Research On Chinese Named Entity Recognition Technology Based On Neural Networks

Nowadays,data grow rapidly on Internet,people creates large amount of texts everyday in social media like Twitter,Facebook,Weibo.These texts have greate potential value to form knowledge network,to build a big knowledge base.Recognizing entity names is an import step in these applications.Named Entity Recognition(NER)is an essential sub-task of IE.In a document,entity names are units that tells important information.NER task aims to find and classify these names,and results of a NER system have greate impact on the next tasks like entity linking(EL),entity ralationship extraction.NER is the very first step of these works;And Automatic Abstraion task,in a way,is about to answer the blank question like ”who”,”doing what”,”when”,”where”,etc,which NER aims to find;In bilingual translation task,entity names are important parts that requires special skills to deal with.The idea of a NER system is similar for different languages,which requires a welldesigned model to modeling context and attributes.But dealing with Chinese is harder than English since the vocabulary in Chinese is much bigger than English.In this thesis,We researched and sumarized the key tech problems,and discussed details of implementation of NER system for Chinese texts:(1)we compared several popular neraul network structures for NER systems,concluded the appropriate parameters of the networks;(2)we analyzed the characteristics of Chinese texts,and adjusted the word representation model that can maximize the utilization of Chinese characters' semantic features;(3)we compared different decoding algorithms,and maked a little adjustmenet on the loss funtions,which significantly improved the speed of training process.Overall speaking,We implemented a NER system for Chiense texts,that aggregating lastest neural networks and character-word joint word representation model.and comparing to other similar NER structure,we can train a model in a much faster speed.