Datasets:

luoojason
/

muddle-eval-bundle-md

a W rit i y se potential as general-purpose models for various vision-language 𝐻,-./012 𝐼& tasks.MLLMsinvolvesignificantexternalknowledgewithintheir ContrastiveLearning 𝐻["#$] 𝐻& 𝐻'()* …𝐻!"#$%&' 𝐼) 𝐼* 𝐼+ …𝐼& parameters; however, it is challenging to continually update these models with the latest knowledge, which involves huge Text Image Text Image Encoder Encoder Encoder Encoder computationalcostsandpoorinterpretability.Retrievalaugmen- tation techniques have proven to be effective plugins for both LLMsandMLLMs.Inthisstudy,weproposemultimodaladap- Arockband Arockband isplaying isplaying tive Retrieval-Augmented Bootstrapping Language-Image Pre- training (RA-BLIP), a novel retrieval-augmented framework for (c) Multimodal Knowledge Semantic Space various MLLMs. Considering the redundant information within vision modality, we first leverage the question to instruct the extractionofvisualinformationthroughinteractionswithoneset of learnable queries, minimizing irrelevant interference during retrieval and generation. Besides, we introduce a pre-trained multimodal adaptive fusion module to achieve question text-to- MultimodalAdaptiveRA-BLIP multimodalretrievalandintegrationofmultimodalknowledgeby projectingvisualandlanguagemodalitiesintoaunifiedsemantic … … … … space.Furthermore,wepresentanAdaptiveSelectionKnowledge Generation (ASKG) strategy to train the generator to au- tonomously discern the relevance of retrieved knowledge, which Which streetwas paved with realizes excellent denoising performance. Extensive experiments boards; Little Champlain Street, Quebec City, 1916 or o th n at op R e A n -B m L u I l P tim a o ch d i a e l ve q s ue si s g ti n o i n fi - c a a n n s t w p er e i r n fo g rm da a t n a c s e ets an d d em su o r n p s a tr s a se te s Quebec C w i i t n y t e R r u 2 e 0 S 1 a 0 int-Louis N Q aN w t u i Q ao i N e w t nu n b i Q a t o i e a e e wn tn b lu c i rt oai ee eH 2 n C ln cbr 0 ti Ha i e es2 C1 t l ctr y 0 i 0 o i Hs 21 tC. r R t y i00 oi Lic s u 1. r tR t oy i e S0 o L c uu . r iR S o e i tSi L s e a u c u i S i iot Seo S n s e a u t f t i iSrS i - ot ne Cs e L at f t er i - aS o o en C tL n - u t f eta ra o a- i te C L n s -d u ea aa a io t s dn - u aaa i s d a Question Image Image-Text Document the state-of-the-art retrieval-augmented models. QuestionPrompt Inter-Modal Intra-Modal Relevant Index Terms—Retrieval-augmented model, vision-language Instruction Interaction Interaction Irrelevant pre-training, multimodal retrieval, open question answering. Question Image Image-Text Text Embeddings Embeddings Embeddings Embeddings I. INTRODUCTION Fig. 1. Illustration of different multimodal retrieval approaches. (a) Cross- THE birth of the Internet has triggered an unprecedented modality retrieval. (b) Late-interaction retrieval. (c) RA-BLIP multimodal information revolution, catapulting humanity into the era adaptive retrieval. For RA-BLIP, questions, documents, images, and image- textpairsareprojectedintoaunifiedmultimodalspace. of information explosion. It is a great challenge to efficiently find answers from a vast amount of information based on our questions. Open Multimodal Multihop Question Answer- BLIP-2 [10], GPT-4 [11], etc., have been notably explored to ing (MMQA) [1]–[7] can help alleviate this problem of in- enhancetheirperformancebyimplicitlyencodingasubstantial formationoverloadbyretrievingexternalknowledgebasedon amount of external knowledge within their parameters, which questionsandgeneratingcorrectanswers.Inrecentyears,sev- now scale into the hundreds of billions [12]. While these eraladvancedLLMsandMLLMslikeFlanT5[8],LLaMA[9], models have yielded exciting results on various multimodal ThisworkwassupportedinpartbytheNationalNaturalScienceFoundation tasks, they have also encountered high computational costs ofChinaunderGrant62376069,inpartbyYoungEliteScientistsSponsorship and significant challenges in terms of interpretability. ProgrambyCASTunderGrant2023QNRC001,andinpartbyGuangdongBa- To alleviate the challenge, many researchers proposed re- sicandAppliedBasicResearchFoundationunderGrant2024A1515012027. Muhe Ding, Jianlong Wu and Liqiang Nie are with the School trievalaugmentationtechniquesthatdividethemodelintotwo of Computer Science and Technology, Harbin Institute of Technology keycomponents:theretrieverandthegenerator[13]–[16].The (Shenzhen),Shenzhen518055,China(e-mail:dmh1216380870@gmail.com, retriever accesses relevant knowledge from a knowledge base jlwu1992@pku.edu.cn,nieliqiang@gmail.com). Yang Ma is with the School of Computer Science, University of Sydney, based on the posed question, while the generator leverages Sydney,NSW2006,Australia(e-mail:yama5878@uni.sydney.edu.au). this information to create textual output in response. In ear- Pengda Qin and Yuhong Li are with the Security Department, Al- lier stages, text-modality retrieval-augmented models, such as ibaba Group, Hangzhou 311121, China (e-mail: qinpengda0406@163.com, daniel.yuhong@gmail.com). REALM [13], RAG [17], and so on [18], [19], have been 4202 tcO 81 ]MM.sc[ 1v45141.0142:viXra

IEEETRANSACTIONSONMULTIMEDIA															2
proposed to solve text-only question answering. They build an adaptive selection knowledge generation strategy, which
the dense index as a non-parametric document memory from gives the generator the capability of selecting knowledge by
extensive textual sources like Wikipedia for effective knowl- data enhancement to make the model automatically judge the
edge retrieval, and the generator produces answers based on relevance of knowledge. ASKG strategy allows the generator
the retrieved knowledge. More recently, multimodal retrieval- tonotsimplyrelyonthewordsimilaritybetweenthequestion
augmented models, such as MuRAG [14], SKURG [20], and and knowledge, but to understand the semantic information
soon[16],[21],haveemergedoneafteranother.Thesemodels of question and know which knowledge contains the answer.
extend the knowledge memory across various modalities, em- Furthermore, the parameters of the image encoder and LLM
ployingpre-trainedvisuallanguagemodelstoretrieverelevant of our framework are frozen, significantly reducing computa-
evidence and support reasoning for answers. tionalcosts.ExtensiveexperimentsonthreerepresentativeQA
However, existing methods exhibit certain limitations. The datasets demonstrate the effectiveness of our methods.
first limitation is the insufficient integration and interaction Overall, our key contributions are as follows:
between	vision	and	language.	On	the one	hand,	existing
-------	------	--------	-----------	---	----------	------------	--------	---------	-------	----------	----------	--------	--------	--------	----------
								• We	propose	a novel	multimodal		adaptive		retrieval-
methods	lack	explicit	integration	of	multimodal	information,
								augmented		framework,	which	achieves		question	text-to-
hindering the alignment of questions and multimodal knowl- multimodalretrievalandknowledge-intensivemultimodal
edge in	the semantic		space.	As shown	in	Fig.	1(a), some
-------	------------	---	------	--------	---	----	----------	---	--------------	---	------	------------	---	----------	---
								QA	by integrating		visual	and language		modalities	and
methods [14], [22], [23] have employed separate visual en- projecting them into a unified semantic space.
coder and	text	encoder	for individual		modality	encoding	and
---------	----	-------	--------------	---	--------	--------	---	---	---	---	---	---	---	---	---
• Weintroduceanadaptiveselectionknowledgegeneration
adopted	contrastive	learning		[24] for	multimodal		alignment
-------	-----------	--------	---	--------	----------	---	---------	--------	----	---------	---	--------	------------	---	-------
								strategy	that	leverages	the	powerful	capabilities		of LLMs
to retrieve. This may lead to an unbalanced and biased to select the relevant retrieved knowledge for answer
multimodal	retrieval		and reasoning		process	towards	specific
----------	---------	---	-------------	---	-------	-------	--------	---------	---	-------------	---	---	---	---	---
								reasoning		autonomously.
modalities.Besides,late-interactionretrievalapproaches[25]– We conduct extensive experiments on various multi-
•
[27] in Fig. 1(b), retain dual-encoder independent encoding modal and multihop datasets (i.e., WebQA [4], Multi-
architecture	and	perform	token-wise		interactions	only	in the
------------	---	-------	----------	---	------------	----	------	-------	---	--------	------	-----	-------	---	------
								modalQA		[5], and	MMCoQA	[6]).	RA-BLIP		demon-
late scoring stage, which sacrifices the retrieval efficiency for strates superiority over the existing state-of-the-art
thebenefitsoffine-grainedfeaturelearning.Ontheotherhand,
								retrieval-augmented			models.
