III: Small: Collaborative Research: Explainable Natural Language Inference

自然语言推理(NLI)可以支持使用自然语言文本中包含的信息进行决策(例如，检测病历中未诊断的疾病，从科学文献中寻找替代治疗方法)。这需要收集从文本中提取的事实并对其进行推理。目前的NLI自动化解决方案在很大程度上无法为他们的推理产生解释，但这种能力对于用户在科学发现和医学等犯错成本较高的领域信任他们的推理是必不可少的。这个项目开发了既准确又可解释的自然语言推理方法。它们之所以准确，是因为它们建立在最先进的深度学习框架上，这些框架使用强大的、自动学习的文本表示法。它们是可解释的，因为它们将信息聚合在单位中，既可以用人类可读的解释表示，也可以用机器可用的矢量表示表示。该项目将推进可解释自然语言推理的方法，使自动推理方法能够应用于关键领域，如医学知识提取。该项目还将与领域专家合作评估推理决策的可解释性。该项目将自然语言推理重新定义为构建和推理胜过解释的任务。特别是，推理将较小的组件事实组装到一个图(解释图)中，它对该图进行推理以做出决策。这一观点认为，作出解释是推理过程的一个组成部分，而不是一个单独的事后机制。该项目有三个主要目标：(A)开发能够有效和高效地探索解释图空间的多智能体强化学习模型；(B)开发基于深度学习的聚集机制，以防止推理结合语义不相容的证据；以及(C)建立基于超图的文本表示的连续体，将结构化知识的离散形式与基于连续嵌入的表示相结合。这些技术将在三个应用领域进行评估：复杂问题解答、医疗关系提取和从病历中检测临床事件。该项目的结果将通过项目网站、学术场所传播，软件和数据集将向公众开放。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。

Natural language inference (NLI) can support decision-making using information contained in natural language texts (e.g, detecting undiagnosed medical conditions in medical records, finding alternate treatments from scientific literature). This requires gathering facts extracted from text and reasoning over them. Current automated solutions for NLI are largely incapable of producing explanations for their inferences, but this capacity is essential for users to trust their reasoning in domains such as scientific discovery and medicine where the cost of making errors is high. This project develops natural language inference methods that are both accurate and explainable. They are accurate because they build on state-of-the-art deep learning frameworks which use powerful, automatically learned, representations of text. They are explainable because they aggregate information in units that can be represented in both a human readable explanation and a machine-usable vector representation. This project will advance methods in explainable natural language inference to enable the application of automated inference methods in critical domains such as medical knowledge extraction. The project will also evaluate the explainability of the inference decisions in collaboration with domain experts.This project reframes natural language inference as the task of constructing and reasoning over explanations. In particular, inference assembles smaller component facts into a graph (explanation graph) that it reasons over to make decisions. In this view, generating explanations is an integral part of the inference process and not a separate post-hoc mechanism. The project has three main goals: (a) Develop multiagent reinforcement learning models that can effectively and efficiently explore the space of explanation graphs, (b) Develop deep learning based aggregation mechanisms that can prevent inference from combining semantically incompatible evidence, and (c) Build a continuum of hypergraph based text representations that combine discrete forms of structured knowledge with their continuous embedding based representations. The techniques will be evaluated on three application domains: complex question answering, medical relation extraction, and clinical event detection from medical records. The results of the project will be disseminated through the project website, scholarly venues, and the software and datasets will be made available to the public.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.