{"id":"arxiv:2112.07899","title":"Large Dual Encoders Are Generalizable Retrievers","date":"2021-12-15","section":"Retrieval, embeddings, benchmarks","quality":0.297587,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/","source_url":"https://arxiv.org/abs/2112.07899","text_available":true,"authors":["Jianmo Ni","Chen Qu","Jing Lu","Zhuyun Dai"],"outgoing":[],"incoming":[{"source":"arxiv:2212.09741","target":"arxiv:2112.07899","evidence":"id","id":"arxiv:2212.09741","title":"One Embedder, Any Task: Instruction-Finetuned Text Embeddings","url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/b162d4a59ad6b25fd180e566615cf3dc2cd3a6c505f0421e15fd036236ad2ecd/"}],"text_parts":[{"start":0,"end":15694,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html"},{"start":15694,"end":29748,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html"}],"contents":[{"title":"Large Dual Encoders Are Generalizable Retrievers","offset":0,"level":1,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"Abstract","offset":52,"level":6,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"1 Introduction","offset":1232,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"2 Background","offset":4652,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"2.1 Dual Encoder and dense retrieval","offset":4669,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"2.2 BEIR generalization task","offset":7010,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"3 Generalizable T5 Retriever","offset":7767,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"3.1 T5 dual encoder","offset":7800,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"3.2 Multi-stage training","offset":9958,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"4 Experimental setup","offset":11424,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"4.1 Training Data","offset":11449,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"Community QA.","offset":11472,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"MS Marco.","offset":11784,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"Natural Questions.","offset":12367,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"4.2 Configurations","offset":12684,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"4.3 Models for comparison","offset":14361,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"5 Evaluation Results","offset":15442,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"5.1 Results on MS Marco","offset":15665,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-1.html#text"},{"title":"5.2 Results on BEIR generalization tasks","offset":16097,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"5.3 Data efficiency for large retrievers","offset":18045,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"6 Ablation Study and Analysis","offset":19047,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"6.1 Effect of scaling up for different training stages","offset":19295,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"6.2 Importance of the fine-tuning dataset","offset":19954,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"6.3 Document length vs model capacity","offset":21474,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"7 Related Work","offset":22940,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"Neural information retrieval.","offset":22959,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"Dual encoders for neural retrieval.","offset":23744,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"Zero-shot neural retrieval.","offset":24487,"level":4,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"8 Inference latency","offset":25304,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"9 Conclusion","offset":25975,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"Acknowledgments","offset":26887,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"Appendix A More results","offset":26973,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"A.1 Comparisons on MS Marco","offset":27001,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"A.2 Comparison of different dual encoder pre-training strategies","offset":27310,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"A.3 Recall on BEIR","offset":28565,"level":3,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"},{"title":"Instructions for reporting errors","offset":28784,"level":2,"url":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/part-2.html#text"}],"text_code_points":29748,"source_markdown":"https://vinnik-dmitry07.github.io/favorite-papers/agents/papers/a98e608b4e34094fed75a118c0663f5f02f485081981cecc2bf3350181d61a51/source.md"}