DPR Dense Passage Retrieval (Karpukhin 2020年)（ディーピーアール）
2020年Karpukhin et al. (Meta FAIR)発表DPR・Industry-leading dense passage retrieval LLM RAG + Industry-leading dual-encoder + Industry-leading dense vector embedding + Industry-leading open-domain QA pioneer。
概要
DPR Dense Passage Retrieval は、2020年Karpukhin et al. (Meta FAIR)発表DPR・Industry-leading dense passage retrieval LLM RAG 2020年 + Industry-leading open-domain QA pioneer position確立。DPR specifications = Industry-leading dense passage retrieval LLM RAG (Industry-leading DPR 2020 + Industry-leading dense passage retrieval + Industry-leading DPR flagship) + Industry-leading dual-encoder (Industry-leading dual-encoder query+passage architecture) + Industry-leading dense vector embedding (Industry-leading dense vector embedding 768-dim BERT-based) + Industry-leading open-domain QA pioneer (Industry-leading open-domain QA pioneer)。
主な特徴・仕組み
- Authors: Industry-leading Vladimir Karpukhin + Barlas Oguz + Sewon Min + Patrick Lewis + Ledell Wu + Sergey Edunov + Danqi Chen + Wen-tau Yih Meta FAIR + UW + Princeton
- Year: 2020年 (arXiv 2020年4月発表 + EMNLP 2020)
- Paper: Industry-leading "Dense Passage Retrieval for Open-Domain Question Answering" arXiv 2004.04906
- Method Type: Industry-leading dense passage retrieval LLM RAG
- Architecture: Industry-leading dual-encoder (query encoder + passage encoder separately trained)
- Embedding: Industry-leading dense vector embedding 768-dim BERT-based
- Training: Industry-leading contrastive learning (positive vs in-batch negatives)
- Open-Domain QA: Industry-leading open-domain QA Wikipedia retrieval
- Speed: Industry-leading FAISS index 10ms retrieval over millions
- Performance: Industry-leading 9-19% absolute improvement over BM25 on QA benchmarks
- Open Source: Industry-leading DPR GitHub facebookresearch/DPR
- Industry Adoption: Industry-leading RAG ecosystem foundation (HuggingFace + LangChain)
- Industry-Leading: Industry-leading dense passage retrieval LLM RAG first pioneer
スペック比較表
| LLM RAG Method | Year | Author/Org | Key Feature | Industry Position |
|---|---|---|---|---|
| DPR | 2020 | Karpukhin Meta FAIR | Dual-encoder dense passage retrieval | Industry-leading RAG pioneer |
| ColBERT | 2020 | Khattab+Zaharia Stanford | Late interaction token-level | Industry-leading late interaction |
| RETRO | 2022 | DeepMind | Retrieval-enhanced transformer pre-training | Industry-leading DeepMind pretrained retrieval |
| Atlas | 2022 | Meta FAIR | Few-shot RAG with continuous retrieval | Industry-leading few-shot RAG |
| Self-RAG | 2023 | Asai UW | Self-reflective retrieval decision | Industry-leading self-reflective |
具体例・対応製品
- DPR (2020年, Karpukhin et al. Meta FAIR): Industry-leading dense passage retrieval
- Industry-leading dual-encoder query+passage architecture: Industry-leading dual-encoder
- Industry-leading dense vector embedding 768-dim BERT-based: Industry-leading 768-dim BERT
- Industry-leading contrastive learning positive vs in-batch negatives: Industry-leading contrastive
- Industry-leading 9-19% absolute improvement over BM25: Industry-leading 9-19% vs BM25
- 競合 ColBERT + RETRO + Atlas + Self-RAG: Industry-leading RAG competitors
自作PCでの選び方・注意点
DPR Dense Passage Retrieval は「Industry-leading dense passage retrieval LLM RAG + Industry-leading dual-encoder」「Industry-leading dense vector embedding + Industry-leading open-domain QA pioneer」用途のIndustry-leading dense passage retrieval 2020年Karpukhin et al. Meta FAIR発表product。Industry-leading dual-encoder (Industry-leading dual-encoder query encoder + passage encoder separately trained + Industry-leading DPR dual-encoder signature) で Industry-leading dual-encoder + Industry-leading query+passage separately + Industry-leading DPR signature。Industry-leading dense vector (Industry-leading dense vector embedding 768-dim BERT-based + Industry-leading dense vs sparse BM25 + Industry-leading DPR dense embedding signature) で Industry-leading 768-dim BERT + Industry-leading dense vs BM25 + Industry-leading DPR dense signature。