One Embedder, Any Task: Instruction-Finetuned Text Embeddings. Su, H., Shi, W., Kasai, J., Wang, Y., Hu, Y., Ostendorf, M., Yih, W., Smith, N. A., Zettlemoyer, L., & Yu, T. May, 2023. 74 citations (Semantic Scholar/arXiv) [2024-02-01] 74 citations (Semantic Scholar/DOI) [2024-02-01] arXiv:2212.09741 [cs]![One Embedder, Any Task: Instruction-Finetuned Text Embeddings [link]](https://bibbase.org/img/filetypes/link.svg "link")
Paper doi abstract bibtex We introduce INSTRUCTOR, a new method for computing text embeddings given task instructions: every text input is embedded together with instructions explaining the use case (e.g., task and domain descriptions). Unlike encoders from prior work that are more specialized, INSTRUCTOR is a single embedder that can generate text embeddings tailored to different downstream tasks and domains, without any further training. We first annotate instructions for 330 diverse tasks and train INSTRUCTOR on this multitask mixture with a contrastive loss. We evaluate INSTRUCTOR on 70 embedding evaluation tasks (66 of which are unseen during training), ranging from classification and information retrieval to semantic textual similarity and text generation evaluation. INSTRUCTOR, while having an order of magnitude fewer parameters than the previous best model, achieves state-of-the-art performance, with an average improvement of 3.4% compared to the previous best results on the 70 diverse datasets. Our analysis suggests that INSTRUCTOR is robust to changes in instructions, and that instruction finetuning mitigates the challenge of training a single model on diverse datasets. Our model, code, and data are available at https://instructor-embedding.github.io.
@misc{su_one_2023,
title = {One {Embedder}, {Any} {Task}: {Instruction}-{Finetuned} {Text} {Embeddings}},
shorttitle = {One {Embedder}, {Any} {Task}},
url = {http://arxiv.org/abs/2212.09741},
doi = {10.48550/arXiv.2212.09741},
abstract = {We introduce INSTRUCTOR, a new method for computing text embeddings given task instructions: every text input is embedded together with instructions explaining the use case (e.g., task and domain descriptions). Unlike encoders from prior work that are more specialized, INSTRUCTOR is a single embedder that can generate text embeddings tailored to different downstream tasks and domains, without any further training. We first annotate instructions for 330 diverse tasks and train INSTRUCTOR on this multitask mixture with a contrastive loss. We evaluate INSTRUCTOR on 70 embedding evaluation tasks (66 of which are unseen during training), ranging from classification and information retrieval to semantic textual similarity and text generation evaluation. INSTRUCTOR, while having an order of magnitude fewer parameters than the previous best model, achieves state-of-the-art performance, with an average improvement of 3.4\% compared to the previous best results on the 70 diverse datasets. Our analysis suggests that INSTRUCTOR is robust to changes in instructions, and that instruction finetuning mitigates the challenge of training a single model on diverse datasets. Our model, code, and data are available at https://instructor-embedding.github.io.},
urldate = {2024-02-01},
publisher = {arXiv},
author = {Su, Hongjin and Shi, Weijia and Kasai, Jungo and Wang, Yizhong and Hu, Yushi and Ostendorf, Mari and Yih, Wen-tau and Smith, Noah A. and Zettlemoyer, Luke and Yu, Tao},
month = may,
year = {2023},
note = {74 citations (Semantic Scholar/arXiv) [2024-02-01]
74 citations (Semantic Scholar/DOI) [2024-02-01]
arXiv:2212.09741 [cs]},
keywords = {Computer Science - Computation and Language},
}
Downloads: 0
{"_id":"nyeBHPmP7YC72hC7F","bibbaseid":"su-shi-kasai-wang-hu-ostendorf-yih-smith-etal-oneembedderanytaskinstructionfinetunedtextembeddings-2023","author_short":["Su, H.","Shi, W.","Kasai, J.","Wang, Y.","Hu, Y.","Ostendorf, M.","Yih, W.","Smith, N. A.","Zettlemoyer, L.","Yu, T."],"bibdata":{"bibtype":"misc","type":"misc","title":"One Embedder, Any Task: Instruction-Finetuned Text Embeddings","shorttitle":"One Embedder, Any Task","url":"http://arxiv.org/abs/2212.09741","doi":"10.48550/arXiv.2212.09741","abstract":"We introduce INSTRUCTOR, a new method for computing text embeddings given task instructions: every text input is embedded together with instructions explaining the use case (e.g., task and domain descriptions). Unlike encoders from prior work that are more specialized, INSTRUCTOR is a single embedder that can generate text embeddings tailored to different downstream tasks and domains, without any further training. We first annotate instructions for 330 diverse tasks and train INSTRUCTOR on this