Learning and Inference with Schemas and Analogies

这个项目将扩展和测试人类推理的理论和计算机模型。人们经常使用模式、描述一般情况、事件、规则或关系的知识结构进行推理。例如，餐馆模式可以指定在餐馆中发生的事件的类型，以及这些事件中的参与者(例如，厨师、服务员、顾客)之间的关系；家庭关系模式可以指定定义各种家庭成员的关系的类型，例如姐妹、兄弟、父母和叔叔。模式既是通用的，因为它指的是许多特定的情况；也是高度结构化的，因为它指定了对象、事件之间的关系，甚至是其他关系之间的关系。心理学家和计算机科学家早就认识到图式作为推理基础的效用：如果一个人可以将一个对象或情景识别为一般图式的一个实例，那么他就可以使用该图式来推理该对象或情景。尽管模式的通用性和结构化特性使它们非常有用，但这些功能也使模式很难建模。计算机建模的传统符号方法擅长表示结构化信息，但它们难以灵活地将特定实例与一般事实匹配；相比之下，连接主义模型擅长将实例与一般类别灵活匹配，但它们很难表示结构化信息(如规则和关系)。我们最近开发了一种基于人类图式的学习和推理理论，并将该理论作为一个工作的计算机模型来实现。该模型称为LISA(利用模式和类比进行学习和推理)，将连接主义体系结构的灵活性与表示和学习关系结构的能力结合在一起。所得到的系统可以(A)从特定示例中学习一般模式，(B)将新实例与所学习的模式相匹配，以及(C)使用模式来对新示例进行归纳推理。丽莎对人类的学习和推理做出了几个新奇的预测。NSF资助的研究的一部分将是一系列实验来验证这些预测。这项研究的另一部分将扩展Lisa模型，以解释人类故事和事件理解的各个方面(例如，如果有人告诉我们‘每个盘子有四个盘子’，我们如何知道这意味着每个盘子将容纳或容纳四个盘子？)，以及空间推理的各个方面(例如，如果我们被告知Bill比Charles高，Abe比Bill高，我们如何计算出Abe比Charles高？)。重要的是，扩展的Lisa模型将解释这些看似不同的能力(以及其他能力，如类比推理)，其基本机制与它用来解释我们使用图式推理的能力相同。从该项目中获得的知识将有助于开发增强人类学习和推理能力的方法，也将有助于开发人工智能系统。

This project will extend and test a theory and computer model of human reasoning. People routinely reason using schemas, knowledge structures that describe general situations, events, rules, or relationships. For example, a Restaurant schema might specify the kinds of events that take place in restaurants, and the relationships between the actors in those events (e.g., chefs, waiters, customers); a Family Relation schema might specify the kinds of relationships that define various family members, such as sisters, brothers, parents, and uncles. A schema is both general, in that it refers to many specific situations, and highly structured, in that it specifies the relationships between objects, events, or even between other relationships. Psychologists and computer scientists have long recognized the utility of schemas as a basis for reasoning: If one can recognize an object or situation as an instance of a general schema, then one can use the schema to reason about that object or situation. Although the generality and structured nature of schemas make them extremely useful, these features also make schemas very difficult to model. Traditional symbolic approaches to computer modeling are good at representing structured information, but they have difficulty in flexibly matching specific instances to general facts; by contrast, connectionist models are good at flexibly matching instances to general categories, but they have great difficulty representing structured information (such as rules and relationships). We have recently developed a theory of human schema-based learning and reasoning, and implemented this theory as a working computer model. The model, called LISA (Learning and Inference with Schemas and Analogies), combines the flexibility of a connectionist architecture with the capacity to represent and learn relational structures. The resulting system can (a) learn general schemas from specific examples, (b) match new instances to learned schemas, and (c) use the schemas to make inductive inferences about the new examples. LISA makes several novel predictions about human learning and reasoning. Part of the NSF-supported research will be a series of experiments to test these predictions. Another part of the research will extend the LISA model to account for aspects of human story and event comprehension (e.g., If someone tells us `There are four plates per tray,` how do we know that this means each tray will hold, or contain, four plates?), and aspects of spatial reasoning (e.g., If we are told that Bill is taller than Charles and Abe is taller than Bill, how do we figure out that Abe is taller than Charles?). Importantly, the extended LISA model will account for these seemingly different capacities (as well as other capacities, such as reasoning by analogy) in terms of the same basic mechanisms as it uses to account for our ability to reason using schemas. The knowledge gained from the project will contribute to the development of methods for enhancing human learning and reasoning abilities, and also to the development of artificially intelligent systems.