EAGER: Grounding Natural Language Inference in Cognitive Processes

能够检测文本相似度对许多应用都很重要，包括机器翻译、剽窃检测、文本生成、事实检查等。在单词层面上，当一个单词与另一个单词交换几乎没有后果时，两个单词的意思往往是相同的。但是这种方法是如何扩展到句子层面的呢？根据以推理为中心的观点，一个句子的意义的重要部分可以通过每个句子的“推理晕”来理解。“推论光环”是指一个句子的所有推论，即隐含意义。比较两个句子或文本的语义相似性可以通过比较每个句子或文本所暗示的所有推论来完成。然而，目前用于检测句子和文本相似度的自然语言处理方法仅限于基于单词或子字符串相似度的度量，这不能充分捕捉文本的含义。该项目解决了以前方法的局限性：a)丰富推理的概念和表示；b)学习人们如何自然地对文本的语义关系进行推理。我们方法的新颖之处在于认知心理学的既定工作。为了丰富推理的概念，我们引入了a)快速和自动推理（称为类型1）和b)较慢和更深思熟虑的推理（称为类型2）之间的区别。例如，第一种推理适用于我们识别一张脸，第二种推理适用于我们计算账单的小费。为了了解人们如何自然推理，我们使用认知心理学中建立的数据收集协议来收集数据，用于快速和慢速推理。对于数据收集，我们尝试了一种新的粒度水平，以更好地捕获人类所做的推理范围，并训练新的检测语义推理的计算模型。所提出的研究将产生结果、指导方针和新的计算模型，这将导致a)研究语言处理中的非正式推理的新方法；b)改进文本相似度的度量标准。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

Being able to detect textual similarity is important for many applications including machine translation, detection of plagiarism, text generation, fact checking etc. At the word level, two words tend to mean the same thing when one can be swapped with the other with little or no consequence. But how does this approach extend to the sentence level and beyond? According to an inference-centered view, a significant part of a sentence’s meaning can be understood in terms of the “inferential halo” of each sentence. The “inferential halo” is all the inferences, i.e., implied meanings, that a sentence has. Comparing the semantic similarity of two sentences or text would then be accomplished by comparing all the inferences that each sentence or text implies. However, current Natural Language Processing approaches to detecting sentence and text similarities are limited to measures based on word or substring similarities which do not capture adequately the meaning of a text. This project addresses the limitations of previous approaches a) enriching the notion and representation of inference and b) learning how people naturally reason about semantic relations of texts. The novelty of our approach draws on established work in Cognitive Psychology. For the enrichment of the notion of inference, we introduce the distinction between a) quick and automatic reasoning, known as Type 1, and b) slower and more deliberate reasoning, known as Type 2. Type 1 reasoning applies when for example we recognize a face and Type 2 when we calculate the tip for a bill. To learn how people naturally reason, we collect data using data collection protocols established in Cognitive Psychology for quick and slow reasoning. For data collection, we experiment with a novel level of granularity to better capture the range of inferences made by humans and train new computational models of detecting semantic inferences. The proposed research will yield results, guidelines, and new computational models that will lead to a) a novel way of studying informal reasoning in language processing and b) improved metrics of textual similarity.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.