基於檢索的中文幽默對話系統建置與評估

Abstract

幽默對話是人際溝通中一項重要的元素，也是人機互動的重要進程之一。本研究透過實作中文幽默對話系統—「破冰機器人」。設置情境，讓使用者查詢相關的笑話並說出，以打破尷尬、僵硬的氣氛並評估其成效。透過系統開發研究法的循環步驟，經過回饋後加入Word2Vec的查詢擴展、關鍵字查詢提示，以及好笑笑話的隨機推薦等功能，讓使用者找不到笑話的比例從25.4%降低到16.7%，而系統達到的破冰效果從27.9%提升到39.9%。可以知道系統優化後確實可以有效的提升使用者的使用率以及破冰效果。實驗後進行語料庫的一致性評估，研究發現： 1. 破冰機器人確實可達到其成效。 2. 語料庫中的好笑程度與使用者的認知接近一致性的臨界值：使用者認為越好笑的笑話，越能達到破冰效果。 綜合而言，本研究的貢獻，不僅進行了幽默語料庫的應用，也建置中文幽默對話系統。並且在研究過程與結果中，提供了實證經驗與意涵：笑話語料的豐富程度與品質（收集更多笑話並標註好笑程度）、以及普遍使用者已經習慣推薦功能大於自己查詢的趨勢。後續的各類對話系統，建議應運用類似的推薦功能，以符合現今使用者的習慣與期待。Humorous dialogue is an important element in interpersonal communication and is also one of the important processes for human-computer interaction. The purpose of this research is to develop related technologies, implement a retrieval-based "icebreaker robot" system which allows users to find relevant jokes for use in relaxing an unduly formal atmosphere when interacting with people, and evaluate its effectiveness. Through the iterative steps of the information system development methodology, query expansion based on Word2Vec technology, frequent keyword prompts, and random recommendation of good jokes are added after user feedback. The results are that the proportion of user queries that fail to find jokes is reduced from 25.4% to 8.0% and that the icebreaker effect achieved has been increased from 25.9% to 40.9%. System optimization can accurately increase the usage rate and effectiveness. By the conformance assessment, get the conclusion of the research below: 1. Icebreaker robot has an effect on relaxing an unduly formal atmosphere. 2. The humor level in the corpus and users cognition are conformance but close to critical value：the funnier jokes that user thinks, better the effect of icebreaker can be achieved. Empirical experience and implications of this study include: the richness and quality of joke corpus (collecting more jokes and identifying their humor level) and the automatic recommendation relative to passive search are important R& D efforts to improve the effectiveness of such services.