透過多輪對談互動方式紀錄使用者需求，因此對話中途亦可變換銀行、本金與儲蓄時間值。系統可根據定存時間長度比較本利和最高方案，也可針對特定銀行進行利息與中途解約試算，幫使用者找到最適合的定存儲蓄方案！

以小學數學試題為範例此系統以結構化方式抽取文句中人、事、物之人類既有知識，並利用事件框架的角色互動關係，協助機器進行自然語言深度理解。小學數學如同真實世界的知識縮影，事件框架的分析使用，有助於機器以更準確掌握人類的思考及語言理解過程中很重要的邏輯、基本常識等非常關鍵的知識。

本計畫為小學數學應用題（Math Word Problem, MWP）開發了智慧型學習助手（Intelligent Learning Assistant, ILA），建立一個可解釋的人工智慧（Explainable AI, XAI）案例。ILA能夠自動解出MWP並提供逐步解釋。當ILA犯了錯誤，我們可以很容易地修正；當ILA正確解決問題時，我們可提供解釋讓學生理解其背後的邏輯。本計畫採用統計準則式方法（Statistical Principle-Based Approach, SPBA），保留了統計機器學習（SML）和規則（Rule-Based, RB）方法的優點，同時避免了它們的缺陷。機器人學會了以類似於人類學習的方式解答MWP。MWP解題需要的自然語言理解，除了知識範圍有限，句型較為單純外，一般自然語言研究會遇到的困難，如：指代、省略、倒裝句、文本蘊含(entailment)、剖析、知識本體、推論等等幾乎無所不包。因此，在我們解決了小學數學的自然語言理解後，多輪對談系統所需要的理解模組幾乎已全數完成。目前，我們已實現80％的目標，正在努力實現半自動的知識本體構建和自然語言生成。

20190122-PrimaryMathPoster

Recent advances in virtual assistants point to new ways for human-machine interaction, enabling us to communicate effortlessly with different software and hardware-based agents using natural language. By means of interactive chatting, an intelligent agent can better understand our intentions so as to provide seamless integration of technology into our lives. Currently, voice assistants or chatbots are of high demand in both academics and industries in Taiwan as well as the world. In the research community, the mainstream model relies on statistical machine learning. Although such an approach can be automated in the learning phase, the labor involved in labeling the training data is intensive. Moreover, it is difficult to interpret the parameters in the trained model, which poses a problem when conducting error analysis. Therefore, large commercial systems are constructed via rule-based (RB) methods instead. However, RB methods operate on numerous hand-crafted rules which still require humans to construct. Besides, adding more rules can sometimes create conflicts that require extra cost to resolve, making it difficult to scale RB methods to larger domains. In light of these problems, we propose this research project to bridge the gap between the two models by creating a novel chatbot framework that can understand domain knowledge as well as assist in the human learning process. It can provide commendable conversation capabilities that will boost our prevalence in the world. In our framework, a domain expert can quickly compose scripts for various scenarios. Furthermore, the agent is able to understand and extract important knowledge from documents to generate sentences in order to provide a more natural conversation experience with the user. Ultimately, we hope to provide a performance that surpasses the state-of-the-art technologies in Europe, US, and Japan. In addition, we propose extensions of the chatbot that aim to assist in solving math word problems and essay composition. For math problems, we will design an agent that not only solves but also explains the “how” and “why” of the solution in natural language. Besides, it is able to locate where the user may have difficulty in understanding the reasoning, and supply adequate suggestions. For essay composition, our framework can exceed the level of correcting grammatical errors and present paraphrasing suggestions during composing, so as to broaden the knowledge of the user.

20190122-PrimaryMathPoster