PLiTS-Net: A PLM-Conditioned Liquid Thread Network for Social Media Sentiment Analysis

This study proposes a PLM-conditioned Liquid Thread Sentiment Network (PLiTS-Net) for modeling the long and noisy conversational threads commonly found on social media, while dynamically adapting to their rapidly evolving, informal, and emerging expressions. Each post is first encoded by a frozen, domain-matched pretrained language model (PLM), with normalized structural cues embedded and fused to explicitly indicate the post’s position within the conversation and its contextual relations. A liquid time-constant recurrent encoder then processes the linearized thread with input-dependent memory, enabling the model to retain distant context without quadratic attention cost. Experiments on four threaded datasets from Weibo and Twitter, including a newly collected Weibo-2025 dataset containing emerging expressions, show that PLiTS-Net consistently outperforms both flat PLM baselines and strong context-aware RNN/tree models, achieving best Macro-F1 score on Chinese threads while using only three-quarters of their computational resources. The model also exhibits smaller performance degradation under domain drift and stronger transferability across platforms and languages, proving that liquid neural units offer a practical and resource-efficient alternative to long-context Transformers for conversation-level sentiment analysis.