Research On Answer Selection Algorithm Of Question Answering System

With the rapid development of the Internet,large amounts of free text continue to accumulate on the network.The question answering system has become a very important research direction in the field of natural language processing.It could analyze and understand the probelm and quickly return accurate answers according to the user's query input which is in the form of natural language.Due to the development of deep learning and the continuous release of more practical large-scaled datasets,the challenge of answer selection task in the question answering system are rising.As a key supporting technology,answer selection task has also become a subject of great theoretical research value and application prospect.The definition of answer selection is to select the most suitable answer from the query's candidate answer set,which is essentially a typical text matching and ranking task.The basis of ranking is to compute the semantic matching score between question and answer.The attention mechanism could provide an effective way of text interaction,focusing on important parts of the sentence,thus becoming an indispensable key module in the answer selection algorithm.This article mainly studies answer selection technology based on the attention mechanism.The main research contributions of this paper are as follows:Firstly,the related technology of question answering system is investigated according to project requirements,and several key algorithms are compared theoretically and experimentally.With the analysis of task definition and several specific task types of question answering system,the main research content of this paper is determined;then several related technologies are investigated in accordance with the process of question answering system;finally,three basic deep neural frameworks(Siamese architecture,Attentive architecture and Compare-Aggregate architecture)for answer selection algorithm are analyzed in detail,and comparative experimental analysis is performed.In addition,baseline systems are also implemented to lay the foundation for the analysis and improvement of QA tasks.Secondly,an algorithm of answer selection based on multi-view attention mechanismis is proposed.This paper considers that the core of the question-answer pair matching algorithm is the accurate encoding of text semantic and multiple attention mechanisms is a way to enhance the representation of semantic features.With the use of multiple attention types(co-attention,self-attention)and multiple attention variants(maximum pooling,average pooling,soft alignment)to model multi-perpective semantic views,the algorithm improves the completeness and accuracy of semantic encoding.In addition,in order to improve the computational efficiency of the algorithm,this paper re-imagines attention as a form of feature augmentation method,achieving multiple attention casts.The model returns scalar feature using compressed fuction after soft attention operations,and re-attaches it to the original word representation,providing hints with global knowledge and cross-sentence knowledge for subsequent encoding layers,which could improve representation learning.Experiments and ablation studies on the factual based question answering dataset(TrecQA),open-domain dataset(WikiQA),and community question answering dataset(SemEval-2016 CQA and YahooCQA)prove the effectiveness of the multi-view attention mechanism.Thirdly,a co-stack residual matching network(CSRMN)based on multi-layered attention refinement is further proposed,thus achieving deeper and finer-grained semantic relevance matching between QA pairs.On the one hand,the model introduces a new co-stacking bidirectional alignment mechanism which integrates shortcut connections into neural models for sequence pair matching,and calculates bidirectional matching score by leveraging all feature hierarchies between text sequence pairs.On the the other hand,the model integrates the multi-view attention mechanism proposed before,successfully expanding it in a multi-layered fashion and repeatedly refining representation at each level of the stacked recurrent encoder,which could fully leverage stacked recurrent architecture.In order to prove the versatility of this.multi-layerd stacked recurrent architecture,extensive experiments on four commonly-used question answering datasets are performed,proving that the model not only performs well on short-text question answering datasets,but also beats other models on long-text community question answering datasets,achieving state-of-the-art performance.Moreover,ablation experiments are studied,demonstrating the effectiveness of the co-stack bidirectional alignment mechanism and the multi-level attention refinement module on the stacked structure.Finally,a community question answering(CQA)system based on CSRMN model is built up for a practical landing project on Japanese tourism.The system follows a pipeline structure form:firstly,the Lucene search engine is used to construct inverted index and initially retrieve several similar questions and corresponding answers for input query,thus improving the efficiency and the response speed of the system;secondly,the input query and similar questions are scored using question similarity matching algorithm;then the answer selection algorithm implemented by CSRMN model performs scoring between the input query and the corresponding answers of similar questions;finally two scores are combined,and the highest-ranked answer will be returned to the user.Experimental analysis shows that this design could not only significantly improve the experimental accuracy,but also greatly shorten the system response time,thus proving the great performance of the CSRMN model in practical projects.Moreover,it also proves that the designed pipeline system structure is very effective for the practical landing project of the community question answering system.