Büyük Dil Modellerini Bağlam Düzenleyiciler Olarak Kullanma: Yabancı Dil Olarak İngilizce Öğrenci Geri Bildirimlerini Analiz Etmek İçin Bir Derlem Destekli Çerçeve

Özet (abstract)

Student feedback is central to English as a Foreign Language (EFL) instruction; however, scaling qualitative interpretation while preserving contextual meaning remains a methodological challenge. This study proposes the T-CADS-LLMs framework, a qualitative text topic analysis framework integrated into an Agent Skill. Rather than treating large language models as autonomous coders, the framework positions them as context organizers constrained by corpus evidence and human verification. Its feasibility was evaluated using interview data from 27 students across two majors in university EFL courses. Across ten course grouped analytical units, 46 manually generated topics were compared with 63 Skill-based topics and 42 GPT-based topics using Topic Coverage, Weighted Coverage, MATS, WMATS, and qualitative analyses of matched topics. Codex+Skill consistently outperformed the Bare-model in descriptive mean scores for coverage (0.672 vs. 0.502), weighted coverage (0.721 vs. 0.600), MATS (0.778 vs. 0.700), and WMATS (0.816 vs. 0.772). Nevertheless, both workflows showed limited ability to capture relational and identity related meanings, performing substantially below human analysis. These findings provide initial evidence for the feasibility and traceability of the framework within a specific EFL context, while further validation with larger and more diverse datasets remains necessary.

Yazarların özeti; kaynağından alınmıştır. Teaching English as a Second or Foreign Language--TESL-EJ, 2026 · DOI ↗