From Form(s) to Meaning: Probing the Semantic Depths of Language Models Using Multisense Consistency
Proposed a multisense consistency framework based on Fregean sense to evaluate GPT-3.5’s semantic stability across five languages, revealing significant inconsistencies.
Key Findings
Methodology
This study employs a Fregean-inspired multisense consistency approach, generating alternative expressions via translation and paraphrasing across five languages (English, German, Italian, Dutch, Swedish). Responses are collected for factual and benchmark tasks, with consistency measured through exact match and semantic similarity metrics. The framework assesses whether models respond equivalently to different expressions of the same meaning, reflecting their grasp of semantic stability. The process involves: 1) generating multiple senses using model capabilities; 2) defining response similarity metrics; 3) comparing responses across senses; 4) analyzing response variance to evaluate semantic understanding. This approach emphasizes the importance of semantic invariance in assessing true language understanding.
Key Results
- Across simple factual questions, GPT-3.5’s multisense consistency averaged only 65%, significantly below the baseline consistency of 90% for identical expressions, indicating notable instability across different forms.
- On four NLP benchmarks (PAWS, XNLI, COPA, Belebele), the average response consistency across senses was 70%, while accuracy remained at 85%, suggesting that inconsistency could impair genuine understanding.
- Further analysis revealed that responses involving metaphors and polysemous words showed greater divergence, especially in complex semantic tasks, highlighting the model’s limited grasp of deep meaning and contextual nuances.
Significance
This research exposes critical limitations in current LLMs’ ability to maintain semantic stability across different expressions, challenging assumptions about their genuine understanding. The findings underscore the necessity for models to develop deeper semantic representations, beyond surface-level text processing, to improve robustness in multilingual and multi-task settings. Such insights are vital for advancing AI systems that can truly comprehend human language, fostering more reliable applications in translation, dialogue, and knowledge retrieval.
Technical Contribution
The paper introduces a novel framework grounded in Frege’s sense-reference distinction, operationalized through multilingual datasets and response similarity measures. It shifts evaluation focus from accuracy alone to semantic invariance, providing a new lens to assess model understanding. The methodology combines generation of alternative senses with quantitative consistency metrics, enabling systematic analysis of semantic stability. This approach offers a theoretical and practical advancement over existing benchmarks, facilitating deeper insights into the nature of language understanding in neural models.
Novelty
This is the first systematic application of Fregean sense theory to evaluate language models’ semantic stability across multiple languages and expressions. Unlike traditional benchmarks that focus solely on task accuracy, this work emphasizes the importance of semantic invariance, filling a critical gap in understanding whether models truly grasp meaning. Its innovative use of model-generated senses and cross-lingual testing sets a new standard for evaluating deep language understanding.
Limitations
- The framework relies on model-generated senses, which may introduce bias or errors, limiting the generalizability of results. Further validation with human-annotated senses is needed.
- Metrics like exact match and semantic similarity may not fully capture nuanced semantic shifts, especially in metaphorical or context-dependent expressions.
- Resource limitations restrict the inclusion of low-resource languages, potentially biasing results toward high-resource settings.
Future Work
Future research will extend this framework to include multimodal data, such as images and audio, to evaluate cross-modal semantic consistency. Additionally, developing more sophisticated sense-generation techniques and semantic metrics will improve robustness. Long-term, the goal is to create models capable of maintaining deep semantic invariance across diverse languages and contexts, approaching human-like understanding, with applications spanning translation, conversational AI, and knowledge reasoning.
AI Executive Summary
The rapid advancement of large language models (LLMs) like GPT-3.5 has revolutionized natural language understanding (NLU), yet fundamental questions about their genuine comprehension remain. Traditional benchmarks, focusing on accuracy, often overlook whether models truly understand the underlying meanings of language or merely mimic surface patterns. Inspired by Frege’s philosophical distinction between sense and reference, this study introduces a multisense consistency framework to evaluate whether models maintain semantic stability across different expressions of the same meaning.
Using a carefully designed multilingual dataset spanning five languages, the research probes GPT-3.5’s responses to factual and benchmark tasks, generated through translation and paraphrasing. The core idea is to compare responses to different senses—expressions that convey the same meaning but differ in form—and assess their consistency. Results reveal that the model’s multisense consistency averages only 65%, significantly below the baseline of 90% for identical expressions, indicating a notable gap in deep semantic understanding.
Further analysis shows that responses involving metaphors and polysemous words exhibit greater divergence, especially in complex semantic tasks. These findings suggest that despite high task accuracy (up to 85%), the model’s responses are often inconsistent across different expressions, challenging assumptions about its genuine understanding. The implications are profound: models may perform well on benchmarks but lack a robust, form-independent grasp of meaning.
This work highlights the importance of evaluating semantic invariance, not just task accuracy, for advancing AI systems toward human-like language comprehension. It underscores the need for developing models that can maintain deep semantic stability across languages and expressions, essential for reliable translation, dialogue, and reasoning applications. While promising, the framework faces limitations, such as reliance on model-generated senses and metrics that may not fully capture nuanced semantics. Future directions include integrating multimodal data and refining sense-generation techniques, aiming to bridge the gap toward truly understanding machines.
Deep Dive
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.