search

gdscewha-3rd / Study-PaperReview

🤖 인공지능 논문읽기 스터디
6 stars 0 forks source link

Seq2SQL : Generating structured queries from natural language using reinforcement learning #26

Closed m0oon0 closed 2 years ago

m0oon0 commented 2 years ago

Model > Seq2SQL : DNN for translating Natural Language Questions to SQL Queries Dataset > WikiSQL : 80654 questions, SQL queries corpus from wikipedia.

  1. aggregation operator classifier Compute attention score over aggregation operations (COUNT, MIN, MAX, ...) for input tokens.

  2. SELECT column pointer Produce the representations for the columns using LSTM. Compute attention score for each column and make a distribution over possible columns. 1,2 can be trained using cross entropy loss.

  3. WHERE clause generator Apply reinforcement learning
    Instead of teacher forcing at each step of query generation, sample from the output distribution to obtain the next token.
    Execute the generated SQL query against the database to obtain a reward. (compute by the result)

minha62 commented 2 years ago

쿼리를 생성하는 자연어처리도 있군요! 흥미로운 주제네요! 좋은 내용 감사합니다:)

SEOJIN-Lab commented 2 years ago

SQL로 자연어 처리를 할 수 있다니...! 좋은 논문 공유 감사합니다:>

Si-jeong commented 2 years ago

잘 읽었습니다! :)