大语言模型时代壮语情感词典构建的方法路径

Methodological Pathways for Constructing a Zhuang Sentiment Lexicon in the Era of Large Language Models

摘要: 壮语作为我国使用人口最多的少数民族语言，在自然语言处理领域长期处于低资源状态，情感词典等细粒度语义资源至今缺失。文章梳理情感词典构建方法的演进脉络，重点考察大语言模型时代涌现的三种新兴技术路径——人机协同主动学习、少样本上下文学习和参数高效微调，并结合壮语的资源现状逐一分析各路径的方法学要求与落地条件。研究表明，预训练语料覆盖薄弱、双语资源结构有限和专业人才稀缺三方面因素相互交织，使壮语情感词典建设难以依靠单一技术路径解决，需要根据资源条件灵活组合传统方法与新兴方法。

Abstract: Zhuang, the most widely spoken ethnic minority language in China, has long remained a low-resource language in natural language processing, with sentiment lexicons and other fine-grained semantic resources still absent. This paper traces the methodological evolution of sentiment lexicon construction and focuses on three emerging technical pathways enabled by large language models: human-in-the-loop active learning, few-shot in-context learning, and parameter-efficient fine-tuning. Each pathway is examined in light of the current resource conditions of Zhuang. The analysis reveals that three intertwined challenges, namely insufficient pretraining coverage, the limited structure of bilingual resources, and the scarcity of specialized expertise, prevent any single technical pathway from independently meeting the demands of Zhuang sentiment lexicon construction. A flexible combination of traditional methods and emerging approaches, calibrated to actual resource conditions, is therefore necessary. Among the three new pathways, few-shot in-context learning offers the strongest fit for current Zhuang research, owing to its lower technical threshold and reduced reliance on pretraining coverage.