Sergey Feldman

Senior Applied Research Scientist at Allen Institute for AI (AI2). Built the ML reranker for Semantic Scholar — a 190M-paper academic search engine — taking evaluation score from 0.7 to 0.93 and delivering +8% clicks per query.

Articles

• Building a Better Search Engine for Semantic Scholar — LightGBM reranker on Elasticsearch, data filtering as key lever

Key Contributions

• Training data filtering: removing ~1/3 of training data via “sensibility checks” yielded 10-15% evaluation improvement — data quality > data quantity
• Custom evaluation metric: 250-query manually-annotated benchmark with component-level criteria; more predictive than generic NDCG
• Rules + ML: post-hoc rule corrections (exact year/phrase boosts) on top of learned model
• Result: 0.7 → 0.93 custom metric, +8% clicks per query, +9% MRR

Related

• Semantic Scholar (AI2) · Learning to Rank · Judgment Lists · NDCG