用于方面情感三元组抽取的双向级联网络

Bidirectional Cascade Network for Aspect Sentiment Triplet Extraction

方面情感三元组抽取(ASTE)是基于方面的情感分析子任务之一,其目标是从给定文本序列中抽取所有提及的方面及其对应的观点、观点表达的情感倾向,构成三元组。基于跨度级交互和端到端两个前提,提出双向级联网络。在编码块里应用条件层归一化让方面跨度和观点跨度的特征进行深层交互,实现两者间的级联;在“方面到观点”和“观点到方面”两个方向上进行三元组抽取,并设计解码策略聚合两个方向上的结果;为缓解类别不平衡问题,在多标签交叉熵损失中加入稀疏因子来提高训练环节对稀疏正类的惩罚力度;为缓解曝光偏差问题,设计跨度漂移和输出采样两种策略来构造负样本加入训练。在ASTE-Data-V2-EMNLP2020数据集上进行实验,结果表明,所提模型在4个子数据集14LAP、14RES、15RES、16RES上较SBN-ASTE模型的F1值分别提高0.91、0.17、2.0、1.56个百分点。

Aspect Sentiment Triplet Extraction(ASTE)is a subtask of aspect-based sentiment analysis.The goal is to extract all mentioned aspects,their corresponding opinions,and the sentiment polarity expressed,forming triplets from a given text sequence.To achieve this,a bidirectional cascade network is proposed,leveraging span-level interactions and end-to-end extraction.Conditional layer normalization is applied within the encoding blocks to facilitate deep-level interactions between the aspect and opinion spans,creating a cascading effect.Triplets are extracted in both″aspect-to-opinion″and″opinion-to-aspect″directions,with decoding strategies that merge results from both directions.To address class imbalance,a sparse factor is introduced into the multilabel cross-entropy loss,increasing the penalty for sparse positive classes during training.To mitigate exposure bias,span drifting and output sampling used to construct negative training samples.Experimental results on the ASTE-Data-V2-EMNLP2020 dataset show improvements in F1 scores over the SBN-ASTE model across four sub-datasets:14LAP,14RES,15RES,and 16RES,with increases of 0.91,0.17,2.0,and 1.56 percentage points,respectively.