設計與實作基於可控制大型語言模型的新聞意見萃取與分析系統

Abstract

新聞意見探勘旨在蒐集與處理新聞中各受訪者與組織發布的訊息，進而客觀地綜整並分析各立場的異同。為了實現更多元、平等、透明且完整的意見分析結果，我們設計並實作了一套基於可控制大型語言模型的新聞意見萃取與分析系統，具有以下三項主要效益： 透明過程：將多筆非結構化資訊經由結構化表格轉換，使系統不僅顯示最終的摘要結果，還能透明化整體運算過程，方便專家檢視、分析、介入與調整作業流程。 設計簡化：利用大型語言模型連結多個自然語言處理任務，簡化分析系統程式的演算法邏輯設計，提升整體運算效率。 高效抽取：通過上下文學習、模型微調及引導語言模型的高效程式範本，讓小型模型在引述抽取任務中也能達到與SOTA大型語言模型相媲美的效能。 這套系統不僅提升了新聞意見萃取與分析的準確性和效率，還為專家提供了更直觀和可操作的分析工具，為新聞領域帶來更全面和客觀的意見分析成果。

News opinion mining aims to collect and process information released by interviewees and organizations in the news, objectively summarizing and analyzing the differences and similarities in various stances. To achieve more diverse, equal, transparent, and comprehensive opinion analysis results, we have designed and implemented a news opinion extraction and analysis system based on controllable large language models (LLMs), which offers the following three main benefits: Transparent Process: By converting multiple pieces of unstructured information into structured tables, the system not only displays the final summarized results but also transparently reveals the entire computational process, making it easier for experts to review, analyze, intervene, and adjust the workflow. Simplified Design: Utilizing LLMs to connect multiple natural language processing tasks simplifies the algorithmic logic design of the analysis system, improving overall computational efficiency. Efficient Extraction: Through contextual learning, model fine-tuning, and efficient programming templates for guiding language models, even small models can achieve performance comparable to state-of-the-art (SOTA) LLMs in citation extraction tasks. This system not only enhances the accuracy and efficiency of news opinion extraction and analysis but also provides experts with more intuitive and operable analytical tools, bringing more comprehensive and objective opinion analysis results to the field of news.