LiveChat: A Large-Scale Personalized Dialogue Dataset Automatically Constructed from Live Streaming

TL;DR

LiveChat presents a large-scale Chinese live-streaming dialogue dataset with fine-grained persona profiles, addressing domain gaps between video-based conversations and text corpora. It introduces an automatic dialogue-constructing pipeline (ASR transcription and reply-to-whom matching) and defines two benchmarks: response modeling and addressee recognition. Through retrieval-based baselines and transfer-learning analyses, the work demonstrates the value of explicit persona profiles and longer average sessions per persona, while highlighting domain-specific challenges for generation models and large language models in this live domain. The dataset and findings offer a foundation for developing practical, personalized, live-domain dialogue systems and point to domain-adaptive strategies for LLMs in video-sourced data contexts.

Abstract

Open-domain dialogue systems have made promising progress in recent years. While the state-of-the-art dialogue agents are built upon large-scale text-based social media data and large pre-trained models, there is no guarantee these agents could also perform well in fast-growing scenarios, such as live streaming, due to the bounded transferability of pre-trained models and biased distributions of public datasets from Reddit and Weibo, etc. To improve the essential capability of responding and establish a benchmark in the live open-domain scenario, we introduce the LiveChat dataset, composed of 1.33 million real-life Chinese dialogues with almost 3800 average sessions across 351 personas and fine-grained profiles for each persona. LiveChat is automatically constructed by processing numerous live videos on the Internet and naturally falls within the scope of multi-party conversations, where the issues of Who says What to Whom should be considered. Therefore, we target two critical tasks of response modeling and addressee recognition and propose retrieval-based baselines grounded on advanced techniques. Experimental results have validated the positive effects of leveraging persona profiles and larger average sessions per persona. In addition, we also benchmark the transferability of advanced generation-based models on LiveChat and pose some future directions for current challenges.

Figures (9)

• Figure 1: A session example of LiveChat. A streamer will respond to one audience's comment from the comments area.
• Figure 2: The whole construction process of LiveChat.
• Figure 3: Our retrieval-based architecture.
• Figure 4: In-context learning results of GLM and GPT3 on different shots.
• Figure 5: A conversation between one streamer and several audiences in LiveChat.