Research on Emotional Feature Recognition of Streamer Discourse and Conversion Effect Prediction in Live Streaming E-commerce

Abstract

This study develops a window-level framework for recognizing emotional features in streamer discourse and predicting conversion effects in live streaming e-commerce. The research examines how emotional intensity, persuasive tone, affective fluctuation, discourse rhythm, and acoustic arousal are associated with product click, add-to-cart behavior, and completed purchase. The dataset consists of 120 live-streaming sessions collected from the beauty and personal-care category over 60 days. It includes live audio, speech transcripts, product exposure records, interaction logs, and order-conversion logs. All records are synchronized into 30-second discourse windows and matched with 3-minute post-exposure conversion outcomes. Whisper-large-v3 is used for transcription, BERT-Base Chinese and an emotion lexicon are used for textual emotion extraction, openSMILE 3.0 is used for acoustic features, and XGBoost 2.0 is used for multi-label conversion prediction. The results show that multimodal emotional discourse features outperform single-modality models, with purchase-conversion AUC reaching 0.873±0.012* and macro-F1 reaching 0.782±0.016*.