Computational natural language processing and the neuro-cognition of language

对人类语言系统的科学理解是认知科学面临的主要挑战之一，也是语言学、心理学、人类学、哲学、生物学和计算机科学等不同领域的兴趣所在。一个关于这个复杂系统的充分的理论需要整合来自几个领域的科学知识，其中许多领域仍处于快速发展的状态。其中一个领域是认知神经科学，它研究人类大脑中的语言功能，目的是发展人类动态理解和语言产生能力的神经生物学和认知上的合理解释。这项研究的一个关键输入是关于语言核心属性的语言信息。这些信息通常是从传统的资源（字典和语法）中获得的，这些资源提供了关于语言的有用概括，但不包括语言使用的统计信息，也没有捕捉到语言项目在时间和年龄、数据类型和类型方面的巨大变化。这些信息将是神经认知实验的宝贵资源，增加语言神经生物学模型的可信性，但它只能通过分析特定人类语言数据（例如患者数据，口语语料库）中的语言模式及其频率来获得。手工分析语言数据的成本高得令人望而却步。使用计算自然语言处理（NLP）的自动语言分析现在是一种可行的替代方案。在过去的几十年里，统计和机器学习方法在自然语言处理中的应用得到了大规模的扩展。这项工作使人类语言数据的大规模处理成为可能，并在语音和语言处理任务中取得了令人印象深刻的成果，包括语音识别、形态分析、解析和语义解释。虽然同样的方法可以用来为涉及语言的神经认知研究提供现实的、数据驱动的语言输入，但还没有系统地尝试这样做。基本的NLP技术是可用的，但对于没有相当计算技能的研究人员来说，它是不可接近的，并且需要进一步发展以优化与神经认知研究的整合。在这个新的跨学科项目中，我们将整合认知神经科学、实验心理语言学和自然语言处理的研究，目的是为更现实的语言结构模型提供基础设施，为思想和大脑中语言的理论驱动的实证研究提供输入。我们将进行一系列神经认知实验，重点从形态学、句法和语义层面对语言的核心成分进行处理，使用从人类相关语言数据中自动提取的语言输入。NLP技术将根据需要得到改进和扩展，以处理更广泛的结构、领域和文本类型。然后将设计一个易于使用的工具，该工具将能够有效地搜索、提取和总结注释数据中的语言信息，并与神经认知实验进行最佳整合。我们期望这个项目(i)通过在更现实的语言分析中扎根来提高神经认知实验的质量，（ii）通过扩展现有技术来推进NLP的研究，从而实现更丰富、更深入的分析，以及（iii）为将NLP整合到认知科学的关键实验研究中提供一个重要的案例研究。这项研究的长期目标是提高对人类语言处理的科学理解，这可以使多个学科受益，并使研究人员能够更好地开发更有用的语言模型，NLP技术，以及未来各种语言障碍的治疗和康复。

Scientific understanding of the human language system is one of the main challenges facing cognitive science, and of interest to fields as diverse as linguistics, psychology, anthropology, philosophy, biology and computer science. An adequate theory of this complex system would need to integrate scientific knowledge from several fields, many of which are still in a state of rapid development. One such field is cognitive neuroscience, which investigates language function in the human brain, with the aim of developing neurobiologically and cognitively plausible accounts of the human capacity for dynamic comprehension and production of language. A key input to this research is linguistic information about the core properties of language. This information is typically obtained from conventional resources (dictionaries and grammars) which provide useful generalisations about language, but which do not include statistical information about language use or capture the considerable variation that linguistic items undergo across time and age, data type, and genre. Such information would be an invaluable resource for neuro-cognitive experiments, increasing the plausibility of neurobiological models of language, but it can only be obtained by analysing linguistic patterns and their frequencies in the specific human language data (e.g. patient data, spoken language corpus) of experimental interest. Manual analysis of linguistic data is prohibitively expensive. Automated language analysis using computational Natural Language Processing (NLP) is now a viable alternative. The last decades have seen a massive expansion in the application of statistical and machine learning methods to NLP. This work has made large-scale processing of human language data possible and yielded impressive results in speech and language processing tasks, including e.g. speech recognition, morphological analysis, parsing, and semantic interpretation. Although the same methods could be used to provide realistic, data-driven linguistic input to neuro-cognitive studies involving language, there have been no systematic attempts to do this. The basic NLP technology is available, but it is inaccessible for researchers without considerable computing skills and requires further development for optimal integration with neuro-cognitive research. In this new interdisciplinary project we will integrate research in cognitive neuroscience, experimental psycholinguistics and NLP with the aim of providing the infrastructure for more realistic models of language structure for input into theoretically-driven empirical studies of language in the mind and brain. We will conduct a series of neuro-cognitive experiments which focus on the processing of the core components of language at the levels of morphology, syntax and semantics, using linguistic input automatically extracted from relevant human language data. NLP techniques will be improved and extended to deal with a wider range of constructions, domains and text types as required. An easy-to-use tool will then be designed which will enable effective search, extraction and summarisation of the linguistic information in the annotated data and optimal integration with neuro-cognitive experiments.We expect this project (i) to improve the quality of neuro-cognitive experiments by rooting them in a much more realistic linguistic analysis, (ii) to advance research in NLP by extending existing techniques enabling richer and deeper analysis, and (iii) to provide an important case study for the integration of NLP into critical experimental research in cognitive sciences. The long term goal of this investigation is improved scientific understanding of human language processing which can benefit several disciplines and place researchers in a better position to develop more useful language models, NLP technology, as well as treatments and rehabilitation of various language disorders in the future.