Research On Multi-view Feature Fusion For Answer Selection Based On Knowledge Reasoning And Text Understanding

The information explosion caused by the rapid development of computer science and internet bring new challenge to traditional information retrieval technology.It also brings new development opportunities for intelligent question-answering(QA),which understands users' natural language questions and provides accurate natural language answers.So,QA systems become an inevitable development trend of search engines,also become the research focus of natural language processing and artificial intelligence,and it is of great significance for both research and application.Answer selection module in a QA system is to directly return answers to askers,so its performance directly impacts the answer result.To better fuse structural knowledge and un-structural text into answer selection process,in this paper,we deeply study answer selection methods.The main work and novelty are listed as follows:(1)Representation Learning-based Hidden Relation InferenceTraditional knowledge reasoning methods are difficult to construct corresponding knowledge base queries and may also have cascading failure problem,when dealing with complex questions.To solve the problem,we propose a sentence and knowledge joint representation method.It represents Subject-Verb-Object sentences into a homogenous structure and learns a mapping function to represent the sentence in the knowledge space.Then,according to the main idea of Trans E,a translation based text-knowledge joint representation is obtained.In the reasoning stage,the knowledge reasoning process is simplified to the translation operation by leveraging the topic words and the question representation in the knowledge space.In this way,complex questions can be effectively resolved.Experiments on Wikipedia and Freebase dataset have demonstrated that,in the task of complex question reasoning,compared with N-gram based method,the improvement of our method is over 30% in H@10;compared with traditional knowledge based QA methods,the improvement is over 100%;and compared with existing representation learning methods,the improvement is over 40%.(2)Feature Attention Mechanism-based Text UnderstandingTraditional text understanding models can hardly distinguish text information importance according to given questions,so this information cannot be effectively utilized.To solve this problem,we propose the question-oriented feature attention mechanism.In this method,the question representation is first obtained by RNN according to the question content.Then the representation is mapped to a feature weight distribution.At last,features are weighted by the weight distribution and fed into the answer selection model.Since the weight distribution contains rich information of the question,the features importance can be distinguished according to questions.In our experiments,we applied the feature attention mechanism into the reading comprehension QA,and verified the model on MCTest dataset.Experimental results demonstrate that,compared with the model without feature attention,our method enhanced more than 5% accuracy;compared with traditional feature selection methods without considering question content(e.g.,PCA and L1 Norm),the accuracy enhancement is more than 3%.(3)Multi-view-Oriented Feature Fusion and Answer SelectionTraditional answer selection method lacks of the feature fusion ability and faces data deficiency problem during the model learning process.In the light of this,we proposed a multi-view feature fusion method.It separates features into different views according to their sources,and describe each data point with their correlations according to the spectral graph theory.The model learns a common space with both labeled and unlabeled data to fuse these multi-view features,and learns data representation in the common space.Finally,the data representation in the common space is used for answer selection.Compared with traditional answer selection model,the proposed method can leverage agreement information among different views and the unlabeled data to strengthen the regression performance.Experiments on c QA dataset demonstrate that,compared with traditional regression model,the proposed method decrease more than 2% mean error on auto evaluation,and increased more than 5%correlation on manual evaluation.In this paper,we deeply studied the answer selection methods in question answering systems.The text understanding and knowledge reasoning methods were utilized to model external unstructured and structured knowledge,respectively.And the multi-view learning method is applied to incorporate these information into answer selection models,so that the performance of answer selection and question answering is significantly improved.