Automated Sentiment-Controlled Chat Generation in Live Streaming

Abstract Live streaming chat is a real-time, fast-paced communication environment dominated by short reactions, rapid topic shifts, and platform-specific conversational norms. Prior research on automated text generation has focused almost exclusively on static artifacts such as social media posts, product reviews, and news articles, leaving real-time live streaming chat largely unexplored. Standard dialogue models perform poorly under live-chat latency constraints, and no system-level architecture has been proposed that addresses this domain mismatch on local hardware. This thesis collects chat data from SOOP, a Korean live streaming platform, through a long-running Selenium-based acquisition pipeline. Domain analysis using a Gaussian Mixture Model identifies a 34-second active-component gap and a 507-second session boundary, while keyword analysis confirms that laughter-type fragments account for 88.6 percent of all special expressions. Building on these characteristics, this thesis proposes a multi-track architecture that separates two functions with incompatible computational profiles. Track 1 provides rule-based instant responses anchored to the empirical distribution of live chat. Track 3 provides probabilistic sentiment-controlled interventions using EXAONE 3.5 2.4B Instruct for constrained sentence-ending post-processing. The entire system runs on a single local machine without cloud dependency. On 110 input messages, architecture achieves an average response latency of 1.51 seconds, approximately 253 times faster than EXAONE 3.5 7.8B Instruct full generation, while producing zero malformed outputs. Multi-rater human evaluation with 20 participants and 880 ratings per method finds no statistically significant difference in naturalness between the proposed system and the 7.8B baseline. A sub-experiment confirms 2.4B as the optimal Track 3 model size. Contributions include a long-running chat acquisition pipeline, a multi-track architecture addressing the latency-quality trade-off, and a live-chat-specific evaluation protocol. Keywords: Live streaming chat, Multi-track architecture, Sentiment-controlled generation, Low-latency NLP, EXAONE, Opinion steering, Automated chat generation