Scholarlink

<p>The rise of digital information requires high-performing methods for organizing and structuring knowledge. Topic modeling is a widely adopted technique for identifying topics and patterns in several documents. However, current topic models have certain limitations, especially when dealing with noisy data. The impressive advances in large language models (LLM) could help enhance current topic models. In this work, we propose a strategy that leverages LLMs to perform content extraction and content augmentation to enhance current clustering-based topic modeling. First, we obtained explanations of the original texts through zero-shot prompting using the Llama-3.1-8B, Llama-3.1-70B, and DeepSeek-R1-Llama-8B models. Subsequently, the embeddings of these explanations were combined with those of the original texts through both concatenation and averaging. Finally, these embeddings were fed into BERTopic, a clustering-based flexible mechanism that has shown strong performance. Our results demonstrate that this technique not only improves topic coherence but also brings the output closer to golden annotations when facing user-created data. Additionally, the performance of the selected LLMs in this task is compared, providing insights into their effectiveness.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Leveraging large language models to enhance clustering-based topic modeling

  • 期刊论文

  • 发表日期：

    2025年10月30日

查看全文
  • Ines Hernandez-Camero,
  • Antonio Garcia-Cabot,
  • Eva Garcia-Lopez,
  • Sergio Caro-Alvaro,
  • Antonio Moreno-Cediel

摘要

The rise of digital information requires high-performing methods for organizing and structuring knowledge. Topic modeling is a widely adopted technique for identifying topics and patterns in several documents. However, current topic models have certain limitations, especially when dealing with noisy data. The impressive advances in large language models (LLM) could help enhance current topic models. In this work, we propose a strategy that leverages LLMs to perform content extraction and content augmentation to enhance current clustering-based topic modeling. First, we obtained explanations of the original texts through zero-shot prompting using the Llama-3.1-8B, Llama-3.1-70B, and DeepSeek-R1-Llama-8B models. Subsequently, the embeddings of these explanations were combined with those of the original texts through both concatenation and averaging. Finally, these embeddings were fed into BERTopic, a clustering-based flexible mechanism that has shown strong performance. Our results demonstrate that this technique not only improves topic coherence but also brings the output closer to golden annotations when facing user-created data. Additionally, the performance of the selected LLMs in this task is compared, providing insights into their effectiveness.