Word Embeddings - An Alternative & Efficient Approach to Search for Docs
summary
Argues that using fastText subword embeddings (n-grams) yields dense vectors in minutes of training and handles unseen tokens and spelling variants (demo: financial‑crime name variations), then stores vectors in dense_vector in Elasticsearch (7.3+) and ranks with cosine similarity at query time. Pipeline converts documents and queries to vectors, filters first, then scores; it improves recall on fuzzy matches and speeds deployment versus deep nets, but performs worse than TF/IDF for exact matches and currently scores after filtering.
AI-written summary · about
chapters (9)
- 0:00Introduction to talk
- 1:03Elasticsearch basics
- 10:05Traditional search vs embeddings
- 20:16What are word embeddings
- 31:55Generating embeddings with FastText
- 41:55FastText in action
- 48:37Building search pipelines
- 56:42Advanced embedding use cases
- 1:00:14Future of embeddings in search
related talks
34mWord Embeddings for fun and profit in Gensim
26mImplementing a Search Index
39mNext generation of word embeddings in Gensim
16mDense Vectors in Elasticsearch
38mSearch: A New Era
31mAdvanced search for your e-commerce website
13mA Beginner’s Guide to Semantic Search
31mTaking AI/ML to the next level: Text!