[Paper Review] Utilizing BERT for Aspect-Based Sentiment Analysis via Constructing Auxiliary Sentence

Chi Sun, Luyao Huang+1 more|arXiv (Cornell University)|Mar 22, 2019

Sentiment Analysis and Opinion MiningComputer Science20 references349 citations

TL;DR

The paper converts (T)ABSA into a sentence-pair classification task by constructing auxiliary sentences and fine-tuning BERT, achieving new state-of-the-art on SentiHood and SemEval-2014 Task 4.

ABSTRACT

Aspect-based sentiment analysis (ABSA), which aims to identify fine-grained opinion polarity towards a specific aspect, is a challenging subtask of sentiment analysis (SA). In this paper, we construct an auxiliary sentence from the aspect and convert ABSA to a sentence-pair classification task, such as question answering (QA) and natural language inference (NLI). We fine-tune the pre-trained model from BERT and achieve new state-of-the-art results on SentiHood and SemEval-2014 Task 4 datasets.

Motivation & Objective

• Motivate ABSA and TABSA as needing fine-grained target-aspect polarity understanding in user comments.
• Propose transforming (T)ABSA into a sentence-pair classification task using auxiliary sentences.
• Demonstrate that BERT-pair models outperform single-sentence BERT baselines on ABSA tasks.
• Provide empirical evidence of state-of-the-art results on SentiHood and SemEval-2014 Task 4 datasets.

Proposed method

• Construct an auxiliary sentence from a target-aspect pair to convert TABSA into a sentence-pair classification task.
• Define four construction methods: QA-M, NLI-M, QA-B, NLI-B for creating the auxiliary sentence.
• Fine-tune pre-trained BERT on the sentence-pair input using a standard classification head with softmax over categories.
• Compare BERT-single (single-sentence) versus BERT-pair (sentence-pair) configurations across the four construction methods.
• Evaluate on SentiHood and SemEval-2014 Task 4 datasets using accuracy, Macro-F1, and AUC as metrics.

Experimental results

Research questions

• RQ1 Can converting TABSA/ABSA into a sentence-pair classification task via auxiliary sentences improve performance over single-sentence BERT fine-tuning?
• RQ2 Which auxiliary-sentence construction method (QA vs NLI; with or without label incorporation) yields the best results for TABSA and ABSA?
• RQ3 Do BERT-pair models consistently outperform BERT-single models across datasets and subtasks (aspect detection and polarity)?

Key findings

• BERT-pair models outperform all baselines on the SentiHood dataset for both aspect detection and sentiment classification.
• Among BERT-pair variants, QA-based constructions tend to excel in sentiment classification, while NLI-based constructions often excel in aspect detection.
• On SemEval-2014 Task 4, BERT-pair configurations achieve state-of-the-art results, with NLI-B delivering strong performance for aspect detection and QA-B for polarity across settings.
• The improvement comes from representing TABSA as a sentence-pair task, leveraging BERT’s strengths in QA/NLI-style inputs, and effectively expanding the data via auxiliary sentences.