RI: Small:Comp Cog: Broad-coverage semantic models of human sentence processing

人类之所以是一个成功的物种，在很大程度上是因为他们可以通过语言解释相互传递有关世界的知识。这些解释可能非常复杂，涉及关于多个对象和事件类的嵌套泛化。这些关系是如何从自然语言中解码出来的准确模型可以加深我们对大脑如何工作的理解，并可能允许非程序员用户解释他们想要的产品、目标和对机器的限制。句子加工实验可能为理解语言理解中的概念形成机制提供了一个重要的窗口，但人类的大脑对实验控制的研究设计中使用的构造刺激的陌生性非常敏感，从而产生由意想不到的单词或句子结构引起的潜在混淆效果。一种常见的替代方法是使用带有统计控制的自然发生刺激的设计，通常在句子处理过程中使用一种或多种概率的惊喜措施。遗憾的是，现有的突击概率测量是基于过于简单的句子处理模型，这些模型没有与语言解释可以描述的泛化的嵌套结构相联系，因此在预测这类频率效应方面存在严重限制。因此，该项目将开发一个句子处理模型，使用类似人类的递增概率过程将句子解码为含义。然后，这个模型将被用来控制神经激活、血液充氧和阅读时间数据中的频率效应，以便分离出可归因于语言理解过程中构建和存储复杂概念的机械过程的影响。该项目构建了一个句子加工模型，其加工决策基于对意义的心理表征，而不仅仅是单词。这意味着，当重复的名词或代词指的是一个在语篇中突出的共同实体时，模型将不会那么惊讶。该项目最初侧重于开发一个统计句子处理模型，该模型在每个词被处理后对一个句子进行若干可能的分析，每个分析都包含句子意义中涉及的每个话语所指的明确表示，作为一组逻辑谓词，在该意义的图形表示中与该所指相邻。该模型的后续版本将这些上下文集合压缩成向量，这些向量通过递归神经网络传递。将这些模型的预测与自然语言处理应用中使用的现有神经网络语言模型进行比较，以确保它们的语言预测是准确的。然后，由这些模型生成的增量概率被用来估计概率惊讶，作为预测功能磁共振(FMRI)、脑电(EEG)、眼睛跟踪和现有数据集中的阅读时间观察的频率控制，以隔离由于内存使用和其他机械因素造成的影响。这一奖项反映了NSF的法定使命，并通过使用基金会的智力优势和更广泛的影响审查标准进行评估，被认为值得支持。

Humans are a successful species in large part because they can pass knowledge about the world to one another using linguistic explanations. These explanations can be quite complex, involving nested generalizations about multiple classes of objects and events. Accurate models of how these relationships are decoded from natural language could further our understanding of how the brain works, and may allow non-programmer users to explain their desired products, goals and constraints to machines. Sentence processing experiments may provide an important window into the mechanisms of idea formation in language comprehension, but the human mind is extraordinarily sensitive to the strangeness of constructed stimuli used in experimentally controlled research designs, yielding potentially confounding effects arising from unexpected words or sentence structures. A common alternative is to use designs employing naturally-occurring stimuli with statistical controls, usually using one or more probabilistic measures of surprise during sentence processing. Unfortunately, existing probabilistic measures of surprise are based on overly simple models of sentence processing that are not connected to the nested structure of generalizations that a linguistic explanation may describe, and thus have severe limits as predictors of these kinds of frequency effects. This project will therefore develop a sentence processing model that decodes sentences into meanings using a human-like incremental probabilistic process. This model will then be used to control for frequency effects in neural activation, blood oxygenation and reading time data in order to isolate effects that can be attributed to the mechanical process of constructing and storing complex ideas during language comprehension.This project constructs a model of sentence processing that bases its processing decisions on mental representations of meanings rather than on words only. This means that the model will be less surprised by repeated nouns or pronouns when these words refer to a common entity which is prominent in a discourse. The project initially focuses on the development of a statistical sentence processing model which maintains several possible analyses of a sentence after each word is processed, each of which contains explicit representations of each discourse referent involved in a sentence meaning as a set of logical predicates adjacent to that referent in a graphical representation of the meaning. A subsequent version of the model compresses these context sets into vectors, which are passed through a recurrent neural network. The predictions of these models are compared against existing neural network language models used in natural language processing applications to ensure that their linguistic predictions are accurate. Incremental probabilities generated by these models are then used to estimate probabilistic surprise as a frequency control in predicting functional magnetic resonance (fMRI), electroencephalographic (EEG), eye-tracking, and reading-time observations in existing datasets, in order to isolate effects due to memory usage and other mechanistic factors.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.