The Neural Search Frontier #bbuzz
summary
Claims that neural search is a silver bullet give way to measured results: experiments on an ACM corpus of 3,000 documents compare Lucene/BM25 and TF‑IDF to word2vec and paragraph vectors (doc2vec) via Anserini/listen for IR. Classic models beat embeddings when retrieving top‑100; when restricting to top‑10 and mixing features (embeddings + language model + BM25) performance improves. A Flink demo continuously indexes tweets and shows document embeddings adapt faster than a Dirichlet language model. Conclusion: embeddings help as LTR features but need enough data.
AI-written summary · about
chapters (28)
- 0:00Intro
- 1:03Deep Learning and Search
- 10:30Lowhanging fruits
- 13:10The problem for everything
- 13:30The Tyvek model
- 19:35The lowhanging fruit
- 22:55The frontier
- 24:50Pretraining
- 25:15Average query vectors
- 25:39Learning to rank
- 26:30Language models
- 27:21Markov language model
- 28:07Pizza
- 28:34Embedding
- 29:33Context
- 30:04Old embedding
- 32:46Not a silver bullet
- 33:09Sessions documents
- 33:28Neural Translation
- 35:14Grading judgments
- 35:49Skip thought vectors
- 36:37The Neural Search Frontier
- 37:15The Core Idea
- 37:45The Demo
- 38:02Tools Used
- 39:13Results
- 42:16Demo
- 45:02Tokenizing Text
related talks
13mA Beginner’s Guide to Semantic Search
30mSemantic Search and Neural Information Retrieval
39mAI-powered Semantic Search; A story of broken promises?
01h 30mBuilding a Search Engine
42mResearch in the Age of AI
31mTaking AI/ML to the next level: Text!
46mNatural Language & Search, Improve your Search w/ Vector capabilities
38mSearch: A New Era