融合回答者排序得分的CQA答案摘要方法  

CQA answer summarization method integrating respondents ranking scores

作　　者：丁邱 严馨[1,2] 刘艳超 徐广义 邓忠莹 DING Qiu;YAN Xin;LIU Yanchao;XU Guangyi;DENG Zhongying(Faculty of Information Engineering and Automation,Kunming University of Science and Technology,Kunming 650504,China;Yunnan Key Laboratory of Artificial Intelligence,Kunming University of Science and Technology,Kunming 650504,China;The Information Technology Center,Hubei Engineering University,Xiaogan 432000,China;Yunnan Nantian Electronic Information Industry Co.,Ltd.,Kunming 650040,China)

机构地区：[1]昆明理工大学信息工程与自动化学院,云南昆明650504 [2]昆明理工大学云南省人工智能重点实验室,云南昆明650504 [3]湖北工程学院信息技术中心,湖北孝感432000 [4]云南南天电子信息产业股份有限公司,云南昆明650040

出　　处：《陕西理工大学学报（自然科学版）》2023年第5期38-46,共9页Journal of Shaanxi University of Technology:Natural Science Edition

基　　金：国家自然科学基金项目(61562049,61462055)。

摘　　要：针对现有答案摘要方法对句子建模不够充分,并且忽略了回答者相关信息在摘要过程中的作用,提出了融合回答者排序得分的CQA答案摘要方法。首先,使用RoBERTa-wwm结合平均池化对句子进行编码获取句子深层语义表示;然后,使用DUM专家推荐方法依据回答者相关信息对回答者进行排序,依据排序结果求得回答者排序得分;最后,综合句子相关性、句子新颖性、回答者排序得分计算句子综合得分,根据句子综合得分使用MMR思想迭代选取句子构成答案摘要。实验表明,使用RoBERTa-wwm结合平均池化能更好地获取到答案句的深层语义信息,综合3个评分选取摘要句既能考虑答案与问题、答案与答案间的交互,也能较好地融入回答者的信息,有效地改善了答案摘要的质量。Aiming at the inadequacy of sentence modeling in existing answer summarization methods,and ignoring the role of respondent related information in the summarization process,a CQA answer summarization method integrating respondent ranking scores is proposed.First,we use RoBERTA-wwm combined with average pooling to encode sentences to obtain deep semantic representation of sentences;Then,the DUM expert recommendation method is used to rank the respondents according to the relevant information of the respondents,and the ranking scores of the respondents are obtained according to the ranking results;Finally,the sentence comprehensive score is calculated by integrating sentence relevance,sentence novelty and respondent ranking scores.According to the sentence comprehensive score,the MMR idea is used to iteratively select sentences to form the answer summarization.The experiment shows that using RoBERTA-wwm combined with average pooling can better obtain the deep semantic information of answer sentences.The summarization sentence is selected by synthesizing the three scores,which not only considers the interaction between the answer and the question,the answer and the answer,but also better integrates the information of the respondent,effectively improving the quality of the answer summarization.

关 键 词：问答社区 答案摘要 RoBERTa-wwm 句子相关性 句子新颖性 回答者排序得分 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]