以提示學習提供時間感知新聞推薦

Abstract

隨著新聞資訊快速增長，個人化新聞推薦系統已成為使用者獲取資訊的重要輔助工具。近年來，提示學習（Prompt Learning）於自然語言處理任務中展現出良好表現，也被應用於新聞推薦領域。提示學習能有效結合預訓練語言模型的語意理解能力，降低訓練成本並提升推薦效能。然而，現有的提示學習新聞推薦方法尚未考量時間資訊。本研究提出一個時間感知新聞推薦架構（TPrompt4NR），於提示學習新聞推薦架構中引入時間資訊，對推薦效用描述、使用者行為描述、語意相關性及類別資訊四種新聞推薦提示模板進行擴展，加入時間單位資訊。系統性分析不同提示模板加上時間單位資訊的效果，並進一步探討在已提供類別資訊的情境下，加入時間資訊對新聞推薦效能的影響。實驗採用MIND新聞資料集，並以Hit Rate與NDCG作為評估指標。結果顯示，加入時間資訊可有效提升模型推薦效能，特別是在推薦效用描述模板中，並且採用「1小時」為時間單位時表現最佳。即使新聞內容已含類別資訊或缺乏標題資訊，時間資訊亦能提供穩定效益。此外，時間資訊的描述方式也影響模型理解效果。綜合而言，本研究驗證了時間資訊於提示學習新聞推薦中的效果，顯示合理設計的時間資訊能提升推薦準確性。With the rapid growth of news content, personalized news recommendation systems have become essential tools for helping users access relevant information. In recent years, prompt learning has demonstrated strong performance in natural language processing (NLP) tasks and has also been applied to the domain of news recommendation. By leveraging the semantic understanding capabilities of pre-trained language models, prompt learning can reduce training costs and improve recommendation effectiveness. However, existing prompt-based news recommendation methods have yet to incorporate temporal information. This study proposes a time-aware prompt-based news recommendation framework (TPrompt4NR), which integrates temporal information into prompt-based recommendation. We extend four types of prompt templates—recommendation utility, user action, semantic relevance, and category information—by incorporating time-unit representations. We systematically analyze how different prompt templates interact with time information and further examine the impact of adding temporal context when category information is already available. Experiments were conducted on the MIND dataset, using Hit Rate and NDCG as evaluation metrics. Results show that incorporating time information significantly improves model performance, especially for utility-based prompt templates. The one-hour time unit yields the best overall performance. Even when news content already contains category information or lacks headlines, temporal information provides stable benefits. Moreover, the way time is expressed in natural language also affects model comprehension—formats closer to natural language yield better results. In summary, this research confirms the effectiveness of temporal information in prompt-based news recommendation, demonstrating that well-designed temporal cues can enhance recommendation accuracy.