Faithful Chart Summarization with ChaTS-Pi

The paper presents CHATS-CRITIC and CHATS-PI, reference-free chart summarization tools that address the issue of factual accuracy in generating text from charts. CHATS-CRITIC uses an image-to-text model and a tabular entailment model to evaluate faithfulness, outperforming existing metrics in aligning with human ratings. CHATS-PI, a pipeline, integrates CHATS-CRITIC to refine and rank summaries by removing unsupported sentences. The study showcases the effectiveness of these models on two datasets, with CHATS-PI achieving state-of-the-art results. The research also highlights the limitations of reference-based metrics and the importance of evaluating factual correctness in chart summarization tasks.