Learning to rank with the Transformer
summary
Applies the Transformer encoder with self‑attention to score whole item lists and reports measurable NDCG gains: on Web30k (30,000 queries) the model beats comparable MLPs and XGBoost at NDCG@5, and on an in‑house 1M‑query binary dataset weighted RankNet/LambdaRank variants perform best where pointwise losses fail. Releases the code as allRank implemented in PyTorch and runs offline simulations showing uplift; neural ranker is being prepared for production deployment.
AI-written summary · about
chapters (12)
- 0:00Learning to rank with Transformers
- 1:54Data sources
- 3:43Scoring functions & metrics
- 7:39Transformer basics
- 9:27Self-attention
- 13:04Transformers for ranking
- 14:46Cascade ranking trade-offs
- 16:33Datasets & evaluation
- 19:48Results on Web30k
- 20:49OL Rank framework
- 22:14Future research
- 23:37Q&A
related talks
30mJourney Towards Deep Learning for Ranking
29mFrom Trees to Transformers: Our Journey Towards Deep Learning for Ranking
24mLearning "Learning to Rank"
29mBuilding Learning to Rank models for search using LLMs
25mBetter search relevance using Learning to Rank at mobile.de
55mLearning to Rank, a Little
23mUnveiling the Journey of NLP
38mUnderstanding and Applying Self-Attention for NLP