The evaluation results on the Charades Captions dataset are shown in Table 2. Figure 1: Qualitative results of the generated tree structure and sentences on the MSR-VTT dataset. Table 1: The average sentence lengths of ground-truth captions and the captions generated by "w/o Lower average edit distance is better. In this section, we illustrate more details of human evaluation. We recruited 10 annotators to carry out the human evaluation process. The user interface for human evaluation is shown in Figure 4. To ensure Do the main claims made in the abstract and introduction accurately reflect the paper's Did you discuss any potential negative societal impacts of your work?