Of Rules and Probabilities: Computational Linguistics Methods
summary
Builds a title grammar from ~100k profile titles to auto-label short search queries, induce a skills dictionary, add grammar-derived features to a CRF and apply rule-based post‑processing. Queries under four words are 98% of traffic so the grammar focuses there. Skills F1 rises from 63% by about 13 percentage points after augmented training data, features and post-filtering, post-processing adds ~2 points, and rare long queries climb from 65% to about 70% with targeted long‑tail examples.
AI-written summary · about
chapters (11)
- 0:00Introduction to talk
- 1:21Problem with labeled data
- 7:04Building title grammar
- 12:19Grammar construction steps
- 14:50Inferring skills from titles
- 17:38Labeling queries with grammar
- 19:23Feature engineering with grammar
- 20:50Results on frequent queries
- 21:54Results on rare long queries
- 22:36Conclusion and applications
- 25:55Q&A
related talks
01h 00mHuman and machine inductive biases for compositional linguistic generalization
36mEnhanced and Efficient Reasoning in Large Language Models
15mAccelerating Rule-matching Systems with Learned Rankers
01h 16mGeneralization in natural language processing
37mCreating Knowledgebases from unstructured text
42mAdvancements in Evaluating Large Language Model Applications
38mPlaying by the Rules-Based-Systems
41mGetting Started with Query Understanding #bbuzz