基于语义对齐与模板可靠性的方面级情感四元组预测
雷子锟 , 陈彦江 , 张凯 , 刘淇
天津大学学报(自然科学与工程技术版) ›› 2026, Vol. 59 ›› Issue (11) : 1187 -1196.
基于语义对齐与模板可靠性的方面级情感四元组预测
Aspect Sentiment Quad Prediction via Semantic Matching and Reliable Template Selection
方面级情感四元组预测(ASQP)任务旨在从文本中同时抽取方面项、方面类别、观点项和情感极性,是情感分析领域中的热门问题.当下主流生成式方法通过多模板引导预训练语言模型生成四元组,然而在进行模板的选取时通常忽视了模板之间潜在的语义关联或各模板本身的性能差异.为此,本文提出一种基于预训练语言模型 T5 的模板可靠性与语义对齐平衡(SMARTS)策略.该策略在模板选择阶段,结合交叉熵和 Jensen-Shannon(JS)散度,构建多个兼顾生成性能与语义关联性的高质量模板分组.首先,以交叉熵作为衡量指标,评估各候选模板在预训练语言模型下的生成性能,筛选出若干生成质量最优的模板;接着,考虑不同模板之间的结构和语义关联,模型进一步引入 JS 散度对模板间的分布差异进行建模,从而为每个性能最优模板匹配若干语义最相关的模板,形成多个高一致性的模板分组.在推理阶段,为充分对齐 SMARTS 策略,本文提出了双层投票聚合策略,先在每组内选出最优结果,再在组间进行全局投票,提升预测稳定性和准确性.在 2 个公开数据集 Rest15 和 Rest16 上实验结果表明了本文所提方法的有效性与优越性,且为了验证各个部分的有效性及更深入了解模型工作原理,文中进行了详细的超参数研究、消融实验以及案例研究.
Aspect sentiment quad prediction(ASQP) involves the simultaneous extraction of aspect terms,aspect categories,opinion terms,and sentiment polarities from text,presenting a significant challenge in sentiment analysis. Current mainstream approaches typically employ generative methods that utilize multiple templates to guide pre-trained language models in generating sentiment quads. However,these methods often neglect the semantic relationships between templates and the performance discrepancies among individual templates during the selection process. To overcome these limitations,this study introduces a strategy called semantic matching and reliable templates selection(SMARTS),which is based on the pretrained language model T5 and aims to balance template reliability with semantic alignment. SMARTS first evaluates the generation performance of candidate templates using cross-entropy to filter out those with optimal generation quality. Subsequently,considering the structural and semantic correlations among different templates,it further introduces Jensen-Shannon(JS) divergence to model their distribution differences. This allows SMARTS to match each optimal template with several of the most semantically related ones,thereby forming multiple highly consistent template groups. To align with the SMARTS strategy during inference,a two-stage voting aggregation mechanism is introduced. First,an intra-group voting process identifies the best candidate within each group. Then,a global voting stage aggregates results across groups to determine the final prediction,thereby enhancing both stability and accuracy. Experimental results on two widely-used benchmarks,Rest15 and Rest16,demonstrate the effectiveness and superiority of the proposed model. In addition,extensive experiments,including hyperparameter analysis,ablation studies,and case studies,are conducted to validate the contributions of individual components and to provide deeper insights into the model's behavior.
| [1] |
郭贤伟, 赖 华, 余正涛, |
| [2] |
|
| [3] |
|
| [4] |
|
| [5] |
|
| [6] |
|
| [7] |
|
| [8] |
鲍小异, 姜晓彤, 王中卿, |
| [9] |
|
| [10] |
|
| [11] |
|
| [12] |
|
| [13] |
|
| [14] |
林 煌, 李弼程. 基于 BERT 模型和图注意力网络的方面级情感分析[J]. 计算机科学, 2024, 51(增 2): 240400018-7. |
| [15] |
|
| [16] |
|
| [17] |
|
| [18] |
|
| [19] |
|
| [20] |
|
| [21] |
|
| [22] |
|
| [23] |
|
国家自然科学基金资助项目(62337001)
/
| 〈 |
|
〉 |