簡易檢索 / 詳目顯示

回結果列表
研究生： 葉家華
Ye, Jia-Hua
論文名稱： 建立露營評論情緒分析管理系統： 利用LLM進行資料增生和深度學習模型評估
A Sentiment Analysis System for Camping Reviews: Leveraging LLMs for Data Augmentation and Deep Learning Evaluation
指導教授： 陳牧言
Chen, Mu-Yen
學位類別： 碩士
Master
系所名稱： 工學院 - 工程科學系
Department of Engineering Science
論文出版年： 2025
畢業學年度： 113
語文別： 中文
論文頁數： 90
中文關鍵詞： 情緒分析 、大型語言模型 、深度學習 、資料增生
外文關鍵詞： Sentiment Analysis, Large Language Models, Deep Learning, Data Augmentation
相關次數： 點閱：21  下載：0 
分享至:
查詢本校圖書館目錄 查詢臺灣博碩士論文知識加值系統 勘誤回報

    • 隨著新冠肺炎疫情改變旅遊型態，露營因其戶外特性與低接觸風險，成為國內旅遊熱門選擇，帶動露營場地數量激增與線上評論快速累積。然而，露營評論常呈現資料不平衡問題，正向評論占多數，負向與中立評論相對稀少，影響情緒分析的全面性與模型訓練效果。同時，環境、社會、治理（ESG）原則逐漸成為旅遊產業永續發展的指導框架，現有露營評論分析系統卻鮮少從ESG視角出發，難以滿足永續經營需求。本研究旨在建構露營評論情緒分析管理系統，結合大型語言模型（LLM）與深度學習技術，解決資料不平衡問題並提升情緒分類準確性，進而協助露營地經營者實現永續發展與顧客滿意度的雙重目標。
    本研究首先透過Google Maps蒐集2018年至2024年間的傳統露營與豪華露營評論，進行資料清理、斷詞與情緒標記（正向、中立、負向）。為解決資料不平衡問題，採用四種LLM（GPT-4o mini、TAIDE、Breeze、Taiwan LLM）生成負向與中立評論，增補稀少類別資料，並比較其生成品質與在地化適用性。隨後，運用五種深度學習模型（BERT、ALBERT、RoBERTa、Multilingual BERT、DistilBERT）進行情緒分類訓練，透過多組實驗評估不同LLM與深度學習模型組合的效能。實驗結果顯示，資料增生顯著提升模型準確度，其中Taiwan LLM因其在地化語料優勢，表現尤為突出。傳統露營的最佳組合為RoBERTa（Batch Size=16），準確度達93.75%；豪華露營則以BERT（Batch Size=8）表現最佳，準確度達97.17%。
    基於最佳模型，本研究開發一套情緒分析管理系統，具備每日評論即時分類、負向評論Line Notify通知、文字雲分析顧客關注焦點，以及基於ESG的每週改善建議生成等功能。該系統不僅協助業者快速回應顧客需求、優化營地設施，還能透過ESG導向建議促進環境保護與經營透明度，實現永續經營目標。本研究展示LLM與深度學習在露營評論分析中的應用潛力，未來可擴展至其他旅遊領域，並進一步優化資料蒐集與模型效能，為永續旅遊產業提供更全面的技術支持。

    As the COVID-19 pandemic reshaped travel patterns, camping has emerged as a popular domestic tourism choice due to its outdoor nature and low-contact risk, driving a surge in campsite numbers and rapid accumulation of online reviews. However, camping reviews often exhibit data imbalance, with positive reviews dominating while negative and neutral reviews remain scarce, impacting the comprehensiveness of sentiment analysis and model training effectiveness. Meanwhile, Environmental, Social, and Governance (ESG) principles have increasingly become a guiding framework for sustainable development in the tourism industry, yet existing camping review analysis systems rarely adopt an ESG perspective, making it challenging to meet sustainable management needs. This study aims to develop a sentiment analysis management system for camping reviews, integrating large language models (LLMs) and deep learning techniques to address data imbalance and enhance sentiment classification accuracy, ultimately assisting campsite operators in achieving both sustainable development and customer satisfaction.