---	---	---	---	---	---	---	---	-------------------	---	---	-------	---	---	---	---
existingmethods[14],[20],[28],[29]donotutilizequestions
to instruct	the	image	encoder	in selectively		extracting	visual
-------------	-----------	--------------	-------	--------------	-------------	----------	----------	------------------	---	------------	-----------	-----	-------	----------------	------
										II.	RELATEDWORK
features,	and	thus suffer	from	interference		and noise	caused
								A. Vision-Language		Pretraining
by redundant	information		in	images.	Moreover,	such	methods
lack mutual	instruction		when	encoding	different		modal fea-
								Vision-language		pre-training		(VLP)	aims	to train	models
tures, making		it challenging		to model	relationships		between
								on large-scale		image-text	datasets	to	capture	the relationship
multiple sources. The second limitation is that existing ap- between these two modalities. Broadly, VLP methodologies
proachesdonotinspectthecorrectnessoftheretrievedrelevant
								fall into	two categories		based	on their	training	approach:	1)
---	---	---	---	---	---	---	---	---------	--------------	---	-----	--------	--------	---------	---
knowledge at the generation stage. However, the retrieved End-to-end Methods: This category includes methods [35]–
knowledge	contains	significant		noise,	resulting	in poor	model
---------	--------	-----------	---	------	---------	-------	-----	---	---	---	---	---	---	---	---
[38]thattrainmodelsend-to-end,backpropagatinglearnedsig-
anti-interference		and	robustness.	The	generator	assumes	all
-----------------	---	---	-----------	---	---------	-------	---	---------------	---	------	--------	-------	---------	---	-----------
								nals to achieve		mutual	learning	between	different		modalities.
retrieved relevant knowledge is correct, potentially leading to 2)ModularMethods:Incontrast,modularmethods,asseenin
the utilization		of irrelevant	or	confusing	information.
---------------	---	-------------	---	---------	------------	---	---	---	---	---	---	---	---	---	---
worksby[39]–[43],involvekeepingtheparametersofspecific
To address the above issues, we propose multimodal adap- pre-trainedcomponents(likeimageencodersorlargelanguage
tiveRetrieval-AugmentedBootstrappingLanguage-ImagePre- models) fixed while focusing on refining other aspects of the
training	(RA-BLIP).		RA-BLIP	consists	of	two key	compo-
--------	----------	---	-------	--------	---	-------	------	------	-------------	---	----	--------	-------------	---	------
								model.	For instance,	LiT	[44]	utilizes	a pre-trained		frozen
nents: a multimodal adaptive retrieval-augmented framework image encoder from CLIP, while Flamingo [45] and BLIP-2
and an	adaptive	selection		knowledge	generation		(ASKG)
------	--------	---------	---	---------	----------	---	------	---	---	---	---	---	---	---	---
[10]freezethelanguagemodeltointegrateLLMsintovision-
strategy. To tackle the first limitation, RA-BLIP is based on language tasks better. Besides, instruction tuning is also an
the InstructBLIP architecture [30] and adopts Q-Former to effective approach during VLP. InstructBLIP [30] represents
implementinstruction-awarevisualfeatureextractionthatuses
arecentadvancementinthisarea,achievinginstruction-aware
questions as instructions. The question instruction interacts visual feature extraction and instruction-guided LLM genera-
withthequeryembeddingsthroughsharedself-attentionlayers
								tion. The	unique	capability	of	InstructBLIP	to	extract	features
--------------	---	--------------	---	--------------------	---	---	-----------	---------	------	-------------	---	------------	----	-------	-----------
and encourages		the extraction		of question-relevant			visual fea-
								based on	prompt	instructions,		combined	with	its	utilization
tures. Additionally, we incorporate a pre-trained multimodal of frozen LLMs and image encoders, positions it as an ideal
adaptive	fusion	module	to fuse	vision	and	text information,
-------------	----------	--------	------------------	----------	---	-----------------	---------	--------	-------	--------	-------------------	---	---	----------	---
								backbone	for our	proposed	retrieval-augmented			framework.
obtaining	multimodal		features	[31]–[33].		As a result,	RA-
BLIP achieves		question	text-to-multimodal			retrieval	by align-
ing the questions and multimodal knowledge bases in the B. Text-modality Retrieval-Augmented Models
semantic space of three modalities: text, image, and image- Retrieval-augmented techniques have proven to be ef-
text [34], as shown in Fig. 1(c). For the second limitation, fective plugins for both LLMs and MLLMs in academia.
we leverage the implicit capabilities of LLMs and introduce These techniques extract pertinent world knowledge from

IEEETRANSACTIONSONMULTIMEDIA 3 extensive databases, subsequently integrating this information A. Problem Formulation to formulate answers. Pioneering methods like ORQA [3] This paper presents a multimodal adaptive retrieval- have employed inverse cloze tasks for retriever pre-training, augmentedframeworkcalledRA-BLIPforopenmultihopand showcasing their efficacy on open-ended question answering multimodalQA,integratingretrievalandgenerationfunctions. datasets. Following suit, REALM [13] extends this by retriev- For knowledge-intensive QA, we deconstruct the task into ing and processing documents from comprehensive sources two stages: retrieval and generation, which are implemented like Wikipedia for logical reasoning, which is pre-trained to by the retriever and generator respectively. The goal of our reason over a large corpus of knowledge on the fly during model training is to learn the distribution P(y|x ) to generate q inference. Furthermore, RAG [17] adopts a pre-trained model a textual output y conditioned on input question x and q andnon-parametricmemoryforlanguagegeneration.FiD[18] multimodal knowledge base KB (KB = k ,...,k ). Firstly, 1 n leveragesencoder-decodertransformermodelsforknowledge- the retriever encodes questions, images, and texts from the intensivetasks,settingnewbenchmarksinQA.Morerecently, knowledge base KB. It identifies the most relevant retrieved RETRO [19] has advanced these methods by handling longer knowledge, K ⊂KB (K is the retrieved knowledge) for ret ret sequences and accessing diverse documents for segmented each question x , which is modeled as p(K |x ). Secondly, q ret q sequences from expansive retrieval datasets. Text-modality the generator utilizes an LLM to generate answers y, condi- retrieval-augmented models construct dense indices as non- tionedonboththequestionandtheretrievedknowledge,which parametricdocumentmemories,usingextensivetextualknowl- is modeled as p(y|x ,K ). We treat multimodal knowledge q ret edge bases to align the retrieval process with specific queries. K asalatentvariablefromtheexternalknowledgebaseand ret Despite these advancements, a notable limitation remains the marginalize it to increase the overall likelihood of the answer needforthesemethodstoeffectivelyleveragevastmultimodal y. The overall process is encapsulated in the equation: knowledge, thereby constraining their applicability in the (cid:88) domain of open multimodal question answering. p(y |x q )= p(K ret |x q )·p(y |x q ,K ret ). (1) (cid:124) (cid:123)(cid:122) (cid:125) (cid:124) (cid:123)(cid:122) (cid:125) Kret⊂KB Retrieval Generation C. Multimodal Retrieval-Augmented Models This dual-stage framework effectively addresses the complex- To overcome the limitations of text-modality retrieval- ities of open multimodal QA by balancing the retrieval of augmented models, recent research [10], [15], [30], [46], [47] multimodal data and knowledge-based generation, and has has made strides in integrating multimodal knowledge. No- been validated by extensive experiments and ablation studies. tableefforts,includingAutoRouting[5]andMAE[6],involve trainingdistinctmodelsforeachmodalityandusingclassifiers B. Model Architecture for task-specific routing, though this approach often hampers RA-BLIP is built on a simple backbone model that is pre- cross-modal reasoning. MuRAG [14] seeks to overcome this trained to encode image-text pairs so that they are suitable limitation by employing separate encoders for visual and for both knowledge base retrieval and answer generation. textualmodalities,followedbyajointencoderformultimodal The overall framework of RA-BLIP is shown in Fig. 2. The fusion. However, this approach lacks integrated guidance for backbone model consists of a multimodal encoder f (·) and θ different modalities and cannot model the relations between decoderg (·),whichareusedascomponentsoftheRA-BLIP θ knowledge sources during retrieval. SKURG [20] attempts to model to implement retrieval and generation. The multimodal bridge this gap by using an entity-centered fusion encoder encoder f (·) contains a frozen image encoder ViT [48], θ to align modalities, yet faces challenges in computational Q-Former architecture [30], and the pre-trained multimodal efficiency and limited interpretability. Besides, Solar [21] adaptive fusion module. The decoder g (·) is composed of θ transforms multimodal inputs into a unified language format a LLM FlanT5 [8]. Querying Transformer (Q-Former) [10] but falls short in handling complex tasks and generalizing is a lightweight Transformer consisting of two modules that visual information. REVAL [16] leverages large-scale knowl- sharethesameself-attentionlayer:oneisanimagetransformer edgegraphstoassistvisuallanguagepre-training,butitbrings that interacts with the frozen image encoder ViT for visual a lot of calculations. In contrast, RA-BLIP distinguishes itself featureextraction,andtheotherisatexttransformercanactas by seamlessly integrating visual and language modalities into both text encoder and text decoder for text feature. The visual a cohesive semantic space, enabling the autonomous selection encoder, composed of ViT and Q-Former image transformer, of relevant knowledge for reasoning, thus addressing these hasinstruction-awarevisualfeatureextractioncapabilitiesand limitations more effectively. can extract visual information based on question instructions. We input N learnable query embeddings into the Q-Former III. METHODOLOGY imagetransformer,whichinteracts withfrozenimagefeatures In this section, we first formulate the research problem through cross-attention layers to obtain visual representation and subsequently elaborate on the model architecture of our f (I) ∈ RN×D, where D is the hidden dimension of the Q- θ retrieval-augmented framework. Then, we describe the learn- Former.Additionally,weusetheQ-Formertexttransformerto ablequeryinteractionapproach,followedbymultimodaladap- encode text, taking the [CLS] token as the text representation tive fusion module. Subsequently, we show how to train the f (T)∈R1×D.Toobtainmultimodalfeaturescombiningboth θ retrieverandranktherelevantknowledge.Lastly,weintroduce imageandtext,weintroduceapre-trainedmultimodaladaptive the adaptive selection knowledge generation strategy. fusion module M (·) to obtain the multimodal representation θ

