Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

BERT-ASC: Auxiliary-Sentence Construction for Implicit Aspect Learning in Sentiment Analysis

Murtadha Ahmed, Bo Wen, Shengfeng Pan, Jianlin Su, Luo Ao, Yunfeng Liu

TL;DR

This work tackles implicit aspect learning in ABSA and TABSA by constructing an auxiliary sentence that encodes the semantic neighborhood of an aspect's seeds and training BERT to respond to this cue rather than the explicit aspect label. Seeds are extracted via L-LDA, and the auxiliary sentence combines semantic candidates with syntactic modifiers to guide aspect-specific representation learning. Across SemEval ABSA and SentiHood TABSA benchmarks, BERT-ASC achieves state-of-the-art results in aspect categorization and strong performance in sentiment tasks, with ablations underscoring the importance of semantic information. The approach reduces dependence on large labeled datasets for implicit aspects and demonstrates robustness across multiple pre-trained language models, offering a practical pathway to improved fine-grained sentiment analysis in real-world scenarios.

Abstract

Aspect-based sentiment analysis (ABSA) aims to associate a text with a set of aspects and infer their respective sentimental polarities. State-of-the-art approaches are built on fine-tuning pre-trained language models, focusing on learning aspect-specific representations from the corpus. However, aspects are often expressed implicitly, making implicit mapping challenging without sufficient labeled examples, which may be scarce in real-world scenarios. This paper proposes a unified framework to address aspect categorization and aspect-based sentiment subtasks. We introduce a mechanism to construct an auxiliary-sentence for the implicit aspect using the corpus's semantic information. We then encourage BERT to learn aspect-specific representation in response to this auxiliary-sentence, not the aspect itself. We evaluate our approach on real benchmark datasets for both ABSA and Targeted-ABSA tasks. Our experiments show that it consistently achieves state-of-the-art performance in aspect categorization and aspect-based sentiment across all datasets, with considerable improvement margins. The BERT-ASC code is available at https://github.com/amurtadha/BERT-ASC.

BERT-ASC: Auxiliary-Sentence Construction for Implicit Aspect Learning in Sentiment Analysis

TL;DR

This work tackles implicit aspect learning in ABSA and TABSA by constructing an auxiliary sentence that encodes the semantic neighborhood of an aspect's seeds and training BERT to respond to this cue rather than the explicit aspect label. Seeds are extracted via L-LDA, and the auxiliary sentence combines semantic candidates with syntactic modifiers to guide aspect-specific representation learning. Across SemEval ABSA and SentiHood TABSA benchmarks, BERT-ASC achieves state-of-the-art results in aspect categorization and strong performance in sentiment tasks, with ablations underscoring the importance of semantic information. The approach reduces dependence on large labeled datasets for implicit aspects and demonstrates robustness across multiple pre-trained language models, offering a practical pathway to improved fine-grained sentiment analysis in real-world scenarios.

Abstract

Aspect-based sentiment analysis (ABSA) aims to associate a text with a set of aspects and infer their respective sentimental polarities. State-of-the-art approaches are built on fine-tuning pre-trained language models, focusing on learning aspect-specific representations from the corpus. However, aspects are often expressed implicitly, making implicit mapping challenging without sufficient labeled examples, which may be scarce in real-world scenarios. This paper proposes a unified framework to address aspect categorization and aspect-based sentiment subtasks. We introduce a mechanism to construct an auxiliary-sentence for the implicit aspect using the corpus's semantic information. We then encourage BERT to learn aspect-specific representation in response to this auxiliary-sentence, not the aspect itself. We evaluate our approach on real benchmark datasets for both ABSA and Targeted-ABSA tasks. Our experiments show that it consistently achieves state-of-the-art performance in aspect categorization and aspect-based sentiment across all datasets, with considerable improvement margins. The BERT-ASC code is available at https://github.com/amurtadha/BERT-ASC.
Paper Structure (20 sections, 2 equations, 3 figures, 12 tables, 1 algorithm)

This paper contains 20 sections, 2 equations, 3 figures, 12 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. The Proposed Solution
  4. Task Definition
  5. Aspect Seed Extraction
  6. Auxiliary-Sentence Construction
  7. Semantic candidates
  8. Syntactic information
  9. BERT-based classifier
  10. Empirical Evaluation
  11. Datasets
  12. Experimental Settings
  13. Exp-I: ABSA
  14. Results
  15. Ablation Study
  16. ...and 5 more sections

Figures (3)

  • Figure 1: illustrative example of BERT-ASC. It consists of two steps: constructing an auxiliary sentence and jointly training for aspect and sentiment classification. Consider sentence $s_2$ from the running example in response to "service." The generated auxiliary sentence includes the semantic candidate waiters and its syntactic modifier friendly. The aspect categorization layer processes the output of the aspect sentiment layer and applies binary classification to determine whether the discussed aspect is on-target in the input sentence. The label none denotes $0$, while any other label denotes $1$.
  • Figure 2: An illustrative example of the food's coherent distribution in the embedding space.
  • Figure 3: An example of the dependency parser for sentence $s_1$ from the running example. POS abbreviations used include: VBD (past tense of a verb), PRP (pronoun), VBP (present tense, non-third person singular form of a verb), IN (preposition), DT (determiner), NN (noun), VBZ (present tense, third person singular form of a verb), JJ (adjective). The dependency links represented are: nsubj (subject of a verb), aux (auxiliary verb), ccomp (complement clause), det (determiner), mark (clause marker), cop (copula verb mark).