🧑‍🔬 Zifan Li

Zifan Li works on semantic parsing, in particular the text-to-SQL problem: turning a question written in ordinary language into an SQL query that can be run against a database. A recurring concern in the papers here is generalisation. Systems should handle databases and tables they never saw during training, rather than succeeding by memorising query patterns. The work combines building carefully annotated benchmarks with designing neural models that read both the question and the database schema.

The most cited of these papers is Spider. It introduced a large, manually written corpus that pairs questions with complex SQL over many multi-table databases. The queries cover joins, grouping, ordering, nesting and set operations. Spider also set up an evaluation in which the test databases are kept apart from those used for training, so that models are tested on domains new to them. The second paper, TypeSQL, addresses SQL generation on WikiSQL. It builds on the sketch-and-slot-filling approach of SQLNet and splits the task across models with shared encoders. It also labels each question word with a type, such as a column match, a number or date, or a named entity, to cope with rare entities and values that pre-trained word embeddings capture poorly.

Both papers date from 2018 and come from work associated with Yale University. Co-authors include Tao Yu, Rui Zhang, Dragomir Radev, Zilin Zhang, Irene Li and Michihiro Yasunaga. The papers sit within semantic parsing and meaning representation and code and programming languages.

Zifan Li 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.