💬 NLP

Finding Meaning in Embeddings: Concept Separation Curves

This paper introduces Concept Separation Curves, a novel, classifier-independent evaluation method that assesses the conceptual stability of sentence embeddings by systematically quantifying their responses to syntactic noise and semantic negations across multiple languages and domains.

Original authors: Paul Keuren, Marc Ponsen, Robert Ayoub Bagheri

Published 2026-04-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Paul Keuren, Marc Ponsen, Robert Ayoub Bagheri

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a magical translator that turns sentences into invisible "concept maps" (vectors). If you say "I love pizza," the translator draws a dot on a map. If you say "I hate pizza," it draws a different dot. The big question is: Does this translator actually understand the difference between love and hate, or is it just guessing based on the shape of the words?

Most current tests for these translators rely on hiring humans to grade them or using other complex AI tools to check the answers. This is like trying to see if a student understands math by asking them to solve a problem, but then having a calculator check the answer for them. You don't know if the student learned the math or just copied the calculator.

This paper introduces a new, clever way to test these translators without needing humans or extra tools. They call it Concept Separation Curves.

Here is how it works, using a simple analogy:

The "Noise" vs. The "Meaning" Test

Imagine you have a sentence written on a piece of paper: "The cat sat on the mat."

The researchers perform two different experiments on this sentence to see how the translator reacts:

The "Static" Test (Fuzzing):
They add a tiny bit of "static" or noise to the sentence, like adding an extra word that doesn't change the meaning.
- Original: "The cat sat on the mat."
- Fuzzed: "The big cat sat on the mat."
- Goal: The meaning is still exactly the same (a cat is sitting). The translator should treat these two sentences as very similar.
The "Flip" Test (Negation):
They change one tiny word that completely flips the meaning.
- Original: "The cat sat on the mat."
- Negated: "The cat did not sit on the mat."
- Goal: The meaning is now the opposite. The translator should treat these two sentences as very different.

The Concept Separation Curve (The Graph)

After running this test on thousands of sentences, they plot the results on a graph.

The "Static" Line: This line should be bunched up on the right side, showing that the translator thinks the noisy sentences are almost identical to the original.
The "Flip" Line: This line should be far away on the left side, showing that the translator realizes the meaning has changed completely.

The Ideal Result: The two lines are far apart, like two distinct groups of people at a party who don't mix. This means the translator truly understands concepts.

The Bad Result: The two lines overlap or mix together. This means the translator is confused. It might think "The cat did not sit" is almost the same as "The cat sat," or it might get confused by the extra word "big."

What They Discovered

The researchers tested many different "translators" (AI models) and found some surprising things:

Some Models are Great: Certain models (like GroNLP for Dutch) clearly separated the "noise" from the "meaning." They acted like wise old sages who know exactly what a sentence means.
Some Models are Confused: Other models (like some older ones) couldn't tell the difference. They treated the "Flip" test almost the same as the "Static" test. It's like a student who thinks "I love pizza" and "I hate pizza" are the same thing because they both contain the word "pizza."
The "Position" Trap: Some models acted like a barcode scanner rather than a thinker. They cared more about where a word was in the sentence than what the word meant. If you moved a word slightly, the model got completely confused, even if the meaning stayed the same.
The Length Problem: The longer the sentence, the harder it was for the models to spot the difference. It's like trying to hear a whisper in a crowded stadium; if the sentence is too long, the tiny change gets lost in the noise.

Why This Matters

This method is like a lie detector test for AI. It doesn't need a human to say "yes, that's right." It just looks at how the AI reacts to small changes.

If the AI reacts to a tiny change in meaning (like "not"), it's smart.
If the AI reacts to a tiny change in noise (like adding "the"), it's just memorizing patterns.

The authors made this test open-source so anyone can use it to check if their AI models are actually "thinking" or just "parroting." It's a simple, visual way to ensure our digital translators are truly understanding the world, not just playing word games.

1. Problem Statement

Current evaluation methods for sentence embedding quality largely rely on downstream tasks (e.g., classification) or human annotations. These approaches suffer from two main limitations:

Confounding Variables: Good performance in a downstream task may stem from the classifier's behavior rather than the quality of the embedding itself.
Annotation Costs: Human evaluation is expensive, time-consuming, and often infeasible for sentence encoders where outputs are vectors rather than discrete labels.
Lack of Conceptual Insight: Existing methods often fail to determine if a model genuinely understands the concept of a sentence or merely recognizes surface-level patterns.

The authors argue for a classifier-independent method to objectively assess whether an embedding model can distinguish between semantic changes (concept shifts) and surface-level noise.

