[Paper Review] Multimodal Interactions Using Pretrained Unimodal Models for SIMMC 2.0

Joosung Lee, Kijong Han|arXiv (Cornell University)|Dec 10, 2021

Topic Modeling4 citations

TL;DR

This paper proposes a multimodal framework that fine-tunes pretrained unimodal models (RoBERTa and DeIT) for multimodal understanding in the SIMMC 2.0 challenge. It introduces two pretraining tasks—Image-to-Text Matching (ITM) and Background-to-Text Matching (BTM)—to align visual and textual representations, achieving 3rd place in subtasks #1 and #2, and runner-up in subtask #4 generation with 68.2 F1 and 0.297 BLEU-4 scores.

ABSTRACT

This paper presents our work on the Situated Interactive MultiModal Conversations 2.0 challenge held at Dialog State Tracking Challenge 10. SIMMC 2.0 includes 4 subtasks, and we introduce our multimodal approaches for the subtask \#1, \#2 and the generation of subtask \#4. SIMMC 2.0 dataset is a multimodal dataset containing image and text information, which is more challenging than the problem of only text-based conversations because it must be solved by understanding the relationship between image and text. Therefore, since there is a limit to solving only text models such as BERT or GPT2, we propose a multimodal model combining image and text. We first pretrain the multimodal model to understand the relationship between image and text, then finetune our model for each task. We achieve the 3rd best performance in subtask \#1, \#2 and a runner-up in the generation of subtask \#4. The source code is available at https://github.com/rungjoo/simmc2.0.

Motivation & Objective

• To address the challenge of multimodal understanding in task-oriented dialogues where both image and text are essential for accurate reasoning.
• To overcome limitations of unimodal models (e.g., BERT, GPT2) in handling visual-linguistic relationships in complex, real-world scenarios like fashion/furniture shopping.
• To improve performance on multimodal dialog tasks—disambiguation, coreference resolution, dialog state tracking, and response generation—by leveraging joint visual-textual representations.
• To validate the effectiveness of pretraining on image-text alignment and multi-task learning for downstream multimodal NLP tasks.

Proposed method

• Pretrain a multimodal model using two contrastive learning objectives: Image-to-Text Matching (ITM) for object-level visual-textual alignment and Background-to-Text Matching (BTM) for scene-level context understanding.
• Fine-tune the pretrained multimodal model on each of the four SIMMC 2.0 subtasks using task-specific heads and shared representations.
• Integrate visual features from DeIT (Deformable Image Transformer) and textual features from RoBERTa-Large to enable cross-modal attention and interaction.
• Use multi-task learning in subtask #2 to jointly optimize for object coreference and utterance classification, improving generalization and filtering of candidate objects.
• Incorporate slot values and object image representations into GPT2 for response generation, enhancing factual consistency and relevance.
• Apply ablation studies to evaluate the contribution of ITM, BTM, system utterance data, and multi-task components in each subtask.

Experimental results

Research questions

• RQ1 Can pretrained unimodal models (RoBERTa and DeIT) be effectively combined to improve multimodal reasoning in dialog systems?
• RQ2 How effective is pretraining on image-text matching (ITM) and background-text matching (BTM) in enhancing downstream multimodal dialog performance?
• RQ3 To what extent do multi-task learning components (e.g., utterance classification, system matching) improve coreference resolution and disambiguation?
• RQ4 How do visual features (object and background images) and metadata (slot_values) contribute to response generation quality in multimodal dialog?
• RQ5 What is the relative impact of different training components (e.g., system utterance data, object images) on model performance across subtasks?

Key findings

• The proposed model achieved 3rd place in subtask #1 (Multimodal Disambiguation) with a test accuracy of 93.1%, significantly outperforming the GPT2 baseline (73.5%).
• In subtask #2 (Multimodal Coreference Resolution), the model achieved a test F1 score of 68.2, a substantial improvement over the baseline (44.1 F1) and the best-performing baseline (32.2 F1 for 'all S' strategy).
• The ablation study showed that using system utterance data as training input improved performance by +1.2 F1, indicating strong data-augmentation benefits.
• The inclusion of slot_values in response generation improved BLEU-4 by +0.01, highlighting their importance in grounding responses to visual and dialog context.
• Multi-task learning improved performance by +1.6 F1 in subtask #2, with utterance classification achieving 99.5% accuracy and system matching 93.24 F1.
• The model achieved a BLEU-4 score of 0.297 in subtask #4 (Response Generation), ranking runner-up, with significant gains from combining object representations and slot values.