Scholarlink

<p>Aspect-based sentiment analysis (ABSA) has emerged as an important task in natural language processing, as it offers an in-depth understanding of the sentiments related to specific aspects in the text. The present research work investigates novel methodologies used in ABSA, by addressing issues like implicit aspect identification, sentiment context alignment, and domain adaptability. A comprehensive examination reveals that current methods face important limitations: Transformer-based models such as GPT provide strong contextual representations but often fail to capture aspect-level sentiment dependencies, while recurrent neural network-based models handle local sequential context but lack broad semantic generalization. To address these shortcomings, a competitive and comprehensive ABSA pipeline is proposed, leveraging the sequential learning strengths of bidirectional long short-term memory (BiLSTM) networks alongside the contextual understanding provided by large language models (LLMs) such as GPT−3.5-turbo and GPT-4o-mini. In a zero-shot setting, GPT is employed to extract aspect terms, which are concatenated with their respective sentences, embedded into aspect–sentence pairs and categorized using a BiLSTM. Datasets such as SemEval 2014 Task 4, which include both restaurant and laptop reviews, were employed to enable cross-domain analysis. This approach proved effective for tasks involving implicit aspect detection and fine-grained sentiment classification, where our hybrid model achieved F1 scores of 87.20% (merged dataset), 83.81% (Restaurant 14), and 86.98% (Laptop 14). To further validate the adaptability across heterogeneous domains, we also conducted experiments on the SemEval 2022 Task 10 dataset, focusing on the English subsets Opener_en, MPQA, and their merged version. The model achieved F1 scores of 69.5%, 77.27%, and 72.57%, showing competitive performance on complex multi-aspect datasets such as MPQA, while lower performance on Opener_en highlights limitations in simpler settings with implicit aspects. This gap is likely due to increased linguistic variability and more complex structures in the SemEval 2022 Task 10 dataset, which make zero-shot aspect extraction more challenging. Collectively, these results demonstrate that the combination of LLM-based aspect extraction and supervised BiLSTM modeling provides a competitive aspect-aware sentiment classification pipeline.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A robust ABSA pipeline integrating large language models and BiLSTM networks

  • 期刊论文

  • 发表日期：

    2026年6月16日

查看全文
  • Nadia Smairi,
  • Houda Abadlia,
  • Moemen Said

摘要

Aspect-based sentiment analysis (ABSA) has emerged as an important task in natural language processing, as it offers an in-depth understanding of the sentiments related to specific aspects in the text. The present research work investigates novel methodologies used in ABSA, by addressing issues like implicit aspect identification, sentiment context alignment, and domain adaptability. A comprehensive examination reveals that current methods face important limitations: Transformer-based models such as GPT provide strong contextual representations but often fail to capture aspect-level sentiment dependencies, while recurrent neural network-based models handle local sequential context but lack broad semantic generalization. To address these shortcomings, a competitive and comprehensive ABSA pipeline is proposed, leveraging the sequential learning strengths of bidirectional long short-term memory (BiLSTM) networks alongside the contextual understanding provided by large language models (LLMs) such as GPT−3.5-turbo and GPT-4o-mini. In a zero-shot setting, GPT is employed to extract aspect terms, which are concatenated with their respective sentences, embedded into aspect–sentence pairs and categorized using a BiLSTM. Datasets such as SemEval 2014 Task 4, which include both restaurant and laptop reviews, were employed to enable cross-domain analysis. This approach proved effective for tasks involving implicit aspect detection and fine-grained sentiment classification, where our hybrid model achieved F1 scores of 87.20% (merged dataset), 83.81% (Restaurant 14), and 86.98% (Laptop 14). To further validate the adaptability across heterogeneous domains, we also conducted experiments on the SemEval 2022 Task 10 dataset, focusing on the English subsets Opener_en, MPQA, and their merged version. The model achieved F1 scores of 69.5%, 77.27%, and 72.57%, showing competitive performance on complex multi-aspect datasets such as MPQA, while lower performance on Opener_en highlights limitations in simpler settings with implicit aspects. This gap is likely due to increased linguistic variability and more complex structures in the SemEval 2022 Task 10 dataset, which make zero-shot aspect extraction more challenging. Collectively, these results demonstrate that the combination of LLM-based aspect extraction and supervised BiLSTM modeling provides a competitive aspect-aware sentiment classification pipeline.