RI: Small: A Bayesian Approach to Dynamic Lexical Resources for Flexible Language Processing

该项目使用统计模型和人类判断来确定单词可扩展用法的动态概率表示；这些表示适合合并到vernet中，这是一个在自然语言处理（NLP）社区中广泛使用的词汇资源。现有的词汇资源反映了用法是否合乎语法的二元概念。然而，在实际语言使用中，形式的可接受性各不相同；此外，强制的过程使词汇超出了它们的标准用法。例如，一个严格的不及物动词，如“sneeze”，可以用在“She sneeze the foam off the cappuccino”中，表达运动的方式。这项研究有一个双管齐下的方法，包括广泛使用机器学习和从根本上改变vernet的开发和使用。具体而言，该研究开发了概率方法：(1)分析大型语料库中的动词用法，并将得到的概率信息整合到动词网络类中；(2)表示关于潜在构造强制的可能性和这种扩展的生产力的信息。这些发展使用了Parisien和Stevenson的分层贝叶斯模型，这是一个理想的框架，可以将复杂的真实世界数据的概率推理与分层组织的vernet词典结合起来。除了统计模型之外，这些表示也通过人类对这些结构的使用的判断来告知。因此，本研究以概率分布信息丰富了动词网络中现有的符号动词表征，这在构式语法的影响下变得更加突出。动词知识的概率编码为表示外延结构提供了必要的灵活性，并支持NLP系统对其进行适当的解释。这对于新领域和新体裁的翻译尤其有用，导致了NLP技术的进步，例如问答和机器翻译，从而改善了信息访问。此外，通过本研究获得的关于结构统计特性的见解对语言习得和第二语言学习的心理语言学模型有价值。

This project uses statistical models and human judgment to determine dynamic, probabilistic representations of extensible usages of words; these representations are suitable for incorporation into VerbNet, a lexical resource widely used in the Natural Language Processing (NLP) community. Existing lexical resources reflect a binary notion of usages as grammatical or not. However, in actual language use, forms vary in acceptability; moreover, the process of coercion extends words beyond their standard usages. For example, a strictly intransitive action verb such as 'sneeze' may be used as in 'She sneezed the foam off the cappuccino', expressing manner of motion. This research has a two-pronged approach involving extensive use of machine learning and a fundamental shift in the development and use of VerbNet. Specifically, the research develops probabilistic methods for: (1) analyzing usages of verbs in large corpora and incorporating the resulting probabilistic information into VerbNet classes; and (2) representing information about the likelihood of potential constructional coercions and the productivity of such extensions. These developments use the Hierarchical Bayesian Model of Parisien and Stevenson, which are an ideal framework for marrying probabilistic reasoning about complex, real-world data within the hierarchically-organized VerbNet lexicon. In addition to statistical models, the representations are also informed by human judgments with respect to the use of such constructions. Thus, this research enriches the current symbolic verb representations in VerbNet with probabilistic distributional information, which becomes salient through the influence of construction grammar. Encoding verb knowledge probabilistically provides the necessary flexibility to represent extensional constructions and support their appropriate interpretation by NLP systems. This is especially useful for interpretation in new domains and genres, leading to advances in NLP technologies, such as question answering and machine translation, thus improving information access. Additionally, insights into statistical properties of constructions gained through this research are valuable for psycholinguistic models of language acquisition and second language learning.