Industry-leading contrastive (Industry-leading contrastive learning positive vs in-batch negatives + Industry-leading DPR training signature + Industry-leading in-batch negatives efficient) で Industry-leading contrastive + Industry-leading DPR training + Industry-leading in-batch negatives。Industry-leading 9-19% vs BM25 (Industry-leading 9-19% absolute improvement over BM25 on QA benchmarks + Industry-leading FAISS index 10ms retrieval over millions + Industry-leading open-domain QA Wikipedia retrieval) で Industry-leading 9-19% improvement + Industry-leading FAISS 10ms + Industry-leading open-domain QA。Industry-leading Meta FAIR open source (Industry-leading Vladimir Karpukhin Meta FAIR authors + Industry-leading DPR GitHub facebookresearch/DPR + Industry-leading RAG ecosystem foundation HuggingFace+LangChain + Industry-leading dense passage retrieval LLM RAG first pioneer) で Industry-leading Karpukhin Meta FAIR + Industry-leading facebookresearch/DPR + Industry-leading RAG foundation + Industry-leading first pioneer。但しIndustry-leading ColBERT + RETRO + Atlas + Self-RAG competition (Industry-leading ColBERT late interaction 2020 + RETRO DeepMind pretrained 2022 + Atlas Meta few-shot RAG 2022 + Self-RAG Asai self-reflective 2023 vs DPR Meta dual-encoder 2020 trade-off) で Industry-leading 4-RAG method competitors vs DPR + Industry-leading dual-encoder + dense embedding 768-dim BERT + contrastive in-batch negatives + 9-19% vs BM25 + open-domain QA + FAISS index + Meta FAIR pioneer 2020 + RAG ecosystem foundation unique advantage adoption alignment必須。
関連用語との違い
- vs ColBERT (Stanford 2020): DPRはdual-encoder + 768-dim単一ベクトル・ColBERTはlate interaction + token-level multi-vector
- vs RETRO (DeepMind 2022): DPRはretrieval post-hoc + open-domain QA・RETROはretrieval-enhanced pre-training transformer
- vs Atlas/Self-RAG (Meta/UW 2022-2023): DPRはfoundation pioneer + 2020・Atlas/Self-RAGはfew-shot/self-reflective進化形
よくある質問（FAQ）
Q1: DPR vs ColBERT 違いは？ A: DPR (Industry-leading dual-encoder query+passage architecture + dense vector embedding 768-dim BERT-based + contrastive learning positive vs in-batch negatives + FAISS index 10ms retrieval + open-domain QA Wikipedia + 9-19% vs BM25 + Meta FAIR + Karpukhin et al. 2020) vs ColBERT (Industry-leading late interaction architecture + token-level multi-vector + MaxSim scoring + 100x faster than cross-encoder + open-domain QA + Stanford + Khattab+Zaharia 2020)・Industry-leading dual-encoder + 768-dim単一 + contrastive + FAISS + Meta = DPR + Industry-leading late interaction + token-level + MaxSim + 100x faster + Stanford = ColBERT preference judgment。
Q2: Industry-leading dual-encoder + dense embedding value は？ A: Industry-leading dual-encoder + dense (Industry-leading dual-encoder query encoder + passage encoder separately trained + Industry-leading dense vector embedding 768-dim BERT-based + Industry-leading contrastive learning positive vs in-batch negatives + Industry-leading DPR architecture signature)。
Q3: Industry-leading open-domain QA pioneer + 9-19% vs BM25 value は？ A: Industry-leading open-domain QA + 9-19% (Industry-leading open-domain QA Wikipedia retrieval + Industry-leading 9-19% absolute improvement over BM25 on QA benchmarks + Industry-leading FAISS index 10ms retrieval over millions + Industry-leading dense passage retrieval LLM RAG first pioneer)。
まとめ
DPR = 2020年Karpukhin et al. Meta FAIR発表のdense passage retrieval LLM RAG。Industry-leading dual-encoder query encoder + passage encoder separately trained + Industry-leading dense vector embedding 768-dim BERT-based + Industry-leading contrastive learning positive vs in-batch negatives + Industry-leading FAISS index 10ms retrieval over millions + Industry-leading 9-19% absolute improvement over BM25 on QA benchmarks + Industry-leading open-domain QA Wikipedia retrieval + Industry-leading Vladimir Karpukhin + Barlas Oguz + Sewon Min + Patrick Lewis Meta FAIR + UW + Princeton + Industry-leading DPR GitHub facebookresearch/DPR open source + Industry-leading RAG ecosystem foundation HuggingFace + LangChain + Industry-leading dense passage retrieval LLM RAG first pioneer 2020 position確立。