My Journey in Text2SQL Research: From Paper Reading to Winning a Global Competition

Last May, while traveling in Luoyang and having already received an offer from Tencent, the author was called back by their supervisor to resume research in June, abandoning the planned internship.

The research focus was Text2SQL—translating natural language questions into SQL queries. Initially the author had only a vague idea, limited code experience, and had read fewer than ten papers.

After returning to the lab at the end of June, the author spent over half a year deepening their understanding of Text2SQL, completing a thesis, publishing a paper, filing a patent, and achieving second place in Yale University's global Text2SQL competition in October.

The experience is summarized in three aspects:

1. Reading Recent Top‑Conference Papers (3‑5 Years)

Systematic literature review is essential to grasp the field’s landscape, avoid duplicated ideas, and inspire new concepts. Efficient paper collection methods include: (1) studying top solutions from public competitions or leaderboards (e.g., WikiSQL, TableQA, Spider, CoSQL); (2) gathering 2‑3 survey papers; (3) searching Google Scholar with keywords and filtering by citations and venue; (4) exploring curated GitHub repositories such as https://github.com/yechens/NL2SQL that compile background, papers, datasets, and solutions.

2. Standing on the Shoulders of Giants to Strengthen Engineering Skills

After gaining academic insight, the author quickly implemented ideas by referencing state‑of‑the‑art (SOTA) solutions rather than building everything from scratch. For Text2SQL, data preprocessing is extensive, so reusing proven pipelines allowed focus on model design and post‑processing. The author recommends deep‑learning books like "Deep Learning with Python" by the Keras creator and "Dive into Deep Learning" by Li Mu.

3. Fully Utilizing School and Lab Resources

Resources include academic (senior lab members and the supervisor), human (collaborating with peers who have complementary strengths), and hardware (servers with Tesla V100 GPUs, 24‑hour lab access, and other equipment). Effective communication with supervisors and leveraging available infrastructure are crucial.

The author concludes with a personal productivity strategy—setting deadlines for literature review, coding, and iteration—and recommends several useful tools for AI research: arXiv, PaperwithCode, DBLP, Connected Paper, NLPIndex, DeepL, and diagrams.net.