基于Span-ASTE的可解释性政府多对象态度识别模型研究  

Research on Interpretable Government Multi-Target Attitude Identification Model Based on Span-ASTE

作　　者：曾江峰 程征 黄泳潼 段尧清[1,2] 

机构地区：[1]华中师范大学信息管理学院,湖北武汉430079 [2]湖北省数据治理与智能决策研究中心,湖北武汉430079

出　　处：《情报理论与实践》2024年第5期175-182,共8页Information Studies:Theory & Application

基　　金：国家自然科学基金青年项目“情感感知的可解释虚假新闻检测研究”(项目编号:62102159);教育部人文社会科学研究青年基金项目“情境大数据驱动的社交媒体虚假信息识别模型与治理策略研究”(项目编号:21YJC870002);湖北省自然科学基金一般面上项目“基于多层语义融合的多模态社交媒体虚假信息检测研究”(项目编号:2023AFB1018)的成果。

摘　　要：[目的/意义]在政府态度识别研究中,针对没有考虑到新闻文本中可能存在多个评价对象、政府态度识别结果可解释性不强等问题,基于Span-ASTE构建政府态度识别模型,提升态度识别效果。[方法/过程]首先,采用BERT提取词级别特征,基于跨度转化为跨度特征表示;然后,联合方面术语和观点术语提取任务提取评价对象和态度描述语,利用双通道跨度剪枝策略筛选得到对象和态度描述语候选池;最后,结合候选池中的对象和态度描述语计算得到态度极性结果。[结果/结论]以美国国务院新闻文本为实验数据进行实验验证。实验结果表明:Span-ASTE在进行政府态度识别时具有一定优越性。相比于效果较好的对比模型,其精确率、召回率和F1值分别提升了约15.68%、19.37%和17.48%;在进行多对象态度识别时同样具有良好的性能表现;态度描述语可为态度极性的判断提供解释依据。[局限]政府态度识别效果还有待进一步提升,且数据规模较为有限,未来可尝试对数据集进行扩充。[Purpose/significance]In the field of government attitude identification,to address the issues of not considering the potential presence of multiple evaluation targets in news texts and the limited interpretability of government attitude identification results,we build a government attitude identification model based on Span-ASTE to enhance attitude identification performance.[Method/process]First,we employ BERT to extract word-level features and transform them into span-level feature representations.Next,in the joint task of aspect term and opinion term extraction,we extract evaluation targets and attitude de-scriptions.Utilizing a dual-channel span pruning strategy,we select candidates for evaluation targets and attitude descriptions.Fi-nally,we calculate the attitude polarity results by combining the targets and attitude descriptions selected from the candidate pool.[Result/conclusion]Experimental validation was conducted using the U.S.Department of State news texts as experimental data,and the results demonstrate the superior performance of Span-ASTE in government attitude identification.In comparison to the best-performing contrastive models,precision,recall,and F1 score have improved by approximately 15.68%,19.37%,and 17.48%,respectively.Moreover,it exhibits commendable performance in multi-target attitude identification.Additionally,attitude descrip-tors serve as explanatory evidence for attitude polarity determination.[Limitations]The identification of government attitudes re-quires further improvement,and the data scale is relatively limited.In the future,it is advisable to consider expanding the dataset for enhanced performance.

关 键 词：Span-ASTE 政府态度 态度识别 多对象 可解释性 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术] D035[自动化与计算机技术—计算机科学与技术]