Improving the Domain Adaptation of Retrieval Augmented Generation RAG Models for Open Domain Question Answering Shamane Siriwardhana Rivindu Weerasekera Elliott Wen

2025-05-08 0 0 340.13KB 15 页 10玖币

侵权投诉

Improving the Domain Adaptation of Retrieval Augmented Generation

(RAG) Models for Open Domain Question Answering

Shamane Siriwardhana , Rivindu Weerasekera, Elliott Wen,

Tharindu Kaluarachchi, Rajib Rana†,and Suranga Nanayakkara4

Augmented Human Lab, Auckland Bioengineering Institute, The University of Auckland

firstname@ahlab.org

4Department of Information Systems & Analytics, National University of Singapore

†University of Southern Queensland

Rajib.Rana@usq.edu.au

Abstract

Retrieval Augment Generation (RAG) is a

recent advancement in Open-Domain Ques-

tion Answering (ODQA). RAG has only

been trained and explored with a Wikipedia-

based external knowledge base and is not

optimized for use in other specialized do-

mains such as healthcare and news. In this

paper, we evaluate the impact of joint train-

ing of the retriever and generator compo-

nents of RAG for the task of domain adap-

tation in ODQA. We propose RAG-end2end,

an extension to RAG, that can adapt to a

domain-speciﬁc knowledge base by updat-

ing all components of the external knowl-

edge base during training. In addition, we

introduce an auxiliary training signal to in-

ject more domain-speciﬁc knowledge. This

auxiliary signal forces RAG-end2end to re-

construct a given sentence by accessing the

relevant information from the external knowl-

edge base. Our novel contribution is unlike

RAG, RAG-end2end does joint training of

the retriever and generator for the end QA

task and domain adaptation. We evaluate

our approach with datasets from three do-

mains: COVID-19, News, and Conversa-

tions, and achieve signiﬁcant performance

improvements compared to the original RAG

model. Our work has been open-sourced

through the Huggingface Transformers li-

brary, attesting to our work’s credibility and

technical consistency. 1

1 Introduction

Open Domain Question Answering (ODQA) (Lee

et al.,2019;Lewis et al.,2020c) is an important

This paper is awaiting publication at TACL and this is a

pre-MIT Press publication version

task in natural language understanding. ODQA

methods generally feature a two-stage pipeline: a

retriever that selects passages relevant to a given

question and a reader that generates the answers

from selected passages. Conventionally, these two

components are trained separately using ground

truth context passages relevant to question-answer

(QA) pairs. However, for many real-world scenar-

ios, it is hard to ﬁnd explicitly annotated context-

question-answer triplets (Lee et al.,2019;Lewis

et al.,2020b;Guu et al.,2020).

Recently, Retrieval Augmented Models (RAGs)

have drawn considerable attention from researchers.

RAG consists of a state-of-the-art-neural retriever

called Dense Passage Retrieval (DPR) (Karpukhin

et al.,2020) and BART seq2seq language

model (Lewis et al.,2020a). Compared to the

conventional two-staged ODQA pipelines, RAG

merges the retriever and reader stages into one

architecture. Moreover, unlike expensive lan-

guage models with billions of parameters (e.g.,

GPT-3 (Brown et al.,2020) and Megatrone-

LM (Narayanan et al.,2021)) where the model’s

parametric memory represents the complete knowl-

edge, RAG can also extract knowledge from an ex-

ternal knowledge base. Using both parametric and

non-parametric memory generally leads to reduced

hallucinations and higher interpretability in tasks

like question answering and summarization (Xu

et al.,2021;Komeili et al.,2021;Guu et al.,2020;

Lewis et al.,2020b).

In this work, we focus on exploring retrieval

augmented architectures for the task of domain-

speciﬁc open-domain question answering. Al-

though there are several similar retrieval augmented

architectures, such as REALM (Guu et al.,2020)

and RETRO (Borgeaud et al.,2021), we used Re-

arXiv:2210.02627v1 [cs.CL] 6 Oct 2022

trieval Augmented Generation (RAG) in our exper-

iments due to its excellent open-source documenta-

tion and availability.

When the RAG model is ﬁnetuned for down-

stream QA tasks, the original implementation keeps

the encoding of passages and the external knowl-

edge base ﬁxed. This is because re-encoding the

external knowledge base is computationally ex-

pensive and relies on a sophisticated implementa-

tion. Despite not ﬁnetuning the passage encodings,

the RAG model performs well for datasets with

Wikipedia-like knowledge bases because the DPR

retriever components have already been trained

on Wikipedia-based datasets (Kwiatkowski et al.,

2019;Joshi et al.,2017). However, the feasibility

of adapting RAG to speciﬁc ODQA domains such

as research papers and news is not well understood.

