TikTalk
收藏资源简介:
TikTalk是由中国人民大学创建的一个视频基础的多模态对话数据集,旨在模拟现实世界中的多模态闲聊场景。该数据集包含从流行的视频分享平台收集的38,703个视频和367,670个用户对话。用户基于观看视频的多模态体验进行自发对话,从而重现真实的闲聊环境。与以往的多模态对话数据集相比,TikTalk中的更丰富的上下文类型导致更多样化的对话,但也增加了从复杂的多模态信息中捕捉人类兴趣以生成个性化响应的难度。此外,外部知识在我们的数据集中更频繁地被唤起。这些事实揭示了多模态对话模型的新挑战。
TikTalk is a video-based multimodal dialogue dataset developed by Renmin University of China, designed to simulate real-world multimodal casual conversation scenarios. This dataset contains 38,703 videos and 367,670 user dialogues collected from popular video-sharing platforms. Users engage in spontaneous dialogues grounded in their multimodal experiences of viewing these videos, thus replicating authentic casual chat environments. Compared with prior multimodal dialogue datasets, the richer contextual types in TikTalk result in more diverse dialogues, but also raise the difficulty of capturing human interests from complex multimodal information to generate personalized responses. Furthermore, external knowledge is invoked more frequently in this dataset. These findings unveil new challenges for multimodal dialogue models.

- 1TikTalk: A Video-Based Dialogue Dataset for Multi-Modal Chitchat in Real World中国人民大学 · 2023年



