[Paper Review] Rethinking Language Models as Symbolic Knowledge Graphs

Vishwas Mruthyunjaya, Pouya Pezeshkpour+2 more|arXiv (Cornell University)|Aug 25, 2023

Topic ModelingComputer Science3 citations

TL;DR

This paper re-evaluates whether large language models (LMs) can replicate the topological and semantic properties of symbolic knowledge graphs (KGs) by introducing nine novel benchmarks targeting attributes like symmetry, hierarchy, and compositionality. Despite strong factual recall, LMs—including GPT-4—achieve only 23.7% hit@1 on these benchmarks, revealing significant limitations in capturing structured relational knowledge, with smaller models sometimes outperforming larger ones.

ABSTRACT

Symbolic knowledge graphs (KGs) play a pivotal role in knowledge-centric applications such as search, question answering and recommendation. As contemporary language models (LMs) trained on extensive textual data have gained prominence, researchers have extensively explored whether the parametric knowledge within these models can match up to that present in knowledge graphs. Various methodologies have indicated that enhancing the size of the model or the volume of training data enhances its capacity to retrieve symbolic knowledge, often with minimal or no human supervision. Despite these advancements, there is a void in comprehensively evaluating whether LMs can encompass the intricate topological and semantic attributes of KGs, attributes crucial for reasoning processes. In this work, we provide an exhaustive evaluation of language models of varying sizes and capabilities. We construct nine qualitative benchmarks that encompass a spectrum of attributes including symmetry, asymmetry, hierarchy, bidirectionality, compositionality, paths, entity-centricity, bias and ambiguity. Additionally, we propose novel evaluation metrics tailored for each of these attributes. Our extensive evaluation of various LMs shows that while these models exhibit considerable potential in recalling factual information, their ability to capture intricate topological and semantic traits of KGs remains significantly constrained. We note that our proposed evaluation metrics are more reliable in evaluating these abilities than the existing metrics. Lastly, some of our benchmarks challenge the common notion that larger LMs (e.g., GPT-4) universally outshine their smaller counterparts (e.g., BERT).

Motivation & Objective

• To assess whether language models (LMs) can replicate the topological and semantic attributes of symbolic knowledge graphs (KGs), such as symmetry, hierarchy, and bidirectionality.
• To address the gap in existing evaluation frameworks, which focus on isolated triples and overlook structural and relational properties critical for reasoning.
• To develop new, attribute-specific benchmarks and evaluation metrics that better capture the complexity of KG representations.
• To challenge the assumption that larger LMs universally outperform smaller ones in capturing structured knowledge.
• To provide a comprehensive evaluation of LMs across diverse relational patterns to guide future model development and integration with external knowledge.

Proposed method

• Constructed nine new qualitative benchmarks based on the T-REx dataset, sampling triples from Wikidata to represent key KG attributes: symmetry, asymmetry, hierarchy, bidirectionality, compositionality, paths, entity-centricity, bias, and ambiguity.
• Designed novel, attribute-specific evaluation metrics tailored to each benchmark, moving beyond standard precision@1 to assess relational understanding.
• Evaluated a range of LMs—including BERT, LLaMA, and GPT-4—across all benchmarks using paraphrased prompts to ensure robustness in knowledge retrieval.
• Analyzed GPT-4’s performance by entity popularity, comparing distributions of correctly predicted examples to overall example distributions to assess reliance on factual recall vs. structural understanding.
• Used a controlled evaluation setup where each example includes multiple related facts, requiring models to reason over relational patterns rather than isolated triples.
• Applied statistical analysis to compare model performance across benchmarks and identify patterns in failure modes, especially for high-popularity entities.

Experimental results

Research questions

• RQ1 Can language models accurately capture symmetric and asymmetric relations in knowledge graphs, and how do they perform compared to symbolic KGs?
• RQ2 To what extent can LMs represent hierarchical and compositional structures essential for complex reasoning in KGs?
• RQ3 Does model size consistently correlate with better performance on relational and topological attributes of KGs?
• RQ4 How do entity popularity and lexical ambiguity affect the reliability of LM predictions in structured knowledge tasks?
• RQ5 Can novel evaluation metrics better capture the structural understanding of LMs than standard precision-based metrics?

Key findings

• GPT-4 achieves only 23.7% hit@1 on the proposed benchmarks, significantly lower than its 50% precision@1 on standard LAMA benchmarks, indicating poor capture of relational structure.
• Smaller LMs such as BERT and LLaMA outperform GPT-4 on certain benchmarks like symmetry and compositionality, challenging the assumption that larger models universally excel.
• The distribution of entity popularity in correctly predicted examples closely mirrors the overall distribution, suggesting GPT-4’s success is not solely due to memorization of popular facts.
• GPT-4 struggles with examples involving very high-popularity entities, indicating a performance ceiling beyond a certain popularity threshold.
• The proposed attribute-specific metrics are more reliable and informative than standard precision@1 for evaluating relational and topological knowledge in LMs.
• Models show inconsistent behavior across different attributes, with strong performance on simple relations but significant failure on complex, path-based or bidirectional reasoning tasks.