Semantic Retrieval in Composite Data Structures: Fine-Tuning with LLM-Assisted Synthetic Data Bileşik Veri Yapilarinda Anlamsal Çikarim: Büyük Dil Model Destekli Sentetik Veri ile ?Ince Ayar


Donmezbilek H. I., Duruoglu G., Gursoy H., TORAMAN Ç.

34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)

  • Yayın Türü: Bildiri / Tam Metin Bildiri
  • Doi Numarası: 10.1109/siu71813.2026.11636863
  • Basıldığı Şehir: İstanbul
  • Basıldığı Ülke: Türkiye
  • Anahtar Kelimeler: dense retrieval, hard negative mining, modality gap, semantic search, structured data, synthetic data
  • Orta Doğu Teknik Üniversitesi Adresli: Evet

Özet

This study addresses information retrieval challenges in composite data structures such as Network Equipment Identifier (NE ID) in the X-Y-Z format. The modality gap between natural language queries and technical codes, combined with high character-level similarity among records (hard negatives), causes fundamental problems in search system. To overcome these issues, various models are fine-tuned using LLM-generated synthetic queries and hard negative mining, yielding notable improvements in semantic matching over their baseline counterparts.