A span-level Interactive Network for Aspect Sentiment Triplet Extraction Based on Learning Automated Concatenation of Embeddings

The ASTE task aims to extract sentiment triplets, where word embeddings are crucial for semantic expression. Existing models typically rely on a single pre-training model for context embeddings. However, recent research indicates that better word representations can be obtained by combining various types of embeddings. At the same time, most existing task models do not take into account the full use of the hidden semantic associations in the word representations of aspect items and opinion items. To address these issues, the Span-ACE (Automated Concatenation of Embedding) model is introduced. Span-ACE employs different pre-trained models for initial word embedding and dynamically selects embedding connections using reinforcement learning. It enhances the interaction between aspect and opinion items by employing self-information and dot product fusion for embedding. The model also screens and minimizes low-scoring embeddings based on predefined thresholds, resulting in improved performance. The fianl experimental results demonstrate that the Span-ACE model surpasses the state-of-the-art (SOTA) model on a subset of the ASTE-Data-V2-EMNLP2020 dataset for the ASTE task.