透過以LLM實現的新聞監控與分析揭露ESG之媒體輿情;Disclosing Media Sentiment in ESG Through LLM-Enabled News Monitoring and Analytics

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/95444

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/95444

題名:	透過以LLM實現的新聞監控與分析揭露ESG之媒體輿情;Disclosing Media Sentiment in ESG Through LLM-Enabled News Monitoring and Analytics
作者:	劉學逸;Liu, Hsueh-Yi
貢獻者:	資訊工程學系
關鍵詞:	ESG;大型語言模型;自動文本摘要;輿情分析;ESG;LLM;automatic text summarization;sentiment analysis
日期:	2024-07-10
上傳時間:	2024-10-09 16:51:47 (UTC+8)
出版者:	國立中央大學
摘要:	近年來環境、社會、治理(Environmental, Social and Government, ESG)的議題越來越受世界各國的重視，例如歐盟將在2027年正式收取碳稅，其相關的媒體報導與輿情也將影響各家企業的形象甚至是市場價值，因此企業在ESG上的表現以及採取的相關行動在現代是很重要的議題。本研究將使用大型語言模型(Large Language Model, LLM)為基礎的方法對大量ESG相關新聞進行多段式摘要生成以及建議生成。其中多段式摘要生成可以解決一些LLM的輸入內容長度限制導致無法直接總結大量新聞的問題。在摘要生成，我們測試了BERT、Pegasus、GPT-3.5-Turbo以及Llama-2對各個文章內的內容進行初步過濾，在Prompt的設計我們使用了簡易Prompt、複雜Prompt、以及使用Directional Stimulus Prompting(DSP)應用在我們的多段式摘要生成，我們選取了最具代表性的GPT-3.5-Turbo以及公開的Llama-2作為最終階段的摘要生成模型，並且透過Multi-News資料集衡量不同方法的優劣。在建議生成上，我們採用了情感分析的Distil-RoBERTa模型以及ESG分類模型和多段式摘要生成產生的摘要作為大型語言模型生成的輸入加以引導生成的內容方向。本研究的結果展示了在多段式摘要生成的任務上使用不同方法的優劣，以及驗證了在不同摘要生成的大型語言模型上使用這些方法的一致性。另外，在Prompt設計實驗的DSP環節以及建議生成的實驗，展示了小型的模型可以如何進一步加強大型語言模型在不同任務上的表現。本研究提出的自動化工具也可以使企業能夠快速掌握ESG相關媒體輿情，並且得到相關建議能即時做出對應的決策。;Issues related to Environmental, Social, and Governance (ESG) have gained increasing attention from countries worldwide in recent years. For instance, the European Union will officially start implementing carbon taxes in 2027. Media reports and public opinion surrounding ESG issues can significantly impact the image and market value of companies. Therefore, a company′s performance and actions in ESG have become crucial topics in contemporary society. This study applies methods based on Large Language Models (LLMs), using multi-stage summary generation and suggestion generation powered by LLMs on a vast amount of ESG-related news. The multi-stage summary generation addresses the problem of input length limitations of LLMs, which hinder direct summarization of large volumes of news. In the summary generation process, we tested BERT, Pegasus, GPT-3.5-Turbo, and Llama-2 to initially filter the content within each article. For prompt design, we utilized simple prompts, complex prompts, Directional Stimulus Prompting (DSP) on our multi-stage summary generation. We selected the most representative models, GPT-3.5-Turbo and the publicly available Llama-2, as the final models for summary generation and measured the performance of different methods using the Multi-News dataset. For suggestion generation, we employed a sentiment Distil-RoBERTa model and an ESG classification model. These models, along with the summaries generated by multi-stage summary generation, guided the content direction generated by the LLM. The results of this study demonstrate the advantages and disadvantages of using different methods for multi-stage summary generation and validate the consistency of these methods across various summary-generating LLMs. Additionally, experiments involving DSP in prompt design and suggestion generation showcase how smaller models can further enhance the performance LLMs in different tasks. The automated tools proposed in this study enable companies to quickly grasp ESG-related media sentiment and receive relevant suggestions to make timely and informed decisions.
顯示於類別:	[資訊工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	65	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....