[Paper Review] The Two Word Test: A Semantic Benchmark for Large Language Models

Nicholas Riccardi, Rutvik H. Desai|arXiv (Cornell University)|Jun 7, 2023

Topic Modeling4 citations

TL;DR

This paper introduces the Two Word Test (TWT), a semantic benchmark evaluating large language models (LLMs) on their ability to judge the meaningfulness of 1,768 noun-noun phrases rated by 150 human raters. Despite strong performance on other tasks, GPT-4, GPT-3.5, and Bard all underperform humans, with GPT-4 showing notable but still insufficient improvement in binary discrimination, challenging claims of human-level understanding or AGI in LLMs.

ABSTRACT

Large Language Models (LLMs) have shown remarkable abilities recently, including passing advanced professional exams and demanding benchmark tests. This performance has led many to suggest that they are close to achieving humanlike or 'true' understanding of language, and even Artificial General Intelligence (AGI). Here, we provide a new open-source benchmark that can assess semantic abilities of LLMs using two-word phrases using a task that can be performed relatively easily by humans without advanced training. Combining multiple words into a single concept is a fundamental aspect of human language and intelligence. The test requires meaningfulness judgments of 1768 noun-noun combinations that have been rated as meaningful (e.g., baby boy) or not meaningful (e.g., goat sky). by 150 human raters. We provide versions of the task that probe meaningfulness ratings on a 0-4 scale as well as binary judgments. We conducted a series of experiments using the TWT on GPT-4, GPT-3.5, and Bard, with both versions. Results demonstrated that, compared to humans, all models perform poorly at rating meaningfulness of these phrases. GPT-3.5 and Bard are also unable to make binary discriminations between sensible and nonsense phrases as making sense. GPT-4 makes a substantial improvement in binary discrimination of combinatorial phrases but is still significantly worse than human performance. The TWT can be used to understand the limitations and weaknesses of current LLMs, and potentially improve them. The test also reminds us that caution is warranted in attributing 'true understanding' or AGI to LLMs. TWT is available at: https://github.com/NickRiccardi/two-word-test

Motivation & Objective

• To develop a simple yet robust benchmark for assessing semantic understanding in large language models using minimal linguistic input.
• To challenge claims of human-level or AGI-level understanding in LLMs by testing fundamental semantic compositionality.
• To provide an open-source, human-annotated dataset of 1,768 noun-noun combinations rated for meaningfulness on a 0–4 scale and as binary sensible/nonsense pairs.
• To evaluate the performance of leading LLMs—GPT-4, GPT-3.5, and Bard—on this semantic task and compare them to human baseline performance.
• To highlight persistent limitations in LLMs' ability to grasp semantic compositionality, even in simple linguistic constructions.

Proposed method

• The TWT constructs 1,768 noun-noun phrases, including both meaningful combinations (e.g., 'baby boy') and nonsensical ones (e.g., 'goat sky').
• 150 human raters independently rated each phrase on a 0–4 meaningfulness scale, establishing a gold-standard human-annotated dataset.
• The benchmark includes two task variants: a 5-point meaningfulness rating (0–4) and a binary classification (sensible vs. nonsensical).
• The authors evaluated GPT-4, GPT-3.5, and Bard on both variants using zero-shot prompting, measuring agreement with human ratings.
• Performance was quantified using correlation metrics (e.g., Spearman’s rho) and accuracy for binary classification, comparing model outputs to human annotations.

Experimental results

Research questions

• RQ1 Can large language models accurately judge the semantic meaningfulness of simple two-word noun-noun phrases as humans do?
• RQ2 How do GPT-4, GPT-3.5, and Bard compare to human performance on binary discrimination between sensible and nonsensical noun-noun combinations?
• RQ3 To what extent do LLMs demonstrate true semantic compositionality in handling common linguistic constructions?
• RQ4 Does improved performance on complex benchmarks imply deeper semantic understanding, or are models still relying on statistical patterns?
• RQ5 What insights does the Two Word Test provide into the limitations of current LLMs regarding semantic generalization and common sense?

Key findings

• All tested LLMs—GPT-4, GPT-3.5, and Bard—performed significantly worse than humans on the 0–4 meaningfulness rating task, with GPT-4 showing the highest correlation but still below human levels.
• GPT-3.5 and Bard failed to reliably distinguish between sensible and nonsensical phrases in the binary classification task, indicating poor semantic discrimination.
• GPT-4 achieved a notable improvement in binary discrimination, but its accuracy remained substantially below human performance, suggesting persistent semantic limitations.
• The benchmark revealed that even state-of-the-art models struggle with fundamental semantic compositionality, challenging claims of human-level understanding or AGI.
• The results underscore the need for caution in attributing 'true understanding' or AGI capabilities to current LLMs, despite their strong performance on other benchmarks.