Multi-modal Program Inference: a Marriage of Pre-trained Language Models and Component-based Synthes
summary
NLX combines GPT-3 candidates with a component-based synthesizer to turn ambiguous English plus examples into precise code. It targets regexes and CSS selectors, extracts frequent nonredundant seed components, uses beam search biased by PTM operator frequencies and semantic condensation to keep semantic diversity, and ranks consistent programs by Euclidean operator vectors and Levenshtein distance. GPT-3 alone hits ~15% on benchmarks vs a ~60% specialist; NLX improves over all baselines by at least 20% on 150 tasks.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:13PRE-TRAINED NATURAL LANGUAGE MODELS (PTM)
- 2:39FIRST HAND EXPERIMENTS WITH (NL→CODE)
- 4:45NOT THE END OF THE STORY!
- 5:33NLX PROGRAM SYNTHESIS FRAMEWORK
- 6:48COMPONENT-BASED SYNTHESIS (CBS)
- 7:57CHALLENGES WITH SEARCH
- 8:41SEED COMPONENTS
- 10:44ITERATIVE EXPANSION
- 12:39FINAL RANKING
- 13:45EXPERIMENTAL EVALUATION
- 14:47RECAP
related talks
43mKeynote: Applied NLP in the age of Generative AI
23mPanel
51mLarge Language Models for Real-World Applications - A Gentle Intro
43mSupercharging your transformers with synthetic query generation and lexical search
01h 00mUnleashGPT: Mastering the Implementation of Large Language Models
01h 50mChatGPT & LLMs: A Practical Starter's Guide to Language Chains
01h 02mA hacker's guide to open source LLMs
12mRevolutionizing Cloud-Native Applications with NLP