ChartThinker: A Contextual Chain-of-Thought Approach to Optimized Chart Summarization

ChartThinker: A Contextual Chain-of-Thought Approach to Optimized Chart Summarization

25 Apr 2024 | Mengsha Liu, Daoyuan Chen, Yaliang Li, Guian Fang, Ying Shen
The paper introduces ChartThinker, an innovative method for chart summarization that integrates deep analysis based on chains of thought and context retrieval strategies. To address the limitations of existing approaches in visual-language matching and reasoning ability, the authors construct a large-scale dataset of comprehensive chart-caption pairs and fine-tune instructions on each chart. ChartThinker leverages a Context-Enhanced CoT Generator module to fuse thought chains with context retrieval, enhancing the model's reasoning ability and logical coherence. The method is evaluated using extensive automatic and human assessments, demonstrating superior performance over 8 state-of-the-art models across 7 evaluation metrics. The dataset and codes are publicly available to facilitate further research.The paper introduces ChartThinker, an innovative method for chart summarization that integrates deep analysis based on chains of thought and context retrieval strategies. To address the limitations of existing approaches in visual-language matching and reasoning ability, the authors construct a large-scale dataset of comprehensive chart-caption pairs and fine-tune instructions on each chart. ChartThinker leverages a Context-Enhanced CoT Generator module to fuse thought chains with context retrieval, enhancing the model's reasoning ability and logical coherence. The method is evaluated using extensive automatic and human assessments, demonstrating superior performance over 8 state-of-the-art models across 7 evaluation metrics. The dataset and codes are publicly available to facilitate further research.
Reach us at info@study.space