Towards Generalized Natural Language Generation with Distributional Semantics

点击翻译按钮获取中文摘要

From smartphones to smart cars, robotic assistants and beyond, the common goal of many emerging technologies is to extract useful information about the world in order to give users greater control over their surroundings. The field of natural language processing provides techniques to extract meaning from text, reason with it, and finally produce text and speech in order to relay information back to the user. Recent successes in personal assistant applications and news summarization systems have raised the demand for natural language generation (NLG) systems that can interactively produce feedback to the user. Yet many existing methods only work well on highly restricted domains and tasks, as typified by a GPS navigation system with pre-programmed templates for generating driving directions. My research program aims to produce a generalized computational account of NLG that is sensitive to differences in the topics, language types, and goals of an application. This will require an analysis of the parameter space of NLG systems in order to characterize their diversity, as well as a correspondingly adaptive and expressive semantics that can support the reasoning and inferences that are necessary for NLG. For the former requirement, my research group will develop techniques for NLG that take into account the desired usage of the system in its broader context. For example, a news text summarization system might emphasize factors such as brevity and formality, whereas an interactive educational game might emphasize other factors such as simplicity and interest. For the latter requirement, my research program will investigate distributional semantics (DS), a data-driven approach to modelling meaning that can be trained without human annotation effort. My research will focus on using DS to model the meaning of entities such as Quebec or Android, as well as events such as the Olympic Games of 2016, which is required in order to reason about the important information that should be expressed by NLG. The impact of this research will be to move NLG from being confined to highly restricted and domain-specific scenarios to being integrated with interactive systems in a natural and pervasive manner. The tools, techniques, and framework contributed by this research program will demonstrate that NLG systems that are supported by expressive, data-driven semantics can be developed and tailored for education, entertainment, or business.