Traditional toxicity detection models have focused on the single utterance\nlevel without deeper understanding of context. We introduce CONDA, a new\ndataset for in-game toxic language detection enabling joint intent\nclassification and slot filling analysis, which is the core task of Natural\nLanguage Understanding (NLU). The dataset consists of 45K utterances from 12K\nconversations from the chat logs of 1.9K completed Dota 2 matches. We propose a\nrobust dual semantic-level toxicity framework, which handles utterance and\ntoken-level patterns, and rich contextual chatting history. Accompanying the\ndataset is a thorough in-game toxicity analysis, which provides comprehensive\nunderstanding of context at utterance, token, and dual levels. Inspired by NLU,\nwe also apply its metrics to the toxicity detection tasks for assessing\ntoxicity and game-specific aspects. We evaluate strong NLU models on CONDA,\nproviding fine-grained results for different intent classes and slot classes.\nFurthermore, we examine the coverage of toxicity nature in our dataset by\ncomparing it with other toxicity datasets.\n