Research On Answer Optimization Methods For Question Answering System

With the arrival of the big data era and the development of computer hardware,artificial intelligence has gained rapid progress in both research and industry fields,e.g.game playing and automation.Automatic question answering(QA)is an important branch of artificial intelligence.It is based on natural language processing.It has the ability of providing short and precise feedbacks towards queries raised by real-world users so as to satisfy their information needs.The performance of a QA system relies on high quality knowledge bases(KBs)that of accurate contents and proper expressions,the main source of which is the large-scale heterogenous information around the Internet.However,due to the complexity in user group and the incompleteness on knowledge extraction method,the candidate answers that stored in the KB of a QA system can often be expressed improperly.For example,the answer has grammatical errors,conveys incorrect semantic relations or considers insufficiently on contextual information.In addition,the complex environment that a QA system runs in makes the existence of problematic answers inevitable.Based on the above,this paper proposes answer optimization methods,aims at checking the potentially improper expressions in answers,so that the quality of answers and the answering ability of the system can be improved.The main research contents include:1.Grammatical error correction for answers based on deep convolutional neural networks.Grammatical errors are inevitable due to the complexity on user group on the Internet.And as a primary source of the KBs of QA systems,the improper language expressions from the Internet will directly affect the quality of the answers and then the experience of users.This section is from the perspective of grammar which aims at automatically detecting and correcting the improper expressions of texts in answers,so that to guarantee the quality of answers on grammar.To achieve this,we introduce a deep convolutional neural network-based model to automatically correct the grammatical errors in English sentences.The model takes word embedding and POS embedding as the distributive input,and learns the context information for a candidate sample through a convolutional neural network.The target function is to maximize the difference between the prediction and the original word.Experiments show that,compared with traditional statistical methods,the deep models have superior ability for contextual feature learning and improves the performance of the baselines.2.Weakly supervised relation extraction for answers.Large-scale KBs can support the QA system to answer a series of factual questions.However,due to the insufficiency of knowledge extraction methods,there may exist incorrect,improper or out-of-date semantic information,the result of which are the incorrect semantic expressions in answers.This section is from the perspective of semantic which aims at learning novel semantic relations based on the existed KBs and new free texts,so that the semantic KB of the system can be renewed and complemented.Weakly supervised relation extraction is one of the private routines for relation extraction and is an important way of renewing and complementing KBs.We propose a semantic relation extraction model based on high-quality samples.During the process of parameter learning,the model applies multiple evaluation strategies to select high-quality bags so as to lower down the negative effect from unreliable training samples and optimize relation extractors.Experiments show that the training noise obviously decreases by applying our model.Compared with baselines,this model can achieve better F1 and P-R curves.3.Context-dependent answer selection in community question answering(cQA).CQA websites offer abundant resources for KB complete for QA systems,and how to efficiently make use of the context information to do automatic answer selection is a challenging task.This section is from the perspective of context which aims at efficiently learning and utilizing the context information to do answer selection(namely questionanswer pair extraction).Through analysis,we discover that the quality of answers is closely related to content correlation and label dependency.Based on this,we propose a deep context-dependent answer selection model.The model integrates deep convolutional neural networks for question/answer encoding,attention mechanism and long short term memory for answer correlation learning and conditional random fields for label dependency learning.Experiments show that both the two context-dependent factors are indispensable and the model enhances the baselines on F1.4.Problematic answer detection based on user feedbacks.User feedbacks are remedies for the growth and evolution of a QA system.Besides explicit evaluation scores from users,in the human-computer interaction process,the action of a human user also reflect the answer quality to some extent.This section is from the perspective of context which aims at efficiently recognizing the reliability of answers from user feedbacks during the online running process and providing evidence for the answer optimization.Through analysis,we find that the feedbacks include user intent and sentiment which can implicitly show the problematic situation in dialogues.In order to study the relationship between problematic situation and user feedback,we firstly construct a dialogue corpus which comes from a real-time dialogue system,and then analyze the problematic situations.Secondly,through defining syntactic and semantic features,a supervised learning scheme is proposed.Experiments validate that problematic answer detection can be benefited from considering both user intent and sentiment,and it can provide more valuable clues for system optimization.