Unified Instance and Knowledge Alignment Pretraining for Aspect-Based Sentiment Analysis

The goal of aspect-based sentiment analysis (ABSA) is to determine the polarity towards an aspect. Because expensive and limited amounts labelled data, pretraining strategy has become de facto standard for ABSA. However, there always exists a severe domain shift between downstream ABSA datasets, which hinders effective knowledge transfer when directly fine-tuning, making task suboptimal. To mitigate this shift, we introduce unified alignment framework into vanilla pretrain-finetune pipeline, that both instance- knowledge-level alignments. Specifically, first devise novel coarse-to-fine retrieval sampling approach select target domain-related instances from large-scale dataset, thus aligning domains ( First Stage ). Then, guidance-based further bridge gap at level. In practice, formulate model pretrained on sampled guidance learner model. On design on-the-fly teacher-student joint fine-tuning progressively xmlns:xlink="http://www.w3.org/1999/xlink">Second Therefore, can maintain more domain-invariant learning new dataset. xmlns:xlink="http://www.w3.org/1999/xlink">Third Stage, finetuned better adapt its learned Extensive experiments analyses several benchmarks demonstrate effectiveness universality our proposed framework.