoder structure.The encoder extracts the feature of input text from the four aspects,they are the target entity itself,the whole text,the contextual entities and the correlation between the target entity and contextual entities.The four feature vectors joint the coding of the target entity.The decoder transforms the coding of target entity into a type distribution vector through the fully connected layer,then obtaining the predicted entity type.The Bi-LSTM network with Attention is used to extract the whole text feature,and the CNN network with Attention is used to extract the features between the entities.The characteristic of this paper is that the feature extraction of contextual entities and correlation between entities.The entities except target entity in the text are considered separately,and the correlation between the target entity and these entities is modeled.Making full use of rich semantic information of the text to get more accurate entity coding.The validity of the model is verified by experiments on FIGER and OwnWiki the two datasets.