2. Methodology: Concept Separation Curves (CSCs)

The core contribution is a novel evaluation framework called Concept Separation Curves (CSCs). This method evaluates embeddings by contrasting how they react to two types of perturbations:

A. Perturbation Strategy

The method applies two parallel modifications to a corpus of sentences:

Fuzzing (Surface-level Perturbation): Introduces non-semantic noise while preserving the original meaning.
- Implementation: Insertion of articles (e.g., "a", "the" in English; "de", "het" in Dutch).
- Goal: The embedding should remain highly similar to the original vector.
Negation (Semantic Perturbation): Introduces a minimal surface change that fundamentally alters the meaning.
- Implementation: Insertion of a negative particle (e.g., "not" in English; "niet" in Dutch).
- Goal: The embedding should show significant dissimilarity to the original vector.

Constraint: To ensure a fair comparison, the number of tokens added in both Fuzzing and Negation is strictly controlled to be equal.

B. Data Sources

The study utilizes a diverse set of corpora across two languages (Dutch and English) to test generalizability and sentence length effects:

CompetentNL (CNL): Short Dutch sentences (skill descriptions).
ESS Questionnaire: English questions of varying lengths.
Paracrawl (PC): Large web-crawled datasets in both languages, filtered to match specific sentence lengths.

C. Evaluation Metrics

Vector Comparison: For each sentence, the cosine similarity is calculated between the original embedding and the perturbed embeddings (fuzzed and negated).
Visualization (CSCs): The distributions of these similarity scores are plotted using Gaussian Kernel Density Estimation.
- A "good" model shows a Fuzzing curve peaking near 1.0 (high similarity) and a Negation curve shifted significantly to the left (lower similarity).
Quantitative Overlap: To measure the separation objectively, the authors calculate the shared surface area between the normalized Fuzzing and Negation density curves.
- Score Range: 0 (perfect separation) to 1 (perfect overlap/failure).
- Formula: $\sum \min(\text{norm}(\text{fuz}), \text{norm}(\text{neg}))$ over the range $[-1, 1]$ .

3. Key Contributions

Concept Separation Curves (CSCs): A classifier-independent, annotation-free visualization and quantification method for evaluating the conceptual stability of sentence embeddings.
Automated Perturbation Framework: A rule-based system for generating controlled semantic (negation) and non-semantic (fuzzing) variations without manual intervention.
Geometric Overlap Measure: A specific metric to quantify the separation between semantic and non-semantic effects in the embedding space.
Empirical Analysis: A comprehensive study across multiple models, languages, and sentence lengths, identifying specific failure modes (e.g., token position sensitivity).

4. Results

The study evaluated various models, including TF-IDF, FastText, GroNLP, MPNet, RobBERTa, and LaBSE.

Model Performance:
- High Performers: Models like GroNLP (Dutch) and LaBSE (English) demonstrated clear separation, with low overlap scores (e.g., GroNLP on CNL: 0.0221). This indicates they successfully distinguish between meaning changes and noise.
- Low Performers: Models like FastText and MPNet showed high overlap scores (e.g., MPNet: ~0.98). In these cases, the embeddings reacted similarly to both noise and negation, suggesting they function more like positional hash functions than conceptual encoders.
Sentence Length Effect:
- As sentence length increases, the ability of the method to detect concept separation decreases. In longer sentences, the addition of a single token represents a smaller percentage of the total content, diluting the impact of the perturbation on the vector.
Token Position Sensitivity:
- Some models (e.g., MPNet, RobBERTa) exhibited drastic vector changes based purely on token insertion position, regardless of whether the change was semantic or non-semantic. This suggests a reliance on positional encoding rather than semantic content.
Language Differences:
- Dutch models generally showed better separation on Dutch data, and English models on English data, though the spread of overlap scores was wider in Dutch datasets.

5. Significance and Implications

Objective Evaluation: CSCs provide a reproducible, automated way to assess embedding quality without the bias or cost of human annotation or downstream classifiers.
Diagnostic Tool: The method reveals specific failure modes, such as a model's inability to handle sentence length variations or its over-sensitivity to token position.
Domain Applicability: The approach is language-agnostic and can be applied to any domain where conceptual stability is critical (e.g., legal text, medical records), provided the perturbation rules can be adapted.
Limitations & Future Work:
- The current method relies on token insertion, which may not work for all languages (e.g., agglutinative languages or those where negation requires structural changes).
- The metric does not directly correlate with performance in specific downstream tasks (e.g., search retrieval might prioritize exact matches over conceptual understanding).
- Future work could explore using synonyms/antonyms instead of insertions or refining the similarity metric to account for high-dimensional vector spaces.

In conclusion, the paper establishes that Concept Separation Curves are a robust, interpretable tool for diagnosing whether sentence embeddings truly capture semantic meaning or merely surface-level statistical patterns.