Build Belonging and Trust Proactively: A Humanized Intelligent Streamer Assistant with Personality, Emotion and Memory

Abstract

Live streaming has become a prevalent form of online entertainment and commerce, where real-time interactions occur between streamers and their audiences. Currently, streamer assistants have some shortcomings in terms of personality and emotional expression. These shortcomings undermine the live streaming effect and audience experience, thereby damaging the streamer’s popularity and income. In this paper, we present the Intelligent Streamer Assistant with Personality, Emotion, and Memory (ISAPEM) framework, which aims to utilize playful animal avatars to establish a sense of belonging and trust proactively with the audience. Firstly, we determine the assistant’s personality. Subsequently, the assistant determines its emotions according to its personality and the danmaku (bullet chats/comments) context analysis, ranging from trust and joy to sadness. Next, the assistant displays matched expressions and actions, and then generates consistent dialogue using large language models (LLMs). For example, when faced with a challenging question in the danmaku, the assistant might appear perplexed, then reach for the corresponding danmaku, catch it, and swallow it. Finally, the assistant stores and analyzes danmaku interaction data to remember and understand the audience’s needs and preferences. Preliminary experimental findings indicate that the ISAPEM framework can create a warmer experience for the audiences and enhance their willingness to interact, which has the potential to foster a sense of belonging and trust among the audiences. This study proposes a novel design framework for streamer assistants that integrates cutting-edge anthropomorphic design cues (ADCs) with a danmaku-based physical interaction mode, expanding the application and interaction modes of novel ADCs and LLMs.