multitask mixture with a contrastive loss. We evaluate INSTRUCTOR on 70 embedding evaluation tasks (66 of which are unseen during training), ranging from classification and information retrieval to semantic textual similarity and text generation evaluation. INSTRUCTOR, while having an order of magnitude fewer parameters than the previous best model, achieves state-of-the-art performance, with an average improvement of 3.4% compared to the previous best results on the 70 diverse datasets. Our analysis suggests that INSTRUCTOR is robust to changes in instructions, and that instruction finetuning mitigates the challenge of training a single model on diverse datasets. Our model, code, and data are available at https://instructor-embedding.github.io.","urldate":"2024-02-01","publisher":"arXiv","author":[{"propositions":[],"lastnames":["Su"],"firstnames":["Hongjin"],"suffixes":[]},{"propositions":[],"lastnames":["Shi"],"firstnames":["Weijia"],"suffixes":[]},{"propositions":[],"lastnames":["Kasai"],"firstnames":["Jungo"],"suffixes":[]},{"propositions":[],"lastnames":["Wang"],"firstnames":["Yizhong"],"suffixes":[]},{"propositions":[],"lastnames":["Hu"],"firstnames":["Yushi"],"suffixes":[]},{"propositions":[],"lastnames":["Ostendorf"],"firstnames":["Mari"],"suffixes":[]},{"propositions":[],"lastnames":["Yih"],"firstnames":["Wen-tau"],"suffixes":[]},{"propositions":[],"lastnames":["Smith"],"firstnames":["Noah","A."],"suffixes":[]},{"propositions":[],"lastnames":["Zettlemoyer"],"firstnames":["Luke"],"suffixes":[]},{"propositions":[],"lastnames":["Yu"],"firstnames":["Tao"],"suffixes":[]}],"month":"May","year":"2023","note":"74 citations (Semantic Scholar/arXiv) [2024-02-01] 74 citations (Semantic Scholar/DOI) [2024-02-01] arXiv:2212.09741 [cs]","keywords":"Computer Science - Computation and Language","bibtex":"@misc{su_one_2023,\n\ttitle = {One {Embedder}, {Any} {Task}: {Instruction}-{Finetuned} {Text} {Embeddings}},\n\tshorttitle = {One {Embedder}, {Any} {Task}},\n\turl = {http://arxiv.org/abs/2212.09741},\n\tdoi = {10.48550/arXiv.2212.09741},\n\tabstract = {We introduce INSTRUCTOR, a new method for computing text embeddings given task instructions: every text input is embedded together with instructions explaining the use case (e.g., task and domain descriptions). Unlike encoders from prior work that are more specialized, INSTRUCTOR is a single embedder that can generate text embeddings tailored to different downstream tasks and domains, without any further training. We first annotate instructions for 330 diverse tasks and train INSTRUCTOR on this multitask mixture with a contrastive loss. We evaluate INSTRUCTOR on 70 embedding evaluation tasks (66 of which are unseen during training), ranging from classification and information retrieval to semantic textual similarity and text generation evaluation. INSTRUCTOR, while having an order of magnitude fewer parameters than the previous best model, achieves state-of-the-art performance, with an average improvement of 3.4\\% compared to the previous best results on the 70 diverse datasets. Our analysis suggests that INSTRUCTOR is robust to changes in instructions, and that instruction finetuning mitigates the challenge of training a single model on diverse datasets. Our model, code, and data are available at https://instructor-embedding.github.io.},\n\turldate = {2024-02-01},\n\tpublisher = {arXiv},\n\tauthor = {Su, Hongjin and Shi, Weijia and Kasai, Jungo and Wang, Yizhong and Hu, Yushi and Ostendorf, Mari and Yih, Wen-tau and Smith, Noah A. and Zettlemoyer, Luke and Yu, Tao},\n\tmonth = may,\n\tyear = {2023},\n\tnote = {74 citations (Semantic Scholar/arXiv) [2024-02-01]\n74 citations (Semantic Scholar/DOI) [2024-02-01]\narXiv:2212.09741 [cs]},\n\tkeywords = {Computer Science - Computation and Language},\n}\n\n","author_short":["Su, H.","Shi, W.","Kasai, J.","Wang, Y.","Hu, Y.","Ostendorf, M.","Yih, W.","Smith, N. A.","Zettlemoyer, L.","Yu, T."],"key":"su_one_2023","id":"su_one_2023","bibbaseid":"su-shi-kasai-wang-hu-ostendorf-yih-smith-etal-oneembedderanytaskinstructionfinetunedtextembeddings-2023","role":"author","urls":{"Paper":"http://arxiv.org/abs/2212.09741"},"keyword":["Computer Science - Computation and Language"],"metadata":{"authorlinks":{}},"html":""},"bibtype":"misc","biburl":"https://bibbase.org/zotero/ifromm","dataSources":["N4kJAiLiJ7kxfNsoh"],"keywords":["computer science - computation and language"],"search_terms":["one","embedder","task","instruction","finetuned","text","embeddings","su","shi","kasai","wang","hu","ostendorf","yih","smith","zettlemoyer","yu"],"title":"One Embedder, Any Task: Instruction-Finetuned Text Embeddings","year":2023}