| Maha Albayati, Oğuz Fındık Ensemble Strategies for Turkish Legal Extractive Summarization Using Semantic Voting and ROUGE-Weighted Averaging |
|---|
| Abstract. The growing volume of Turkish legal documents poses an important challenge for legal professionals who must efficiently review complex, domain-specific texts. Although extractive summarization has shown promise across many languages, Turkish remains difficult due to its agglutinative morphology and the rigid structure of legal discourse. Existing approaches mainly rely on single methods and often fail to capture both semantic relevance and contextual structure. This study examines whether ensemble strategies can improve the extractive summarization of Turkish legal documents beyond traditional baseline methods. TF-IDF, TextRank, and a BERTurk-Legal-based extractive ranker serve as baseline methods, whereas Semantic Voting and ROUGE-guided weighting are evaluated as ensemble extensions built on these baselines. A curated dataset of 100 Turkish court decisions was compiled and annotated with manually prepared extractive reference summaries. Performance was assessed with ROUGE metrics. Semantic Voting obtained ROUGE-1, ROUGE-2, and ROUGE-L scores of 0.4681, 0.3632, and 0.3968, respectively. ROUGE-Weighted Averaging obtained the highest scores, reaching 0.6423, 0.6377, and 0.6423 on ROUGE-1, ROUGE-2, and ROUGE-L. Overall, the findings suggest that the ensemble configuration, particularly ROUGE-Weighted Averaging, improves summary quality relative to the traditional baselines in this experimental setting. |
| Keywords: Ensemble methods, Extractive summarization, ROUGE evaluation, Semantic similarity, Turkish legal summarization |
 Download PDF |
| DOI: https://doi.org/10.54381/itta2026.3.01 |