Functional Sentence Classification in Legal Decision Summaries: Inter-Task Dependency and Label Derivation Hukuki Karar Özetlerinde Islevsel Cümle Siniflandirmasi: Görevler Arasi Bagimlilik ve Etiket Türetme
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Turkey, 7 - 10 July 2026, (Full Text)
- Publication Type: Conference Paper / Full Text
- Doi Number: 10.1109/siu71813.2026.11636372
- City: İstanbul
- Country: Turkey
- Keywords: legal argument mining, legal sentence classification, multi-task deep learning, Natural Language Processing (NLP)
- Middle East Technical University Affiliated: Yes
Abstract
This study addresses sentence-level functional classification in Constitutional Court decision summaries. A dataset of 10,213 sentences compiled from 876 decision summaries was used. Each sentence was evaluated in terms of its function within the summary, its content type, and a composite label derived from the combination of these two tasks. Relevant sentence segments were also annotated with linguistic markers that make the surface traces of legal reasoning visible. Methodologically, a TF-IDF-based SVM and single-task, multi-task, and two-task label-deriving BERT models were compared. The findings show that the marker layer improves classification performance, that the TF-IDF SVM provides a strong but more overfitting-prone baseline, and that BERT-based models achieve higher macro-F1 scores under the marked main-data setting. The two-task BERT approach reaches the highest 128-token test macro-F1 for the composite task and shows a limited train-test gap at the derived-output level.