IEEETRANSACTIONSONMULTIMEDIA 4 Q:The Marina Bay Sands in Singapore is made of what building material? Retrieval Knowledge Template: {Q}. Which item is related to this question? Positive & Negative Knowledge Pair Answer Generation Stage: Q-Former Q-Former Q-Former Fusion Multimodal LLM Retrieval Stage: Learnable queries 🔥 🔥 Image 𝑓!(𝑘"#) … Encoder 𝑓!(𝑘") Marina Bay Q K1 🔥Q-Former Sands Singapore 🔥 Q:The Marina Bay Sands 𝑓!(𝑘#) in Singapore is made of what building material? It is close to some of Singapore's famous In Asia, the Marina 🔥 S l u a n n t d e m c a C r i k t s y , s u M c a h r i a n s a S B i a n y g a S p a o n r d e s w lo a c s a a te d d d e in d 𝑓!(𝑘#) Bay Sands, Esplanade, to the company's portfolio in 2010… … ModalityRepresentation … Text Question QA Data Format Image Relevant Knowledge Input Learnable queries Image-Text concat Q K2 Image Encoder Instruction Top-K Q K3 Q K4 Question& Caption Marina Bay ASKG Data Format Sands Singapore FeatureEncoder In 2010, when it opened, at total cosTt he Downtown MRInT Asia, the Marina 🔥 Classifier inc o lu f d S in $ g 8 l b a i n ll d io c n o stm, st e a t t e io rs n t i o s t b h u e i l W t a e s fe St B w oi a nf y g a S p a o n r d e s w lo a c s a a te d d d e in d Marina Bay Sands… the building. to the company's portfolio in 2010… Maximum Inner Product Search Ranking Fig. 2. The overall workflow of RA-BLIP consists of a retrieval stage and a generation stage. We utilize multimodal encoder f θ (·) to project questions and multimodal knowledge into a unified semantic space to achieve question text-to-multimodal retrieval, and exclude confusing knowledge by ranking. Additionally, we employ ASKG strategy to filter out invalid knowledge, enabling precise reasoning. The parameters of LLM and image encoder are fixed, onlytheQ-Formerandclassifieraretrainable. f (IT) ∈ R(N+1)×D. Consequently, the multimodal encoder Q-Former has aligned visual and text feature representation, θ can simultaneously encode image, text, and image-text fea- we use Image-Text Matching loss and Image-grounded Text tures. Questions and knowledge are encoded into multimodal Generation loss for pre-training. As shown in Fig. 3, our information through the multimodal encoder and then input approach is to fix the parameters of the image encoder and into the decoder for answer generation. In the generation Q-Former, and solely fine-tune the parameters of the multi- stage, compared with the retriever, the multimodal encoder modal adaptive fusion module. The module concatenates the discards the multimodal adaptive fusion module to reduce the visual embedding and text embedding with a dimension of computational cost. R(N+L)×D, where N is the learnable query embeddings and L is the length of text tokens. Image-text matching loss is used to fuse image and text representations, and the results C. Learnable Query Interaction for Multi-images of the fusion module are fed into a binary linear classifier Questions and image captions are used as instructions to foreachoutputquerytoobtainthelogitsandtaketheaverage extract visual features to get learnable query embeddings logitsofallqueriesasthematchingscore.Givenapre-training and input them together with text knowledge to LLMs for dataset X = {I ,T }n , we randomly sample negative texts i i i=1 generation. We employ a novel approach for extracting visual foreachimageandrandomlysamplenegativeimagesforeach information to alleviate the burden of LLMs in distinguishing text, to generate negative training data. Therefore, we denote knowledge. In the original Q-Former in [30], multiple images the ground truth label as y ∈ {1,0} for each image-text pair are processed by employing multiple separate sets of queries (I ,T ), indicating if the input image-text pair is relevant or i i for each, with each set of queries independently extracting not.Weusethemultimodalencoderf (·)toencodeimage-text θ visualfeatures.Thisresultsinusingmultiplequeryfeaturesfor pairs and input it into the multimodal adaptive fusion module generation, which can be computationally intensive and less M (·). The objective function is defined as follows: θ efficientincapturingtheinterrelationsamongdifferentimages. 1 (cid:88) In contrast, our method innovates by succinctly utilizing one L =− ylog(ρ(M (f (I );f (T )))), (2) itm n θ θ i θ i set of learnable queries to directly interact with and extract Ii,Ti∈X features from multiple images in a unified manner. This pro- whereρ(·)isthesoftmaxfunction.Image-groundedTextGen- cessoccursduringtheQ-Formerstage,enablingmoreefficient eration loss trains the fusion module to generate texts, given and integrated interaction among multiple visual references. input images as the condition [10], [30]. For the image-text By employing one set of query interaction approach, RA- pairs in the pre-training dataset, each image I corresponds to BLIP not only simplifies the feature extraction process but a text sentence y ={y ,...,y } of length T. We employ a 1:T 1 T also enhances the efficiency of information extraction. This multimodalcausalself-attentionmaskformultimodalencoder unified interaction allows the model to understand better and f (·)andmultimodaladaptivefusionmoduleM (·)tocontrol θ θ represent the collective information presented in multiple im- the interaction between queries and text. The visual query ages, enabling more effective and cohesive feature utilization, functions as a prefix causal, ensuring that queries can attend especially when dealing with complex scenes or subjects to each other while excluding text tokens. Similarly, each text across multiple images. token y can attend to all visual queries and preceding text tokens. The loss function is defined as: D. Multimodal Adaptive Fusion Module T (cid:88) L =− logM (f ((y |y ,I)). (3) The pre-trained multimodal adaptive fusion module M (·) itg θ θ t <t θ t=1 consists of a 3-layer BERT network [31], [32]. Since the

IEEETRANSACTIONSONMULTIMEDIA															5
								at	the feature	level,	yet it	fails to	understand	the	question
Image-grounded Text
unmasked	masked		Image-Text Matching
--------	------	---	-------------------	---	---	----------	---	---	------	-------	---------	------	--------	-----	-----
						Generation		and	cannot	provide	an answer	in the	semantic	space	[14].
Q: query embeddings
								For	instance,	consider	the	question “The	Marina		Bay Sands
--------------	---	--------------------	---	---	---	----------------	---	---	---------	--------	-------	-------------	-----------	---	---------
T: text tokens	Q	T 🔥🔥MultimodalFusion				MultimodalFusion
					🔥🔥			in	Singapore	is made	of what	building	material?”,		and the
Q
			Feed Forward			Feed Forward
---	---	---	------------	---	---	------------	---	---	---	---	---	---	---	---	---
ITM Attention relevanttextknowledge“Singaporeisalsothenewdowntown
		T			share
---	---	---	---	---	-----	---	---	---	---	---	---	---	---	---	---
Self Attention Self Attention of Singapore, built on reclaimed land.”. The question is about
Q
× N × N the building materials of a Singapore hotel, but this text is
ITG Attention
T about the location of a Singapore hotel. Despite their token-
··· ··· wise similarity in the words, they cannot answer this question
Image andcauseconfusion.Toaddressthisissue,weintroducearank
Q-Former
Encoder strategy to sort the Top-K candidates and exclude confusing
								samples,whereK		isthemaximumnumberofpositivesamples
---	---	---	-----------------	---	---	-----------------------	---	--------------	---	-----------------------------------	---------	------------	-------	-----	--------
				···		Redfoxsittinginthegrass
								corresponding		to the	question.	We select	the	Top-K	samples,
			Learnable Queries			Image Caption
								categorize		them into	positive	and negative	samples		based on
Fig.3. Schematicdiagramofthepre-trainingprocessofmultimodaladaptive thegroundtruth,andthenperformranktraining.SincetheQ-
fusionmodule. Former does not have a classifier, we input the multimodal
(kI;kT;kIT)
								features		output f θ		by	the multimodal		encoder
---	---	---	---	---	---	---	---	--------	---	----------	---	---	--------------	---	-------
intothefixed-parameterLLMsencoderandtrainableclassifier
E. Retrieval-Augmented			Retriever	Training
----------------------------------------------------	-------------	---	---------	-------------	--------	---	------------	---	--------	-------------	----------	----------------------	---	---	-------
								z.	The loss	function	is defined	as:
During	the retrieval		stage,	the retriever	utilizes		the question
												(cid:0) (kI,kT,kIT)),y			(cid:1)
									L	=CrossEntropy		z(f			, (6)
x toretrieverelevantknowledgefrommultimodalknowledge									cls			θ
q
base KB. To achieve this, we apply the multimodal encoder where y is the ground truth about the knowledge is relevant
f θ (·), which	encodes		the question		x q along	with	all latent	or	not.
