Research On Key Technologies Of Question Answering Systems

Artificial intelligence(AI)is an important field and hotspot of current academic circles,where NLP(Natural Language Processing)is a significant component of AI,however,question answering system is a major branch of NLP technology.With the spread of Internet,the requirements of information that use words as input way like question answering system and intelligent client system is increasingly,especially 5G brings about more information demands to the public's mobile life.The key technologies of question answer system have made great progress and improvement in recent years,which owes to the successful application of GPU(Graphics Processing Unit)and data and deep learning.However,the interactive mode of question-and-answer that returns the exactness answer sentence in question answering systems is hard to meet the demands.It mainly shows that there are a few deviation and inaccuracy of question sentence understanding and analysis in semantic information,which leads to incorrect answer sentence returned.To solve the above problems,research on the concerning fundamental technology is how to obtain the semantic information or feature of question text and classify text on this basis,and how to build semantic matching model of QA(Question Answer)pairs.At present,machine learning and deep learning are important method of NLP technology,related methods of deep learning and traditional machine learning are thoroughly studied in this dissertation,and discussing the key technologies of representation and classification of question sentence and answer selection in QA systems based on machine learning and deep learning is proposed.The main contributions and innovations of this dissertation are as follows.(1)In view of the problems of the classifier of question text needs repeated training in small-scale unlabeled corpus,and could not make full use of updating feature space,which would affect the performance and effect of question text classifier,this dissertation proposed a novel incremental algorithm of question text classification based on the double threshold Naive Bayes(DTNB).Firstly,an improved incremental algorithm of updating and mapping into feature space was proposed via the traditional incremental algorithm,this not only updated the corresponding parameters of the classifier,but also optimized the classifier property.Secondly,for the incremental sample selection of the unlabeled corpus,a minimum posterior probability was designed as the double threshold method of sample selection by using the traditional class confidence,it not only solved new text feature that was selected and the insufficiency of the prior probability of various categories that was only modified,but also could get better the incremental sample classification result.Finally,evaluation indexes of the incremental algorithm and the incremental sample selection way and the evaluation method of reasonable quantification were presented.The experimental results show that the proposed DTNB algorithm not only significantly outperforms the incremental learning ability of the traditional incremental algorithms and the single way of new text feature,but also the advantage and effect of this algorithm were promoted,namely,the classification accuracy of the proposed DTNB algorithm greatly increased by at least 3%,which is superior to the compared algorithms.(2)Aiming at solving the problems that the information of question text representation is incomplete and could not sufficiently reflect the relationship between forward and backward word sequence,which could result in the semantic information without context and inaccuracy of semantic information,this dissertation proposed a new method of question text representation and classification algorithm based on BiLSTM(Bidirectional Long Short Term Memory)and double word embedding(DWEL).Firstly,a method of generating the dual index word-embedding was proposed by using the dictionary index and the corresponding part of speech index based on obtaining the word embedding of question text,and then concatenating into real vectors.Secondly,feature vectors of question text was extracted further from these acquired word-embeddings by building the recurrent neural network model of BiLSTM and double word-embedding,this model not only does the feature representation of question text itself,but also captures the feature information with context of the relation between forward and backward sequences of question text.So,text feature vector with high-quality context information was acquired and the train efficiency and performance of the proposed model were further improved.Finally,the sentence vectors were processed by mean-pooling layer and then question text categorization was classified by softmax layer.The experimental results show that not only the validity of the proposed DWEL method was verified and the power and accuracy of semantic feature information with context were improved,but also the classifi cation effect performed better and the classification accuracy increased by at least 5%,2%and 1%,compared with the traditional algorithms based on machine learning,LSTM and LSTM+context window models,BiLSTM models,respectively.(3)In view of the problems of the shortage of feature learning and key information of question text representation,and the effective deviation of feature extraction and abstract,such as key information of word sequence is extracted from polysemy,which could give rise to the optimum validity of question text representation,this dissertation proposed a novel method of term and pooling based on deep BiLSTM and CNN(Convolutional Neural Network)for question text representation and classification(DBCTP).Firstly,a way of term-based combination operation is first proposed in CNN and then made pooling operation.That is to combine the convolution results of several convolution kernels and then make pooling operation by this approach.Secondly,this research describes a new method of horizontal convolution and vertical convolution of pooling the convolution results of CNN,which takes text n-gram feature as the object of study,and on this basis the feature extraction uses different combination ways and comparative analysis of their effects.Once again,three sorts of deep CNN models(We named TB-CNN,MCT-CNN and MMCT-CNN,respectively)and one deep BiLSTM model of combination term and pooling via CNN are constructed for the first time(DBCNN).Finally,the parameters and train performance of DBCNN model are optimized,and the sound effects of three key parameters in three deep CNN models are analyzed.Therefore,this method not only enhanced the ability of feature extraction and learning,but also improved the abstract effect of semantic feature and the accuracy of extracting key information for question text representation.The experimental results show that the proposed DBCTP method not only performs well both the availability of model and the best effective power of text feature learning and extraction,but also the classification accuracy outperforms the traditional CNN models and one obviously increased by 5%,2%and 2%,compared with the traditional algorithm by machine learning,LSTM and BiLSTM,and LSTM or BiLSTM+CNN models,respectively.(4)Aiming at solving the problems that most of text representation of question and answer sentence were limited to part of semantic information,and could not make full use of the relative feature weight between question and answer sentence,that is to neglect the relationship between question and answer sentence,which would lead to the inaccurate or deflection of answer sentence,it is essentially a semantic similarity calculation and efficiency issue of question answer pairs,this dissertation proposed a novel method of answer selection based on BiLSTM and attention mechanism(BiAM).In order to obtain the rich features vector and the information of various features in question and answer sentence.Integrating and utilizing the research results of updating feature space of DTNB,DWEL model and BiLSTM+CNN combination model in DBCTP,respectively,the semantic representation model of question answer was firstly created by BiLSTM and attention mechanism,and a way of contaminating pooling and adding extra text feature,where the rich and rational weight information of related features between sentences was obtained by attention mechanism and pooling operation.For improving the similarity calculation and efficiency between question and answer sentence,secondly,a kind of model based on semantic matching for answer selection is proposed,and computing the semantic similarity between question and answer sentence is exercised by Manhattan distance.Finally,the experimental results show that not only the validity of the proposed model was proved and the enrichment key semantic features were captured,and but also outperform the compared algorithms,where the indicators of MAP and MRR are increased by 1%.As a result,the quality of the answer sentence and the efficiency of semantic similarity computation were improved by the proposed method for the task of answer selection.