Utilizing BERT for Aspect-Based Sentiment Analysis via Constructing Auxiliary Sentence

• ABSA identifies fine-grained polarity toward specific aspects, enabling more granular evaluation of product or service quality.Source: §1 · p. 1
• TABSA handles comments involving multiple objects by identifying aspects associated with each target and resolving their polarity.Source: §1 · p. 1
• Pre-trained BERT has been effective in QA and NLI, but its direct use has produced little improvement on (T)ABSA.Source: §1 · p. 1
• The paper constructs an auxiliary sentence and transforms (T)ABSA into a sentence-pair classification task.Source: §1 · p. 2
• Fine-tuning pre-trained BERT in this formulation achieves new state-of-the-art results on SentiHood and SemEval-2014 Task 4 datasets.Source: §1 · p. 2

Task formulation

• TABSA predicts positive, negative, or none for each target-aspect pair, while ABSA predicts aspect detection and polarity jointly.Source: §2.1 · p. 2

Auxiliary sentence construction

• The paper considers four auxiliary-sentence constructions: QA-M, NLI-M, QA-B, and NLI-B.QA-M uses questions; NLI-M uses pseudo-sentences; QA-B and NLI-B append candidate polarity labels.Source: §2.2 · pp. 2–3
• QA-M generates a question from each target-aspect pair, such as asking about the safety of LOCATION1.Source: §2.2 · p. 2
• QA-B creates three labeled sequences for positive, negative, and none, then selects the category with the highest yes-matching score.Source: §2.2 · pp. 2–3
• The constructed auxiliary sentence transforms single-sentence classification into sentence-pair classification, which significantly improves TABSA results.Source: §2.2 · p. 3

BERT fine-tuning

• BERT represents the original and auxiliary sentences as a token sequence using token, segment, and position embeddings, with [CLS] as the classification representation.Source: §2.3.1 · p. 3
• Fine-tuning feeds the final hidden state of [CLS] through a classification layer and softmax to obtain category probabilities.The pooled vector is C ∈ R^H, the classifier matrix is W ∈ R^{K×H}, and P = softmax(CW^T).Source: §2.3.2 · p. 3

Model variants

• BERT-single trains separate classifiers for target-aspect or aspect sentiment problems, whereas BERT-pair uses the four auxiliary-sentence variants.Source: §2.3.3 · p. 3

SentiHood

• SentiHood evaluation covers target-aspect detection and sentiment polarity for four frequent aspects, using strict accuracy, Macro-F1, AUC, accuracy, and macro-average AUC.The dataset contains 5,215 sentences and includes both single-target and multiple-target cases.Source: §3.1 · p. 3, §3.3 · p. 4

SemEval-2014 Task 4

• SemEval-2014 evaluation jointly addresses aspect category detection and polarity using Micro-F1 and accuracy, respectively.Source: §3.1 · p. 3, §3.4 · p. 4

SentiHood

• BERT-single detects aspects better than Dmu-Entnet but trails SenticLSTM and Dmu-Entnet in sentiment-classification accuracy by 3.8 and 5.5, respectively.Source: §3.3.1 · p. 4
• 9.4 macro-average F1 and 2.6 accuracies improvement over Dmu-Entnet are obtained by BERT-pair on SentiHood aspect detection and sentiment analysis.BERT-pair-NLI models perform relatively better on aspect detection, while BERT-pair-QA models perform better on sentiment classification.Source: §3.3.1 · p. 4

SemEval-2014 Task 4

• BERT-single improves both SemEval-2014 subtasks, while BERT-pair improves further; BERT-pair-NLI-B leads detection and BERT-pair-QA-B leads polarity across all settings.Source: §3.4.1 · p. 5

Discussion

• Auxiliary sentences expand each original sentence into target-aspect-specific sentence pairs, increasing the effective corpus examples for classification.Source: §4 · p. 5
• BERT’s strength on QA and NLI supports the sentence-pair formulation, drawing on masked language modeling and next sentence prediction.Source: §4 · p. 5
• Direct BERT fine-tuning does not improve TABSA, whereas separating target and aspect into an auxiliary sentence enables the model’s pre-trained advantage to be utilized.Source: §4 · p. 5
• BERT-pair-QA-B and BERT-pair-NLI-B achieve better sentiment-classification AUC values, probably because they model label information.Source: §4 · p. 5

Conclusion

• The method transforms (T)ABSA from single-sentence classification into sentence-pair classification through an auxiliary sentence.Source: §5 · p. 5
• Fine-tuned BERT produces new state-of-the-art results, while comparisons between single-sentence and sentence-pair classification verify the conversion method’s validity.Source: §5 · p. 5
• The authors propose applying the conversion method to other similar tasks in future work.Source: §5 · p. 5