Embarrassingly Simple Unsupervised Aspect Based Sentiment Tuple Extraction

TL;DR

This work tackles the challenge of ABSA in low-resource settings by proposing a simple unsupervised pipeline for AOOSPE, capable of extracting opinion terms and assigning sentiment to aspect terms without labeled data. The method combines domain adaptation, compound phrase extraction, attention-based candidate weighting, and semantic polarity assignment, with $AttScore_{[a_i^k]} = \text{softmax}\left(\frac{Q_{[a_i^k]} \cdot K^T}{\sqrt{d_k}}\right)$ and $sp_i^k = \arg\max_{c \in \{positive, negative, neutral\}} \cos\left(h_{op}, \vec{c}\right)$. Evaluations on SemEval 2014–2016 datasets (L14, R14–R16) show strong unsupervised performance across AOOE, ATSC, and AOOSPE, and ablation analyses reveal benefits from domain adaptation, massive domain finetuning, and joint-domain adaptation. The approach establishes a practical benchmark for unsupervised opinion word mining in ABSA and demonstrates notable cross-domain generalization, suggesting utility in low-resource domains where labeled data are scarce. The work also highlights the positive impact of labeled data and larger model scales on unsupervised ABSA performance, and the authors provide public code to facilitate further research.

Abstract

Aspect Based Sentiment Analysis (ABSA) tasks involve the extraction of fine-grained sentiment tuples from sentences, aiming to discern the author's opinions. Conventional methodologies predominantly rely on supervised approaches; however, the efficacy of such methods diminishes in low-resource domains lacking labeled datasets since they often lack the ability to generalize across domains. To address this challenge, we propose a simple and novel unsupervised approach to extract opinion terms and the corresponding sentiment polarity for aspect terms in a sentence. Our experimental evaluations, conducted on four benchmark datasets, demonstrate compelling performance to extract the aspect oriented opinion words as well as assigning sentiment polarity. Additionally, unsupervised approaches for opinion word mining have not been explored and our work establishes a benchmark for the same.

Figures (3)

• Figure 1: Approach overview for ATSC subtask
• Figure 2: An overview of our approach for AOOE subtask
• Figure 3: Accuracy across tasks when different % of labelled instances are available by the ELECTRA model.