The Interactive Cookbook

该项目的目标是开发一种“互动食谱”：一种指导用户烹饪菜肴过程的口语对话系统。该系统将离线分析给定菜肴的单个食谱文本，并将它们聚合到一个符号食谱图表中，该图表捕捉到解释和执行食谱的其他方式。在对话时，自然语言理解(NLU)组件将处理关于菜肴的用户请求，而自然语言生成(NLG)组件将基于食谱图生成英语或德语的逐步说明。对话系统将根据用户的需要调整呈现食谱的详细程度。在开发这样一本交互式食谱的过程中，我们将面临许多研究挑战。首先，烹饪食谱是一种相当具体的文本类型，在主流语料库上培训的NLU方法必须适应食谱领域。我们还将把同一道菜的不同食谱合并到一个单一的表示中，这样我们就可以根据用户的要求切换到不同的细节级别。第二，我们将不得不改进神经单词的嵌入，以捕捉特定领域的含义差异：“盐”和“糖”在分布上非常相似，但将食谱中的一个与另一个混淆将极大地改变菜肴。最后，我们将不得不超越NLG中的现有方法：基于语法的方法在解释任意食谱时遇到覆盖问题，而神经方法，如最近抽象总结方法中的方法，难以产生传达给定含义的语言。因此，该项目的一个重点将是通过将神经NLG系统限制在符号食谱图上，来生成在语义上真实于潜在食谱的指令。

The goal of this project is to develop an "Interactive Cookbook": a spoken dialogue system that guides users through the process of cooking a dish. The system will analyze individual recipe texts for a given dish offline and aggregate them into a symbolic recipe graph, which captures alternative ways in which the recipe can be explained and carried out. At dialogue time, a natural language understanding (NLU) component will process user requests regarding the dish, and a natural language generation (NLG) component will generate step-by-step instructions in English or German based on the recipe graph. The dialogue system will adapt the level of detail at which it presents the recipe to the user's needs.In developing such an Interactive Cookbook, we will face a number of research challenges. First, cooking recipes are a rather specific type of text, and NLU methods trained on main- stream corpora will have to be adapted to the recipe domain. We will also consolidate different recipes for the same dish into a single representation, so we can switch to different levels of detail as required for the user.Second, we will have to refine neural word embeddings to capture domain-specific meaning distinctions: “salt” and “sugar” are distributionally very similar, but confusing one for the other in a recipe will drastically change the dish.Finally, we will have to move beyond current methods in NLG: grammar-based methods run into coverage problems when explaining arbitrary recipes, and neural methods such as those in recent approaches to abstractive summarization struggle to produce language which conveys a given meaning. One focus of the project will therefore be to generate instructions which are semantically true to the underlying recipes, by conditioning a neural NLG system on symbolic recipe graphs.