When human experts evaluate complex, high-stakes scenarios, they rarely speak in precise numbers. A landing signal officer assessing an aircraft carrier landing might describe the risk as "high" or "unacceptable" rather than assigning it a probability of 0.83 (Jiang et al.). A procurement team evaluating a new supplier might rate their technical capability as "good" but express "hesitation" about their financial stability (Chong et al.).
Traditional mathematical frameworks struggle with this input. To process qualitative language, standard engineering and decision-making models typically force humans to map their subjective feelings to rigid, discrete numerical scales (such as a 1-to-5 Likert scale). However, this compression introduces artificial precision, discards the inherent hesitation of human judgment, and creates significant noise when different experts interpret the same qualitative terms differently (Luo et al.; Lu et al.).
The field of Linguistic Uncertainty Quantification addresses this mismatch. Instead of forcing natural language into rigid scalar values, it builds mathematical structures that natively preserve the fuzziness, hesitation, and probabilistic distribution of human words.
Mathematical Frameworks for Words
To understand why this works, we must look at how a qualitative word can be transformed into a rigorous mathematical object without losing its nuance.
From Fuzzy Sets to Intuitionistic Logic
The foundation of this approach relies on fuzzy set theory, which replaces binary true/false logic with degrees of membership (Lu et al.). If an expert describes a risk as "medium," a fuzzy set defines a membership function across a spectrum, meaning the situation can be 70% "medium" and 20% "high" simultaneously.
To capture human hesitation more accurately, researchers utilize Intuitionistic Fuzzy Numbers (IFNs). Traditional fuzzy sets only define a membership degree ($\mu$). Intuitionistic fuzzy sets add a non-membership degree ($\nu$) and an explicit hesitation margin ($\pi$), where:
$$\mu + \nu + \pi = 1$$
This allows an expert to state that they are confident a supplier is "good" to a degree of 0.6, confident they are not good to a degree of 0.2, and entirely hesitant or uncertain to a degree of 0.2 (Li).
Distributed and Probabilistic Representations
When dealing with complex preferences, a single fuzzy term is often still too restrictive. If an expert is torn between two categories, forcing them to select one causes information loss.
To solve this, Distributed Linguistic Representations (DLRs) allow experts to distribute their symbolic assessments across an entire vocabulary index (Wu et al.). For instance, an evaluation can be mathematically represented as:
$$\{\text{0.2 low}, \text{0.5 medium}, \text{0.3 high}\}$$
Similarly, Probabilistic Linguistic Term Sets (PLTSs) formalize this by assigning explicit probability distributions to linguistic terms chosen by a group of experts (Li et al.). This preserves the exact spread of opinions across a panel, ensuring that minority dissent or internal hesitation is carried directly into the final algorithmic computation.
Cloud Models: Separating Fuzziness from Randomness
A core challenge in linguistic uncertainty is that words suffer from two distinct types of ambiguity: fuzziness (the boundary of what "high risk" means is vague) and randomness (the specific value changes stochastically depending on the individual and the context).
The Linguistic Multi-Criteria Group Decision-Making Method Combining Cloud Model and Evidence Theory handles this via Cloud Models. A cloud model represents a linguistic term using three numerical characteristics:
- Expectation ($Ex$): The central value that best represents the concept.
- Entropy ($En$): The uncertainty width, measuring how blurry the concept's boundaries are (fuzziness).
- Hyper-entropy ($He$): The dispersion of the entropy itself, measuring how much different experts disagree on the blurriness (randomness).
By generating a cloud of mathematical points based on these three parameters, the system can perform algebraic operations on words while respecting both their conceptual vagueness and interpersonal variance.
Aggregating Group Consensus
Once individual linguistic statements are mapped to these advanced mathematical structures, the next challenge is aggregation: how do we combine multiple vague, conflicting expert opinions into a single, actionable decision? This is the domain of Multi-Attribute Group Decision Making (MAGDM).
[Expert 1: "High Risk"] --------> [Cloud Model / IFN] ---\
[Expert 2: "Medium-High"] ------> [Cloud Model / IFN] ----+---> [Similarity Aggregation / Evidence Theory] ---> Consensus Ranking
[Expert 3: "Low-Medium"] -------> [Cloud Model / IFN] ---/
When experts provide highly divergent qualitative feedback, simple averaging fails because it flattens critical variance. Instead, researchers deploy specialized aggregation mechanisms:
- Similarity Aggregation Methods (SAM): As demonstrated by Liu et al., systems can calculate a "bidirectional similarity degree" between the trapezoidal intuitionistic fuzzy numbers provided by different experts. Opinions that share high mutual agreement are assigned higher weights, preventing extreme, uncalibrated outliers from distorting the consensus.
- Evidence Theory and Game Theory: To model strategic interactions where manufacturers or stakeholders have conflicting interests, Jiang et al. combine Dempster-Shafer evidence theory—which handles incomplete and uncertain data structures—with game theory to mathematically identify Nash equilibrium points across qualitative evaluations.
- Three-Way Decisions (3WD): Instead of forcing a premature consensus when information is highly uncertain, Krejcar et al. integrate a three-way decision framework into group dynamics. This allows the system to output three distinct states: acceptance, rejection, or deferral (non-commitment), mirroring the natural human tendency to gather more data when linguistic feedback is too polarized to reconcile safely.
Linguistic Uncertainty in Modern AI and LLMs
The problem of linguistic uncertainty has taken on a new dimension with the rise of Large Language Models (LLMs). While traditional fuzzy logic focuses on structured decision matrices, modern NLP must deal with the unconstrained linguistic uncertainty present in raw text.
Bayesian Reasoning over Vague Text
When an LLM encounters text containing words of estimative probability (such as "it is highly likely the valve will fail if pressure rises"), performing formal risk calculations is notoriously difficult. To evaluate this capability, Lukas Lange introduced the QUITE benchmark, which tests how models perform Bayesian inference on natural language text.
The research shows that standard LLMs struggle with backward probabilistic reasoning when forced to process these terms directly. Instead, the most robust solution is a neuro-symbolic approach: using the LLM as a translation layer that maps qualitative words of estimative probability into precise numerical values or formal logic code, which is then executed by a rigorous, symbolic Bayesian backend.
Learning from Verbal Feedback vs. Scalar Rewards
Linguistic uncertainty quantification also offers an alternative to traditional reinforcement learning alignment methods like RLHF (Reinforcement Learning from Human Feedback). Standard alignment pipelines squash rich, nuanced human critiques down into a single scalar reward number. This compression introduces significant noise, as a single scalar cannot capture why a human disliked an output or which parts they were hesitant about.
To bypass this limitation, Liu et al. developed the Feedback-Conditional Policy (FCP) framework.
Standard RLHF: [Rich Human Critique] ------------> [Scalar Reward: 0.4] ------> [Loss Function] (Nuance Lost)
FCP Framework: [Rich Human Critique] ------------> [Conditioning Signal] ------> [Language Model] (Nuance Preserved)
Instead of converting text into a number, FCP treats verbal critiques directly as conditioning signals for the language model. By training the model to predict text conditioned on the qualitative feedback itself, the system learns directly from the rich, uncertain, and emotional nuances of human language without the lossy intermediate step of scalar quantization.
Key Foundational Work
Syllogistic Reasoning as a Basis for Combination of Evidence in Expert Systems (1985) — contributed 1 concept elements to this direction.
PRUF - A Language for the Representation of Meaning in Natural Languages (1977) — contributed 1 concept elements to this direction.