Abstract
Out-of-distribution (OOD) detection in multimodal contexts is essential for identifying deviations in combined inputs from different modalities, particularly in applications like open-domain dialogue systems or real-life dialogue interactions. This paper aims to improve the user experience that involves multi-round long dialogues by efficiently detecting OOD dialogues and images. We introduce a novel scoring framework named Dialogue Image Aligning and Enhancing Framework (DIAEF) that integrates the visual language models with the novel proposed scores that detect OOD in two key scenarios (1) mismatches between the dialogue and image input pair and (2) input pairs with previously unseen labels. Our experimental results, derived from various benchmarks, demonstrate that integrating image and multi-round dialogue OOD detection is more effective with previously unseen labels than using either modality independently. In the presence of mismatched pairs, our proposed score effectively identifies these mismatches and demonstrates strong robustness in long dialogues.
Introduction
As AI conversational agents evolve from simple text-based bots into multimodal assistants capable of processing both visual and textual information, ensuring their reliability becomes paramount. A significant challenge in this evolution is the "reliability gap": models trained on closed datasets often struggle when encountering inputs outside their training distribution. When presented with unfamiliar topics or contradictory information, these models frequently "hallucinate" confident but incorrect responses. This paper addresses the critical need for robust Out-of-Distribution (OOD) detection in long, multimodal dialogues to enhance safety and maintain user trust.
Defining the Multimodal OOD Problem
In the context of multimodal dialogue, OOD detection is more complex than in single-modality systems. We identify two primary types of OOD instances:
- Semantic OOD (Label OOD): The input pair belongs to a category or label entirely unseen during training (e.g., a model trained on domestic animals is shown a professional sports match).
- Mismatched OOD: Both the image and the text belong to known categories, but they are contextually inconsistent with one another (e.g., a dialogue about a cat occurring while the system is presented with an image of a car).
Methodology: The DIAEF Framework
To address these challenges, we propose the Dialogue Image Aligning and Enhancing Framework (DIAEF). The core of DIAEF is a novel scoring function designed to synthesize signals from both modalities and their relationship.
Scoring Mechanism
The framework utilizes a composite scoring function:
$$g(x_T, x_I, y) = s(x_T, x_I)^\gamma (\alpha s_I(x_I, y) + (1 - \alpha)s_T(x_T, y))$$
- Alignment Term ($s(x_T, x_I)$): Uses Vision-Language Models (such as CLIP) to measure the similarity between the image and the dialogue history. A low alignment score indicates a mismatch.
- Enhancing Term: Combines individual OOD scores from the image ($s_I$) and text ($s_T$), balanced by a hyperparameter $\alpha$.
- Strictness Control ($\gamma$): This parameter determines how heavily the system penalizes misalignments between sight and speech.
By multiplying label-based confidence by the alignment score, DIAEF effectively identifies inputs that are either unrecognizable or contextually nonsensical.
Experimental Evaluation
We evaluated DIAEF across several benchmarks, including Visdial QA (120,000 images with multi-round dialogues) and Real MMD (combining COCO images with DailyDialog and Persona-Chat). To test robustness against mismatched pairs, we specifically introduced 10,000 mismatched examples into the Visdial QA set. Furthermore, we employed WordNet-based selection to ensure OOD labels were semantically distant from the training data.
Key Results
Our framework was compared against standard OOD scoring methods such as MSP, Energy, and Mahalanobis distance. Key findings include:
- Reduced False Alarms: On the Visdial QA dataset, DIAEF significantly lowered the FPR95 (False Positive Rate at 95% True Positive Rate). For instance, using the "Prob" scoring method, DIAEF achieved an FPR95 of 55.3%, outperforming image-only (60.0%) and dialogue-only (67.9%) baselines.
- Cross-Modal Synergy: The integration of image and text data proved far more effective than single-modality detection, as the two modalities provide a vital cross-check for each other.
- Dialogue Robustness: The alignment term helps the system remain effective even as dialogues grow longer and noise typically accumulates.
Conclusion
DIAEF demonstrates that in multimodal dialogue, the relationship between modalities is as important as the content of the modalities themselves. By prioritizing alignment, we can create conversational agents that are more aware of their own limitations. Future work will focus on refining score distributions and extending the framework to completely open-domain scenarios where pre-defined labels are unavailable.
Summary updated