基于大语言模型的用户目标导向小众词汇主题挖掘方法OA

To address challenges in niche topic word mining,including implicit slang semantics,scarce corpora for emerging terms,and significant semantic drift,the SeedLLM framework was proposed.It introduced cross-context con-sistency constraints to perform global reasoning over slang terms and capture their deep semantics.Retrieval-augmented generation was further employed to supplement external semantics for emerging words.Meanwhile,an emerging-word evaluation model and an online reinforcement learning mechanism were designed to suppress excessive retrieval and dy-namically adapt to semantic evolution.The experimental results on multiple cross-domain datasets demonstrate that the proposed method outperforms existing methods in niche topic word identification and topic aggregation,showing strong generalization and practical value.