KnowledgePrompts: Exploring the Abilities of Large Language Models to Solve Proportional Analogies via Knowledge-Enhanced Prompting

TL;DR

This work addresses the challenge of solving proportional analogies with large language models by introducing a large 15K MCQA dataset spanning 236 semantic relations. It systematically compares zero-shot, exemplar-based, and knowledge-enhanced prompting, introducing Structured Knowledge Prompting (SKP) and Targeted Knowledge Prompting (TKP) to augment reasoning. The key finding is that TKP provides the largest gains, with GPT-3.5-Turbo achieving up to 55% EMA, while structured knowledge alone can underperform, underscoring the nuanced role of knowledge in complex reasoning. The results highlight the potential and limits of knowledge-guided prompting for cognitive tasks, and point to future work in automating targeted knowledge acquisition and prompting strategies for robust generalization.

Abstract

Making analogies is fundamental to cognition. Proportional analogies, which consist of four terms, are often used to assess linguistic and cognitive abilities. For instance, completing analogies like "Oxygen is to Gas as <blank> is to <blank>" requires identifying the semantic relationship (e.g., "type of") between the first pair of terms ("Oxygen" and "Gas") and finding a second pair that shares the same relationship (e.g., "Aluminum" and "Metal"). In this work, we introduce a 15K Multiple-Choice Question Answering (MCQA) dataset for proportional analogy completion and evaluate the performance of contemporary Large Language Models (LLMs) in various knowledge-enhanced prompt settings. Specifically, we augment prompts with three types of knowledge: exemplar, structured, and targeted. Our results show that despite extensive training data, solving proportional analogies remains challenging for current LLMs, with the best model achieving an accuracy of 55%. Notably, we find that providing targeted knowledge can better assist models in completing proportional analogies compared to providing exemplars or collections of structured knowledge. Our code and data are available at: https://github.com/Thiliniiw/KnowledgePrompts/

Figures (11)

• Figure 1: Knowledge-enhanced Prompting. An illustration of our knowledge-enhanced prompting approach with types of knowledge and prompting techniques. The question consists of two terms ("Oxygen" and "Gas"), and answer choices consist of term pairs that are analogous to the question term pair. Each model is queried using the prompting techniques illustrated.
• Figure 2: Distribution of Semantic relations. The distribution of the top 59 semantic relations (these are the frequencies of semantic relations between the question word pair )
• Figure 3: An illustration of the knowledge filtering approach. "Random" indicates Random Filtering and "Semantic" indicates Semantic Filtering.
• Figure 4: Perfromance with structured knowledge. Performance of each model when Structured Knowledge Prompting with semantic filtering (SKP[semantic]) is used. All indicates the prompt is enhanced with all three types of knowledge (Wikidata, ConceptNet and WordNet). EMA values are reported on 20% of the 15K dataset where all three knowledge types available.
• Figure 5: Best and least performing models for each prompting technique.