CONDA: a CONtextual Dual-Annotated dataset for in-game toxicity understanding and detection

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 2406–2416 August 1–6, 2021. ©2021 Association for Computational Linguistics 2406 CONDA: a CONtextual Dual-Annotated dataset for in-game toxicity understanding and detection Henry Weld Guanghao Huang Jean Lee Tongshu Zhang Kunze Wang Xinghong Guo Siqu Long Josiah Poon Soyeon Caren Han∗ School of Computer Science, The University of Sydney, NSW, Australia {hgua3108,tzha6458,kwan4418,xguo2796,slon6753}@uni.sydney.edu.au {henry.weld,jean.lee,josiah.poon,caren.han}@sydney.edu.au Abstract Traditional toxicity detection models have fo- cused on the single utterance level without deeper understanding of context. We introduce CONDA, a new dataset for in-game toxic lan- guage detection enabling joint intent classifi- cation 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. We propose a robust dual semantic-level toxicity framework, which handles utterance and token-level pat- terns, and rich contextual chatting history. Ac- companying the dataset is a thorough in-game toxicity analysis, which provides comprehen- sive understanding of context at utterance, to- ken, and dual levels. Inspired by NLU, we also apply its metrics to the toxicity detection tasks for assessing toxicity and game-specific aspects. We evaluate strong NLU models on CONDA, providing fine-grained results for dif- ferent intent classes and slot classes. Further- more, we examine the coverage of toxicity na- ture in our dataset by comparing it with other toxicity datasets.1 1 Introduction As the popularity of multi-player online games has grown, the phenomenon of in-game toxic behav- ior has taken root within them. Toxic behavior is strongly present in recent online games and is problematic to the gaming industry (Adinolf and Turkay, 2018). For instance, 74% of US players of such games report harassment with 65% experienc- ing severe harassment. (ADL, 2019). In the past few years, Natural Language Process- ing (NLP) researchers have proposed several on- line game/community toxicity analysis frameworks ∗Corresponding author (caren.han@sydney.edu.au) 1The dataset and lexicons are available at https:// github.com/usydnlp. Figure 1: An example intent/slot annotation from the CONDA (CONtextual Dual-Annotated) dataset. (Kwak et al., 2015; Murnion et al., 2018; Wang et al., 2020) and datasets (M¨artens et al., 2015; Stoop et al., 2019). However, existing datasets (1) focus only on the single utterance level with- out deeper understanding of context in the whole conversation/chat, and (2) do not explicitly use se- mantic clues from the words within the utterance. The chat in online games and communities is similar in nature to spoken language, an area stud- ied by Natural Language Understanding (NLU). NLU research aims to best represent human com- munication by extracting semantic structure in the form of intent and slot analysis. Intent detection is the classification of the desired outcome of an utter- ance (or sentence), and slot filling is the labeling of each token (or word) in the utterance with the type of semantic information it carries. In recent litera- ture, these two tasks are trained jointly to capture synergies between them, and these jointly trained models give better results (Zhang …

Content truncated to 500 words