🧑‍🔬 Zilin Zhang

Zilin Zhang works on text-to-SQL generation and semantic parsing, where a question in natural language must be turned into an executable database query. Their papers here sit where semantic parsing meets code and programming languages. The recurring concern is how well models generalize to databases and tables they never saw during training. The approach combines building evaluation resources with designing neural models.

The most cited of these papers is Spider (2018). Earlier benchmarks either had complex queries over a single small database or covered many tables with only simple queries. Spider responds with a hand-annotated corpus that pairs questions with complex SQL over many multi-table databases, using joins, grouping, nesting and set operations. It also sets up a task in which the test databases are kept apart from the training ones. TypeSQL, from the same year, takes up text-to-SQL on WikiSQL. It builds on the sketch-based slot-filling design of SQLNet, splitting prediction across models with shared encoders. It also gives each question token a type, such as a column match, a number or date, or a named entity, to cope with rare words that pretrained embeddings handle poorly.

Both papers date from 2018, so neither marks a later shift in direction. Together they show an interest in cross-domain evaluation and in using schema and type knowledge to guide query generation. Zhang is currently at Zhejiang University. Earlier bylines include Yale University, where the 2018 work appeared, and Beijing Institute of Technology. Frequent co-authors are Tao Yu, Dragomir Radev, Rui Zhang and Zifan Li.

Zilin Zhang has 2 papers in our library, in 2018, most of them on Code & programming languages, Semantic parsing & meaning representation; often with Dragomir Radev, Rui Zhang, Tao Yu.