Chart question answering has gained significant attention recently due to the growing need for effective data visualization interpretation. This paper presents a novel two-step model architecture for chart question answering, which includes a pre-training phase and a fine-tuning phase based on the UniChart framework, specifically applied to the ChartQA dataset. Our approach leverages the strengths of large-scale pretrained models to enhance the understanding of chart types and their associated queries, followed by a targeted fine-tuning process to adapt the model for improved performance on human-generated questions. Experimental results demonstrate substantial improvements in answering accuracy and relevance for human questions, indicating that our proposed method effectively bridges the gap between charts and interpretative questions. The findings underscore the model’s ability to generalize across diverse chart types and its practical utility in real-world data analysis contexts. Despite the inherent challenges of the ChartQA and OpenCQA datasets, our method achieves compelling results (66.88 RA and 14.95 BLEU, respectively), surpassing the performance of previous approaches, particularly on complex queries.