DeepCS: a Code Search Tool Powered by Deep Learning
summary
DeepCS builds a joint embedding that maps English queries and Java methods into the same vector space using a CODE-NN architecture, API‑call and name token embeddings, and k‑nearest code vectors search. The team trains on 18 million Java methods from GitHub, indexes unseen repos, and benchmarks against Lucene and CodeHow on the top 50 Stack Overflow Java questions and ~10,000 Java projects; DeepCS returns the first relevant result much earlier and a higher top‑N relevance rate.
AI-written summary · about
chapters (10)
- 0:00Introduction to DeepCS
- 2:18Problem of code search
- 4:27Vector representations of text
- 10:49Word2Vec and document embeddings
- 14:21Vector representations of code
- 19:59Combining text and code signals
- 21:46Joint embedding training
- 23:52DeepCS evaluation results
- 25:14Summary and acknowledgments
- 25:35Q&A session
related talks
01h 01mDeep Learning and Java
37mMachine Learning on Source Code
16mMachine Learning on Source Code
43mMachine Learning for Developer Productivity
22mThe Capabilities of Semantic Search
28mDelivering AI on Code: Live Demo of source{d}
46mCode + AI: Will Robots Take Our Coding Jobs? ML Applied to Programming
27mTeaching a Machine to Code