MIntRec2.0 Multimodal Intent Recognition Dialogue Dataset

MIntRec2.0 is a large-scale multimodal multi-party benchmark dataset proposed by Tsinghua University and others, which is specifically used to identify intent in conversations and detect non-intent content. Compared with the previous MIntRec, the data volume of MIntRec2.0 has increased to 15K, covering 30 intent categories, and contains about 9.3K in-intent and 5.7K out-of-intent annotated sentences, involving multiple modalities such as text, video and audio.

The dataset consists of 1,245 conversations, with an average of 12 sentences per conversation. Each sentence is equipped with an intent label, and each conversation involves at least two speakers, and all sentences are labeled with the speaker's identity. In addition, in response to the needs of open-world scenarios, MIntRec2.0 introduces OOS tags to identify sentences that do not belong to known intent categories to enhance the robustness of the system. This dataset aims to promote research related to multimodal intent understanding and lay a solid foundation for achieving more natural human-computer interaction and the road to AGI.