This is a critical research gap to address, as im-

proved domain adaptation can further improve the

ODQA performance of RAG.

This paper explores the feasibility of using RAG

in specialized domains for ODQA. In particular,

we propose two modiﬁcations to the original RAG

to improve its domain adaptability. Motivated

by recent end2end retrieval augmented mecha-

nisms (Guu et al.,2020;Sachan et al.,2021;Singh

et al.,2021), we ﬁrst propose a method to ﬁne-

tune the RAG model with its neural retriever and

update its knowledge encodings asynchronously

during training. We refer to this as RAG-end2end

since it allows us to update all RAG components

during training, including the external knowledge

base, the DPR model, and the BART model. Sec-

ondly, we propose an auxiliary training signal to

help our model learn more domain-speciﬁc knowl-

edge. This took the form of generating a concise

and factual statement about a document using a

self-retrieved set of passages from the provided

domain-speciﬁc knowledge base. These two mod-

iﬁcations offer a unique feature to RAG-end2end

over RAG: joint training of the retriever and gen-

erator for the end QA task and domain adaptation.

Although asynchronous updates to the knowledge

encoder have been proposed before in the REALM,

previous work has not evaluated the effects of joint

training of the RAG’s retriever and the generator

for the domain adaptation in ODQA.

We evaluate our proposed approach on three dif-

ferent datasets from three domains: COVID-19

research (Wang et al.,2020), Conversations (Wu

et al.,2021b), and News (Trischler et al.,2016).

The major ﬁnding of our work is that the adap-

tation of the retriever component plays a critical

role in overall domain adaptation performance in

RAG-like architectures. Updating only the ques-

tion encoder without updating the knowledge base

encoding could degrade performance. Instead of

ﬁnetuning the DPR retriever separately, our ex-

periments show that ﬁnetuning it as a part of the

RAG-end2end mechanism gives better overall re-

sults. Our results also show that using the auxiliary

signal improves both the retriever component and

the overall accuracy.

In addition, we open-source the implementation

of RAG-end2end with the HuggingFace Transform-

ers (Wolf et al.,2019) Library

providing the oppor-

tunity for the scientiﬁc community to use/test/build

on our work.

2 Background and Related Work

Open-domain QA systems (Yang et al.,2015;

Kwiatkowski et al.,2019) generally have a two-

stage pipeline: passage retrieval (i.e., ﬁnding rele-

vant text chunks related to an input question from

a knowledge base) and machine comprehension

(i.e., generating an answer from a set of selected

documents). Traditionally sparse vector methods

such as TF-IDF and BM25 are used for document

retrieval (Robertson and Zaragoza,2009). Re-

searchers have recently moved to use dense text

representations, which allows modeling textual sim-

ilarity more semantic level. A recent example is the

‘Dense Passage Retriever (DPR)’ (Karpukhin et al.,

2020), which generates embeddings for questions

and text passages using two BERT (Devlin et al.,

2018) models. The dot product of the embeddings

is used as a similarity score between a question

and a passage. DPR has demonstrated that higher

retrieval precision results in a higher end-to-end

QA accuracy. For the answer generation compo-

nent of QA systems, recent studies have used either

extractive language models like BERT or genera-

tive language models like BART/GPT-2 (Min et al.,

2021;Lewis et al.,2021).

2.1 Retrieval Augmented Architecture

Recently, Retrieval Augmented Architectures

(Lewis et al.,2020b;Guu et al.,2020) have drawn

a lot of attention due to their explainable, scalable,

and adaptable nature. Unlike other open-domain

2Huggingface Transformers implementation

QA architectures, RAG (Lewis et al.,2020b) com-

bines the information retrieval stage and answer

generation stage in a differentiable manner. It uses

a combination of parametric and non-parametric

memory, where the parametric memory consists of

a pre-trained seq2seq BART (Lewis et al.,2019)

