Probabilistic Models of Semantic and Pragmatic Acquisition and Processing

尽管Siri和Alexa这样的工具——这可能意味着《她》和《机械姬》中的类人系统即将问世——但与计算机进行真实的自然语言交互仍然是一个难以实现的目标。最近的“大数据”方法可以在文本和语音处理中实现令人印象深刻的行为，但类似人类的语言处理仍然是人工智能（AI）的重大挑战之一。捕捉词义的复杂性和可扩展性是实现这一目标的主要障碍，然而人类学习词义并使其适应新情况几乎毫不费力。只有将语言能力建模为一个计算认知系统，我们才有希望在未来的人工智能应用中实现类似人类的表现。该提案的具体目标是设计反映前沿语言学和心理学理论的最先进的计算表示和算法，目标是在表示、学习和解释单词的含义方面实现类似人类的行为，从而导致三组相关的研究项目。******第一组项目将提供基于自然语言处理（NLP）的对齐和图方法的新算法，这些算法可以有效地产生跨语言语义数据结构。我们将证明这些表示更准确地捕获人类语义知识的属性，从而可以支持改进的自动语言分析工具。第二组项目将展示采用最先进的机器学习技术将认知影响整合到单词学习中的好处。这些技术将使我们能够将对单词学习的科学理解扩展到对结构化和抽象意义的习得。基于前沿的计算机科学方法，这些发现可以指导NLP方法提取丰富的语义关系，并利用自上而下和自下而上的信息来获取意义。我们的第三组项目将通过整合反映认知因素的增量概率学习方法来扩展最先进的认知参考模型。通过开发捕捉实验证明的影响并适应对话过程的模型，我们将有助于增加对NLP系统中必须反映的因素的理解，以匹配人类在对话中的期望。总的来说，这里的项目展示了多学科研究计划的双重好处：将计算机科学的进步用于提高我们对人类认知作为计算系统的理解，并通过在计算框架内适应最新的语言学和心理语言学见解来为NLP提供信息。

Despite tools such as Siri and Alexa – which may suggest that the human-like systems in Her and Ex Machina are just around the corner – realistic natural language interaction with a computer remains an elusive goal. Recent “big data” approaches can achieve impressive behaviours in text and speech processing, but human-like processing of language remains one of the grand challenges of artificial intelligence (AI). Capturing the complexity and extensibility of word meanings is a major obstacle to achieving that goal, yet humans learn word meanings and adapt them to new situations almost effortlessly. Only by modeling language abilities as a computational cognitive system can we hope to achieve human-like performance in future AI applications. The specific objective of this proposal is to devise state-of-the-art computational representations and algorithms that reflect leading-edge linguistic and psychological theories, with the goal of achieving human-like behaviour in representing, learning, and interpreting the meaning of words, leading to three related sets of research projects.******The first set of projects will contribute novel algorithms drawing on alignment and graph methods from natural language processing (NLP) that can efficiently yield crosslinguistic semantic data structures. We will demonstrate that these representations more accurately capture properties of human semantic knowledge, and can thus support improved automatic language analysis tools. The second set of projects will demonstrate the benefit of adapting state-of-the-art machine learning techniques to integrate cognitive influences on word learning. Such techniques will enable us to extend the scientific understanding of word learning to acquisition of structured and abstract meanings. Given their basis in leading-edge computer science approaches, these findings can inform NLP methods for extracting rich semantic relations and exploiting top-down and bottom-up information in acquiring meaning. Our third set of projects will extend the state-of-the-art in cognitive models of reference by integrating incremental probabilistic learning methods that reflect cognitive factors. By developing models that capture experimentally-demonstrated influences and that adapt over the course of a conversation, we will contribute to increased understanding of the factors that must be reflected in NLP systems to match human expectations in conversation. Overall, the projects here demonstrate the dual benefits of the multidisciplinary research program: Bringing advances in computer science to bear on improving our understanding of human cognition as a computational system, and informing NLP by adapting recent linguistic and psycholinguistic insights within a computational framework.