Research On Multi-granular And Fragmented Knowledge Acquisition From Human-computer Interaction Data

The development of artificial intelligence enables computers to understand the natural inputs of users,such as action,language.Thus,human-computer interaction becomes more and more close to the natural interaction between human.However,it requires systems understand user generated contents which are heterogeneous and multi-modal to find the real intention of user.This paper stands on the theory of granular computing.Two streaming feature selection methods are proposed,named CIE-OSFS and ML-OSMI.CIE-OSFS uses conditional information entropy as uncertainty measure.ML-OSMI adopts label granulating to solve multi-label streaming feature selection problem.These two methods could cope with data generated in real time,such as operational data in human-computer interaction.Besides,this paper proposes a joint extraction method named JMC which learns multi-granular features automatically based on neural network structure.JMC can extract entities and relations between entities to find the intetion of users.These methods could help systems to understand users.Main contributions of this thesis are follows:1.Firstly,a streaming feature selection framework based on uncertainty measures is proposed.The framework uses uncertainty measures in granular computing field to determine the importance of features.Then,based on the framework,a streaming feature selection method named CIE-OSFS based on conditional information entropy is implemented.To verify the effectiveness of the propsed method,comparsions to state-of-the-art streaming feature selection methods such as fast-OSFS,Alpha-investing and Grafting are conducted.Results show that CIE-OSFS can select fewer features while ensuring accuracy.2.From the perspective of granular computing,a multi-label streaming feature selection method based on label granulation is proposed,namely ML-OSMI.It first granulates labels by clustering and transforms them into a lower-dimensional space.Then,the correlation and redundancy of features are redefined under the multi-label scenrio based on mutual information to guide the feature selection procedure.Experimental results show that ML-OSMI is effective in traditional multi-label feature selection scenarios and streaming feature selection scenarios.3.Introducing the prior knowledge of multi-granular features to design the neural network structure,this paper proposes a neural-based joint extraction method named JMC.It learns multi-granular features automatically.To verify the effectiveness of the JMC,experiments on NYT,a distant supervision dataset,are conducted.Copmparsions to state-of-the-art methods on the entity extraction task,the relationship classification task and the joint extraction task show that JMC produced better results.