generator, and the non-parametric memory consists

of dense vector representations of Wikipedia arti-

cles indexed with the FAISS library (Johnson et al.,

2017). RAG ﬁrst encodes a question into a dense

representation, retrieves the relevant passages from

an indexed Wikipedia knowledge base, and then

feeds them into the generator. The loss function

can ﬁnetune both the generator and the question en-

coder at the same time. Lewis et al. (Lewis et al.,

2020b) highlight RAG’s ability to perform well

in Wikipedia-based general question-answering

datasets like Natural Questions (Kwiatkowski et al.,

2019). Other recent work also highlights how the

outputs generated from RAG models are much

more factual due to RAG being conditioned on

the retrieved documents, possibly providing an an-

swer to the hallucination problem of generative

language models. Shuster, Kurt, et al. (Shuster

et al.,2021) also highlight how RAG reduces hal-

lucinations in knowledge-grounded conversational

tasks, where the task is to generate responses to

dialogues based on a large Wikipedia knowledge

base. Xu et al. (2021) illustrate the effectiveness

of RAG in chat-bot frameworks and highlight how

RAG models are able to recall and summarize con-

versations compared to standard seq2seq models

with only parametric memory. This paper aims

to understand how RAG could be extended to an

end2end model and adapted to speciﬁc domains.

To the best of our knowledge, this is the ﬁrst time

RAG is being investigated on domain adaptation

for the task of ODQA systems.

2.2 REALM-like end2end Retrieval Augment

Architectures

REALM (Guu et al.,2020) is a similar Retrieval

Augmented model to RAG. REALM introduced

a novel masked language pre-training step that in-

volves an end-to-end trainable retriever. In the

REALM work, the authors ﬁrst train the entire

model on the masked language prediction task

and then ﬁne-tune it on question-answering tasks

(keeping the retriever frozen). In comparison to

REALM, the original RAG model uses an already

trained DPR retriever and conducts partial end-to-

end training with a BART reader model. Com-

pared to REALM, RAG is less computationally

expensive, and its code is available open-source.

We explore and extend the original RAG archi-

tecture for domain adaptation in our work. We

adapted some concepts of our RAG-end2end ex-

tension from REALM. REALM only updates its

retriever during the pre-training process that uses

the masked language modeling (MLM) (Devlin

et al.,2018) task. Then during the downstream

ﬁne-tuning task, REALM keeps its retriever ﬁxed.

However, the REALM end-to-end training code

is not open-sourced, possibly due to its computa-

tional complexity. Compared to REALM, RAG

is a combination of already pre-trained language

models where the users do not need to go through a

heavy pre-training stage. Due to these engineering-

friendly features and high availability, we con-

ducted our experiments with RAG and extended

RAG into an end-to-end trainable retrieval augmen-

tation model. It is also important to highlight that

none of the prior work has explored the domain

adaptation of retrieval augment models for question

answering; instead, most focus on general question

answering with Wikipedia-based knowledge bases.

Similar to REALM’s end2end architecture, re-

cent work (Sachan et al.,2021) extended RAG

and highlighted that the retriever training could

improve the overall performance in question-

answering datasets like Natural Questions. Com-

pared to our work, the authors did not focus on

the domain adaptation of retrieval augment mod-

els. The authors mainly explore the ability to train

neural retrievers in an end-to-end way using re-

trieval augment models. Similarly, another related

work (Singh et al.,2021) extended retrieval aug-

mented architectures to an end-to-end model and

illustrated that it could improve the question an-

swering accuracy. Singh et al. (2021) mainly fo-

cused on improving the document reading ability

and answer generation rather than domain adapta-

tion.

3 Model Architecture and Training

Procedure

In this work, we extend RAG to ﬁnetune all compo-

nents, including the DPR retriever, and dynamically

update the external knowledge base during train-

ing. We hypothesize that the use of asynchronous

updates helps with domain adaptation. Figure 1

demonstrates the main workﬂow of our model. In

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

ImprovingtheDomainAdaptationofRetrievalAugmentedGeneration(RAG)ModelsforOpenDomainQuestionAnsweringShamaneSiriwardhana,RivinduWeerasekera,ElliottWen,TharinduKaluarachchi,RajibRana,andSurangaNanayakkara4AugmentedHumanLab,AucklandBioengineeringInstitute,TheUniversityofAucklandfirstname@ahlab.or...

展开>> 收起<<

Improving the Domain Adaptation of Retrieval Augmented Generation RAG Models for Open Domain Question Answering Shamane Siriwardhana Rivindu Weerasekera Elliott Wen.pdf

共15页,预览3页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

Improving the Domain Adaptation of Retrieval Augmented Generation RAG Models for Open Domain Question Answering Shamane Siriwardhana Rivindu Weerasekera Elliott Wen

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: