Rationalizing Recommendations

在许多情况下，专家和外行都会收到关于如何行动、做出哪些决定或购买哪些产品的建议。然而，导致某一特定建议的情况和论据往往不透明。虽然在大多数情况下有各种备选方案可供选择，但建议的接受者通常在采取行动的压力下，因此缺乏时间和其他资源来系统地将建议的备选方案与所有可用的备选方案进行比较。我们的目标是开发一种计算方法，在给定建议的情况下，能够执行我们称之为合理化建议的任务，这包括通过系统地将推荐选项与其他可用选项进行比较，将推荐选项设置为更全面的上下文中，并生成描述建议的相对优缺点的比较论证摘要。为此，我们将开发新的信息提取解决方案，利用概率图模型从科学文献中提取相关证据。我们将扩展当前的信息检索方法，以检索与从异构文档集合中生成参数相关的证据。基于提取的证据，我们的方法将生成支持或反对某一建议的论据，并将其与特定维度或方面的其他选项进行比较。为此，我们提出了一种新的数据结构，称为分层论证树，它提供了分层结构论证的主干。考虑到不同比较维度之间的偏好，树可以用来推断在哪些假设下给定的建议可以嵌入到有效的论证中。我们将进一步开发一种基于模板的自然语言生成方法，该方法基于这种分层论证树和其中明确的偏好，用自然语言生成复杂而简洁的论证，在哪些条件和假设下，与其他选项相比，推荐的选项是最好的，从而支持决策。作为概念的证明，我们将我们的方法应用于循证医学背景下的治疗建议。事实上，针对患者的最佳治疗方案的医疗决策通常是基于实验室诊断的建议。通过使这些建议的假设和前提明确且可追溯，我们的方法可以支持医生在当前科学文献中描述的治疗证据水平的支持下做出更明智的决定。我们打算与医学专家密切合作开发我们的系统，他们将在需求和用例定义、数据注释以及系统的最终评估方面支持我们。

In many situations, experts and laymen receive recommendations on how to act, which decisions to take, or which products to purchase. Very often, however, the circumstances and arguments that led to a particular recommendation are not transparent. While in most cases there are various alternatives to choose from, receivers of recommendations are usually under pressure to act, thus lacking time and other resources to systematically compare the recommended option against all available alternatives.Our goal is to develop a computational methodology that, given a recommendation, is able to perform a task we call rationalizing recommendations, which consists in setting the recommended option into a more comprehensive context by systematically comparing it to other available options, and generating a comparative argumentative summary describing the relative pros and cons of the recommendation. To this end, we will develop novel information extraction solutions that employ probabilistic graphical models to extract relevant evidence from scientific literature. We will extend current information retrieval methods to retrieve evidence relevant for generating arguments from heterogeneous document collections. Based on the extracted evidence, our method will generate arguments in favour or against a certain recommendation in comparison to other options with respect to a particular dimension or aspect. To this end, we propose a new data structure called Hierarchical Argumentation Tree that provides the backbone of a hierarchically structured argument. Taking preferences between different comparative dimensions into account, the tree can be used to infer under which assumptions the given recommendation can be embedded into a valid argument. We will further develop a template-based natural language generation approach which, based on this hierarchical argumentation tree and the preferences made explicit therein, generates a complex yet concise argument in natural language that makes explicit and transparent under which conditions and assumptions the recommended option is the best one compared to other options, thus supporting decision making. As proof of concept, we apply our methodology to the case of therapy recommendations in the context of evidence-based medicine. In fact, medical decision making aiming at the optimal therapy for a patient is often based on recommendations accompanying a laboratory diagnosis. By rendering the assumptions and premises underlying such a recommendation explicit and traceable, our methodology could support physicians in making more informed decisions that are backed by the level-of-evidence of a therapy as described in the current scientific literature. We intend to develop our system in close cooperation with medical experts who will support us in requirement and use case definition, data annotation as well as the final evaluation of the system.