Bilkent University

Ankara, Turkey

Bilkent University

Ankara, Turkey

The project focuses on the real-world enterprise scenarios, where models need to specialize to a single database schema and organizations require to be able to evaluate their Text-to-SQL systems on their own databases. Proposed a fully automated two-stage framework for generating high-quality synthetic Text-to-SQL data tailored to any database without SQL logs or manual annotations. The method hierarchically partitions schemas into sub-schemas, synthesizes multi-level SQL queries, and applies LLM-as-a-judge validation, executability checks, and column balancing for high coverage. Fine-tuned models trained on this data achieved state-of-the-art in-domain performance, outperforming the best baseline (comparable scale) by +16.21%.

Developed an information retrieval system to compare sparse and dense (embedding-based) retrieval methods on the multilingual Istella22 dataset. Implemented sparse retrieval using PyLucene and BM25 scoring, and dense retrieval via multilingual transformer-based embeddings indexed with HNSW for approximate nearest neighbor search. Evaluated multiple models and embedding aggregation strategies (first-passage vs. average) under various dataset scales. Additionally, implemented the HDCT framework using fine-tuned BERT for context-aware term weighting. Assessed performance using NDCG, MAP, and MRR, highlighting trade-offs between retrieval quality and computational efficiency.

Developing a challenging NLP application utilizing large language models (LLMs) for translating natural language queries into SQL statements (i.e., NL2SQL). This approach significantly lowers the barriers to accessing relational databases by parsing and understanding the structure of natural language to accurately generate corresponding SQL queries.

An NLP application that employs weak-supervision techniques to identify the relation between natural language query tokens and database entities, treating the keyword mapping problem as a sequence tagging problem using neural network.