S2TDR: Structured Multimodal Reconstruction for Sticker-Aware Dialogue Understanding

# S2TDR: Structured Multimodal Reconstruction for Sticker-Aware Dialogue Understanding
Source: [https://arxiv.org/html/2610.41176](https://arxiv.org/html/2610.41176)
Yuanchen ShiEmail:[20227927002@stu\.suda\.edu\.cn](mailto:[email protected])Address:Natural Language Processing Laboratory, School of Computer Science and Technology, Soochow University, Suzhou, ChinaFang KongEmail:[kongfang@suda\.edu\.cn](mailto:[email protected])Corresponding author:Corresponding author\.Address:Natural Language Processing Laboratory, School of Computer Science and Technology, Soochow University, Suzhou, China

###### Abstract

Stickers are widely used in social media dialogues to express affect, stance, and intent beyond literal text\. Their meanings are often non\-literal and context\-dependent, making them difficult to handle with generic visual fusion or standalone captioning\. We propose Sticker\-to\-Text Dialogue Reconstruction \(S2TDR\), a structured multimodal reconstruction framework that converts sticker turns into contextually appropriate textual utterances\. S2TDR decomposes sticker interpretation into visual cue extraction, affective\-intent inference, and context\-aware dialogue reconstruction, using OCR text only as an auxiliary cue when available\. The reconstructed dialogue can then be consumed by existing text\-based dialogue models\. Experiments on sentiment analysis, intent recognition, and dialogue summarization show that S2TDR generally improves multiple downstream models over direct sticker integration methods\. Further analyses on component ablations, OCR\-empty stickers, and reconstruction quality indicate that structured reconstruction provides an interpretable and effective way to integrate sticker semantics into dialogue understanding\.

###### Keywords:

Sticker\-based multimodal understanding; sentiment and intent recognition; dialogue summarization

## 1Introduction

Stickers are pervasive in social media dialogues, where they often express stance, affect, and intent beyond literal text\. Unlike ordinary images, stickers frequently function as pragmatic acts: their meanings are stylized, non\-literal, and strongly shaped by the surrounding turns\. This makes sticker\-aware dialogue understanding challenging\. A model must infer not only what a sticker depicts, but also what role it plays in the conversation\. Existing studies have introduced sticker\-related tasks and datasets, including retrieval\[[1](https://arxiv.org/html/2610.41176#bib.bib1)\], generation\[[2](https://arxiv.org/html/2610.41176#bib.bib2)\], multimodal sentiment analysis\[[3](https://arxiv.org/html/2610.41176#bib.bib3)\], and emotion recognition\[[4](https://arxiv.org/html/2610.41176#bib.bib4)\]\. However, effective and interpretable ways to integrate sticker semantics into downstream dialogue models remain limited\.

Figure[1](https://arxiv.org/html/2610.41176#S1.F1)illustrates the problem\. The crying\-character sticker does not merely add visual content; it conveys anxiety and unease, which affects both sentiment and intent interpretation\. For summarization, the surrounding text provides the factual event, while the sticker contributes the implicit emotional state\. This suggests that sticker understanding should be context\-conditioned and transformed into a representation that downstream dialogue models can reliably use\.

![Refer to caption](https://arxiv.org/html/2610.41176v1/1.png)Figure 1:A dialogue example with text and a sticker\.To address this challenge, we propose Sticker\-to\-Text Dialogue Reconstruction \(S2TDR\), a structured reconstruction framework that converts sticker turns into contextually appropriate textual utterances\. Given a multimodal dialogue, S2TDR interprets each sticker through three components: visual cue extraction, affective\-intent inference, and context\-aware dialogue reconstruction\. OCR text is used only when available; for OCR\-empty stickers, the reconstruction relies on visual cues and dialogue context\. The resulting dialogue is fully text\-based and can be directly used by existing text encoders, sequence\-to\-sequence models, or large language models\.

We evaluate S2TDR on two sticker\-aware dialogue benchmarks covering sentiment analysis, intent recognition, and dialogue summarization\. Results show that structured reconstruction brings broad improvements over direct sticker integration methods such as emotion\-intent replacement and description appending\. To examine why the framework works and where it may fail, we further conduct component ablations, OCR\-empty sticker analysis, reconstruction\-quality evaluation, and qualitative error analysis\.