--------------	---------	---	------------	------------	---------	-----	-----------	---	----	---	---	---	---	---	---
multimodal	knowledge		into	an embedding		space	to identify
the Top-K most relevant candidates, illustrated in Fig. 2 F. Adaptive Selection Knowledge Generation
retrieval stage. We use contrastive learning to construct pos- During the generation stage, the retrieved multimodal
itive and	negative	samples	for	training.	The	knowledge	type
---------	--------	-------	---	---------	---	---------	-------	-----------------------------------	---	---	---	---	---	-------------	---
								knowledgeiscombinedwiththequestionx					q	asanaugmented
							kI, kT,
consists primarily of three components: image text input [k ,...,k ,x ], which is fed to the multimodal encoder
									1	l q
---	---	---	---	---	---	---	---	---	---	---	---	---	---	---	---
and image-text kIT. Thus, the l-th example in the dataset is and LLMs [8] to produce multimodal representation encoding
		,{kˆI,kˆT,kˆIT}			I	T	IT
---	---	---------------	---	---	---	---	---	---	---	---	---	---	---	---	---
represented as (x l ,y l l ,{k ,k ,k } l ), where and generate answers. We observe that existing methods [14],
			i	i i	j	j	j
---	---	---	---	---	---	---	---	---	---	---	---	---	---	---	---
kˆ is the i-th positive (image, text, image-text) sample and [21] directly rely on retrieval results without distinguishing
i
k j represents j-th negative (image, text, image-text) sample. thecorrectnessoftheretrievedknowledge,potentiallyleading
For a batch of knowledge examples, we gather all associ- to the utilization of incorrect, confusing, or irrelevant infor-
ated positive and negative knowledge sources into a batch mation. To address this, we propose an adaptive selection
			I	T IT		I	T IT
---	---	---	---	----	---	---	----	---	---	---	---	---	---	---	---
K = {{kˆI,kˆT,kˆIT} ,{k ,k ,k } ,...,{k ,k ,k } }. knowledge generation (ASKG) strategy based on a question-
B	i	i i	1 j	j j	1	j	j j B
--------------	---	-------	--------------	---	---	--------	--------	----------	---	------------	-----	-------	---	----------	------
								and-answer		formulation,	shown	in Fig.	2	generation	stage.
The multimodal		encoder	is responsible		for	encoding	the mul-
ASKGstrategyenablesthegeneratortogobeyondmereword
timodal	feature	representations		and	aligning	the questions	and
-------	-------	---------------	---	---	--------	-------------	---	---	---	---	---	---	---	---	---
knowledge within the unified semantic space. This alignment similarity between the question and the retrieved knowledge,
								allowing		it to grasp the	semantic	information		of the	question
-----------	-----------	---	-------------	---	-------	---	------------	--------	---	---------------	--------	-----------	---	------	--------
facilitates	identifying		the proximity		between	a	question and
itscorrespondingknowledgethroughcontrastivelearning.The and identify which piece of knowledge contains the answer.
objective function is defined as follows: Specifically, we manually construct question-and-answer data
								to	enable	the model	to discriminate		the relevance		of multi-
---	---	-----	---	---	---------------	---	---	---	------	---------	---------------	---	-------------	---	---------
		exp(f	(x	)·f	(kˆI;kˆT;kˆIT))
θ q θ modalknowledge,therebyutilizingtheimplicitcapabilitiesof
L	=−log	(cid:80)					, (4)
---	-----	--------	---	---	---	---	-----	---	---	---	---	---	---	---	---
con exp(f (x )·f (kI;kT;kIT)) LLMs for knowledge filtering. Based on the original dataset,
				θ q	θ
---	---	---	---	---	---	---	---	---	---	---	---	---	---	---	---
k∈KB
								we	select	relevant knowledge		as positive	examples		and irrel-
---	---	---	---	---	---	---	---	---	------	------------------	---	-----------	--------	---	----------
where f (·) is the multimodal encoder and K is a batch of evant knowledge as negative examples. We create an ASKG
	θ					B
---------	--------	---	---	--------------	---	-------	-------	--------	---	-----------	-------	---	---------	---------	---
								enhanced		dataset and	combine	the	knowledge	according	to
knowledge	sources.	We	use	the multimodal		encoder	trained
to encode text, image, and image-text features, and apply templates, with the identifier of the positive examples serving
								as	the answer.	The	template	for the	question	x	is : “We
---	---	---	---	---	---	---	---	---	-----------	---	--------	-------	--------	---	--------
Maximum Inner Product Search (MIPS) [49] to select Top-K (cid:101)q
from knowledge base KB as the relevant K , as shown in would like to request your feedback on ranking the questions
ret
the following: accordingtotheirrelevancetothereferencesbelow.Relevance
								refers	to	the degree	to which	the reference		can answer	the
------	---	-----------	---	---	-----------------	---	---	------	---	----------	--------	-------------	---	----------	---
TopK(K			x )=TopK{f		(x	)·f (kI;kT;kIT)}.		(5)
ret q θ q θ question. The input format is Question: [content], Reference
k∈KB [knowledge ID]: [content]. The output format is: Related
Although the retriever is more efficient for many retrieval contentis[knowledgeID].”,andtheansweryisintheformof
(cid:101)
tasks, its accuracy is lower on open multimodal question “The most relevant reference is Reference [knowledge ID].”.
answering. There are instances where certain knowledge is We refer to the above enhanced dataset of questions and
confusing and bears token-wise similarity to the question answers as x (cid:101)q and y (cid:101) = {y (cid:101)1 ,...,y (cid:101)M }, where M is the text

IEEETRANSACTIONSONMULTIMEDIA 6 Algorithm 1 The Training pipeline for RA-BLIP. TABLEI Retrieval Training Stage OVERALLDETAILSOFDOWNSTREAMDATASETS. Input: question {x }N , knowledge base KB qi i=1 Dataset Train Dev Test for sampled mini-batch x and K do q B Compute contrastive loss L by Eq. (4) WebQA[4] 34.2K 5K 7.5K con MultimodalQA[5] 23.8K 2.4K 3.6K end for MMCoQA[6] 4.6K 0.6K 0.6K Return retrieval model θ ret Input: question {x }N , ground truth y, Top-K qi i=1 retrieved knowledge from θ ret across text, image, and tabular data types. The perfor- for sampled mini-batch x , and TopK do q B mance of MultimodalQA is measured by F1 score at the Calculate crossentropy loss L by Eq. (6) cls word level and the Exact Match (EM) of the answers. end for • MMCoQA [6] is the first dataset constructed for mul- Return ranking model θ ran timodal conversational QA tasks and aims to answer Generation Training Stage users’ questions with multimodal knowledge sources via Input: question {x }N , answer y, K from θ , qi i=1 ret ran multi-turnconversations.Itcomprisesmultiplesupervised ASKG datasets x and y (cid:101)q (cid:101) signals, including decontextualized questions, answers, for sampled mini-batch x , K and x do q ret (cid:101)q and corresponding evidence. Compute generation loss L by Eq. (7) gen 2) Compared Methods: For WebQA, MultimodalQA, and end for MMCoQA, we make comparisons with different baseline Return generation model θ gen methods. The model parameter quantity comparison is shown in Table II. The number of Solar parameters is not pub- lished, and both Solar and SKURG use other models to length.Giventhedatasetquestionx andground-turthanswer q exploit multimodal information without accounting for the oflengthT,y ={y ,...,y },aswellastheconstructedx 1:T 1 T (cid:101)q parameter counts of other models. We have frozen LLM and and y, the generator g (·) utilizes attention over question x (cid:101) θ q ViT, focusing solely on training Q-Former, which has fewer and relevant knowledge K encoded by multimodal encoder ret trainable parameters and bfloat16 encoding. In order to verify f (·) to generate textual outputs token by token. The final θ thescalinglaw[51],weselectedmorepowerfulFlanT5xxlfor generation loss is defined by: experiment. To compare LLMs with other methods of similar parameter magnitude, we utilized T5-base and T5-large as T L = (cid:88) −logg (y |y ,f (x ,K )) benchmarksforafaircomparison.SinceT5-baseandT5-large gen θ i 1:i−1 θ q ret arenotalignedwiththemodelthroughpre-training,theyneed i=1 (7) M to be fine-tuned during training. (cid:88) +α −logg θ (y (cid:101)i |y (cid:101)1:i−1 ,f θ (x (cid:101)q )), • VLP [4], [52] pre-trains its transformer-based encoder- i=1 decoder with both textual and visual information. They first retrieve knowledge based on the question and feed where α is the hyperparameter which will be discussed in it into the model to generate answers. In addition, VLP section IV-D. To give a clear illustration of RA-BLIP, we integrates VinVL [41] to improve performance. summarize the training pipeline in Algorithm 1. • MuRAG [14] encodes the question and selects Top-K nearest neighbors from multimodal memory. They are IV. EXPERIMENTS then fed into the backbone encoder-decoder to generate textualoutputstokenbytoken.Thebackbonemodeluses A. Experimental Settings T5-base [53] and ViT-large [48], respectively.

Datasets: WeevaluateourmethodonthreeQAdatasets: • SKURG [20] takes multimodal information sources as WebQA [4], MultimodalQA [5], and MMCoQA [6]. The inputandencodesthemseparately,thenutilizesanentity- details of these datasets are showcased in Table I. centered fusion encoder to align the sources of different modalities via the shared entities and structured knowl- • WebQA [4] is a large-scale dataset for multimodal and edge. The method adopts OFA-base [54] and BART-base multihop QA where all questions are knowledge-seeking [55]. Besides, it integrates ELMo-based NER [56] and queries that require two or more knowledge sources. OpenNRE [57] for entity and relation extraction. Evaluation metrics are retrieval F1 and QA for assessing answer generation quality, which is measured as both • Solar [21] first converts multimodal inputs into textual data and then utilizes a T5 [53] to generate answers fluency (QA-FL) and accuracy (QA-ACC). We calculate through retrieval, ranking, and decoding. It retrieves and fluency through BARTScore [50] and evaluate accuracy ranks the information using BERT [32]. Additionally, via F1 and recall. The fluency score and accuracy score are multiplied FL∗Acc to calculate the overall score. it adopts BLIP [58] for image caption generation and VinVL [41] for image-attribute feature extraction. • MultimodalQA [5] is a collection of multihop QA pairs thatnecessitatethefusionofknowledgefromtext,tables, 3) Implementation Details: Our method includes multi- and images. This dataset requires retrieval and reasoning modal fusion pre-training, retrieval, ranking, and generation.

