The TWIML AI Podcast (formerly This Week in Machine Learning & Artificial Intelligence) Generating SQL Database Queries from Natural Language with Yanshuai Cao - #519
17 snips
Sep 16, 2021 
Yanshuai Cao, a Senior Research Team Lead at Borealis AI, discusses his groundbreaking work on Turing, an engine transforming natural language into SQL queries. He compares it with OpenAI's Codex, highlighting the unique challenges of SQL generation. The conversation reveals insights into the crucial role of reasoning and common sense in accurate query creation. They also tackle complexities in multilingual datasets, data augmentation, and the ongoing quest for model explainability, shedding light on fascinating advancements in AI technology.
AI Snips
Chapters
Transcript
Episode notes
Reasoning in Turing
- Reasoning in Turing involves leveraging schema information and common sense.
- It allows the model to resolve ambiguities, like inferring age from "under 30" or joining unrelated tables.
Reasoning Implementation
- Turing uses relational-aware transformers, which incorporate structural priors about the problem.
- These transformers leverage SQL grammar and schema relations to prune the search space and refine links.
Training Relational Layers
- Train relational-aware transformer layers with initial heuristics-based links.
- The model can refine and correct these initial links during training.