Finding Fuzziness in Neural Network Models of Language Processing

Kanishka Misra, Julia Taylor Rayz

Introduction

Human language is inherently vague. We rarely communicate in crisp, binary truths; instead, we use terms like "sort of tall," "quite cold," or "warmish." While humans navigate these imprecise boundaries effortlessly, most Natural Language Processing (NLP) systems are evaluated on "bivalent" settings—where a statement is strictly true or false. This paper investigates whether modern neural network models, such as RoBERTa, implicitly learn the "fuzziness" of human language—the gray areas where a temperature isn't just "hot" or "cold" but somewhere in between.

Probing for Fuzzy Logic

The researchers addressed a fundamental gap: Do Pre-trained Language Models (PLMs), which learn purely from the statistical distribution of words in massive text corpora, actually capture the underlying "fuzzy" membership of concepts? For example, if a model knows that 90°F is "hot," does it also understand that 90°F is "warmer" than 70°F in a way that aligns with human psychological perception and fuzzy logic?

To test this, the authors used a "probing" approach via the Natural Language Inference (NLI) task. In NLI, a model is given a premise and a hypothesis and must decide if the premise entails, contradicts, or is neutral toward the hypothesis.

Examples of NLI premises and hypotheses used to probe fuzzy membership scores.

The researchers used RoBERTa-large, fine-tuned on the Multi-Genre Natural Language Inference (MNLI) dataset. They constructed synthetic sentence pairs where the premise stated a specific temperature (e.g., "It is 70 degrees Fahrenheit in the bedroom") and the hypothesis used a fuzzy category (e.g., "It is warm in the bedroom"). Instead of a binary classification, they analyzed the entailment score—the raw probability assigned to the "entailment" label—treating it as a proxy for a "fuzzy membership function."

Key Insights: The "Hedge" Discovery

The most striking finding is that the model's internal logic surprisingly mirrors classical fuzzy-set theory, specifically mimicking mathematical "hedges." In fuzzy theory, a hedge like "very X" is often modeled by squaring the membership function of "X" (a process called concentration).

Conceptual illustration of fuzzy membership functions and linguistic hedges.

When the authors searched for a power ($\lambda$) to map the "warm" curve onto the "hot" curve, they found the best fit was $\lambda \approx 2$. This aligns almost perfectly with Lotfi Zadeh’s 1972 proposal that "very" should be modeled as $\mu(x)^2$.

Furthermore, the models showed a "noisy-but-sound" grasp of temperature scales. They correctly identified that as temperature increases, the probability of "freezing" drops while "hot" rises. However, an American-Centric Bias was evident: the models performed significantly better and more intuitively with Fahrenheit than with Celsius. When no unit was specified, the model's behavior matched Fahrenheit patterns, suggesting a strong bias toward US-based training data.

Results and Consistency

The entailment curves for freezing, cold, cool, warm, and hot showed overlapping, graded transitions rather than sharp cliffs, mimicking human perception. For the "hot $\approx$ very warm" hypothesis, the Root Mean Square Error (RMSE) was minimized when the exponent $\lambda$ was near 2 across all unit conditions.

However, consistency was not universal. While the "warm/hot" relationship was consistent, the "cold/freezing" relationship was noisier, suggesting that the model doesn't apply fuzzy logic uniformly across all concepts.

Limitations and Future Work

The study's scope was limited to temperature. Whether this "fuzzy" understanding extends to other vague concepts like "tallness" or "beauty" remains to be explored. Additionally, the model's outputs are still noisy; small changes in temperature sometimes led to jittery jumps in entailment probability. The authors suggest that explicitly incorporating multi-valued logic into neural architectures could help models handle imprecise human language more reliably in the future.

Summary updated

BibTeX
@inproceedings{misra2021finding,
  author = {Kanishka Misra and Julia Taylor Rayz},
  title = {{Finding Fuzziness in Neural Network Models of Language Processing}},
  booktitle = {NAFIPS},
  year = {2021},
  doi = {10.1007/978-3-030-82099-2\_25},
}
Figure