Traditional toxicity detection models have focused on the single utterance level without deeper understanding of context. We introduce CONDA, a new dataset for in-game toxic language detection enabling joint intent classification and slot filling analysis, which is the core task of Natural Language Understanding (NLU). The dataset consists of 45K utterances from 12K conversations from the chat logs of 1.9K completed Dota 2 matches.… See the full description on the dataset page: https://huggingface.co/datasets/Matrix430/CONDA.
