Document-Level Sentiment Knowledge Transfer Network for Aspect Sentiment Triplet Extraction

Aspect Sentiment Triplet Extraction (ASTE) task has been widely studied in Natural Language Processing (NLP). Nevertheless, existing ASTE methods are processed on a small public dataset, leading to frequent appeared sentiment words being focused on and neglecting the processing of niche sentiment words with strong sentiment tendencies. Therefore, the model extracts the wrong triples. To address this problem, we combine two attention weights of the Document-Level Sentiment Classification (DSC) and the ASTE based on the attention mechanism to obtain the ASTE representation with enhanced sentiment knowledge. Meanwhile, an attention masking strategy is introduced in the DSC model to ignore noisy words irrelevant to sentiment classification and ensure the effective extraction of sentiment triples. Experiments on four benchmark datasets from ASTE-Data-V1 show that our proposed method can acquire sentiment knowledge from large-scale document-level corpora, outperforming existing baseline models.