NeuroComparatives: Neuro-Symbolic Distillation of Comparative Knowledge

TL;DR

NeuroComparatives introduces a neuro-symbolic distillation pipeline to convert language-model capabilities into a large, high-quality comparative knowledge base. By collecting comparable entities from Wikidata, overgenerating with both open-source and proprietary LMs under NeuroLogic constraints, and applying multi-stage filtering, the authors build NCs containing up to 8.8 million comparisons across 1.74 million entity pairs. The NC-XL corpus is roughly 10x larger than WebChild and, with discriminative filtering, achieves up to 90% human acceptance while preserving diversity, demonstrating strong downstream benefits on tasks like Elephant, VPR, ComparativeQA, and COPEN benchmarks. The results highlight that combining smaller neuro-symbolic models with extreme-scale prompts can outperform single-source approaches, offering a cost-effective, scalable pathway to robust world knowledge on comparative reasoning.

Abstract

Comparative knowledge (e.g., steel is stronger and heavier than styrofoam) is an essential component of our world knowledge, yet understudied in prior literature. In this paper, we harvest the dramatic improvements in knowledge capabilities of language models into a large-scale comparative knowledge base. While the ease of acquisition of such comparative knowledge is much higher from extreme-scale models like GPT-4, compared to their considerably smaller and weaker counterparts such as GPT-2, not even the most powerful models are exempt from making errors. We thus ask: to what extent are models at different scales able to generate valid and diverse comparative knowledge? We introduce NeuroComparatives, a novel framework for comparative knowledge distillation overgenerated from language models such as GPT-variants and LLaMA, followed by stringent filtering of the generated knowledge. Our framework acquires comparative knowledge between everyday objects, producing a corpus of up to 8.8M comparisons over 1.74M entity pairs - 10X larger and 30% more diverse than existing resources. Moreover, human evaluations show that NeuroComparatives outperform existing resources in terms of validity (up to 32% absolute improvement). Our acquired NeuroComparatives leads to performance improvements on five downstream tasks. We find that neuro-symbolic manipulation of smaller models offers complementary benefits to the currently dominant practice of prompting extreme-scale language models for knowledge distillation.

Figures (6)

• Figure 1: Our neuro-symbolic framework to distill NeuroComparatives. (1) We seed entity pairs for comparison from Wikidata, and expand the set with CategoryBuilder to construct templated prompts for a language model. (2) Next, we use these prompts to overgenerate comparatives from different language models to ensure our generations contain valid comparisons between a given pair of entities. (3) Finally, we discard contradictory and otherwise lower quality generations via various clustering and filtering techniques. Our resultant corpus NeuroComparatives contains 8.8 million comparisons over 1.74 million entity pairs.
• Figure 2: Wikidata hierarchical class structure for retrieved entities ' blender' and ' food processor'.
• Figure 3: As our knowledge discriminator gets stricter, human acceptance of NeuroComparatives increases.
• Figure 4: Examples of generated comparatives which satisfy and violate our constraint ordering.
• Figure 5: The distribution of the top-20 relations in WebChild is more skewed than NC-XL.