Rethinking ASTE: A Minimalist Tagging Scheme Alongside Contrastive Learning

TL;DR

This work tackles Aspect Sentiment Triplet Extraction (ASTE) by critiquing traditional 2D tagging schemes and proposing a minimalist full-matrix tagging approach paired with token-level contrastive learning. The authors introduce a Contrastive-learning-enhanced PLM Encoder and a Minimalist Tagging Scheme that together yield improved representation distributions and simpler decoding, achieving state-of-the-art or competitive results on two SemEval-derived ASTE datasets while maintaining superior efficiency versus large language models. Empirical analyses (ablation and efficiency studies) demonstrate the synergy between tagging design and contrastive learning, and results extend to other ABSA subtasks. The framework offers a practical and scalable solution compatible with current LLM-era workflows and provides guidance for tagging-scheme design in fine-grained sentiment extraction.

Abstract

Aspect Sentiment Triplet Extraction (ASTE) is a burgeoning subtask of fine-grained sentiment analysis, aiming to extract structured sentiment triplets from unstructured textual data. Existing approaches to ASTE often complicate the task with additional structures or external data. In this research, we propose a novel tagging scheme and employ a contrastive learning approach to mitigate these challenges. The proposed approach demonstrates comparable or superior performance in comparison to state-of-the-art techniques, while featuring a more compact design and reduced computational overhead. Notably, even in the era of Large Language Models (LLMs), our method exhibits superior efficacy compared to GPT 3.5 and GPT 4 in a few-shot learning scenarios. This study also provides valuable insights for the advancement of ASTE techniques within the paradigm of large language models.

Figures (5)

• Figure 1: An example for the ASTE task illustrating Aspect terms in purple, Opinion terms with Negative sentiment in blue, and Opinion terms with Positive sentiment in orange.
• Figure 2: Schematic diagram of our proposed framework which contains: 1) a contrastive learning mechanism that aligns Aspect and Opinion; and 2) a tagging scheme encompassing 5 label classes: NULL, CTD, POS, NEU, and NEG.
• Figure 3: A more detailed illustration of our contrastive learning mechanism.
• Figure 4: Decomposition of the tagging scheme into two components: 1) a beginning mark matrix with sentiment labels; and 2) a placeholder matrix denoting regions of triplets with "1"s and default regions with "0"s. Remember that each row is taken as candidates for an Aspect and each column is taken as candidates for an Opinion. Naturally, each cell in the square matrix can be seen as an ordered pair for a unique candidate of <Aspect, Opinion>. When we simply sum the two components up, we have the left-hand tagging scheme in Figure \ref{['fig.tagging']}, where the "Sentiment & Beginning Tag" is like a trigger (just like you click your mouse), and the "Place Holder" is like a "continued shift" (continue to hold and drag the mouse to the downright).
• Figure 5: A plot of the hidden word representation, where the dimension is reduced to 2 for convenience of display. "Pretrained" means the model with official released version of weights. "Finetuned" means the model is a finetuned version on ASTE task for certain epochs. "w/o. cl" means the model is trained without contrastive learning loss. "+ cl" means the model is trained with contrastive learning. All the plotted results are from experiment carried on $\mathcal{D}_1$ 14Res.