CRII: RI: What do you mean? -- Automatic identification of inferences drawn from text

在处理语言时，读者/听者所理解的远不止他们所读/听到的单词的字面意思。例如，如果你儿子的老师告诉你“我怀疑你的儿子会通过他的考试”，你会推断你的儿子可能会考试不及格。这个项目研究如何自动推导出人们通常绘制的系统推断。为了接近人类的理解，自然语言处理（NLP）系统必须开发广泛覆盖的模型，以捕获语言中没有明确表达的信息。这个项目关注的是事件是如何被感知的：在上面的例子中，人们相信儿子会通过吗？准确地识别商定并作为事实的事件对于任何需要准确推理过程的NLP任务（例如信息提取）都具有重要意义。该项目的成果包括更好地掌握如何使用语言洞察力来自动接近人类水平的理解，以及公开可用的数据，这些数据将用于开发强大的、广泛覆盖的NLP系统，以及评估和改进语言理论。自动获得共同推理的目标是通过开发分类器来实现的，这些分类器将语义学和语用学中研究的语言见解作为特征，并通过众包技术构建来自不同类型的自然发生示例的数据集，并用人类的直觉进行注释。最近，NLP领域的大量工作都集中在“表面”特征对NLP任务的作用上。但这些功能已经达到了极限。这个项目展示了专业的语言特征如何超越了给定可用数据的表面信息所能接近的范围，并导致了NLP系统的根本进步。从理论语言学的角度来看，语义和语用特征与NLP任务的表现无关。通过定量研究大量自然发生的例子中语言特征的相互作用，该项目不仅对NLP有影响，而且对语义和语用理论也有影响。

When dealing with language, readers/hearers understand far more than only the literal meaning of the words they read/hear. For instance, if your son's teacher tells you "I doubt your son will pass his exam", you will infer that your son will probably fail the exam. This project investigates how to automatically derive systematic inferences that people commonly draw. To approximate human understanding, it is essential for natural language processing (NLP) systems to develop broad-coverage models that capture what is conveyed in language without being explicitly said. The project focuses on how events are perceived: in the example above, do people believe that the son will pass? Accurately identifying events that are agreed upon and taken as facts has implications for any NLP task that require an accurate inference process, such as in information extraction. The outcomes of the project consist of a better grasp of how linguistic insights can be used to automatically approximate human-level understanding, and publicly available data that will serve to develop robust, broad-coverage NLP systems as well as to evaluate and sharpen linguistic theories. The goal of automatically deriving common inferences is pursued by developing classifiers that bring in, as features, linguistic insights studied in semantics and pragmatics, and by constructing a dataset of naturally occurring examples, from different genres, annotated with humans' intuitions via crowdsourcing techniques. A large body of work in NLP is recently focusing on the power of "surface" features for NLP tasks. But such features are reaching a limit. This project demonstrates how specialized linguistic features go beyond what can be approximated with surface-level information given available data and leads to fundamental advances in NLP systems. Independently of performance on NLP tasks, semantic and pragmatic features are of interest from a theoretical linguistics perspective. By quantitatively studying the interactions of linguistic features on a large amount of naturally-occurring examples, this project has an impact not only for NLP but also for semantic and pragmatic theories.