Research And Application Of Key Technologies Of Cross-language Automatic Question Answering System For Vertical Field

The question and answer system in the vertical field,as a way of information query,plays an increasingly important role in each field.However,because most of the vertical question answering systems are in a single language,the corresponding question and answer system needs to be constructed separately for people of different languages.If the same question and answer system is used to serve people in different languages,the workload of building the system will be reduced to a great extent.Therefore,this thesis analyzes the vertical domain text in a specific vertical field,aiming to construct a cross-language question answering system in the vertical domain.The answer extraction of text and the translation model between multiple languages are the premise of constructing cross-language question answering system.This thesis focuses on the research of answer extraction,machine translation and the integration of the two systems.The specific research contents are as follows:1)A method for superimposing attention model of convolutional neural networks is presented.In the process of constructing the model of answer extraction,a two-layer attention mechanism is adopted to extract the features of the text and the problem respectively,and then merge the features.Through the secondary reinforcement learning representation of the text-question,the problem-text in the vertical field,the contextual context is enhanced,the experimental effect has been significantly improved.2)A method for vertical domain machine translation model is presented.This thesis constructs a Chinese-English,English-Chinese translation model for the text in a specific field,and proposes a method of combining the neural network model with the rule-based translation.Firstly,the domain-aligned corpus and user dictionary are constructed.Based on phrases and words in vector layer,construct neural network model by Long Short-Terms Memory(Bi-LSTM)network and attention mechanism.Furthermore,syntactic analysis is used to analyze and classify problems,and to build a basic translation rule base.Complete the machine translation work by merging the two.3)A method based on knowledge representation of knowledge map in vertical domain is presented.The entities,attributes and their relationships in this domain are extracted,and each entity and attribute is connected through the relationship,and a domain knowledge representation platform is constructed by using triplet.Through the platform,the answer extraction and machine translation models are merged,and the merged system will be the final model of the question and answer system in the field.The experimental results prove that the method proposed in this paper is effective.The EM and F1 values of the answer extraction based on the superposition attention model can reach 71.2 and 80.2,respectively.The translation model based on neural network and traditional rules can effectively translate the problem.The question answering system after fusion of the answer extraction and translation model can accurately answer questions in specific fields using different languages.The question answering system after the fusion of the answer extraction and translation models can accurately answer questions in a specific field in different languages.Finally,this thesis describes the existing problems and the further research plans.