Generating Faithful and Salient Text from Multimodal Data
This addresses the challenge of generating accurate and relevant text from multimodal inputs for applications relying on large multimodal models, though it is incremental as it builds on existing methods.
The paper tackles the problem of hallucination and lack of saliency in text generation from multimodal data by developing a framework that uses a vision critic model to identify and correct these issues, resulting in improved generation quality on faithfulness and saliency metrics, outperforming recent techniques.
While large multimodal models (LMMs) have obtained strong performance on many multimodal tasks, they may still hallucinate while generating text. Their performance on detecting salient features from visual data is also unclear. In this paper, we develop a framework to generate faithful and salient text from mixed-modal data, which includes images and structured data ( represented in knowledge graphs or tables). Specifically, we train a small vision critic model to identify hallucinated and non-salient features from the image modality. The critic model also generates a list of salient image features. This information is used in the post editing step to improve the generation quality. Experiments on two datasets show that our framework improves LMMs' generation quality on both faithfulness and saliency, outperforming recent techniques aimed at reducing hallucination.