MSc Thesis Defense: Niyazi Ahmet Metin, SARCASM IN TURKISH INFORMAL ONLINE DISCOURSE: CONTEXT-AWARE DETECTION AND ITS EFFECTS ON HATE SPEECH DETECTION, Date & Time: 14 July, 2026 – 3:30 PM, Place: FENS L065

SARCASM IN TURKISH INFORMAL ONLINE DISCOURSE: CONTEXT-AWARE DETECTION AND ITS EFFECTS ON HATE SPEECH DETECTION

Niyazi Ahmet Metin
Computer Science & Engineering, MSc Thesis, July 2026

Thesis Jury

     Asst. Prof. Dilara Keküllüoğlu (Thesis Advisor)

  Asst. Prof. Onur Varol

  Asst. Prof. Gizem Gezici

Date & Time: July 14th, 2026 – 3:30 PM

Place: FENS L065

Keywords : Sarcasm Detection, Turkish Natural Language Processing, Sarcasm and Hate Speech, Online Discourse, Context Generation

Abstract

Sarcasm is a form of non-literal expression in which intended meaning can diverge from surface wording. This makes it difficult for NLP systems that rely on explicit textual cues. This thesis studies sarcasm in Turkish informal online discourse through two complementary datasets built from Ekşi Sözlük. First, it introduces SarcasTürk, a context-aware Turkish sarcasm detection dataset containing 1,515 entries from 98 titles with binary sarcasm labels and title-level contextual summaries. The dataset enables comparison between entry-only and context-aware sarcasm detection. Title-level contexts are generated through sentence selection, redundancy control, large language model rewriting, and human validation. Experiments with fine-tuned BERTurk and zero-shot large language models show that context-aware BERTurk achieves the best performance, with 0.76 accuracy and balanced class-wise F1 scores.
Second, the thesis introduces SarcasTraps, an evaluation set for analyzing how sarcasm affects Turkish hate speech detection. SarcasTraps contains naturally occurring Ekşi Sözlük entries annotated for both hate speech and sarcasm. Encoder classifiers trained on existing Turkish hate speech datasets and zero-shot large language models are evaluated on sarcastic and non-sarcastic examples. Results show that fine-tuned encoders lose recall on sarcastic split, while zero-shot large language models maintain higher recall but suffer precision drops on sarcastic examples. Overall, the thesis shows that Turkish sarcasm is both a context-dependent linguistic phenomenon and a reliability challenge for intent-sensitive NLP systems.