[Paper Review] Zero-Shot Aspect-Based Sentiment Analysis

Lei Shu, Xu Hu+2 more|arXiv (Cornell University)|Feb 4, 2022

Sentiment Analysis and Opinion Mining9 citations

TL;DR

This paper proposes CORN, a contrastive post-training method on rule-curated review natural language inference (RNLI) data to enable zero-shot aspect-based sentiment analysis (ABSA) across aspect extraction (AE), aspect sentiment classification (ASC), and end-to-end ABSA without any human-annotated ABSA data. By casting ABSA tasks as natural language inference and leveraging domain-invariant representations through contrastive learning on RNLI, CORN achieves strong zero-shot performance, outperforming baselines on AE and E2E tasks while eliminating the need for task-specific annotations.

ABSTRACT

Aspect-based sentiment analysis (ABSA) typically requires in-domain annotated data for supervised training/fine-tuning. It is a big challenge to scale ABSA to a large number of new domains. This paper aims to train a unified model that can perform zero-shot ABSA without using any annotated data for a new domain. We propose a method called contrastive post-training on review Natural Language Inference (CORN). Later ABSA tasks can be cast into NLI for zero-shot transfer. We evaluate CORN on ABSA tasks, ranging from aspect extraction (AE), aspect sentiment classification (ASC), to end-to-end aspect-based sentiment analysis (E2E ABSA), which show ABSA can be conducted without any human annotated ABSA data.

Motivation & Objective

• To address the challenge of scaling aspect-based sentiment analysis (ABSA) to new domains without requiring in-domain human-annotated data.
• To enable zero-shot transfer for three ABSA tasks—aspect extraction (AE), aspect sentiment classification (ASC), and end-to-end ABSA (E2E)—without any labeled examples in the target domain.
• To develop a unified model that learns domain-invariant aspect-opinion and aspect-entity dependencies through self-supervised contrastive post-training on review NLI.
• To reduce reliance on expensive human annotation by replacing human-annotated NLI data with rule-based curation of review NLI examples.

Proposed method

• Casts all ABSA tasks (AE, ASC, E2E) as natural language inference (NLI) by constructing hypotheses from text spans in reviews.
• Pre-processes unlabeled reviews into review NLI (RNLI) format using rule-based heuristics to create hypothesis pairs like 'The product has <S>' for span S.
• Applies contrastive post-training on RNLI data to learn representations that capture aspect-opinion and aspect-entity dependencies across domains.
• Uses NLI model predictions to transfer to ABSA label spaces: entailment for AE (B/I/O), entailment for ASC (positive/negative/neutral), and combined predictions for E2E ABSA.
• Employs a contrastive learning objective during post-training to improve generalization by distinguishing correct aspect-opinion relations from incorrect or non-aspect spans.
• Utilizes a unified BART-based architecture fine-tuned via contrastive post-training on RNLI to support zero-shot inference across all three ABSA tasks.

Experimental results

Research questions

• RQ1 Can a unified model perform zero-shot aspect extraction (AE) without any human-annotated data in the target domain?
• RQ2 Can zero-shot aspect sentiment classification (ASC) be achieved through NLI-based inference without task-specific fine-tuning?
• RQ3 Can end-to-end ABSA be performed in zero-shot fashion using only rule-curated NLI data instead of human-annotated ABSA examples?
• RQ4 How effective is contrastive post-training on rule-based RNLI data compared to standard post-training or fine-tuning on human-annotated NLI datasets?

Key findings

• CORN outperforms all baselines on aspect extraction (AE) and end-to-end ABSA (E2E), achieving the highest F1 scores without any human-annotated ABSA data.
• On AE, CORN achieves higher recall than MNLI-fine-tuned models (BART+MNLI), indicating better coverage of aspect terms despite using rule-based RNLI curation.
• CORN performs comparably to MNLI-fine-tuned models (BART+MNLI) on ASC, with similar accuracy and macro-F1, demonstrating the effectiveness of rule-based RNLI for sentiment classification.
• Contrastive post-training on RNLI leads to significant performance gains: CORN outperforms BART+CE+RNLI, validating the effectiveness of the contrastive objective.
• Post-training on domain-specific RNLI data improves performance over pre-trained models by at least 6.3% on E2E ABSA, highlighting the importance of domain adaptation.
• Qualitative analysis shows CORN correctly identifies 'Windows 8' as an aspect and assigns neutral sentiment, while BART+MNLI misses '8' and BERT+PT misclassifies sentiment, underscoring CORN's robustness.