Reaserch On Answer Fusion Method Based On Deep Learning

Question answering system is an important task in the field of Natural Language Processing.The corpus based on question answering pairs is the main source of answers to the question answering system,and the question answering pairs in the corpus are generally extracted from the question and answer community such as Baidu knowing and so on.However,ask a question in the community usually have multiple answers,from different angles reply to questions,question answering community answer only to select one of the answers as question reply,this leads to the corpus of the answer is not comprehensive.Therefore,this paper studies the answer fusion method to fuse multiple candidate answers to solve the problem of that the answer is incomplete and redundant in question answering system.This paper uses deep learning method and attention mechanism to solve the problem of answer fusionThe fusion method are extracted from multiple candidate document answers,therefore the accuracy of answer extraction,determines the fusion method of accuracy and completeness of the results.At the same time,the answer is that the solution is extracted from multiple documents,and there is a problem of incoherent and unreadable semantics.Therefore,this paper attempts to improve the fusion effect of the answer from the two aspects: automatic extraction and semantic coherence.The answer is to automatically extract answers from multiple candidate documents,making the answer leaner and more comprehensive.Semantic coherence is usually characterized by the sequence of sentences within paragraphs,so use the sentences in order to solve semantic consistency problem,the semantic consistency between the candidate answer,make the answer the fusion results more readable,more coherent semantic.This paper focuses on the automatic extraction of the answers and the sequence of sentences,which are divided into four aspects:1.Automatic extraction model based on word overlap.The mechanism of attention within the sentence to question and answer sentences are extracted,at the same time for the introduction of corpus,word overlap features,document reciprocal feature,word similarity feature,and using random sampling method for corpus processing the data imbalance.Comparing the baseline method,the automatic extraction model based on word overlap can improve the accuracy of the extracted answers.2.Sentence ordering method based on sentence matching.This paper will introduce the deep learning method into the sentence sorting,using the deep learing method to solve the learning sentence ordering problem,the sentence ordering method based on deep learning and sentence matching compared to the baseline method improves the sentence ranking method.3.Sentence ordering method based on attention mechanism.According to the sentence ordering tasks,in order to enhance the ability of the model to capture the semantic logical relation,the attention mechanism introduced into the sentence ordering task,introduce sentence ordering method based on static attention mechanism,sentence ordering method based on word-by-word attention mechanism and sentence ordering methed based on inner-attention mechanism.The sentence ordering method based on attention mechanism can capture the semantic logical relation between sentences effectively and improve the sentence sorting effect.4,answer fusion system design and implementation.The answer automatic extraction module and sentence sorting module are integrated to implement the answer fusion system to solve the problem of semantic incomplete and lengthy in Corpus Construction...