How can I finetune on CSpider about resdsql

[lihaoyang-ruc]

This is an interesting issue.

Currently, our method is fine-tuned for English RoBERTa and T5, since the datasets used in our paper are in English. In fact, RESDSQL is a general method and is not limited by language. Therefore, we also want to know if RESDSQL can achieve promising performance on multilingual Text-to-SQL benchmarks such as CSpider.

You can replace the original English RoBERTa and T5 with multilingual RoBERTa and T5 that support both Chinese and English, and then fine-tune them.

Although it sounds easy, there are still potential problems that may be encountered during implementation. Good luck!

[wikty]

Thank you very much, that sounds very reasonable!

Training on Chinese-English bilingual datasets has great significance. For example, in a real industry scene our user input is in Chinese, but the database schema is in English. I'll try your proposal idea.

I still have a little concern about preprocessing.py, does it bind with English?

[lihaoyang-ruc]

There are two potential problems.

• First, see lines 353-360 in preprocessing.py, where we follow PICARD and use string matching to retrieve some useful values from the database based on the given English question. This step facilitates the generation of predicates in SQL queries (e.g., WHERE country = 'France'). However, this function may not work when you input Chinese questions.
• Second, since the most powerful model RESDSQL-3B+NatSQL uses NatSQL (an intermediate representation of SQL aligned to the Spider dataset) as the seq2seq training target, I do not know if the NatSQL dataset can also be aligned to the CSpider dataset. If not, you can only fine-tune a SQL version of RESDSQL, i.e., RESDSQL-{Base, Large, 3B} in our paper.

[wikty]

Thanks a lot, I'll try it out.