    The study first collects traditional and glamping reviews from Google Maps between 2018 and 2024, performing data cleaning, tokenization, and sentiment labeling (positive, neutral, negative). To address data imbalance, four LLMs (GPT-4o mini, TAIDE, Breeze, Taiwan LLM) are employed to generate negative and neutral reviews, augmenting scarce data categories, with their generation quality and localization applicability compared. Subsequently, five deep learning models (BERT, ALBERT, RoBERTa, Multilingual BERT, DistilBERT) are used for sentiment classification training, with multiple experiments evaluating the performance of different LLM and deep learning model combinations. Results show that data augmentation significantly improves model accuracy, with Taiwan LLM excelling due to its localized corpus advantage. For traditional camping, the best combination is RoBERTa (Batch Size=16), achieving 93.75% accuracy; for glamping, BERT (Batch Size=8) performs best, reaching 97.17% accuracy.
    Based on the optimal model, this study develops a sentiment analysis management system featuring real-time daily review classification, Line Notify alerts for negative reviews, word cloud analysis of customer concerns, and ESG-based weekly improvement suggestions. The system not only enables operators to promptly address customer needs and optimize campsite facilities but also promotes environmental protection and operational transparency through ESG-guided recommendations, achieving sustainable management goals. This study demonstrates the potential of LLMs and deep learning in camping review analysis, with future applications extensible to other tourism sectors and further optimization of data collection and model performance to provide comprehensive technical support for the sustainable tourism industry.

    摘要 I Extended Abstract I 誌謝 I 目錄 II 表目錄 VI 圖目錄 VII 第一章、 緒論 1 1.1 研究背景與動機 1 1.2 研究目的 3 1.3 章節摘要 5 第二章、 文獻探討 6 2.1 露營的起源與演變 6 2.1.1 傳統露營 7 2.1.2 豪華露營 7 2.1.3 傳統露營與豪華露營的比較 8 2.2 深度學習模型 8 2.2.1. BERT（Bidirectional Encoder Representations from Transformers） 9 2.2.2. ALBERT（A Lite BERT for Self-supervised Learning of Language Representations） 10 2.2.3. RoBERTa（Robustly optimized BERT approach） 12 2.2.4. Multilingual BERT 13 2.2.5. DistilBERT 14 2.3 大型語言模型 14 2.3.2. 繁體中文大型語言模型介紹 15 2.4 大型語言模型與深度學習在語意分析的應用 18 2.4.1. 語意分析技術的演進 18 2.4.2. 深度學習在語意分析的挑戰與進展 18 2.4.3. LLM在特定領域語意分析的應用 19 第三章、 研究方法 20 3.1 研究架構 20 3.2 資料蒐集 21 3.3 資料前處理 22 3.4 資料平衡 25 3.4.1 提示工程 (Prompt Engineering) 25 3.4.2 生成評論流程 26 3.5 建立研究模型 29 3.5.1 實驗一：使用原始資料集建立情緒分類模型 29 3.5.2 實驗二：使用LLM增生以平衡資料集並建立情緒分類模型 29 第四章、 實驗設計與結果分析 30 4.1 實驗環境 30 4.2 模型參數設定 31 4.3 評估績效指標 31 4.3.1 準確率（Accuracy） 32 4.3.2 精確率（Precision） 32 4.3.3 召回率（Recall） 32 4.3.4 F1 Score 33 4.4 文字雲 33 4.5 實驗結果與討論 35 4.5.1 傳統露營（實驗一） 35 4.5.2 傳統露營（實驗二） 37 4.5.2.1. GPT-4o mini資料增生 37 4.5.2.2. TAIDE資料增生 40 4.5.2.3. Breeze資料增生 42 4.5.2.4. Taiwan LLM資料增生 45 4.5.3 豪華露營（實驗一） 48 4.5.4 豪華露營（實驗二） 50 4.5.4.1. GPT-4o mini資料增生 50 4.5.4.2. TAIDE資料增生 53 4.5.4.3. Breeze資料增生 56 4.5.4.4. Taiwan LLM資料增生 58 4.5.5 實驗結果討論 61 4.5.6 露營評論情緒分析管理系統 61 第五章、 結論與未來展望 66 5.1 研究貢獻 66 5.2 研究限制 67 5.3 未來展望 67 參考文獻 69

    [1]. 林人芳 (2023, 7月)。疫情趨緩　民眾走向戶外瘋露營. 財團法人海峽交流基金會. 擷取自 https://www.sef.org.tw/article-1-129-10807.
