| As technology continues to develop,more and more people are considering social network platforms an indispensable part of daily life.In the massive amount of text generated by these users,there is often sentiment expressed towards a specific topic.Accurately mining the sentiment expressed in these text messages has become one of the hot research areas in the field of natural language processing(NLP).Among them,aspect sentiment triplet extraction(ASTE)technology is widely studied due to its high theoretical and practical significance.Based on the grid tagging scheme(GTS)model,this study proposes three different methods for aspect sentiment triplet extraction,as follows:(1)To address the issue of limited sentiment knowledge in existing aspect sentiment triplet extraction(ASTE)models due to insufficient training data,this study proposes a Document-Level Sentiment Knowledge Transfer Network for Aspect Sentiment Triplet Extraction(SETE)model.The proposed model utilizes large-scale document-level corpora from Twitter and Amazon to pre-train an attention-based document-level sentiment classification(DSC)model.The attention weights of the DSC model and ASTE model are then fused through a gating mechanism to enhance the sentiment knowledge of the triplet extraction model.Experimental results demonstrate that SETE outperforms current mainstream methods on four datasets: res14,lap14,res15,and res16.(2)To address the issue of lacking sub-task interaction in current end-to-end aspect sentiment triplet extraction(ASTE)models,this study proposes a Multi-Task Learning Framework for Aspect Sentiment Triplet Extraction(MTLTE)model.The proposed model simultaneously performs three sub-tasks: aspect-opinion pair extraction,aspectsentiment pair extraction,and triplet extraction.In the triplet extraction task,a gating mechanism is used to obtain semantic information from the other two tasks,enabling the establishment of word pair vector representations with perception abilities for both aspect-opinion pairs and aspect-sentiment pairs.Experimental results on four publicly available datasets demonstrate the effectiveness of the MTLTE model.(3)To address the issue of lack of word-to-word connections in existing aspect sentiment triplet extraction methods,this study proposes a Dependency Tree for Aspect Sentiment Triplet Extraction(DTTE)model.The proposed model utilizes the dependency tree induced from fine-tuned BERT as input.Considering that the aspect-sentiment pair extraction task requires the use of both aspect and opinion words and has fewer parameters relative to other tasks,this task is chosen as the fine-tuning task.Experimental results on four publicly available datasets demonstrate that the introduction of the dependency tree is highly effective for aspect sentiment triplet extraction. |
