Multimodal systems have highly complex processing pipelines and are pretrained over large datasets before being fine-tuned for specific tasks such as visual captioning. However, it becomes hard to disentangle what the model learns during the fine-tuning process from what it already knows due to its pretraining. In this work, we learn a probabilistic model using Hybrid Markov Logic Networks (HMLNs) over the training examples by relating symbolic knowledge (extracted from the caption) with visual features (extracted from the image). For a generated caption, we quantify the influence of training examples based on the HMLN distribution using probabilistic inference. We evaluate two types of inference procedures on the MSCOCO dataset for different types of captioning models. Our results show that for BLIP2 (a model that uses a LLM), the fine-tuning may have smaller influence on the knowledge the model has acquired since it may have more general knowledge to perform visual captioning as compared to models that do not use a LLM.

Multimodal problems such as caption generation advances AI as a whole since they require integration of several key domains such as computer vision, NLP and knowledge representation. In this paper, we develop a new approach to evaluate captioning models by verifying them using Markov Logic Networks (MLNs). Specifically, we compile an MLN from training data and perform probabilistic inference to estimate uncertainty in a generated caption. To reify the caption, we leverage advances in Natural Language Inference (NLI) models and convert a caption into a query for the MLN. Further, we add visual context into the MLN distribution using an attention-based Multiple Instance Learning model and evaluate a caption based on this augmented distribution. We perform experiments using MSCOCO on several state-of-the-art benchmarks and show that our approach can evaluate captioning models just as effectively as methods that require human-generated captions.