    [2]. 蟬說 (2022, 9月)。露營跟豪華露營差別是什麼？一次認識4種露營種類與露營注意事項. 蟬說. 擷取自https://www.chanshuo.tw/%E9%9C%B2%E7%87%9F%E7%A8%AE%E9%A1%9E/.
    [3]. Google Trends. (n.d.). Google. Retrieved Jan 30, 2025, from https://trends.google.com.tw/trends.
    [4]. 林子硯（2015）。改善多類別不平衡資料之分類準確率。﹝碩士論文。國立交通大學﹞臺灣博碩士論文知識加值系統。 https://hdl.handle.net/11296/3y5482.
    [5]. Chung, J. J. Y., Kamar, E., & Amershi, S. (2023). Increasing diversity while maintaining accuracy: Text data generation with large language models and human interventions. arXiv preprint arXiv:2306.04140.
    [6]. Devlin, J. (2018). Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805.
    [7]. Lan, Z. (2019). Albert: A lite bert for self-supervised learning of language representations. arXiv preprint arXiv:1909.11942.
    [8]. 唐偉展(2024, 2月)。露營新時代 融入綠色永續與ESG概念 政府應積極引導行業發展. 旅報. 擷取自 https://www.ttnmedia.com/?p=98744.
    [9]. De Rosa, G. H., & Papa, J. P. (2021). A survey on text generation using generative adversarial networks. Pattern Recognition, 119, 108098.
    [10]. Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., ... & Bengio, Y. (2014). Generative adversarial nets. Advances in neural information processing systems, 27.
    [11]. Vaswani, A. (2017). Attention is all you need. Advances in Neural Information Processing Systems.
    [12]. Sahoo, P., Singh, A. K., Saha, S., Jain, V., Mondal, S., & Chadha, A. (2024). A systematic survey of prompt engineering in large language models: Techniques and applications. arXiv preprint arXiv:2402.07927.
    [13]. Liu, Y. (2019). Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 364.
    [14]. Muller, B., Elazar, Y., Sagot, B., & Seddah, D. (2021). First align, then predict: Understanding the cross-lingual ability of multilingual BERT. arXiv preprint arXiv:2101.11109.
    [15]. Sanh, V. (2019). DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter. arXiv preprint arXiv:1910.01108.
    [16]. OpenAI. (2024). GPT-4o mini: advancing cost-efficient intelligence. OpenAI Blog. Retrieved from https://openai.com/news.
    [17]. 財團法人國家實驗研究院. (2024). taide/Llama3-TAIDE-LX-8B-Chat-Alpha1. Hugging Face. Retrieved from https://huggingface.co/taide/Llama3-TAIDE-LX-8B-Chat-Alpha1.
    [18]. 聯發創新基地. (2024). MediaTek-Research/Breeze-7B-Instruct-v1_0. Hugging Face. Retrieved from https://huggingface.co/MediaTek-Research/Breeze-7B-Instruct-v1_0.
    [19]. Lin, Y. T. (2024). yentinglin/Llama-3-Taiwan-8B-Instruct. Hugging Face. Retrieved from https://huggingface.co/yentinglin/Llama-3-Taiwan-8B-Instruct.
    [20]. Brown, T., Mann, B., Ryder, N., Subbiah, M., Kaplan, J. D., Dhariwal, P., ... & Amodei, D. (2020). Language models are few-shot learners. Advances in neural information processing systems, 33, 1877-1901.
