| In current days,the Internet has developed rapidly,it creates lots of unstructured content.Recently,how to transform unstructured content into structured content is one of the most important research in nature language tasks.Supervised Method and Semi-Supervised Method are two mainstream methods used in Named Entity Recognition and Relationship Extraction model training.The models trained by using Supervised Method achieve excellent performance,but using Supervised Method training model depends on large scale manually labeled high quality data set;Semi-Supervised Method enhances models portability and it doesn’t depend on large scale manually labeled high quality data set but models still achieve great performance.Recently most mainstream Chinese Named Entity Recognition tasks are using deep learning methods.Deep learning methods could catch features of Chinese characters and Chinese entities by using neural network models,and these feature information plays an important role in NER tasks.While NER task in Chinese contents all ignore the feature of Chinese radical recent days.So Chinese radical feature is added into NER method to enhance Chinese semantic information caught by encoder in this thesis.Meanwhile in this thesis the Weakly Supervised method used to train the NER model,and the results show that comparing with existing open source model in text data sets,the model proposed in this thesis shows best performance.Most RE methods focus on the coded representations of characters,the positions of characters in content and target entity pairs.Based on the polysemy of Chinese entities,this thesis mixed Chinese semantic information with Chinese content features.And using MIL method training RE model to reduce the impact of noise data generated by remote supervised learning method on the performance of relation extraction model.By doing comparative experience on existing models and the model proposed in this thesis on the public test set,the result shows that the model proposed in this thesis could forecast relationship type more accurate.The optimization scheme proposed on NER and RE task in this thesis,it focus on let encoder catches more diverse Chinese content information and rich coded feature representation.And using Weakly Supervised Learning methods to reduce the dependence of manually labeled dataset and reduce the impact of noise data on the performance of model.It also improves learning ability of model,enhance its portability,so it could have great performance in different data sets.To sum up,the NER method based on Chinese radical feature and the RE method based on Chinese semantic information could explore NER and RE tasks in finer grained features and diversity perspectives.They improve deficiencies in existing methods to make full use of content feature information and semantic information,thus these models got great performance in NER and RE tasks.It provides new enlightenment for later researchers in this field. |
