Research On Answer Quality Evaluation Method Of Online Inquiry Service

With the aging of the population,the frequent occurrence of chronic diseases and the improvement of people's health awareness,the demand for high-quality medical services has grown rapidly.However,due to the shortage and unbalanced distribution of medical resources,the demand for medical services by healthcare consumers cannot be effectively met.Fortunately,with the rapid popularization of the mobile internet and smart terminals,an online inquiry service is being rapidly developed.Healthcare consumers can remotely access the online disease and health consulting and guidance that are provided by medical professionals from famous hospitals without leaving their homes.This inexpensive and efficient new online medical approach effectively integrates offline medical resources,thereby not only meeting the general needs of healthcare consumers but also relieving hospital operating pressures to a certain extent.Compared with traditional community question-answering(CQA)systems and search engines,each answer in an online inquiry service is provided by a certified professional medical practitioner,ensuring that most answers are trustworthy.However,although the online inquiry service has the ability to provide high-quality answers,it does not ensure that each answer is high quality.For example,senior doctors may be busy,and they do not always have sufficient spare time to provide online health consumers with detailed and high-quality answers.Some lower-level doctors may just want to promote themselves and their hospitals by using the online inquiry service platform,so they often provide some irrelevant advertisements.These low-quality answers will cause great problems for the knowledge mining and reuse of the massive amount of health questionanswering information that is accumulated in the online inquiry service.The knowledge base for high-quality health question-answering information is the key to establishing an intelligent online inquiry service,a virtual health assistant,and a health recommendation system.Thus,it is crucial to automatically assess the answer quality in the context of the online inquiry service.This thesis will focus on the key questions regarding the automatic evaluation technology of answer quality in an online inquiry service and will carry out a series of basic and systematic research work on the related content.Specifically,the main research work of this thesis includes the following four aspects.(1)For the problem that the online inquiry service lacks knowledge of the crowd attributes and that the algorithms and features in the related research on the existing community question-answering system answer quality evaluations are no longer fully applicable,through an in-depth analysis of the characteristics of the online inquiry service and its similarities and differences with a community question-answering system,the typical non-textual features(including surface language,social,and temporal features)are proposed and introduced to comprehensively characterize the answer quality of the online inquiry service.By deeply analyzing the data of the online inquiry service,this thesis puts forward a set of clear and definite answer quality evaluation standards for the online inquiry service,and based on these standards,the first medical question-answering dataset for academic research is collected and organized.(2)For the problem that the answers in the online inquiry service are mainly short text with sparse features,the first online inquiry service word embeddings for academic research are pretrained.By combining two kinds of convolutional neural networks to model short text from both global and local perspectives,an extension of the medical question-answering short text semantic space is implemented.Next,the collaborative decision strategy is proposed to generate a more precise semantic representation of medical question-answering short text.Subsequently,multimodal learning and factorization machines are introduced on the basis of the above,and a collaborative decision convolutional neural network framework is proposed.The predictive performance of the online inquiry service answer quality is improved by capturing the highly nonlinear relationships between different modalities and capturing the nonindependent interaction relationships in the same modality.(3)For the problem that the online inquiry service data has a multimodal heterogeneous representation and that there are complex highly nonlinear relationships between different data modalities,by combining Bernoulli and Gaussian restricted Boltzmann machines,an extensible multimodal deep learning framework is built to fuse the semantic knowledge from different data modalities,which improves the performance of the answer quality prediction of an online inquiry service.By combining the Bernoulli restricted Boltzmann machines,a deep belief network for mining high-level hidden semantic representations of short texts is established,which effectively overcomes the severe feature sparseness issue that is faced by medical question-answering short text.Through a large number of experiments,the important influence of the imbalanced data problem on the online inquiry service answer quality evaluation research is explored.(4)For the problem that the online inquiry service has many inexpensive unlabeled short text data and expert annotations are too expensive,we propose two kinds of conditionally independent and sufficient deep textual views based on domain-specific word embeddings and introduce factorization machines as the base-level classifiers.A deep co-training framework for mining the highly nonlinear semantic knowledge that is embedded in a great quantity of unlabeled data,capturing the nonindependent interaction relationships among diverse features within the single deep view and capturing the highly nonlinear relationships between different views is built.As a result,the automatic annotation of a large number of unlabeled feature sparse short text answers and improved answer quality prediction performance are realized.