IEEETRANSACTIONSONMULTIMEDIA															7
			TABLEII								TABLEIII
COMPARISONOFPARAMETERQUANTITY.PARAMETERQUANTITIESOF RESULTSOFWEBQAOFFICIALTEST-SET.∗REPRESENTSLLMIS
OTHERMETHODSREFERTO[20],[21]. FLANT5XL,WHILE†REPRESENTSLLMISFLANT5XXL.BOLDAND
UNDERLINEDENOTETHEBESTANDPREVIOUSSOTARESULTS.
Model			#TrainableParams			#TotalParams
------------------	---	---	----------------	-----	---	------------	---	-----------------	---	---	--------	------	-------	----	----
VLP+VinVL[4]				220M		220M		Model			Retr-F1↑	QA-FL↑	QA-Acc↑		QA↑
MuRAG[14]				527M		527M		VLP[52]			0.69	42.6		36.7	22.6
SKURG[20]				447M		447M		VLP+VinVL[41]			0.71	44.2		38.9	24.1
ImplicitDecomp[5]				1310M		1310M		MuRAG[14]			0.75	55.7		54.6	36.1
RA-BLIP(T5-base)				387M		1398M		SKURG[20]			0.88	55.4		57.1	37.7
RA-BLIP(T5-large)				902M		1913M
								Solar[21]			0.89	60.9		58.9	40.9
RA-BLIP(FlanT5xl)				109M		4.1B
								InstructBLIP∗[30]			-	51.7		59.0	31.4
RA-BLIP(FlanT5xxl)				109M		12.1B
								InstructBLIP†[30]			-	53.4		62.5	35.0
								RA-BLIP(T5-base)			-	62.6		59.7	41.6
								RA-BLIP(T5-large)			-	62.9		60.9	42.5
For WebQA and MultimodalQA, we follow all steps as men- RA-BLIP∗ 0.83 65.1 65.3 45.8
								RA-BLIP†			0.89	65.5		68.7	48.5
---	---	---	---	---	---	---	---	--------	---	---	----	----	---	----	----
tionedabove.ForMMCoQA,wemanuallyincludethepositive
clues in	the retrieval		results	without	performing		subsequent
--------	-------------	---	-------	-------	-----------	---	----------	-------	--------	----	-----	------	------------	----	-------
ranking,	following	the	process	used	in previous		work [21].
								used in	previous	work	[21].	We set	the learning	rate	as 1e-5
WeadoptInstructBLIP[30]withfrozenEVA-ViT-g/14[59]as
andbatchsizeas32for10epochsattheretrievalstage.Then,
well as	LLM (FlanT5xl		and	FlanT5xxl)	[8]	for generation.	In
-------	-------------	---	---	----------	---	---------------	---	---	---	---	---	---	---	---	---
order to compare LLMs with fewer than 1 billion parameters, we use a cosine learning rate of 1e-6, warmup of 1K steps,
								and a batch	size	of 16	for 10	epochs	at the generation		stage.
-----------	---	---	--------	----	--------	---	--------	-----------	----	-----	------	------	-----------------	---	------
we replaced	the	LLM	backbone	with	T5-large	and	T5-base.
DuetotheinconsistentdimensionsbetweenT5andQ-Former, We use the standard evaluation protocol for each dataset and
reportthesamemetrics,aswellastherandomseedsarefixed
we added	a linear	layer	to	align the	dimensions		and did not
--------	--------	-----	---	---------	----------	---	-----------	---	---	---	---	---	---	---	---
for reproducibility.
perform	pre-training.		We keep	the	image	encoder	and the
------------	-------------	------	--------	--------	-----	-------	----------	---	---	---	---	---	---	---	---
LLMs frozen,		tuning	only the	Q-Former		and the	multimodal
B. Main Results
fusion module		during	the pre-training		and	retrieval	stage. We
-------------	---	------	----------------	---	---	---------	---------	---	---	---	---	---	---	---	---
adopt the multimodal encoder and LLM encoder as feature Results on WebQA.
											We	show	the WebQA		results in
---	---	---	---	---	---	---	---	---	---	---	---	----	---------	---	----------
encoder at ranking stage. At the generation stage, we froze Table III. We can see that RA-BLIP surpasses all baselines in
the image encoder as well as LLMs and only trained the Q- termsofbothQAandretrievalF1scores.RA-BLIP(FlanT5xl)
Former	with ASKG.	We	froze	the	image encoder		and FlanT5
------	----------	---	-----	---	-------------	---	----------	--------	-----	---------	-----	---	-----	------	--------
								achieves	45.8%	accuracy,	which	is	+4.9%	higher	than the
duringalltrainingprocesses,aswellasusedbfloat16encoding state-of-the-art Solar [21]. Especially the metric QA-Acc is
to achieve RA-BLIP with the fewest trainable parameters. +6.4% higher, proving the model’s powerful generation abil-
For pre-training, we pre-train the multimodal adaptive fu- ity.Besides,RA-BLIP(FlanT5xxl)beatsSOTASolarby7.6%
sion module on the SBU dataset [60]. Our approach is to onoverallQAaccuracy,whichshowsthatRA-BLIPcomplies
fix the parameters		of	the image	encoder		and Q-Former,	and
------------------	---	---	---------	-------	---	-------------	---	---	---	---	---	---	---	---	---
withscalinglaw[51]andcanimprovethegenerationaccuracy
solely fine-tune the parameters of the multimodal adaptive by using more advanced LLM. In order to prove that it is our
fusionmodule.WeusetheAdamWoptimizerandadoptcosine RA-BLIP framework rather than the advanced MLLM back-
learning rate of 1e-5, warmup of 1K steps, and batch size 64 bone that improves the generative performance, we conducted
for 10 epochs. For fine-tuning, we use the AdamW optimizer experiments on InstructBLIP [30] on WebQA. We used RA-
with β	= 0.9,	β	= 0.99,	and	a weight	decay	of 0.01
------	------	---	-------	---	--------	-----	-------	---	---	---	---	---	---	---	---
1 2 BLIP’s optimal 0.89 search result for InstructBLIP generation
uniformly for three datasets. For WebQA retrieval, we use and found that its accuracy was 5.9% lower than Solar, which
a cosine learning rate of 1e-5, warmup of 1K steps, and batch further proves the effectiveness of RA-BLIP framework and
size 4 for 10 epochs. For WebQA ranking, we select the top ASKG.RA-BLIP(T5-base)and(T5-large)arenotpre-trained
10 samples to train the model with a cosine learning rate of to align with Q-Former, but they achieve 41.6% and 42.5%
1e-5, warmupof		1Ksteps	and	batchsize	40	for 5	epochs. For
--------------	---	-------	---	---------	---	-----	-----------	--------	------------	---	-----	-------	---------------	---	----------
								accuracy	respectively		based	on 0.89	search results,		surpassing
generation, we adopt cosine learning rate of 1e-6, warmup of Solar and proving it is the RA-BLIP framework rather than
1K steps, and batch size 4 for 10 epochs. We set learning rate LLM that improves performance. Compared with Solar and
of 5e-5 for T5-large and T5-base. SKURG which require additional model assistance, RA-BLIP
For MultimodalQA [5] and MMCoQA [6], we tested the does not use additional models, but it also achieves very good
results on the dev set of MultimodalQA as well as the dev 14%
								results in	retrieval	and	is	higher	than	MuRAG,	which
---	---	---	---	---	---	---	---	----------	---------	---	---	------	----	------	-----
and test sets of the MMCoQA. For MultimodalQA retrieval, similarly does not use additional models.
