Answering Comparative Questions with Arguments (ACQuA 2.0)

在现代社会，个人每天都面临着选择的问题。在本项目中，我们将自然语言处理和信息检索的方法结合起来，设计了一种新型的比较问答议论机。这样的机器满足了用户无处不在的比较信息需求，这些用户对在丰富的相似对象(例如，数码相机、作为旅游目的地的城市或编程语言)之间的选择感到困惑。也就是说，机器在各自的领域中扮演着人类专家的角色，将问题作为输入，在给定的上下文中建议最佳对象选择，并用论点支持答案。与与人类专家交谈类似，与比较论证机器的交互是使用自然语言进行的。该项目的主要科学挑战是设计一种通用的方法，该方法能够很好地覆盖各个领域的对象及其方面，以及与其他对象相比时支持或反对对象的论点，但不依赖于特定于领域的资源。一些工业系统能够使用特定的手动创建的数据库，在相当窄的领域内比较产品。然而，即使是各自领域中最先进的工业体系，也都有很大的局限性，可以很好地满足跨领域的比较信息需求。目前，没有以合理的覆盖面回答一般比较问题的方法，这正是我们在项目中解决的问题。为了达到这一目标，我们提出了一种新的方法，包括几个阶段：(1)用户理解比较问题，(2)从网络规模的文本语料库中检索与问题相关的比较论证结构，(3)从基于维基百科的知识库中收集对象的方面，其中包括从文本中提取的方面，(4)基于比较论证结构和对象方面的对象比较，(5)支持对象选择的论点的生成，以及(6)答案呈现。我们技术的关键是一种新的论点挖掘方法，它执行与输入问题相关的文档的联合检索，并从文本中提取复杂的论点结构。我们的项目将把论点挖掘放在回答比较问题的复杂任务的背景下，展示论点挖掘如何能够创造新类型的语义技术。提出的比较问答方法的突出应用是对话系统、决策支持系统和网络搜索引擎中的直接答案。

In modern society, individuals are faced with choice problems on a daily basis. In this project, we combine methods from natural language processing and information retrieval to design a new kind of argumentative machine for comparative question answering. Such a machine satisfies the ubiquitous comparative information needs of users puzzled by a choice between abundant analogous objects (e.g., digital cameras, cities as travel destinations, or programming languages). Namely, the machine acts like a human expert in a respective field, taking a question as an input and suggesting the best object choice in the given context, supporting the answer with arguments. Similar to talking to human experts, interaction with the comparative argumentative machine is performed using natural language. The main scientific challenge of the project is to design a generic approach that yields a good coverage of objects and their aspects over various domains, as well as arguments for or against objects when compared with others but not relying on domain-specific resources. Some industrial systems are able to compare products in rather narrow domains, using specific manually created databases. Yet, even the most advanced industrial systems in the respective fields all have significant limitations satisfying comparative information needs robustly across domains. Currently, there is no methodology for answering comparative questions in general with reasonable coverage, which is precisely what we address in our project. To reach this goal, we propose a new approach involving several stages: (1) understanding comparative questions by users, (2) retrieving comparative argumentative structures relevant to a question from a web-scale text corpus, (3) gathering aspects of objects from a Wikipedia-based knowledge base enriched with aspects extracted from text, (4) comparison of objects on the basis of comparative argumentative structures and object aspects, (5) generation of arguments supporting the object choice, and (6) answer presentation. Key to our technology is a novel approach to argument mining that performs joint retrieval of documents relevant to the input question and extraction of complex argumentative structures from text. Our project will put argument mining in the context of the complex task of answering comparative questions, showing how argument mining can enable the creation of new kinds of semantic technologies. Prominent applications of the proposed comparative question answering approach are dialogue systems, decision support systems, and direct answers in web search engines.