Recent reasoning methods (e.g., chain-of-thought, entailment reasoning) help users understand how language models (LMs) answer a single question, but they do little to reveal the LM's overall understanding, or "theory," about the question's topic, making it still hard to trust the model. Our goal is to materialize such theories - here called microtheories (a linguistic analog of logical microtheories) - as a set of sentences encapsulating an LM's core knowledge about a topic. These statements systematically work together to entail answers to a set of questions to both engender trust and improve performance. Our approach is to first populate a knowledge store with (model-generated) sentences that entail answers to training questions and then distill those down to a core microtheory that is concise, general, and non-redundant. We show that, when added to a general corpus (e.g., Wikipedia), microtheories can supply critical, topical information not necessarily present in the corpus, improving both a model's ability to ground its answers to verifiable knowledge (i.e., show how answers are systematically entailed by documents in the corpus, fully grounding up to +8% more answers), and the accuracy of those grounded answers (up to +8% absolute). We also show that, in a human evaluation in the medical domain, our distilled microtheories contain a significantly higher concentration of topically critical facts than the non-distilled knowledge store. Finally, we show we can quantify the coverage of a microtheory for a topic (characterized by a dataset) using a notion of $p$-relevance. Together, these suggest that microtheories are an efficient distillation of an LM's topic-relevant knowledge, that they can usefully augment existing corpora, and can provide both performance gains and an interpretable, verifiable window into the model's knowledge of a topic.

One of the fundamental aspects of cognitive architectures is their ability to encode and manipulate knowledge. Without a consistent, well-designed, and scalable knowledge management scheme, an architecture will be unable to move past toy problems and tackle the broader problems of cognition. In this paper, we document some of the challenges we have faced in developing the knowledge stack for the Companion cognitive architecture and discuss the tools, representations, and practices we have developed to overcome them. We also lay out a series of potential next steps that will allow Companion agents to play a greater role in managing their own knowledge. It is our hope that these observations will prove useful to other cognitive architecture developers facing similar challenges.

To interact with humans, artificial intelligence (AI) systems must understand our social world. Within this world norms play an important role in motivating and guiding agents. However, very few computational theories for learning social norms have been proposed. There also exists a long history of debate on the distinction between what is normal (is) and what is normative (ought). Many have argued that being capable of learning both concepts and recognizing the difference is necessary for all social agents. This paper introduces and demonstrates a computational approach to learning norms from natural language text that accounts for both what is normal and what is normative. It provides a foundation for everyday people to train AI systems about social norms.

Creating systems that can learn to answer natural language questions has been a longstanding challenge for artificial intelligence. Most prior approaches focused on producing a specialized language system for a particular domain and dataset, and they required training on a large corpus manually annotated with logical forms. This paper introduces an analogy-based approach that instead adapts an existing general purpose semantic parser to answer questions in a novel domain by jointly learning disambiguation heuristics and query construction templates from purely textual question-answer pairs. Our technique uses possible semantic interpretations of the natural language questions and answers to constrain a query-generation procedure, producing cases during training that are subsequently reused via analogical retrieval and composed to answer test questions. Bootstrapping an existing semantic parser in this way significantly reduces the number of training examples needed to accurately answer questions. We demonstrate the efficacy of our technique using the Geoquery corpus, on which it approaches state of the art performance using 10-fold cross validation, shows little decrease in performance with 2-folds, and achieves above 50% accuracy with as few as 10 examples.

The majority of work in knowledge representation has dealt with the technicalities of relating predicate calculus to other formalisms and with the details of various schemes for default reasoning. There has almost been an aversion to addressing the problems that arise in actually representing large bodies of knowledge with content. However, deep, important issues must be addressed if we are to ever have a large intelligent knowledge-based program: What ontological categories would make up an adequate set for carving up the universe? What are the important facts and heuristics most humans today know about solid objects? In short, we must bite the bullet.

Authoring interactive stories where the player is afforded a wide range of social interactions results in a very large space of possible social and story situations. The amount of effort required to individually author for each of these circumstances can quickly become intractable. The social AI system Comme il Faut (CiF) aims to reduce the burden on the author by providing a playable model of social interaction where the author provides reusable and recombinable representations of social norms and social interactions. Motivated through examples from an in-development video game, Prom Week, this paper provides a detailed description of the structures with which CiF represents social knowledge and how this knowledge is employed to simulate social interactions between characters.

After explicating the need for a large commonsense knowledge base spanning human consensus knowledge, we report on many of the lessons learned over the first five years of attempting its construction. We have come a long way in terms of methodology, representation language, techniques for efficient inferencing, the ontology of the knowledge base, and the environment and infrastructure in which the knowledge base is being built. We describe the evolution of Cyc and its current state and close with a look at our plans and expectations for the coming five years, including an argument for how and why the project might conclude at the end of this time.