weuseacosinelearningrateof1e-5,warmupof1Ksteps,and Results on MultimodalQA. We demonstrate Multi-
a batch size 4 for 10 epochs. For MultimodalQA generation, modalQA results in Table IV. MultimodalQA contains tables
we adopt a cosine learning rate of 1e-6, warmup of 1K steps, and has many multihop questions that require combining
and a batch size of 8 for 10 epochs. We set learning rate multimodal information. RA-BLIP also improved EM and
of 1e-5 for T5-large and 5e-5 for T5-base. For MMCoQA, F1 by 6.0% and 6.6%, respectively, compared to state-of-
we manually include the positive clues in the retrieval results the-art Solar, which demonstrates the generative ability of
without performing subsequent ranking,following the process our method in incorporating multihop knowledge. In addition,

IEEETRANSACTIONSONMULTIMEDIA 8 TABLEIV MULTIMODALQADEV-SETRESULTS.∗REPRESENTSLLMISFLANT5XL, WHILE†REPRESENTSLLMISFLANT5XXL.SINGLE-MODALAND MUTLI-MODALRESPECTIVELYINDICATEWHETHERREASONINGRELIES ONSINGLE-MODALORMUTLI-MODALKNOWLEDGE.BOLDAND UNDERLINEDENOTETHEBESTANDSOTARESULTS,RESPECTIVELY. Model Single-Modal Mutli-Modal All EM F1 EM F1 EM F1 AutoRouting[5] 51.7 58.5 34.2 40.2 44.7 51.1 ImplicitDecomp[5] 51.6 58.4 44.6 51.2 48.8 55.5 )ODQ7[O )ODQ7[[O SKURG[20] 66.1 69.7 52.5 57.2 59.8 64.0 Solar[21] 69.7 74.8 55.5 65.4 59.8 66.1 RA-BLIP(T5-base) 65.4 71.6 59.7 65.7 63.1 69.3 RA-BLIP(T5-large) 65.2 71.9 62.6 68.4 64.1 70.5 RA-BLIP* 70.1 77.6 59.3 65.5 65.8 72.7 RA-BLIP† 69.9 76.4 59.1 65.6 65.6 72.1 TABLEV MMCOQATEST-DEV-SETRESULTS.∗REPRESENTSLLMISFLANT5XL, WHILE†REPRESENTSLLMISFLANT5XXL.BOLDANDUNDERLINE DENOTETHEBESTANDSOTARESULTS,RESPECTIVELY. Dev Test Model EM F1 EM F1 ORConvQA[61] 1.0 3.0 1.0 1.9 ManyModelQA[62] 0.7 2.3 1.0 1.8 MAE[6] 21.5 30.2 24.9 32.3 Solar[21] 56.8 62.5 57.3 64.6 RA-BLIP* 59.2 67.1 61.0 67.8 RA-BLIP† 58.7 66.7 59.5 66.7 RA-BLIP’s accuracy is ahead of SOTA Solar in both single- modality and multi-modality, demonstrating our model can well combine multiple contextual semantic knowledge for cross-modal reasoning. Both RA-BLIP (T5-large) and RA- BLIP(T5-base)surpassSolar,indicatingthattheperformance improvements are due to the RA-BLIP framework rather than the underlying LLM. Notably, the accuracy of more powerful FlanT5xxl is lower than that of FlanT5xl, probably because the powerful LLM is overfitted. Results on MMCoQA. Our results on MMCoQA are shown in Table V. Compared with WebQA and Multi- modalQA,MMCoQArequiresthemodeltocorrectlyincorpo- ratedialoghistoryanddevelopdeepmultimodalunderstanding and reasoning capabilities across multiple conversations. RA- BLIP achieves a margin of 3.7% enhancement over the best Solar for the EM score and 3.2% for the F1 score in the test split. These results suggest the generalization ability and versatility of our model. Due to the small dataset, RA- BLIP (FlanT5xxl) has resulted in overfitting, which prevents further performance improvement. C. Ablation Study To analyze the effectiveness of our proposed method, we conducted comprehensive ablations on the WebQA dataset in both retrieval and generation stages. As shown in Fig. 4 (a), RA-BLIP (FlanT5xl) and RA-BLIP (FlanT5xxl) with ASKG strategy improved by 1.9% and 2.7% respectively, which \FDUXFF$$4 D$.6* 5$%/,3ZR$6.* 5$%/,3Z$6.* )ODQ7[O )ODQ7[[O \FDUXFF$$4 E,QWHUDFWLRQTXHULHV 0XOWLSOHVHWVRITXHULHV 2QHVHWRITXHULHV 5HWULHYHU5HWULHYHU5HWULHYHU ZRIXVLRQZIXVLRQ Z5DQN )UWH5 F5HWULHYDO Fig.4. AblationstudyforRA-BLIPgenerationandretrievalonWebQA. +\SHUSDUDPHWHU \FDUXFF$$4 )ODQ7[O Fig.5. Influenceofvaryingthehyperparameterαofgenerationloss. validates that employing ASKG strategy can assist LLMs in efficientlydiscerningtherelevanceofretrievedknowledgeand effectively activate the implicit capabilities of more powerful LLMs. In Fig. 4 (b), we compared the effects of one set of queries and multiple sets of queries, where one set of queries has a significant improvement. This proves that compared to multiple sets of queries that simply concatenate visual informationfromdifferentimages,onesetofqueriescanbetter interactwithandextractmixedvisualinformationfrommulti- ple images at the feature level. We demonstrated the ablation results of RA-BLIP at the retrieval stage in Fig. 4 (c). The result of the retriever without the multimodal adaptive fusion module exhibits lower performance than that of the complete retriever, which suggests that fusing vision and language in a unified semantic space is essential for multimodal retrieval. By utilizing a retrieval-rank strategy, the retrieval result is significantly improved 20%, demonstrating the necessity of denoising confusing knowledge through fine-level ranking. The retrieval-rank process employed in our method facilitates precise retrieval among candidate knowledge sources that are primarily relevant but potentially confusing. D. Sensitivity Analysis The α is the trade-off hyperparameter of generation loss with ASKG strategy in Eq. (7). We set the range of α from 0.01 to 5. According to Fig. 5, we can see that even for such a large range, the difference between the best and the lowest resultsislessthan0.04%,indicatingourmethodisrobustand insensitive to this parameter.

IEEETRANSACTIONSONMULTIMEDIA														9
							RA-BLIP.	In the	future,	we	will explore	image-multimodal
(1) Q: What is in the man's mouth in L'hommeà la Tulipe?
Retrieval retrieval and multimodal-multimodal retrieval to realize om-
		Multimodal Knowledge Base					nipotent	retrieval-augmented		models.
---	-------------------	-------------------------	-------------------------------	---	---	---	--------	-------------------	---	-------	---	---	---	---
	L'homme à la tulipe		HerfilmdébutwasinPasdepitiépour
lesfemmes(1951),followedbyFanfan
laTulipe(1952),inwhichsheplayed
			Madame	de Pompadour	alongside					REFERENCES
---	---	---	------	------------	---------	---	---	---	---	----------	---	---	---	---
GérardPhilipeandGinaLollobrigida.
Sincethen,shehasappearedinItalian, [1] S.Antol,A.Agrawal,J.Lu,M.Mitchell,D.Batraetal.,“VQA:visual
French,BritishandAmericanfilms.
							question	answering,”		in Proceedings	of	the IEEE/CVF	International
---	-----------	---	---	-----------	---	---	--------	-----------	---	--------------	---	------------	-------------	---
	Reference 1			Reference 2
ConferenceonComputerVision,2015,pp.2425–2433.
ASKG Output: The relevant reference is Reference 1. [2] Y.Goyal,T.Khot,D.Summers-Stay,D.Batra,andD.Parikh,“Making
SKURG: In L'hommeà la Tulipe, there are flowersin the man's mouth. the V in VQA matter: Elevating the role of image understanding in
visualquestionanswering,”inProceedingsoftheIEEE/CVFConference
RA-BLIP: A cigaris in the man's mouth in L'hommeà la Tulipe. onComputerVisionandPatternRecognition,2017,pp.6325–6334.
							[3] K. Lee,	M. Chang,	and	K. Toutanova,		“Latent	retrieval for	weakly
---	---	---	---	---	---	---	-----------	---------	---	-------------	---	-------	-------------	------
Ground Truth: A cigarette is in the man's mouth in L'hommeà la Tulipe.
							supervised	open	domain	question	answering,”	in	Proceedings	of the
---	---	---	---	---	---	---	----------	----	------	--------	-----------	---	-----------	------
AnnualMeetingoftheAssociationforComputationalLinguistics,2019,
(2)	Q: How many years after Pyramid began airing did Ransom's third						pp.6086–6096.
---	----------------------------------------------------------------	---	---	---	---	---	-------------	---	---	---	---	---	---	---
season begin airing? [4] Y.Chang,G.Cao,M.Narang,J.Gao,H.Suzuki,andY.Bisk,“Webqa:
Retrieval
MultihopandmultimodalQA,”inProceedingsoftheIEEE/CVFCon-
Multimodal Knowledge Base
ferenceonComputerVisionandPatternRecognition,2022,pp.16474–
	Cody Ransom with Yankees		Theshowwasquicklypickedupby
---	------------------------	---	---------------------------	-----	--------------	---	------	---	---	---	---	---	---	---
			ABC and	began	airing on that		16483.
[5] A.Talmor,O.Yoran,A.Catav,D.Lahav,Y.Wangetal.,“Multimodalqa:
networkonMay6,1974.OnJuly16,
2018, CBS and Global announced complexquestionansweringovertext,tablesandimages,”inProceed-
thattheserieshasbeenrenewedfora
ingsoftheInternationalConferenceonLearningRepresentations,2021.
			13-episode	third	season, which
---	---	---	----------	-----	-------------	---	---	---	---	---	---	---	---	---
premieredonFebruary16,2019. [6] Y.Li,W.Li,andL.Nie,“Mmcoqa:Conversationalquestionanswering
Reference 1 Reference 2 overtext,tables,andimages,”inProceedingsoftheAnnualMeetingof
theAssociationforComputationalLinguistics,2022,pp.4220–4231.
ASKG Output: The relevant reference is Reference 2. [7] S.Wu,G.Zhao,andX.Qian,“Resolvingzero-shotandfact-basedvisual
SKURG: It aired 11 years later. questionansweringviaenhancedfactretrieval,”IEEETransactionson
Multimedia,vol.26,pp.1790–1800,2024.
RA-BLIP: It was 45 years after Pyramid began airing that Ransom’s
							[8] H. W.	Chung,	L. Hou,	S.	Longpre,	B. Zoph,	Y. Tay	et al.,
---	---	---	---	---	---	---	---------	------	-------	---	--------	--------	------	-------
third season began airing. “Scaling instruction-finetuned language models,” arXiv preprint
Ground Truth: It aired 45 years after. arXiv:2210.11416,2022.
							[9] H. Touvron,		T. Lavril,	G. Izacard,	X.	Martinet,	M. Lachaux	et al.,
---	---	---	---	---	---	---	---------------	---	----------	-----------	---	---------	----------	-------
“Llama:Openandefficientfoundationlanguagemodels,”arXivpreprint
Fig. 6.	QA Examples.		We demonstrate	ASKG,	RA-BLIP	answers,
-------	------------	---	--------------	-----	-------	--------	---	---	---	---	---	---	---	---
arXiv:2302.13971,2023.
SKURG[20]andgroundtruth.Relevantknowledgeisingreenwindowand [10] J. Li, D. Li, S. Savarese, and S. C. H. Hoi, “BLIP-2: bootstrapping
irrelevantisinredwindow.Relevanttextindocumentisunderlined. language-image pre-training with frozen image encoders and large
							language	models,”	in	Proceedings	of the	International	Conference	on
---	---	---	---	---	---	---	--------	--------	---	-----------	------	-------------	----------	---
MachineLearning,2023,pp.19730–19742.
