From Form(s) to Meaning: Probing the Semantic Depths of Language Models Using Multisense Consistency

TL;DR

This work probes whether large language models truly understand meaning or merely manipulate surface forms by testing multisense consistency across paraphrase and translation. Grounded in Fregean sense and reference, the authors generate meaning-preserving senses for each input and evaluate whether the model’s responses remain stable across senses. Experiments with GPT-3.5 across five languages and four NLU benchmarks reveal substantial form-dependency: consistency across senses is far from perfect, even when accuracy is high, suggesting current LLM semantics are not fully human-like. The findings highlight the value of multisense-consistency as a diagnostic benchmark for semantic understanding and call for caution when interpreting benchmark performance as evidence of true meaning.

Abstract

The staggering pace with which the capabilities of large language models (LLMs) are increasing, as measured by a range of commonly used natural language understanding (NLU) benchmarks, raises many questions regarding what "understanding" means for a language model and how it compares to human understanding. This is especially true since many LLMs are exclusively trained on text, casting doubt on whether their stellar benchmark performances are reflective of a true understanding of the problems represented by these benchmarks, or whether LLMs simply excel at uttering textual forms that correlate with what someone who understands the problem would say. In this philosophically inspired work, we aim to create some separation between form and meaning, with a series of tests that leverage the idea that world understanding should be consistent across presentational modes - inspired by Fregean senses - of the same meaning. Specifically, we focus on consistency across languages as well as paraphrases. Taking GPT-3.5 as our object of study, we evaluate multisense consistency across five different languages and various tasks. We start the evaluation in a controlled setting, asking the model for simple facts, and then proceed with an evaluation on four popular NLU benchmarks. We find that the model's multisense consistency is lacking and run several follow-up analyses to verify that this lack of consistency is due to a sense-dependent task understanding. We conclude that, in this aspect, the understanding of LLMs is still quite far from being consistent and human-like, and deliberate on how this impacts their utility in the context of learning about human language and understanding.

Figures (10)

• Figure 1: Illustration of the relationship between sense and meaning for the classical Fregean example of "morning star" and "evening star" (left) and for the addition task in our experiments (right).
• Figure 2: Illustration of the multisense consistency paradigm. We use a model to generate alternative meaning-preserving senses of the original input, and then evaluate whether the same model gives consistent responses to the original input and alternative sense. In this example, the task is to answer a simple factual question, and the model is asked to generate an alternative sense through translation (from English to German). The example illustrates that accuracy and consistency are distinct. Even though the model's responses are incorrect (Marrakesh/Marrakesch instead of Rabat), they are consistent because they refer to the same city.
• Figure 3: Accuracy (%) for the Simple facts datasets, with 95% confidence intervals. Apart from the arithmetics task, the accuracy scores are generally similar across different senses. Numerical scores can be found in \ref{['tab:simple-facts-accuracy']}.
• Figure 4: Consistency (%) for the Simple facts datasets. None of the senses have a consistency close to the maximum possible given the difference in accuracy between the two senses (indicated by the horizontal blue lines), indicating that the models are inconsistent even beyond those differences. Numerical scores can be found in \ref{['tab:simple-facts-consistency']}.
• Figure 5: Accuracy (%) for the benchmark datasets, with 95% confidence intervals. For Belebele, we have no en$^P$ score, because the model did not provide useable paraphrases. Horizontal lines indicate chance accuracy. Numerical scores can be found in \ref{['tab:benchmarks-accuracy']}.