From Models to Microtheories: Distilling a Model's Topical Knowledge for Grounded Question Answering

TL;DR

This work proposes microtheories, compact, topic-focused NL knowledge representations distilled from language models to support fully grounded question answering. The approach combines prompt-driven raw fact extraction, SBERT-based deduplication, and entailment-based condensation, followed by three optimization strategies to fit a fixed fact budget. Empirical results on ARC and MedQA show that integrating microtheories with a general corpus improves both grounding and QA accuracy, with human and automatic relevance analyses indicating the distilled facts capture core, topically critical knowledge. A new p-relevance metric enables predicting data needs and guiding deployment to new domains, underscoring the practical value of interpretable, verifiable model knowledge.

Abstract

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.

Figures (26)

• Figure 1: Given a set of topical training questions, we construct a microtheory, a set of statements articulating a model's core, reusable knowledge about that topic. These help prove (entail) answers to test questions.
• Figure 2: Comparison of distillation techniques for extracting $n$-Mts from a larger fact pool $\mathcal{C}$ based on their entailments ($L$s of different training hypotheses ($h$s)). The usage approach prioritizes facts based on the number of hypotheses for which they are used. This risks keeping facts that serve the same role in explaining the same $h$s (e.g., $f_5$ and $f_5^*$ are both kept, even though only one is needed to explain $h_2$, $h_3$, and $h_4$). The question coverage (QC) approach maximizes the number of $h$s for which the $n$-Mt contains all of a supporting argument's leaves (here $h_2$, $h_3$, and $h_4$). This risks failing to cover some $h$s at all (e.g. $h_1$). Partial coverage (PC) maximizes the total fraction of the argument covered for all questions (preferring the most covered argument for each $h$, if there is more than one). In this example, coverage = .66+1+1+.5 = 3.16 (of a possible 4).
• Figure 3: (Upper) Dataset details for the two domains. (Lower) Results of microtheory extraction from training data. The last two rows result from the "min-fact" LP.
• Figure 4: Histogram of training question per fact in $\mathcal{C}$ and 650-850 training questions. ARC facts are used more frequently than MedQA.
• Figure 5: Rate at which different microtheory approaches entail training hypotheses. Lighter bars are the sum of fractional coverages for questions not fully covered by the Mt. Total coverage for ARC questions is generally higher than MedQA. The 1000-Mt$_{PC}$s effectively fully cover the training hypotheses for which a basis was found in both domains; since some hypotheses did not have a basis found, 100% coverage is impossible.