    [21]. Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., ... & Liang, P. (2021). On the opportunities and risks of foundation models. arXiv preprint arXiv:2108.07258.
    [22]. Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019).
    Language models are unsupervised multitask learners. OpenAI Blog. Retrieved from https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf.
    [23]. Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M. A., Lacroix, T., ... & Lample, G. (2023). Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971.
    [24]. Abacha, A. B., Yim, W. W., Fu, Y., Sun, Z., Yetisgen, M., Xia, F., & Lin, T. (2024). Medec: A benchmark for medical error detection and correction in clinical notes. arXiv preprint arXiv:2412.19260.
    [25]. Hsu, C. J., Liu, C. L., Liao, F. T., Hsu, P. C., Chen, Y. C., & Shiu, D. S. (2024). Breeze-7b technical report. arXiv preprint arXiv:2403.02712.
    [26]. Lin, Y. T., & Chen, Y. N. (2023). Taiwan llm: Bridging the linguistic divide with a culturally aligned language model. arXiv preprint arXiv:2311.17487.
    [27]. Pires, T., Schlinger, E., & Garrette, D. (2019). How multilingual is multilingual BERT?. arXiv preprint arXiv:1906.01502.
    [28]. Gupta, S., Ranjan, R., & Singh, S. N. (2024). Comprehensive study on sentiment analysis: From rule-based to modern llm based system. arXiv preprint arXiv:2409.09989.
    [29]. Islam, M. S., Kabir, M. N., Ghani, N. A., Zamli, K. Z., Zulkifli, N. S. A., Rahman, M. M., & Moni, M. A. (2024). Challenges and future in deep learning for sentiment analysis: a comprehensive review and a proposed novel hybrid approach. Artificial Intelligence Review, 57(3), 62.
    [30]. Shaikh, S., Daudpota, S. M., Yayilgan, S. Y., & Sindhu, S. (2023, December). Exploring the potential of large-language models (LLMs) for student feedback sentiment analysis. In 2023 International Conference on Frontiers of Information Technology (FIT) (pp. 214-219). IEEE.
    [31]. Google Maps. (n.d.). Google. Retrieved Feb 28, 2025, from https://www.google.com/maps.
    [32]. 蔡俊青（2022）。應用情緒分析於Google Map評論與推薦系統實作。﹝碩士論文。國立成功大學﹞臺灣博碩士論文知識加值系統。 https://hdl.handle.net/11296/ctmsp3.
    [33]. Reimers, N., & Gurevych, I. (2019). Sentence-bert: Sentence embeddings using siamese bert-networks. arXiv preprint arXiv:1908.10084.
    [34]. Grandia (2023, 8月)。古老的露營傳統：人與自然的對話.隨心所欲. 擷取自 https://mysangui.com/journey/%E5%8F%A4%E8%80%81%E7%9A%84%E9%9C%B2%E7%87%9F%E5%82%B3%E7%B5%B1%EF%BC%9A%E4%BA%BA%E8%88%87%E8%87%AA%E7%84%B6%E7%9A%84%E5%B0%8D%E8%A9%B1/.
    [35]. 辣個露營 (2023, 8月)。豪華露營Glamping」與「傳統露營」六大特色差異你知道嗎？來看你適合哪一種？.辣個露營. 擷取自 https://spicycamping.com/camping/.
    [36]. Wikipedia. (2025). Glamping. Wikipedia. Retrieved from https://en.wikipedia.org/wiki/Glamping.
    [37]. Wikipedia. (2025). Camping. Wikipedia. Retrieved from https://en.wikipedia.org/wiki/Camping.
    [38]. Liu, P., Yuan, W., Fu, J., Jiang, Z., Hayashi, H., & Neubig, G. (2023). Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM computing surveys, 55(9), 1-35.
    [39]. Kaggle. (2025). Prompt Engineering. Kaggle. Retrieved from https://www.kaggle.com/whitepaper-prompt-engineering.

    無法下載圖示 校內：2028-08-31公開
    校外：2028-08-31公開
    電子論文尚未授權公開，紙本請查館藏目錄
    QR CODE