Research On Deep Learning Based Interactive Question Answering Techniques

With the increasing popularity of mobile Internet and the rapid development of intelligent mobile devices,interactive question answering(i QA)systems had been widely used in many fields of daily life,such as service online,financial consultation and so on.As an important application,the target of an i QA system is to communicate with people naturally and continuously.For the i QA,one critical issue is to learn and represent the natural sentence associated with context,another one is how to parse the complicate sentence relations among the interactive scenario automatically.Recently,deep neural networks have been shown the strong capability of learning and representing data on speech recognition,image classification,and machine translation.Deep learning has become the mainstream method for Natural Language Processing(NLP).This dissertation aims to study the deep-learning-based technologies for i QA,and explores using deep neural networks to learn the representation of the context and model the semantic relationship existing in i QA.We mainly study the key technical problems involved in the framework of the i QA system in depth.Firstly,we construct one real world oriented knowledge base for the contextual association phenomenon of i QA,it provides a reliable resource for the research of i QA technologies.The contextual association phenomenon is studied at the context,the pragmatics and the semantic respectively.To construct the i QA knowledge base,we annotate the topic and dialogue act for each sentence,and tag the semantic link between sentences.We establish the knowledge framework based on topic category and act category,and obtain the knowledge item from the interactive scene.Each item is one sequence of question and answer sentences,which embedded with the annotated semantic links.The i QA corpus and the knowledge base constructed by us can be used to evaluate the i QA tasks,including the question parsing,the question retrieval,the answer selection,the contextual relation structure parsing and so on.To verify the reliability of the constructed knowledge base,we conduct the experiment of user intent detection on the data of knowledge base.Using the dialogue model based on the Hidden Markov,we present the topic relevance structure and the dialogue act structure in i QA.Secondly,we study to parse the question of i QA,and proposes a novel model based on recurrent neural network to learn the representation of the interactive context.Considering the universal missing phenomenon of semantic information in i QA,our model achieves the end-to-end contextual learning via the encoder-decoder architecture which is based on recurrent neural network.Our model is able to encode and decode the interactive scenario automatically,and parse the question associated with the classifier.For encoding the context,the model implies the interactive encoding mechanism to learn the matching patterns embodied in interactive scenario.Unlike the traditional method,our approach not only removes the step-by-step processing such as the integrity detection of question,context extraction,and question completion,but also reduces the dependencies on manual prior knowledge during the supervised learning process.The experimental results show that the contextual learning model proposed in this thesis has more adaptability in multiple tasks,and performs higher than the state-of-the-art models in parsing the question.Thirdly,we study to address the QA semantic matching of i QA,and propose a recurrent convolution neural network(RCNN)to achieve the context-based answer selection.In the i QA scenario,there is not only the semantic matching relationship between the question and the candidate answer,but also the semantic similarity or relevance among the context-based answers.In view of this,a novel RCNN architecture is proposed to model the Q&A semantic matching and the semantic association among answers jointly by integrating convolution neural network with recurrent neural network.The two-phrases learning algorithm is proposed to train the RCNN,which is able to improve the ability of RCNN in learning the answer correlation.The sufficient experiments are conducted on the datasets of customer service QA and c QA respectively.The experimental results show that the performance of RCNN is higher than the state-of-the-art models significantly,especially effective to distinguish the semantic similar or relevance answers.It also demonstrates that modeling the semantic association among answers is very helpful to improve the performance on the answer selection.Finally,we study to parse the relational structure of i QA,and propose a novel model to tag contextual relationships among i QA sentences via attention mechanism.For the i QA system,one important way of knowledge self-learning is to extract the QA knowledge from the interactive scenario automatically,the key of it is how to correctly parse the complicate sentence relations among the context.To model the sentence relation,this thesis presents a novel attention mechanism based on matching pattern to achieve the feature mapping between relational pattern and sentence semantics,so as to enhance the ability in learning the representation of the complicate sentence relation.In view of the context dependencies,we use recurrent neural network to model the context among the sequence of sentence relations,and achieve to learn and tag the sentence relations which are associated with context.Experiments show that the proposed model is higher than competitor models significantly.It demonstrates that our model has the strong ability in parsing the complicate sentence relations effectively.It also proves that our model can improve the level of the knowledge self-learning,which makes the i QA system to extract the knowledge item embodied more comprehensive semantic links.