Bidirectional Cascade Network for Aspect Sentiment Triplet Extraction
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 Fl 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.
万方数据
维普
