DILiGENt: Domain-Independent Language Generation

我们提出了一个为期两年的项目，通过结合三个领域的最新进展，开发一种新的数据驱动方法，以快速创建新领域的高质量NLG系统：(1)NLG统计模型的进展；(2)自然语言数据收集的众包方法，这些方法在相关领域（如机器翻译）首次显示出有希望的结果；(3)最近开发的用于结构化预测的模仿学习算法。项目团队结合了这些领域的两个领先研究小组的专业知识：在Heriot-Watt大学，我们最近展示了在有限领域中数据驱动的统计NLG的潜力。为了使这个框架独立于领域，我们将利用由伦敦大学学院的研究人员开发的最新机器学习模型。这些模型通过模仿人类专家生成自然语言话语的行为来学习，我们通过紧密集成的众包程序收集这些话语。这项工作的结果是一个框架，它将允许在新领域快速开发NLG系统，从而加速NLG技术对市场的影响。我们将在英国广播公司提供的数据集上展示这个框架，在那里我们解决了为超过20,000个单独地点生成天气报告的问题。目前，BBC网站上只有10篇气象学家撰写的报道。这些报告中的每一个都覆盖了相当大的国家区域（例如英格兰东部），因此对于通常对特定地点（例如诺里奇）的天气感兴趣的用户来说几乎没有兴趣。在第二步中，我们将探索这个框架如何扩展到更复杂的互动对话设置，其中生成必须考虑话语现象，如长距离话语关系或句法协调。这将在Heriot-Watt大学主办的交互式系统生成共享任务挑战中进行评估。总而言之，该项目将进一步加深我们对领域独立语言生成的理解，并提供大量新颖的资源来支持该领域（以代码和数据的形式）的未来研究，以及NLG系统在广泛领域（从天气报告到自然语言接口）的实际实现。

We propose a two year project to develop a novel data-driven methodology to rapidly create high quality NLG systems for new domains, by combining recent advances in three domains: (1) advances in statistical models for NLG, (2) crowdsourcing methods for natural language data collection, which have shown first promising results in related fields, such as Machine Translation, and (3) recently developed imitation learning algorithms for structured prediction. The project team combines expertise of two leading research groups in these areas:At Heriot-Watt University, we recently demonstrated the potential for data-driven statistical NLG in limited domains. In order to make this framework domain-independent we will leverage recent machine learning models, developed by researchers at the University College London. These models learn by imitating the actions a human expert would perform to generate NL utterances, which we collect via a tightly integrated crowdsourcing procedure. The outcome of this work is a framework which will allow the rapid development of NLG systems for new domains, and thus accelerate the impact NLG technology has on the market. We will showcase this framework on a dataset provided by the BBC, where we address the problem of generating weather reports for over 20,000 individual locations. Currently, the BBC website features only 10 reports written by meteorologists. Each of these reports covers a rather large area of the country (e.g. East of England), and thus of little interest to their users who are usually interested in the weather in a particular location (e.g. Norwich).In a second, more ambitious step, we will explore how this framework scales to more complex interactive dialogue settings, where generation has to account for discourse phenomena, such as long-distance discourse relations or syntactic coordination. This will be evaluated in a shared task challenge for generation in interactive systems, hosted by Heriot-Watt University.In sum, this project will further our understanding of domain-independent language generation, as well as deliver substantial and novel resources to support future research in this area (in the forms of code and data), and practical implementations of NLG systems in a wide-range of domains, from weather reports to natural language interfaces.