東華大學圖書館 |

Language: English

Help

回圖書館首頁

手機版館藏查詢

Back

Switch To: Labeled | MARC Mode | ISBD

Designing Accurate Retrieval Systems...

Sachan, Devendra Singh,

Linked to FindBook

Google Book

Amazon

博客來

Designing Accurate Retrieval Systems Using Language Models /

Record Type:	Electronic resources : Monograph/item
Title/Author:	Designing Accurate Retrieval Systems Using Language Models // Devendra Singh Sachan.
Author:	Sachan, Devendra Singh,
Description:	1 electronic resource (180 pages)
Notes:	Source: Dissertations Abstracts International, Volume: 86-04, Section: B.
Contained By:	Dissertations Abstracts International86-04B.
Subject:	Linguistics. -
Online resource:	https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31527183
ISBN:	9798342123006

Designing Accurate Retrieval Systems Using Language Models /
Sachan, Devendra Singh,

Designing Accurate Retrieval Systems Using Language Models /Devendra Singh Sachan. - 1 electronic resource (180 pages)

Source: Dissertations Abstracts International, Volume: 86-04, Section: B.

The success of pre-trained language models (PLMs) in language understanding tasks has attracted attention to these models being applied as building blocks in information retrieval systems. Indeed, due to the application of these PLMs, the retrieval accuracy has improved considerably over traditional approaches. This improved retrieval has led to a big boost in the performance of downstream knowledge-intensive tasks such as question answering (QA). However, the training methods of these systems still suffer from a number of shortcomings, some being --- reliance on thousands of aligned question-document pairs, dependence on the output of BM25 retriever to mine hard-negative examples, multi-stage training, etc. In this thesis, we propose several methods to address these drawbacks and further improve the underlying systems. Specifically, our contributions include: (i) a unifying approach of unsupervised pre-training of dense retrievers followed by supervised finetuning to improve retrieval accuracy, (ii) an end-to-end training approach to jointly train a system consisting of a language model and a dense retriever for QA tasks, (iii) using frozen large language models as unsupervised passage re-rankers to improve retrieval accuracy, and (iv) a simplified yet accurate approach to train a dense retriever without aligned question-document pairs and without using hard-negative examples. Our proposed approaches are more robust and have achieved new state-of-the-art results on multiple retrieval and QA benchmarks.

English

ISBN: 9798342123006Subjects--Topical Terms:

524476
Linguistics.

Designing Accurate Retrieval Systems Using Language Models /
LDR:04675nmm a22003853i 4500 001 2431632
005 20250616055544.5
006 m o d
007 cr|nu||||||||
008 261011s2024 miu||||||m |||||||eng d
020 $a 9798342123006
035 $a (MiAaPQD)AAI31527183
035 $a (MiAaPQD)McGill3197xs14t
035 $a AAI31527183
040 $a MiAaPQD $b eng $c MiAaPQD $e rda
100 1 $a Sachan, Devendra Singh, $e author. $3 3820538
245 1 0 $a Designing Accurate Retrieval Systems Using Language Models / $c Devendra Singh Sachan.
264 1 $a Ann Arbor : $b ProQuest Dissertations & Theses, $c 2024
300 $a 1 electronic resource (180 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 86-04, Section: B.
500 $a Advisors: Pineau, Joelle.
502 $b Ph.D. $c McGill University (Canada) $d 2024.
520 $a The success of pre-trained language models (PLMs) in language understanding tasks has attracted attention to these models being applied as building blocks in information retrieval systems. Indeed, due to the application of these PLMs, the retrieval accuracy has improved considerably over traditional approaches. This improved retrieval has led to a big boost in the performance of downstream knowledge-intensive tasks such as question answering (QA). However, the training methods of these systems still suffer from a number of shortcomings, some being --- reliance on thousands of aligned question-document pairs, dependence on the output of BM25 retriever to mine hard-negative examples, multi-stage training, etc. In this thesis, we propose several methods to address these drawbacks and further improve the underlying systems. Specifically, our contributions include: (i) a unifying approach of unsupervised pre-training of dense retrievers followed by supervised finetuning to improve retrieval accuracy, (ii) an end-to-end training approach to jointly train a system consisting of a language model and a dense retriever for QA tasks, (iii) using frozen large language models as unsupervised passage re-rankers to improve retrieval accuracy, and (iv) a simplified yet accurate approach to train a dense retriever without aligned question-document pairs and without using hard-negative examples. Our proposed approaches are more robust and have achieved new state-of-the-art results on multiple retrieval and QA benchmarks.
520 $a Le succes des modeles linguistiques pre-entraines (PLM) dans les taches de comprehension du langage a attire l'attention sur l'application de ces modeles comme elements de base dans les systemes de recherche d'informations. En effet, grace a l'application de ces PLM, la precision de la recuperation s'est considerablement amelioree par rapport aux approches traditionnelles. Cette recuperation amelioree a conduit a une forte amelioration des performances des taches en aval a forte intensite de connaissances, telles que la reponse aux questions. Cependant, les methodes de formation de ces systemes souffrent encore d'un certain nombre de lacunes, certaines etant : la dependance a l'egard de milliers de paires question-document alignees, la dependance a l'egard du resultat du recuperateur BM25 pour extraire des exemples fortement negatifs, la formation en plusieurs etapes, etc. Dans cette these, nous proposons plusieurs methodes pour remedier a ces inconvenients et ameliorer davantage les systemes sous-jacents. Plus precisement, nos contributions comprennent : (i) une approche unificatrice de pre-entrainement non supervisee des recuperateurs denses suivie d'un reglage fin supervise pour ameliorer la precision de la recuperation, (ii) une approche d'entrainement de bout en bout pour former conjointement un systeme compose d'un modele de langage et d'un recuperateur dense pour les taches d'assurance qualite, (iii) utiliser de grands modeles de langage figes comme reclasseurs de passages non supervises pour ameliorer la precision de la recuperation, et (iv) une approche simplifiee mais precise pour former un recuperateur dense sans paires question-document alignees et sans utiliser d'exemples strictement negatifs. Les approches proposees sont plus robustes et ont permis d'obtenir de nouveaux resultats de pointe sur plusieurs tests de recuperation d'information et de question-reponse.
546 $a English
590 $a School code: 0781
650 4 $a Linguistics. $3 524476
650 4 $a Writing. $3 551664
650 4 $a Neural networks. $3 677449
690 $a 0290
690 $a 0800
710 2 $a McGill University (Canada). $e degree granting institution. $3 3759190
720 1 $a Pineau, Joelle $e degree supervisor.
773 0 $t Dissertations Abstracts International $g 86-04B.
790 $a 0781
791 $a Ph.D.
792 $a 2024
856 4 0 $u https://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31527183