E. Qualitative Results [11] OpenAI, “GPT-4 technical report,” arXiv preprint arXiv:2303.08774,
2023.
Fig. 6 illustrates four examples obtained by RA-BLIP and [12] A. Chowdhery, S. Narang, J. Devlin, M. Bosma, G. Mishra et al.,
the baseline SKURG [20]. The retrieved relevant and con- “Palm:Scalinglanguagemodelingwithpathways,”JournalofMachine
fusing knowledge is listed, where the green boxes indicate LearningResearch,vol.24,pp.240:1–240:113,2023.
[13] K.Guu,K.Lee,Z.Tung,P.Pasupat,andM.Chang,“Realm:Retrieval
the positive clue and the red boxes represent the negative augmentedlanguagemodelpre-training,”inProceedingsoftheInterna-
cue. RA-BLIP outputs the correct answer, while SKURG tionalConferenceonMachineLearning,2020,pp.3929–3938.
							[14] W. Chen,	H.	Hu, X.	Chen, P.	Verga,	and W.	W. Cohen,	“Murag:
---------	---	------------	-----	---	---------	-----------	-------------	---	------	--------	------	------	---------	-------
generates	the	wrong answer	under	the	identical	conditions.
Multimodalretrieval-augmentedgeneratorforopenquestionanswering
This indicates		RA-BLIP	has more	powerful	abilities	to com-
--------------	---	-------	--------	--------	---------	-------	----	----------	------	--------------	---	--------------	------------	---
							over	images and	text,”	in Proceedings	of	the Conference	on Empirical
prehensivelyunderstandtheretrievalinformation,regardlessof MethodsinNaturalLanguageProcessing,2022,pp.5558–5570.
[15] Z.Hu,A.Iscen,C.Sun,Z.Wang,K.Changetal.,“Reveal:Retrieval-
textualorvisualmodality.Besides,throughASKG,ourmodel
							augmented	visual-language		pre-training		with multi-source	multimodal
---	---	---	---	---	---	---	---------	---------------	---	------------	---	-----------------	----------	---
autonomouslyjudgestherelevanceofretrievedknowledgeand
							knowledge	memory,”	in	Proceedings	of	the IEEE/CVF	Conference	on
----------------	---	-------	-------------	--------	--------	---	---------	--------	---	-----------	---	------------	----------	---
selects relevant		ones to	generate more	accurate	answers.
ComputerVisionandPatternRecognition,2023,pp.23369–23379.
							[16] J. Rao,	Z. Shan,	L.	Liu, Y.	Zhou, and	Y. Yang,	“Retrieval-based
---	---	---	----------	---	---	---	------------	---------	---	---------------	---------	--------------	----------------	---
							knowledge	augmented		vision language		pre-training,”	in International
		V.	CONCLUSION
ConferenceonMultimedia,2023,pp.5399–5409.
In this paper, we propose a novel multimodal adaptive [17] P. S. H. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin et al.,
							“Retrieval-augmented			generation	for knowledge-intensive		NLP	tasks,”
---	---	---	---	---	---	---	--------------------	---	---	----------	-----------------------	---	---	-------
Retrieval-Augmented BLIP (RA-BLIP), a general retrieval- Advances in Neural Information Processing Systems,
							in						vol. 33,	2020,
--------------------------------------------------	---	---	---	---	---	---	-------------	---	---	---	---	---	--------	-----
augmentedframeworkforvariousclassicalMLLMs.RA-BLIP							pp.9459–9474.
utilizes questions as instructions to extract visual features [18] G.IzacardandE.Grave,“Leveragingpassageretrievalwithgenerative
modelsforopendomainquestionanswering,”inProceedingsoftheCon-
for less	irrelevant	interference.	It	incorporates	a	pre-trained
--------	----------	-------------	---	------------	---	-----------	---	---	---	---	---	---	---	---
ferenceoftheEuropeanChapteroftheAssociationforComputational
multimodal adaptive fusion module to efficiently integrate Linguistics,2021,pp.874–880.
							[19] S. Borgeaud,		A. Mensch,	J. Hoffmann,		T. Cai,	E. Rutherford	et al.,
-----------	----	----	----------	-------	-----------	-------	-----------------	---	----------	------------	---	-------	-------------	-------
information	from	both	visual and	textual	modalities,	thereby
“Improvinglanguagemodelsbyretrievingfromtrillionsoftokens,”in
achieving	question	text-to-multimodal		retrieval.	Additionally,
---------	--------	------------------	---	----------	-------------	---	---	---	---	---	---	---	---	---
ProceedingsoftheInternationalConferenceonMachineLearning,vol.
weintroduceanadaptiveselectionknowledgegenerationstrat-
162,2022,pp.2206–2240.
egytomakethegeneratorautonomouslydiscerntherelevance [20] Q. Yang, Q. Chen, W. Wang, B. Hu, and M. Zhang, “Enhancing
multi-modalmulti-hopquestionansweringviastructuredknowledgeand
of retrieved	knowledge.		Extensive	experiments	on	multimodal
------------	----------	---	---------	-----------	---	----------	---	---	---	---	---	---	---	---
unifiedretrieval-generation,”inInternationalConferenceonMultimedia,
multihop QA and ablation studies verify the effectiveness of 2023,pp.5223–5234.

IEEETRANSACTIONSONMULTIMEDIA															10
[21] B. Yu, C. Fu, H. Yu, F. Huang, and Y. Li, “Unified language rep- [42] M.Tsimpoukelli,J.Menick,S.Cabi,S.M.A.Eslami,O.Vinyals,and
resentation for question answering over text, tables, and images,” in F.Hill,“Multimodalfew-shotlearningwithfrozenlanguagemodels,”in
ProceedingsoftheAnnualMeetingoftheAssociationforComputational AdvancesinNeuralInformationProcessingSystems,2021,pp.200–212.
Linguistics,2023,pp.4756–4765. [43] Y.Xing,Q.Wu,D.Cheng,S.Zhang,G.Liang,P.Wang,andY.Zhang,
[22] F.Liu,J.Liu,Z.Fang,R.Hong,andH.Lu,“Visualquestionanswering “Dual modality prompt tuning for vision-language pre-trained model,”
withdenseinter-andintra-modalityinteractions,”IEEETransactionson IEEETransactionsonMultimedia,vol.26,pp.2056–2068,2024.
								[44] X. Zhai,	X. Wang,	B.	Mustafa,	A. Steiner,	D.	Keysers	et al., “Lit:
---	---	---	---	---	---	---	---	-------------	--------	---	--------	-----------	---	-------	-------------
Multimedia,vol.23,pp.3518–3529,2021.
Zero-shottransferwithlocked-imagetexttuning,”inProceedingsofthe
[23] Z. Liu,	C. Xiong,	Y.	Lv, Z.	Liu,	and G.	Yu, “Universal	multi-
------------	---------	---	------	----	------	--------------	------	---	---	---	---	---	---	---	---
modality retrieval with one unified embedding space,” arXiv preprint IEEE/CVF Conference on Computer Vision and Pattern Recognition,
arXiv:2209.00179,2022.								2022,pp.18102–18112.
----------------------	---	---	---	---	---	---	---	--------------------	---	---	---	---	---	---	---
[24] T.Chen,S.Kornblith,M.Norouzi,andG.E.Hinton,“Asimpleframe- [45] J. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr et al., “Flamingo: a
								visual	language	model	for few-shot	learning,”	in	Advances	in Neural
---	---	---	---	---	---	---	---	------	--------	-----	------------	----------	---	--------	---------
workforcontrastivelearningofvisualrepresentations,”inProceedings
InformationProcessingSystems,vol.35,2022,pp.23716–23736.
of the	International	Conference		on Machine	Learning,		vol. 119, 2020,
------	-------------	----------	---	----------	---------	---	---------------	---	---	---	---	---	---	---	---
pp.1597–1607. [46] P. Qi, H. Lee, T. Sido, and C. D. Manning, “Answering open-domain
[25] X. Lu, T. Zhao, and K. Lee, “Visualsparta: An embarrassingly simple questions of varying reasoning steps from text,” in Proceedings of the
approach to large-scale text-to-image search with weighted bag-of- Conference on Empirical Methods in Natural Language Processing,
2021,pp.3599–3614.
words,”	in Proceedings		of the	Annual	Meeting	of the	Association for
-------	--------------	---	------	------	-------	------	---------------	---	---	---	---	---	---	---	---
[47] S.Yavuz,K.Hashimoto,Y.Zhou,N.S.Keskar,andC.Xiong,“Mod-
ComputationalLinguistics,2021,pp.5020–5029.
								eling	multi-hop	question	answering	as single	sequence	prediction,”	in
------------	------	------	------------	---	-------	---------	------------	-----	---------	--------	---------	---------	--------	------------	---
[26] H. Liu,	T. Yu,	and P.	Li, “Inflate	and	shrink:	Enriching	and reducing
ProceedingsoftheAnnualMeetingoftheAssociationforComputational
interactionsforfasttext-imageretrieval,”inProceedingsoftheConfer- Linguistics,2022,pp.974–990.
enceonEmpiricalMethodsinNaturalLanguageProcessing,2021,pp.
[48] A.Dosovitskiy,L.Beyer,A.Kolesnikov,D.Weissenborn,X.Zhaietal.,
9796–9809.
								“An	image is worth	16x16	words:	Transformers		for image	recognition
---	---	---	---	---	---	---	---	---	--------------	-----	------	------------	---	---------	-----------
[27] D.Lin,Y.Ma,Y.Li,X.Song,J.Wu,andL.Nie,“OFAR:Amultimodal
								at scale,”	in Proceedings		of the	International	Conference		on Learning
---	---	---	---	---	---	---	---	----------	--------------	---	------	-------------	----------	---	-----------
evidenceretrievalframeworkforillegallive-streamingidentification,”in
Representations,2021.
Proceedings of the International ACM SIGIR Conference on Research [49] R.Guo,P.Sun,E.Lindgren,Q.Geng,D.Simchaetal.,“Accelerating
andDevelopmentinInformationRetrieval,2023,pp.3410–3414. large-scaleinferencewithanisotropicvectorquantization,”inProceed-
[28] J.Yu,W.Zhang,Y.Lu,Z.Qin,Y.Hu,J.Tan,andQ.Wu,“Reasoning ings of the International Conference on Machine Learning, vol. 119,
on the	relation:	Enhancing	visual	representation		for visual	question
------	---------	---------	------	--------------	---	----------	--------	---	---	---	---	---	---	---	---
2020,pp.3887–3896.
answeringandcross-modalretrieval,”IEEETransactionsonMultimedia,
[50] W.Yuan,G.Neubig,andP.Liu,“Bartscore:Evaluatinggeneratedtextas
vol.22,pp.3196–3209,2020. textgeneration,”inAdvancesinNeuralInformationProcessingSystems,
[29] J.Zhuang,J.Yu,Y.Ding,X.Qu,andY.Hu,“Towardsfastandaccurate vol.34,2021,pp.27263–27277.
image-textretrievalwithself-supervisedfine-grainedalignment,”IEEE [51] J. Kaplan, S. McCandlish, and Henighan, “Scaling laws for neural
TransactionsonMultimedia,vol.26,pp.1361–1372,2024.
languagemodels,”arXivpreprintarXiv:2001.08361,2020.
[30] W. Dai,	J. Li,	D. Li,	A. M.	H. Tiong,	J.	Zhao, W.	Wang, B. Li,
------------	------	------	-----	---------	---	--------	------------	-------------	-----------	---	---------	------	-----	------	-----------
								[52] L. Zhou,	H. Palangi,		L. Zhang,	H. Hu,	J. J.	Corso,	and J. Gao,
P. Fung, and S. C. H. Hoi, “Instructblip: Towards general-purpose “Unifiedvision-languagepre-trainingforimagecaptioningandVQA,”
vision-languagemodelswithinstructiontuning,”inAdvancesinNeural inProceedingsoftheAAAIConferenceonArtificialIntelligence,2020,
InformationProcessingSystems,2023.								pp.13041–13049.
----------------------------------	---	---	---	---	---	---	---	---------------	---	---	---	---	---	---	---
[31] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones et al., “At- [53] C.Raffel,N.Shazeer,A.Roberts,K.Lee,S.Narangetal.,“Exploring
tentionisallyouneed,”inAdvancesinNeuralInformationProcessing the limits of transfer learning with a unified text-to-text transformer,”
Systems,2017,pp.5998–6008. JournalofMachineLearningResearch,vol.21,pp.140:1–140:67,2020.
[32] J.Devlin,M.Chang,K.Lee,andK.Toutanova,“BERT:pre-trainingof [54] P.Wang,A.Yang,R.Men,J.Lin,S.Baietal.,“OFA:unifyingarchi-
deepbidirectionaltransformersforlanguageunderstanding,”inProceed- tectures, tasks, and modalities through a simple sequence-to-sequence
ingsoftheConferenceoftheNorthAmericanChapteroftheAssociation learningframework,”inProceedingsoftheInternationalConferenceon
forComputationalLinguistics,2019,pp.4171–4186. MachineLearning,2022,pp.23318–23340.
								[55] M.	Lewis, Y. Liu,	N. Goyal,	M.	Ghazvininejad,		A. Mohamed	et al.,
--------------	--------	---	------	------	------	--------------	-------	-------	--------------	---------	---	--------------	---	----------	-------
[33] X. Zheng,	Z. Wang,	S.	Li, K.	Xu, T.	Zhuang	et al., “MAKE:	vision-
language pre-training based product retrieval in taobao search,” in “BART: denoising sequence-to-sequence pre-training for natural lan-
ProceedingsoftheACMWebConference,2023,pp.356–360. guage generation, translation, and comprehension,” in Proceedings of
[34] L. Yu, J. Chen, A. Sinha, M. Wang, Y. Chen et al., “Commercemm: the Annual Meeting of the Association for Computational Linguistics,
Large-scalecommercemultimodalrepresentationlearningwithomnire-								2020,pp.7871–7880.
--------------------------------------------------------------	---	---	---	---	---	---	---	------------------	----------	---------	---	------------	---	---------	------
								[56] M.	E. Peters,	W. Ammar,	C.	Bhagavatula,	and	R. Power,	“Semi-
trieval,”inProceedingsoftheACMSIGKDDConferenceonKnowledge
								supervised	sequence	tagging	with	bidirectional		language	models,” in
---	---	---	---	---	---	---	---	----------	--------	-------	----	-------------	---	--------	-----------
DiscoveryandDataMining,2022,pp.4433–4442.
ProceedingsoftheAnnualMeetingoftheAssociationforComputational
[35] C. Jia,	Y. Yang,	Y. Xia,	Y. Chen,	Z.	Parekh et	al., “Scaling	up visual
------------	--------	-------	--------	---	---------	-------------	---------	---	---	---	---	---	---	---	---
andvision-languagerepresentationlearningwithnoisytextsupervision,” Linguistics,2017,pp.1756–1765.
[57] X.Han,T.Gao,Y.Yao,D.Ye,Z.Liu,andM.Sun,“Opennre:Anopen
in Proceedings		of the	International	Conference		on Machine	Learning,
--------------	---	------	-------------	----------	---	----------	---------	---	---	---	---	---	---	---	---
andextensibletoolkitforneuralrelationextraction,”inProceedingsof
2021,pp.4904–4916.
theConferenceonEmpiricalMethodsinNaturalLanguageProcessing,
[36] H. Tan	and M.	Bansal,	“LXMERT:		learning	cross-modality	encoder
-----------	------	-------	--------	---	--------	--------------	-------	---	---	---	---	---	---	---	---
2019,pp.169–174.
representations		from transformers,”		in	Proceedings	of	the Empirical
---------------	---	-------------------	---	---	-----------	---	-------------	-----------	------	---------	---	--------	----	------	-------------
								[58] J. Li,	D. Li,	C. Xiong,	and	S. C. H.	Hoi,	“BLIP:	bootstrapping
MethodsinNaturalLanguageProcessing,2019,pp.5099–5110. language-image pre-training for unified vision-language understanding
[37] Z.Wang,J.Yu,A.W.Yu,Z.Dai,Y.Tsvetkov,andY.Cao,“Simvlm: and generation,” in Proceedings of the International Conference on
Simple	visual	language	model	pretraining	with	weak supervision,”		in
------	------	--------	-----	-----------	----	------------------	---	---	---	---	---	---	---	---	---
MachineLearning,2022,pp.12888–12900.
Proceedings	of	the International		Conference		on Learning	Representa-
-----------	---	-----------------	---	----------	---	-----------	-----------	---	---	---	---	---	---	---	---
[59] Y.Fang,W.Wang,B.Xie,Q.Sun,L.Wu,X.Wang,T.Huang,X.Wang,
tions,2022.
andY.Cao,“EVA:exploringthelimitsofmaskedvisualrepresentation
[38] Q. Qi, A. Zhang, Y. Liao, W. Sun, Y. Wang, X. Li, and S. Liu, learning at scale,” in Proceedings of the IEEE/CVF Conference on
“Simultaneously training and compressing vision-and-language pre- ComputerVisionandPatternRecognition,2023,pp.19358–19369.
trainingmodel,”IEEETransactionsonMultimedia,vol.25,pp.8194–
[60] V.Ordonez,G.Kulkarni,andT.L.Berg,“Im2text:Describingimages
8203,2023.
								using	1 million	captioned	photographs,”		in Advances	in	Neural Infor-
---	---	---	---	---	---	---	---	-----	---------	---------	-------------	---	-----------	---	-------------
[39] Y.Chen,L.Li,L.Yu,A.E.Kholy,F.Ahmedetal.,“UNITER:universal
mationProcessingSystems,2011,pp.1143–1151.
image-text representation learning,” in Proceedings of the European [61] C.Qu,L.Yang,C.Chen,M.Qiu,W.B.Croft,andM.Iyyer,“Open-
ConferenceonComputerVision,2020,pp.104–120. retrieval conversational question answering,” in Proceedings of the
[40] X.Li,X.Yin,C.Li,P.Zhang,X.Huetal.,“Oscar:Object-semantics International ACM SIGIR Conference on Research and Development
aligned pre-training for vision-language tasks,” in Proceedings of the inInformationRetrieval,2020,pp.539–548.
EuropeanConferenceonComputerVision,2020,pp.121–137.
								[62] D. Hannan,	A.	Jain, and	M. Bansal,	“Manymodalqa:		Modality	disam-
---	---	---	---	---	---	---	---	---------------	---	---------	----------	-------------	---	--------	------
[41] P. Zhang, X. Li, X. Hu, J. Yang, L. Zhang et al., “Vinvl: Making biguation and QA over diverse inputs,” in Proceedings of the AAAI
visualrepresentationsmatterinvision-languagemodels,”arXivpreprint ConferenceonArtificialIntelligence,2020,pp.7879–7886.
arXiv:2101.00529,2021.