- 學位論文
基於限制性蒙地卡羅搜尋樹之因果故事生成系統
Causal Story Generation Systems Based on Constrained Monte Carlo Tree Search
摘要
現存的故事產生系統多為弱人工智慧，由於電腦本身並不具有語意理解及常識推理的能力，因此自然語言的產生往往建立在填詞模板和訂定大量規則之下，除了極具仰賴人類所給定的專業知識而無法彈性擴增外，大量的規則也讓故事的產生必須耗費冗長的時間。由此，我們建立了一套因果性故事生成系統，在使用者給定生成參數後，能夠在短時間內從大量語料庫中搜索出多樣化、合理且符合使用者需求的故事。在此篇論文中將會提及如何自動的從語料中搜索出因果性劇情結構，並應用限制性蒙地卡羅搜尋樹來有效率的搜索故事序列。此外，我們設計了一套知識庫系統，結合不同深度學習的框架，用於協助多種故事範本以及限制性蒙地卡羅搜尋樹來生成故事。最後，我們設計多種自然語言模版來將我們所生成的故事序列轉換為可讀性故事。在實驗結果中不同故事的生成結果將會分別列出，並且比較限制性蒙地卡羅搜尋樹在調變各種參數上對於故事生成的差別，另外在知識庫系統部分，我們也會展示各種設計的功能性以及差異性，以及對於故事生成結果的影響。
並列摘要
Current story generation systems are in general weak AI systems in the sense that it lacks of semantic understanding and common-sense knowledge that often result in requirements of enormous story templates and narrative rules for computers to generate stories. Consequently, without experts, computers cannot flexibly augment knowledge. Moreover, plenty of time and efforts are taken in create stories by a complicated rule-based generation system. We wish to construct a system of causal story generation capable of searching for diverse, reasonable and user-desired stories quickly from a large knowledge database. In this thesis, we proposed an approach of automatically extracting causal knowledge from existing knowledge base ConceptNet to construct our own database so that we can efficiently search for story sequences using Constrained Monte Carlo Tree Search (cMCTS) algorithm. Furthermore, a Knowledge-Based System is built with deep learning techniques in order to support cMCTS and story frameworks for generating stories. To generate the story in human comprehensive way, we design of various translation templates to convert the formal causal story sequences into natural language sentences. The stories generated under different parameter settings of cMCTS are illustrated as well as comparisons in several simulation experiments along with the evaluations on the functionality of the Knowledge-Based System.