Article

Profiling Sentiment Archetypes of Popular Twitch Emotes: A Comparative Analysis of Rule-Based Segmentation and Unsupervised Clustering Techniques

---

Abstract

Live-streaming platforms like Twitch have fostered unique communicative ecosystems where non-linguistic tokens, or "emotes," are central to expressing collective sentiment. However, conventional sentiment analysis tools, designed for standard text, largely fail to capture the nuanced and context-dependent meaning of these symbols. This paper addresses this gap by profiling sentiment archetypes of popular Twitch emotes through a comparative methodological framework. We contrast a theory-driven, rule-based segmentation approach, where emotes are manually assigned to predefined sentiment categories, with a data-driven, unsupervised clustering approach that groups emotes based on their contextual usage patterns. Applying this dual analysis to a dataset of Twitch chat messages, we constructed a feature set for each emote incorporating both textual context (TF-IDF) and behavioral metrics. A K-Means clustering algorithm was then used to identify emergent archetypes from the data. Our results reveal a profound divergence between the two methods, quantified by a near-zero Adjusted Rand Index (ARI) of 0.0012. This indicates that an emote's prescribed semantic meaning has virtually no correlation with its functional co-usage in practice. The clustering algorithm successfully identified coherent, function-based groups, such as a "Hype/Spam" cluster and a "Nuanced Reaction" cluster, which are not captured by the rule-based taxonomy. We conclude that emote meaning is not static but is an emergent property of community practice, defined primarily by pragmatic function rather than semantic content. This finding highlights the critical limitations of applying traditional sentiment analysis to dynamic digital cultures and underscores the necessity of data-driven, context-aware methods for understanding online communication.

Keywords: Clustering, Computational Linguistic, Emotes, Sentiment Analysis, Twitch

How to Cite:

Izumi, C., Ghaffar, S. & Setiawan, W., (2025) “Profiling Sentiment Archetypes of Popular Twitch Emotes: A Comparative Analysis of Rule-Based Segmentation and Unsupervised Clustering Techniques”, Journal of Digital Society 1(3), 183-200. doi: https://doi.org/10.63913/jds.v1i3.38