DC Field | Value | Language |
---|---|---|
dc.contributor.advisor | 제임스 손 | - |
dc.contributor.author | Song, EuiYul | - |
dc.contributor.author | 송의열 | - |
dc.date.accessioned | 2024-07-25T19:30:44Z | - |
dc.date.available | 2024-07-25T19:30:44Z | - |
dc.date.issued | 2023 | - |
dc.identifier.uri | http://library.kaist.ac.kr/search/detail/view.do?bibCtrlNo=1045720&flag=dissertation | en_US |
dc.identifier.uri | http://hdl.handle.net/10203/320532 | - |
dc.description | 학위논문(석사) - 한국과학기술원 : 김재철AI대학원, 2023.8,[iv, 29 p. :] | - |
dc.description.abstract | Generative retrieval models encapsulate the understanding of queries within their parameters. However, to minimize uncertainty and attain a more comprehensive understanding, it is crucial to thoroughly explore the knowledge derived from decoding and auxiliary indexes. In this regard, we present Re3val, which facilitates a generative retrieval model to reinforce and rerank utilizing limited data. Specifically, we generate questions from our pre-training dataset to counterbalance epistemic uncertainty and bridge the domain discrepancy between the pre-training and fine-tuning datasets. Additionally, we utilize the REINFORCE to infuse information sourced from the undifferentiable constrained decoding algorithm. Furthermore, our page title reranker integrates contexts procured via Dense Passage Retrieval to rerank the retrieved page titles. Subsequently, we extract contexts from the KILT database using the reranked page titles and rerank them by relevance. Upon grounding the top five reranked contexts, Re3val demonstrates the Top 1 KILT scores compared to all other generative retrieval models across five KILT datasets. | - |
dc.language | eng | - |
dc.publisher | 한국과학기술원 | - |
dc.subject | 생성 검색▼a강화 학습▼a질문 생성▼a재랭크 | - |
dc.subject | Generative retrieval▼aReinforcement learning▼aQuestion generation▼aRerank | - |
dc.title | Re3val: reinforced and reranked² data-efficient generative retrieval | - |
dc.title.alternative | 강화 및 재랭크²된 데이터 효율적인 생성 검색 | - |
dc.type | Thesis(Master) | - |
dc.identifier.CNRN | 325007 | - |
dc.description.department | 한국과학기술원 :김재철AI대학원, | - |
dc.contributor.alternativeauthor | James, Thorne | - |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.