Search & Ranking

Definitive reference unifying the RankNet family; LambdaMART remains the industry workhorse for gradient-boosted ranking.

Foundational pairwise neural ranking using cross-entropy loss; won ICML Test of Time Award 2015.

First listwise method using probability distributions over permutations; influenced all subsequent listwise methods.

Foundational listwise LTR with theoretical analysis of listwise loss functions.

Seminal SVM-based pairwise LTR using click data for preference constraints; established click-based learning to rank paradigm.

Seminal taxonomy of query intent (navigational, informational, transactional); foundational framework still used today.

Foundational work on mapping queries to topic categories using intermediary data sources.

Industrial-scale BERT-based intent classification for typeahead and search blending.

Foundational pseudo-relevance feedback method introducing RM3; widely used for query expansion in both traditional and neural retrieval.

Pioneered use of session context for query interpretation; demonstrated significant gains from implicit user feedback.

Modern neural query reformulation using GPT-2 for conversational settings; addresses context carryover in multi-turn search.

Landmark paper on watch-time optimization vs. clicks; explains freshness handling at scale.

Multi-objective ranking balancing engagement with satisfaction using multi-gate mixture-of-experts.

Influential industry paper on balancing business metrics vs. user value in production systems.

Demonstrates empirically that engagement-based ranking underperforms for user satisfaction.

Established context-dependent satisfaction thresholds beyond fixed dwell time cutoffs; showed satisfaction prediction requires query-document context.

Demonstrated satisfaction signals exist beyond clicks; query reformulation patterns predict user satisfaction better than clicks alone.

Seminal empirical study establishing cascade model as best explanation for user examination behavior.

Foundational counterfactual/IPS framework for unbiased LTR; propensity-weighted ranking SVM.

Comprehensive synthesis covering all major click models (PBM, DCM, DBN, UBM); essential reference.

Dual learning algorithm jointly training ranking and propensity models; widely used for production debiasing.

Practical industry case on position bias correction via dropout at inference; shows major production gains.

First rigorous treatment of trust bias in ULTR framework; extends counterfactual work to account for users trusting higher-ranked results more.

Foundational dual-encoder dense retrieval; outperforms BM25 by 9-19%; established the dense retrieval paradigm.

Introduced late interaction achieving near cross-encoder quality with bi-encoder speed.

Transformative demonstration of BERT for passage re-ranking; 27% improvement on MS MARCO.

Best Paper Award. Production embedding system driving 99% of Airbnb conversions.

Made BERT practical for dense retrieval with siamese architecture; widely used for semantic similarity.

Learned sparse representations via MLM head with sparsity regularization; bridges gap between dense and sparse retrieval with interpretable term weights.

Critical training strategies for dense retrieval: cross-batch negatives, denoised hard negatives, knowledge distillation from cross-encoder. Baidu research.

Global hard negative mining via ANN index (ANCE); addresses limitation of in-batch negatives by refreshing hard negatives from evolving model.

Established BERT-based cross-encoder reranking paradigm with MonoBERT and duoBERT; 'Expando-Mono-Duo' design pattern for neural IR pipelines.

Neural document expansion via T5-generated queries; improves BM25 without runtime overhead by enriching documents at indexing time.

Distinguished 'relevance matching' from 'semantic matching' in neural IR; histogram-based matching with term gating mechanism.

Pioneered end-to-end neural ranking with interpretable soft-match kernels (K-NRM); spawned KNRM variants used in production.

Pre-training specifically for dense retrieval via Condenser architecture; coCondenser adds corpus-aware contrastive objective for further gains.

First end-to-end pre-training of retrieval + LM with backpropagation through retrieval; foundational architecture for knowledge-intensive tasks.

THE paper that coined 'RAG'; combines pre-trained retriever with seq2seq generator for open-domain QA. Foundation for modern RAG systems.

Fusion-in-Decoder (FiD) architecture enabling efficient scaling to 100+ passages; backbone of Atlas and subsequent RAG systems.

Generative retrieval paradigm (DSI): model memorizes corpus and generates document IDs directly; alternative to dense/sparse retrieve-then-rank.

Introduced DCG and NDCG; most widely used ranking metric enabling graded relevance evaluation.

Cascade-based metric modeling user stopping behavior; primary TREC Web Track metric accounting for diminishing returns.

Large-scale passage ranking benchmark with 1M queries; enabled neural retrieval research and remains primary leaderboard.

18-dataset benchmark testing out-of-distribution generalization; key finding that BM25 remains robust while dense models struggle on domain shift.

Definitive validation of interleaving as online A/B testing gold standard; demonstrated high agreement with editorial judgments at scale.

Foundational personalization paper establishing paradigm for implicit user modeling from desktop and search history.

Established distinction between session-level and historical personalization signals; showed combination outperforms either alone.

Operationalized session and context signals in production LTR frameworks; showed how to integrate behavioral context into ranking features.

Defining benchmark for conversational IR with 80 dialogues over 38M passages; established evaluation methodology for multi-turn search.

Open-domain retrieval for conversational QA; advances beyond simplified settings with ORConvQA dataset and baselines.

THE seminal diversification paper introducing Maximal Marginal Relevance (MMR) formula; balances relevance with novelty.

Standard evaluation framework for diversity introducing α-nDCG; enabled TREC Web diversity track and systematic diversity research.

Intent-aware diversification with formal coverage guarantees; models query as distribution over intents and optimizes expected coverage.