Research On Multitask Learning For Named Entity Recognition And Semantic Role Labeling

As the information technology rapidly develops and popularizes,the Internet is suffused with vast amounts of information.Information explosion is a double-edged sword for everyone.Owing to this,on the one hand,people can acquire increasingly abundant information.On the other hand,it increases the difficulty to obtain high-quality information.Faced with a large amount of text information,how to accurately and efficiently identify information with semantics has drawn great concern in the field of natural language processing.Academic researchers constructed various kinds of tasks for diverse application scenarios.The traditional research idea is to build a corresponding model for a specific task,and isolates the named entity recognition and semantic role labeling two tasks.This not only leads to a user-oriented system needs to maintain multiple sub-models,which causes system redundancy and difficult maintenance,but also neglects the correlation between tasks.In addition,many existing approaches are designed for English language scenario,resulting in the abnormal phenomenon that the performances on Chinese language are inferior to the English's.Starting from these motivations,this thesis proposes the approach to build a unified neural network for named entity recognition and semantic role labeling two tasks.Besides,we further improve the model's performance by exploring the effective semantic features on Chinese language scenario.The specific contributions are as follows:(1)Design a unified neural network for named entity recognition and semantic role labeling both two tasks.This paper proposes a unified neural network architecture that adopts the Highway-LSTM network to improve traditional stacked networks,which can further mitigate long-distance dependence issues by capturing more contextual semantic features.When applying the conditional random field model to utilize the sentence level semantic features,we predefine the probability matrix with transition constraints,which can prevent an invalid tag sequence.(2)Explore the effective features to improve the performance of neural networks under the Chinese language scenario.Majority of the existing approaches focused on English datasets.The proposed approaches exist the abnormal phenomenon that the performances on Chinese datasets are inferior to that on English datasets.This paper proposes the approach of adopting convolutional neural networks to extract the character-level semantic features.Meanwhile,all human languages obey an inherent set of grammar and part-of-speech tags are the classes of formal equivalents of words in linguistics.Therefore,we adopt the part-of-speech tags as the input semantic feature.(3)Design and implement a unified annotation system for named entity recognition and semantic role labeling two tasks.The best way to demonstrate the research's value is to apply it into an industrial application.Based on this goal,we design and implement a unified information extraction system.