Scholarlink

<p>Arabic is a linguistically rich and morphologically complex language, yet it remains under-resourced, particularly in terms of annotated datasets for subjectivity analysis. Current methods struggle with the fragmentation of available resources and the methodological homogeneity of existing ensembles, which often rely on architecturally similar models. This makes it difficult to create robust, generalized tools. To address this, the present study proposes Dhati+, a novel cross-architectural framework for subjectivity assessment. To address the scarcity of specialized datasets, we constructed a comprehensive dataset, <i>AraDhati+</i>, by integrating and augmenting existing Arabic datasets and collections, including the Arabic Subjectivity/Sentiment Dataset (ASTD), the Large-scale Arabic Book Reviews dataset (LABR), the Hotel Arabic-Reviews Dataset (HARD), and the SANAD dataset. We then fine-tuned state-of-the-art Arabic language models–Cross-lingual Language Model-RoBERTa (XLM-RoBERTa), Arabic Bidirectional Encoder Representations from Transformers (AraBERT), and Arabian Generative Pre-trained Transformer (ArabianGPT)–on the AraDhati+ dataset. Crucially, we implemented a Heterogeneous Ensemble Strategy that fuses the distinct cognitive paradigms of Bidirectional Discriminative Encoders and a Unidirectional Generative Decoder. This approach mitigates individual inductive biases by balancing contextual semantic extraction with causal reasoning. Our proposed approach achieved an impressive accuracy of 97.88% in Arabic subjectivity classification. Moreover, our ensemble undergoes thorough validation via extensive experiments, encompassing benchmark dataset assessments, comparisons with prior work, and evaluations against both Arabic-specific and commercial LLMs, thereby establishing a new state-of-the-art in Arabic subjectivity detection. These results demonstrate the effectiveness of our cross-architectural methodology and the relevance of the <i>AraDhati+</i> dataset in addressing the challenges posed by limited resources in Arabic language processing.</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

Dhati+: fine-tuned large language models for Arabic subjectivity evaluation

  • 期刊论文

  • 发表日期：

    2026年2月26日

查看全文
  • Slimane Bellaouar,
  • Attia Nehar,
  • Soumia Souffi,
  • Mounia Bouameur

摘要

Arabic is a linguistically rich and morphologically complex language, yet it remains under-resourced, particularly in terms of annotated datasets for subjectivity analysis. Current methods struggle with the fragmentation of available resources and the methodological homogeneity of existing ensembles, which often rely on architecturally similar models. This makes it difficult to create robust, generalized tools. To address this, the present study proposes Dhati+, a novel cross-architectural framework for subjectivity assessment. To address the scarcity of specialized datasets, we constructed a comprehensive dataset, AraDhati+, by integrating and augmenting existing Arabic datasets and collections, including the Arabic Subjectivity/Sentiment Dataset (ASTD), the Large-scale Arabic Book Reviews dataset (LABR), the Hotel Arabic-Reviews Dataset (HARD), and the SANAD dataset. We then fine-tuned state-of-the-art Arabic language models–Cross-lingual Language Model-RoBERTa (XLM-RoBERTa), Arabic Bidirectional Encoder Representations from Transformers (AraBERT), and Arabian Generative Pre-trained Transformer (ArabianGPT)–on the AraDhati+ dataset. Crucially, we implemented a Heterogeneous Ensemble Strategy that fuses the distinct cognitive paradigms of Bidirectional Discriminative Encoders and a Unidirectional Generative Decoder. This approach mitigates individual inductive biases by balancing contextual semantic extraction with causal reasoning. Our proposed approach achieved an impressive accuracy of 97.88% in Arabic subjectivity classification. Moreover, our ensemble undergoes thorough validation via extensive experiments, encompassing benchmark dataset assessments, comparisons with prior work, and evaluations against both Arabic-specific and commercial LLMs, thereby establishing a new state-of-the-art in Arabic subjectivity detection. These results demonstrate the effectiveness of our cross-architectural methodology and the relevance of the AraDhati+ dataset in addressing the challenges posed by limited resources in Arabic language processing.