LLM-Enhanced Chinese Morph Resolution in E-Commerce Live Streaming Scenarios

TL;DR

This paper introduces a method using large language models to detect and correct misleading speech in Chinese e-commerce live streams, improving accuracy and efficiency.

Contribution

A novel LLM-enhanced training framework for morph resolution that extracts structured knowledge without fine-tuning the LLM itself.

Findings

The proposed method achieves an F0.5 score of 0.943 in-domain, a 7 pp improvement over the baseline.

Out-of-domain performance reaches 0.799, a 5 pp improvement over the baseline.

LLM-derived signals can be effectively used to train lightweight models for accurate morph resolution.

Abstract

E-commerce live streaming in China has become a major retail channel, yet hosts often employ subtle phonetic or semantic “morphs” to evade moderation and make unsubstantiated claims, posing risks to consumers. To address this, we study the Live Auditory Morph Resolution (LiveAMR) task, which restores morphed speech transcriptions to their true forms. Building on prior text-based morph resolution, we propose an LLM-enhanced training framework that mines three types of explanation knowledge—predefined morph-type labels, LLM-generated reference corrections, and natural-language rationales constrained for clarity and comprehensiveness—from a frozen large language model. These annotations are concatenated with the original morphed sentence and used to fine-tune a lightweight T5 model under a standard cross-entropy objective. In experiments on two test sets (in-domain and out-of-domain), our method achieves substantial gains over baselines, improving F0.5 by up to 7 pp in-domain (to 0.943) and 5 pp out-of-domain (to 0.799) compared to a strong T5 baseline. These results demonstrate that structured LLM-derived signals can be mined without fine-tuning the LLM itself and injected into small models to yield efficient, accurate morph resolution.

Figures