Our contributions are summarized as follows:

- 1\.We formulate sticker\-to\-text dialogue reconstruction as an interpretable intermediate representation for sticker\-aware dialogue understanding\.
- 2\.We propose S2TDR, a structured multimodal reconstruction framework that converts sticker semantics into context\-conditioned textual utterances\.
- 3\.We show that S2TDR improves multiple downstream models on sentiment analysis, intent recognition, and dialogue summarization\.
- 4\.We provide diagnostic analyses on component contributions, OCR\-empty stickers, reconstruction quality, and typical failure cases\.

## 2Related Work

Sticker\-related multimodal understanding has received increasing attention in recent years\. Existing work has introduced datasets and tasks for sticker\-aware sentiment analysis, emotion recognition, sticker tagging, retrieval, generation, and empathetic dialogue modeling\[[3](https://arxiv.org/html/2610.41176#bib.bib3),[4](https://arxiv.org/html/2610.41176#bib.bib4),[5](https://arxiv.org/html/2610.41176#bib.bib5),[1](https://arxiv.org/html/2610.41176#bib.bib1),[6](https://arxiv.org/html/2610.41176#bib.bib6),[7](https://arxiv.org/html/2610.41176#bib.bib7),[8](https://arxiv.org/html/2610.41176#bib.bib8),[9](https://arxiv.org/html/2610.41176#bib.bib9)\]\. These studies show that stickers carry rich affective and semantic information, but most of them focus on recognition, retrieval, or generation\. The context\-dependent pragmatic role of stickers in downstream dialogue understanding remains less explored\.

Our work is also related to multimodal affect modeling in dialogue\. Prior studies on multimodal emotion recognition and vision\-grounded dialogue modeling typically fuse textual, visual, or acoustic features through attention\-based or joint encoders\[[10](https://arxiv.org/html/2610.41176#bib.bib10),[11](https://arxiv.org/html/2610.41176#bib.bib11)\]\. Emoji\-aware and sarcasm\-oriented studies further show that visual paralinguistic symbols can substantially affect pragmatic interpretation\[[12](https://arxiv.org/html/2610.41176#bib.bib12)\]\. However, stickers differ from ordinary images: they are stylized, non\-literal, and often used as conversational acts\. Direct feature fusion or standalone captioning may therefore miss how a sticker functions within the dialogue context\.

Finally, staged reasoning with large multimodal models provides a practical way to decompose complex multimodal inference\. Existing studies have explored multimodal chain\-of\-thought reasoning, perception–decision decomposition, and intermediate scene\-graph\-style reasoning for visual\-language tasks\[[13](https://arxiv.org/html/2610.41176#bib.bib13),[14](https://arxiv.org/html/2610.41176#bib.bib14),[15](https://arxiv.org/html/2610.41176#bib.bib15),[16](https://arxiv.org/html/2610.41176#bib.bib16)\]\. In contrast to proposing a general prompting algorithm, we use staged reasoning as a task\-specific reconstruction mechanism: S2TDR converts sticker turns into context\-conditioned textual utterances, making sticker semantics directly usable by downstream dialogue models\.

![Refer to caption](https://arxiv.org/html/2610.41176v1/method.png)Figure 2:Overview of S2TDR\.![Refer to caption](https://arxiv.org/html/2610.41176v1/prompts.png)Figure 3:Staged sticker\-to\-text reconstruction\.
## 3S2TDR Framework

### 3\.1Task Formulation

Given a multimodal dialogueD=\{xi\}i=1nD=\\\{x\_\{i\}\\\}\_\{i=1\}^\{n\}containing both textual utterances and sticker turns, S2TDR produces a fully text\-based dialogueD′D^\{\\prime\}that preserves the affective and intentional meaning conveyed by stickers\. The reconstructed dialogue is then used by a task\-specific downstream model:

D′=S2TDR⁡\(D\),y=f⁡\(D′\),D^\{\\prime\}=\\mathrm\{S2TDR\}\(D\),\\qquad y=f\(D^\{\\prime\}\),wheref⁡\(⋅\)f\(\\cdot\)denotes a model for sentiment analysis, intent recognition, or dialogue summarization\. This formulation treats reconstruction as an intermediate representation: S2TDR does not replace downstream models, but converts sticker semantics into a textual form that existing dialogue models can directly consume\.

### 3\.2Framework Overview

Figure[2](https://arxiv.org/html/2610.41176#S2.F2)shows the overall framework\. S2TDR converts each sticker turn into a contextually appropriate textual utterance through three components: visual cue extraction, affective\-intent inference, and context\-aware dialogue reconstruction\. OCR text, when available, is used as an auxiliary cue rather than a required input\. For OCR\-empty stickers, the framework relies on visual cues and dialogue context\.

The reconstructed dialogue keeps the original conversational structure while replacing sticker images with textual utterances\. This makes the output interpretable and compatible with text encoders, sequence\-to\-sequence models, and instruction\-tuned language models\.

![Refer to caption](https://arxiv.org/html/2610.41176v1/OEI.png)Figure 4:Context helps infer sticker emotion and intent\.
### 3\.3Structured Sticker\-to\-Text Reconstruction

S2TDR uses a structured decomposition rather than rewriting a sticker in one step\. Sticker interpretation involves three levels of information: what the sticker depicts, what affective or intentional meaning it conveys, and how that meaning should be expressed in the current dialogue\. Separating these levels makes the reconstruction more interpretable and easier to ablate\.

Figure[3](https://arxiv.org/html/2610.41176#S2.F3)illustrates the reconstruction flow\. S2TDR first extracts visual cues from the sticker, then infers the sticker’s overall emotion and intent by combining visual information with OCR text when available, and finally rewrites the sticker as a concise utterance conditioned on the dialogue context\.

#### Visual cue extraction

For each sticker image, a multimodal language model extracts salient visual cues such as expression, posture, objects, and style\. PaddleOCR\[[17](https://arxiv.org/html/2610.41176#bib.bib17)\]extracts embedded text when available, but OCR is treated only as an auxiliary cue\. For OCR\-empty stickers, visual cues provide the main perceptual evidence\.

#### Affective\-intent inference

The visual description and OCR text are used to infer the sticker’s overall emotion and intent, such as self\-mockery, sarcasm, complaint, encouragement, or compromise\. This component summarizes the pragmatic role of the sticker before it is rewritten as dialogue\.

#### Context\-aware reconstruction

S2TDR conditions on the dialogue context, OCR text, visual description, and inferred emotion\-intent information to generate a single utterance that replaces the sticker turn\. The generated utterance is constrained to match the original speaker and remain concise, reducing unrelated expansion\.

Figure[4](https://arxiv.org/html/2610.41176#S3.F4)shows why context is necessary\. A sticker that appears relaxed in isolation may express helplessness or compromise when paired with a negative utterance\. S2TDR therefore reconstructs stickers according to their dialogue function rather than treating them as standalone images\.

### 3\.4Context Construction

Each dialogue turn is represented asxi=\(si,ci,ki\)x\_\{i\}=\(s\_\{i\},c\_\{i\},k\_\{i\}\), wheresis\_\{i\}is the speaker,cic\_\{i\}is the content, andki∈\{text,sticker\}k\_\{i\}\\in\\\{\\mathrm\{text\},\\mathrm\{sticker\}\\\}\. Text turns are kept unchanged\. For a sticker turn, S2TDR constructs a local context and maps the sticker into a textual replacement:

ci′=S2T⁡\(ci,context⁡\(xi\)\)\.c\_\{i\}^\{\\prime\}=\\mathrm\{S2T\}\(c\_\{i\},\\mathrm\{context\}\(x\_\{i\}\)\)\.
The context is constructed sequentially\. Preceding sticker turns are replaced with their reconstructed text, allowing later reconstructions to use an updated dialogue history\. Succeeding turns are included only when they are textual, avoiding circular dependence on future sticker reconstructions\. The final output is a text\-only dialogueD′=\{\(si,ci′\)\}i=1nD^\{\\prime\}=\\\{\(s\_\{i\},c\_\{i\}^\{\\prime\}\)\\\}\_\{i=1\}^\{n\}, which is fed into the downstream model\.

## 4Experiments

### 4\.1Datasets and Tasks

We evaluate S2TDR on two sticker\-aware dialogue benchmarks\. MSAIRS is used for multimodal sentiment analysis and intent recognition\[[18](https://arxiv.org/html/2610.41176#bib.bib18)\], while MCDSCS is used for multimodal dialogue summarization\[[19](https://arxiv.org/html/2610.41176#bib.bib19)\]\. These datasets cover both classification\-style understanding and generation\-style dialogue modeling\.

For MSAIRS, we report accuracy and macro F1 for sentiment analysis and intent recognition\. For MCDSCS, we follow prior work and report ROUGE\-1, ROUGE\-2, ROUGE\-L\[[20](https://arxiv.org/html/2610.41176#bib.bib20)\], BLEU\[[21](https://arxiv.org/html/2610.41176#bib.bib21)\], and METEOR\[[22](https://arxiv.org/html/2610.41176#bib.bib22)\]\. To examine the role of readable sticker text, we report OCR/text coverage in both datasets\. As shown in Table[1](https://arxiv.org/html/2610.41176#S4.T1), OCR text is common but not universal: over half of MSAIRS samples have empty sticker text, and MCDSCS also contains many OCR\-empty stickers\. This supports our treatment of OCR as an auxiliary cue rather than a required input\.

Table 1:OCR/text coverage\.*Note:*For MCDSCS dialogue samples, “Non\-empty” means that at least one sticker has OCR text, while “Empty” means that all stickers are OCR\-empty\.

### 4\.2Implementation Details

In the default setting, visual cue extraction is implemented with LLaVA\-v1\.6\-7B\[[23](https://arxiv.org/html/2610.41176#bib.bib23)\], while affective\-intent inference and dialogue reconstruction use Qwen2\.5\-7B\-Instruct\[[24](https://arxiv.org/html/2610.41176#bib.bib24)\]\. PaddleOCR\[[17](https://arxiv.org/html/2610.41176#bib.bib17)\]is used to extract embedded sticker text when available\. For OCR\-empty stickers, S2TDR relies on visual descriptions and dialogue context\.

The prompt templates are fixed across datasets\. The first component produces a concise visual description, the second summarizes the sticker’s emotion and intent, and the third generates a single utterance to replace the sticker turn\. We use low\-temperature decoding and constrain the reconstructed utterance length to reduce irrelevant expansion\. The prompt flow is illustrated in Figure[3](https://arxiv.org/html/2610.41176#S2.F3); the same templates are used across datasets unless otherwise specified\.

Table 2:Results on MSAIRS\.
### 4\.3Baselines and Sticker Integration Methods

We compare S2TDR with two groups of methods\. For downstream models, MSAIRS includes fine\-tuned text encoders, MMSAIR, and zero\-shot LLMs, while MCDSCS includes encoder–decoder summarization models, an instruction\-tuned LLM, and representative multimodal baselines\.

For sticker integration, we compare four settings: the original input without explicit reconstruction; EI\[[19](https://arxiv.org/html/2610.41176#bib.bib19)\], which replaces stickers with emotion\-intent labels; ITEI\[[31](https://arxiv.org/html/2610.41176#bib.bib31)\], which appends prompt\-derived sticker descriptions; and S2TDR, which rewrites sticker turns into context\-conditioned textual utterances\. This comparison isolates how sticker information is incorporated from the choice of downstream architecture\.

Table 3:Results on MCDSCS\.
### 4\.4Main Results

Tables[2](https://arxiv.org/html/2610.41176#S4.T2)and[3](https://arxiv.org/html/2610.41176#S4.T3)report the main results\. On MSAIRS, S2TDR improves fine\-tuned text encoders, zero\-shot LLMs, and the multimodal MMSAIR model over their original inputs\. Compared with EI and ITEI, S2TDR gives more stable gains, especially for intent recognition, where implicit sticker acts need to be converted into explicit dialogue cues\. This suggests that intent prediction benefits from pragmatic reconstruction rather than standalone sticker descriptions\.

On MCDSCS, S2TDR generally improves summarization models across ROUGE, BLEU, and METEOR metrics\. Although a few individual metrics are better with EI or ITEI, reconstruction\-based integration yields the strongest overall trend across backbones\. This is consistent with the nature of summarization: sticker information should be integrated into the discourse rather than appended as isolated visual annotations\.

### 4\.5Component Ablation

We ablate the main components of S2TDR on MMSAIR for MSAIRS and LLaMA2 for MCDSCS\. Removing visual cues, affective\-intent inference, or dialogue reconstruction consistently weakens performance\. The visual and affective\-intent components provide perceptual and pragmatic evidence, while dialogue reconstruction determines how these signals are expressed as a usable utterance\.

Among the three components, removing dialogue reconstruction causes the clearest drop\. This shows that obtaining sticker descriptions or emotion\-intent labels alone is not sufficient; the key step is converting them into a coherent turn that fits the dialogue\. We report ablations by removed component, which better reflects how each variant changes the reconstruction pipeline\.

### 4\.6Diagnostic Analysis

#### OCR\-empty stickers

Table[1](https://arxiv.org/html/2610.41176#S4.T1)shows that OCR text is unavailable for a substantial portion of stickers\. In MSAIRS, 52\.47% of samples have empty sticker text; in MCDSCS, 32\.30% of unique stickers have empty OCR text\. S2TDR therefore treats OCR as an auxiliary cue rather than a required input\. For OCR\-empty stickers, reconstruction relies on visual descriptions and dialogue context\.

#### Reconstruction quality

We further examine reconstruction quality through preference\-based evaluation\. As shown in Table[4](https://arxiv.org/html/2610.41176#S4.T4), S2TDR receives the highest preference share on faithfulness, fluency, and overall quality under both human and GPT\-4 judgments\. These values are preference shares rather than absolute quality scores\. We therefore use this table as auxiliary diagnostic evidence rather than as a standalone human\-evaluation benchmark\.

Table 4:Preference\-based reconstruction quality\.
#### Case study

Table[5](https://arxiv.org/html/2610.41176#S4.T5)presents representative cases from both MSAIRS and MCDSCS\. The cases show that S2TDR can recover useful sticker semantics for OCR\-empty stickers and make sticker\-derived affect explicit for both classification and summarization\. They also reveal a key limitation: when the context is short or underspecified, reconstruction may add plausible but over\-specific affective details\.

Table 5:Qualitative case study\.
#### Failure cases

The risk cases suggest that S2TDR is not an error\-free semantic parser\. Ambiguous context, indirect sticker text, and noisy OCR may lead to over\-specific reconstructions\. Nevertheless, the intermediate visual, affective\-intent, and reconstructed outputs make such errors easier to inspect than opaque multimodal predictions\.

### 4\.7Discussion and Limitations

S2TDR is designed as a reconstruction module rather than a replacement for multimodal dialogue models\. Its main advantage is that sticker semantics are converted into an interpretable textual form, allowing existing text encoders, sequence\-to\-sequence models, and LLMs to use sticker information without changing their architectures\. The intermediate outputs also make errors easier to inspect: visual descriptions, affective\-intent summaries, and reconstructed utterances can be examined separately\.

The framework still has limitations\. First, reconstruction quality depends on the visual and language models used in the S2T process, so stronger or domain\-adapted MLLMs may further improve performance\. Second, converting stickers into text may lose fine\-grained visual style, especially for highly stylized or culturally specific stickers\. Third, S2TDR may over\-specify the speaker’s affect when the dialogue context is short or ambiguous\. Future work can address these issues through uncertainty estimation, retrieval of similar sticker usages, and human\-in\-the\-loop verification for high\-risk cases\.

## 5Conclusion

We studied sticker\-aware dialogue understanding from the perspective of structured multimodal reconstruction\. Instead of treating stickers as standalone images or auxiliary labels, S2TDR converts sticker turns into context\-conditioned textual utterances through visual cue extraction, affective\-intent inference, and dialogue reconstruction\. Experiments on sentiment analysis, intent recognition, and dialogue summarization show that this reconstruction\-based formulation improves multiple downstream models over direct sticker integration methods\. Diagnostic analyses further show that OCR is only an auxiliary cue, component removals weaken performance, and dialogue\-style reconstructions are generally preferred over description\-style alternatives\. These findings suggest that verbalizing the pragmatic function of stickers is an effective and interpretable way to bridge stickers and text\-based dialogue models\.

## References

- \[1\]B\. Wang, Y\. Du, et al\.,[A new formula for sticker retrieval: Reply with stickers in multi\-modal and multi\-session conversation](https://ojs.aaai.org/index.php/AAAI/article/view/34720), AAAI 39 \(24\) \(2025\) 25327–25335\.[doi:10\.1609/aaai\.v39i24\.34720](https://doi.org/10.1609/aaai.v39i24.34720)\. URL[https://ojs\.aaai\.org/index\.php/AAAI/article/view/34720](https://ojs.aaai.org/index.php/AAAI/article/view/34720)
- \[2\]Z\. Yuan, J\. Zhang, et al\.,[Vsd2m: A large\-scale vision\-language sticker dataset for multi\-frame animated sticker generation](https://arxiv.org/abs/2412.08259)\(2025\)\.[arXiv:2412\.08259](http://arxiv.org/abs/2412.08259)\. URL[https://arxiv\.org/abs/2412\.08259](https://arxiv.org/abs/2412.08259)
- \[3\]F\. Ge, W\. Li, H\. Ren, Y\. Cai,[Towards exploiting stickers for multimodal sentiment analysis in social media: A new dataset and baseline](https://aclanthology.org/2022.coling-1.591), in: Proceedings of the 29th ICCL, International Committee on Computational Linguistics, 2022, pp\. 6795–6804\. URL[https://aclanthology\.org/2022\.coling\-1\.591](https://aclanthology.org/2022.coling-1.591)
- \[4\]S\. Liu, X\. Zhang, J\. Yang, Ser30k: A large\-scale dataset for sticker emotion recognition, in: Proceedings of the 30th ACM MM, 2022, pp\. 33–41\.
- \[5\]B\. Wang, B\. Liang, et al\., Towards real\-world stickers use: A new dataset for multi\-tag sticker recognition, arXiv preprint arXiv:2403\.05428 \(2024\)\.
- \[6\]Y\. Zhang, F\. Kong, P\. Wang, S\. Sun, L\. Wang, S\. Feng, D\. Wang, Y\. Zhang, K\. Song,[STICKERCONV: Generating multimodal empathetic responses from scratch](https://aclanthology.org/2024.acl-long.417/), in: L\.\-W\. Ku, A\. Martins, V\. Srikumar \(Eds\.\), Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics \(Volume 1: Long Papers\), Association for Computational Linguistics, Bangkok, Thailand, 2024, pp\. 7707–7733\.[doi:10\.18653/v1/2024\.acl\-long\.417](https://doi.org/10.18653/v1/2024.acl-long.417)\. URL[https://aclanthology\.org/2024\.acl\-long\.417/](https://aclanthology.org/2024.acl-long.417/)
- \[7\]Z\. Yuan, J\. Zhang, Y\. Deng, Y\. Zhu, J\. Zhou, J\. Zhang, Vsd2m: Large\-scale vision\-language sticker dataset for multi\-frame animated sticker generation, in: 2025 IEEE International Conference on Multimedia and Expo \(ICME\), IEEE, 2025, pp\. 1–6\.
- \[8\]H\. E\. M\. Chee, J\. Wang, Z\. Guo, W\. Ma, M\. Zhang,[Gse: Evaluating sticker visual semantic similarity via a general sticker encoder](https://arxiv.org/abs/2511.04977)\(2025\)\.[arXiv:2511\.04977](http://arxiv.org/abs/2511.04977)\. URL[https://arxiv\.org/abs/2511\.04977](https://arxiv.org/abs/2511.04977)
- \[9\]H\. E\. M\. Chee, J\. Wang, Z\. Guo, W\. Ma, M\. Zhang, Small stickers, big meanings: A multilingual sticker semantic understanding dataset with a gamified approach, in: Proceedings of the 33rd ACM International Conference on Multimedia, 2025, pp\. 13457–13463\.
- \[10\]C\. Wu, Y\. Cai, Y\. Liu, P\. Zhu, Y\. Xue, Z\. Gong, J\. Hirschberg, B\. Ma,[Multimodal emotion recognition in conversations: A survey of methods, trends, challenges and prospects](https://aclanthology.org/2025.findings-emnlp.332/), in: C\. Christodoulopoulos, T\. Chakraborty, C\. Rose, V\. Peng \(Eds\.\), Findings of the Association for Computational Linguistics: EMNLP 2025, Association for Computational Linguistics, Suzhou, China, 2025, pp\. 6257–6274\.[doi:10\.18653/v1/2025\.findings\-emnlp\.332](https://doi.org/10.18653/v1/2025.findings-emnlp.332)\. URL[https://aclanthology\.org/2025\.findings\-emnlp\.332/](https://aclanthology.org/2025.findings-emnlp.332/)
- \[11\]T\. Yu, et al\.,[Vision guided generative pre\-trained language models for multimodal abstractive summarization](https://aclanthology.org/2021.emnlp-main.326), in: EMNLP, Association for Computational Linguistics, 2021, pp\. 3995–4007\.[doi:10\.18653/v1/2021\.emnlp\-main\.326](https://doi.org/10.18653/v1/2021.emnlp-main.326)\. URL[https://aclanthology\.org/2021\.emnlp\-main\.326](https://aclanthology.org/2021.emnlp-main.326)
- \[12\]D\. S\. Chauhan, G\. V\. Singh, A\. Arora, A\. Ekbal, P\. Bhattacharyya, An emoji\-aware multitask framework for multimodal sarcasm detection, Knowledge\-Based Systems 257 \(2022\) 109924\.
- \[13\]Z\. Zhang, A\. Zhang, M\. Li, H\. Zhao, G\. Karypis, A\. Smola, Multimodal chain\-of\-thought reasoning in language models, arXiv preprint arXiv:2302\.00923 \(2023\)\.
- \[14\]T\. Gao, P\. Chen, M\. Zhang, C\. Fu, Y\. Shen, Y\. Zhang, S\. Zhang, X\. Zheng, X\. Sun, L\. Cao, et al\., Cantor: Inspiring multimodal chain\-of\-thought of mllm, in: Proceedings of the 32nd ACM International Conference on Multimedia, 2024, pp\. 9096–9105\.
- \[15\]C\. Mitra, B\. Huang, T\. Darrell, R\. Herzig, Compositional chain\-of\-thought prompting for large multimodal models, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp\. 14420–14431\.
- \[16\]Y\. Xia, R\. Wang, X\. Liu, M\. Li, T\. Yu, X\. Chen, J\. McAuley, S\. Li,[Beyond chain\-of\-thought: A survey of chain\-of\-X paradigms for LLMs](https://aclanthology.org/2025.coling-main.719/), in: O\. Rambow, L\. Wanner, M\. Apidianaki, H\. Al\-Khalifa, B\. D\. Eugenio, S\. Schockaert \(Eds\.\), Proceedings of the 31st International Conference on Computational Linguistics, Association for Computational Linguistics, Abu Dhabi, UAE, 2025, pp\. 10795–10809\. URL[https://aclanthology\.org/2025\.coling\-main\.719/](https://aclanthology.org/2025.coling-main.719/)
- \[17\]Y\. Du, C\. Li, et al\., Pp\-ocr: A practical ultra lightweight ocr system, arXiv preprint arXiv:2009\.09941 \(2020\)\.
- \[18\]Y\. Shi, F\. Kong, L\. Zhang, Impact of stickers on multimodal sentiment and intent in social media: A new task, dataset and baseline, in: Proceedings of the 33rd ACM International Conference on Multimedia, 2025, pp\. 5637–5646\.
- \[19\]Y\. Shi, F\. Kong, Integrating stickers into multimodal dialogue summarization: A novel dataset and approach for enhancing social media interaction, in: Proceedings of the 32nd ACM MM, 2024, pp\. 9525–9534\.
- \[20\]C\.\-Y\. Lin,[Rouge: A package for automatic evaluation of summaries](https://aclanthology.org/W04-1013), in: Text Summarization Branches Out, Association for Computational Linguistics, 2004, pp\. 74–81\. URL[https://aclanthology\.org/W04\-1013](https://aclanthology.org/W04-1013)
- \[21\]Y\. Graham,[Re\-evaluating automatic summarization with bleu and 192 shades of rouge](https://aclanthology.org/D15-1013), in: EMNLP, Association for Computational Linguistics, 2015, pp\. 128–137\.[doi:10\.18653/v1/D15\-1013](https://doi.org/10.18653/v1/D15-1013)\. URL[https://aclanthology\.org/D15\-1013](https://aclanthology.org/D15-1013)
- \[22\]S\. Banerjee, A\. Lavie,[Meteor: An automatic metric for mt evaluation with improved correlation with human judgments](https://aclanthology.org/W05-0909), in: Proceedings of the ACL Workshop on Intrinsic and Extrinsic Evaluation Measures for Machine Translation and/or Summarization, Association for Computational Linguistics, 2005, pp\. 65–72\. URL[https://aclanthology\.org/W05\-0909](https://aclanthology.org/W05-0909)
- \[23\]B\. Li, et al\., Llava\-onevision: Easy visual task transfer, arXiv preprint arXiv:2408\.03326 \(2024\)\.
- \[24\]P\. Wang, S\. Bai, et al\., Qwen2\-vl: Enhancing vision\-language model’s perception of the world at any resolution, arXiv preprint arXiv:2409\.12191 \(2024\)\.
- \[25\]J\. Devlin, M\.\-W\. Chang, K\. Lee, K\. Toutanova, Bert: Pre\-training of deep bidirectional transformers for language understanding \(2019\)\.[arXiv:1810\.04805](http://arxiv.org/abs/1810.04805)\.
- \[26\]Y\. Liu, M\. Ott, et al\., Roberta: A robustly optimized bert pretraining approach, arXiv preprint arXiv:1907\.11692 \(2019\)\.
- \[27\]T\. GLM, A\. Zeng, B\. Xu, et al\.,[Chatglm: A family of large language models from glm\-130b to glm\-4 all tools](https://arxiv.org/abs/2406.12793)\(2024\)\.[arXiv:2406\.12793](http://arxiv.org/abs/2406.12793)\. URL[https://arxiv\.org/abs/2406\.12793](https://arxiv.org/abs/2406.12793)
- \[28\]J\. Yu, J\. Jiang, Adapting bert for target\-oriented multimodal sentiment classification, in: Proceedings of the Twenty\-Eighth International Joint Conference on Artificial Intelligence, IJCAI, 2019, pp\. 5408–5414\.[doi:10\.24963/ijcai\.2019/751](https://doi.org/10.24963/ijcai.2019/751)\.
- \[29\]Y\. Li, R\. Quan, et al\., Efficient multimodal fusion via interactive prompting, in: CVPR, 2023, pp\. 2604–2613\.
- \[30\]A\. Young, et al\., Yi: Open foundation models by 01\. ai, arXiv preprint arXiv:2403\.04652 \(2024\)\.
- \[31\]Y\. Shi, F\. Kong, Msa\-itei: A novel method for multimodal analysis of social media stickers, in: ICASSP, IEEE, 2025, pp\. 1–5\.
- \[32\]M\. Lewis, Y\. Liu, et al\.,[Bart: Denoising sequence\-to\-sequence pre\-training for natural language generation, translation, and comprehension](https://aclanthology.org/2020.acl-main.703), in: ACL, Association for Computational Linguistics, 2020, pp\. 7871–7880\.[doi:10\.18653/v1/2020\.acl\-main\.703](https://doi.org/10.18653/v1/2020.acl-main.703)\. URL[https://aclanthology\.org/2020\.acl\-main\.703](https://aclanthology.org/2020.acl-main.703)
- \[33\]C\. Raffel, N\. Shazeer, et al\., Exploring the limits of transfer learning with a unified text\-to\-text transformer, The Journal of Machine Learning Research 21 \(1\) \(2020\) 5485–5551\.
- \[34\]J\. Zhang, Y\. Zhao, M\. Saleh, P\. Liu, Pegasus: Pre\-training with extracted gap\-sentences for abstractive summarization, in: ICML, PMLR, 2020, pp\. 11328–11339\.
- \[35\]H\. Touvron, L\. Martin, et al\., Llama 2: Open foundation and fine\-tuned chat models, arXiv preprint arXiv:2307\.09288 \(2023\)\.