💬 NLP

Interpretable Semantic Gradients in SSD: A PCA Sweep Approach and a Case Study on AI Discourse

This paper introduces a PCA sweep procedure to systematically determine the optimal number of components in Supervised Semantic Differential (SSD) analysis, thereby reducing researcher degrees of freedom and enhancing the interpretability of semantic gradients, as demonstrated in a case study linking narcissism traits to distinct framings of AI discourse.

Original authors: Hubert Plisiecki, Maria Leniarska, Jan Piotrowski, Marcin Zajenkowski

Published 2026-03-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hubert Plisiecki, Maria Leniarska, Jan Piotrowski, Marcin Zajenkowski

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 are trying to understand how different people feel about a complex topic, like Artificial Intelligence. You have a pile of short essays written by 349 people, and you also have a score for each person that measures how much they crave Admiration (wanting to be seen as great) versus how much they feel Rivalry (wanting to fight against others).

The researchers want to find a "semantic gradient"—a invisible line in the language that shows how the meaning of "AI" shifts as a person's personality changes. Do people who crave admiration see AI as a helpful partner? Do those who feel rivalry see it as a threat?

To do this, they use a method called Supervised Semantic Differential (SSD). Think of SSD as a high-tech compass that tries to map the "mood" of words. But here's the problem: the map is huge and messy. To make sense of it, the researchers have to shrink the map down, like folding a giant world map into a small pocket guide.

The Problem: The "Folding" Dilemma

In the old version of this method, the researchers had to guess how much to fold the map (how many dimensions to keep).

Fold it too much? You lose important details. The map becomes a blurry scribble where "innovation" and "deception" look the same.
Don't fold it enough? The map is so full of tiny, noisy details (like individual dust specks) that the main roads disappear, and the compass spins wildly.

This guesswork was a "researcher degree of freedom." It meant the scientist could accidentally (or on purpose) choose a folding size that made their favorite theory look true, even if it wasn't. It was like trying to find a needle in a haystack, but you get to decide how big the haystack is.

The Solution: The "PCA Sweep" (The Goldilocks Search)

The authors of this paper invented a new tool called the PCA Sweep. Instead of guessing, they built a robot that tests every possible way to fold the map, one by one.

Imagine you are tuning a radio to find a clear station. You turn the dial slowly:

Too low (Static): The signal is weak and the meaning is lost.
Too high (Noise): You hear too much static and background chatter; the music is drowned out.
Just right (The Sweet Spot): The music is clear, the lyrics make sense, and the signal doesn't jump around.

The PCA Sweep does exactly this. It checks three things for every setting:

Capacity: Is the map holding enough information?
Interpretability: Can we clearly see the "Admiration" side and the "Rivalry" side?
Stability: If we nudge the dial just a tiny bit, does the picture stay the same, or does it jump to a completely different channel?

The robot picks the "Goldilocks" setting: the smallest size that gives a clear, stable, and meaningful picture.

The Case Study: AI and Narcissism

The researchers tested this new tool on their AI essays. They looked for a connection between Admiration and how people talked about AI.

What they found with the "Goldilocks" setting (K=15):
The map became crystal clear. They found a strong, stable line:

The "Admiration" Side (Positive): People who wanted to be admired saw AI as a partner. They used words like innovation, collaboration, empower, and cultivate. They saw AI as a tool to build a better future together.
The "Non-Admiration" Side (Negative): People lower in this trait saw AI as a threat. They used words like deception, ridiculous, unfair, and laughable. They saw AI as a trickster or a joke.

What happened with the "Wrong" setting?
To prove their point, they tried a "counterfactual" experiment. They forced the map to be huge (120 dimensions), ignoring the sweep.

Result: The map became a blurry mess. The clusters of words were "diffuse" (scattered everywhere). You couldn't tell if a word belonged to "innovation" or "deception." It was like looking at a photo that was out of focus; you could see colors, but no shapes.

Why This Matters

This paper is a bit like inventing a standardized ruler for measuring language.

Before: Scientists could use a ruler that stretched or shrank depending on what they wanted to measure, leading to confusing or biased results.
Now: The PCA Sweep forces everyone to use the same, scientifically proven "Goldilocks" ruler.

The Takeaway:
By using this new "sweep" method, the researchers showed that people who crave Admiration genuinely view AI through a lens of hope and partnership, while those who don't feel that drive view it with suspicion and mockery.

Most importantly, they proved that this isn't just a fluke of their math. By automatically finding the "stable" setting, they removed the ability to "cherry-pick" results. It makes the science of understanding human language more honest, transparent, and reliable.

In short: They built a smart filter that automatically finds the perfect amount of detail to see the truth in how we talk, without letting the researchers' biases blur the picture.

1. Problem Statement

Supervised Semantic Differential (SSD) is a mixed-method approach used to model how the meaning of text concepts shifts based on continuous individual-difference variables (e.g., personality traits). It operates by:

Aggregating word embeddings into Personal Concept Vectors (PCVs).
Applying Principal Component Analysis (PCA) to reduce dimensionality.
Fitting a linear regression to find a "semantic gradient" (a vector direction in embedding space) that correlates with the outcome variable.
Interpreting the poles of this gradient via clustering and text retrieval.

The Core Issue: The original SSD method requires reducing the dimensionality of PCVs via PCA before regression. However, there is no systematic method for selecting the number of retained components ( $K$ ).

Leaving $K$ to researcher discretion introduces "researcher degrees of freedom."
This increases the risk of overfitting, reduces transparency, and can bias the substantive interpretation of semantic gradients.
Current practices often rely on arbitrary choices or stability across similar ranges without a principled optimization criterion.

2. Methodology: The PCA Sweep Procedure

The authors propose a PCA Sweep procedure to automate and principled the selection of $K$ . Unlike standard model selection which optimizes for predictive accuracy, this sweep optimizes for interpretability and stability, which are central to SSD's qualitative goals.

The procedure iterates through a range of candidate dimensions ( $K$ ) and evaluates three specific diagnostics:

A. Representation Statistics

Computes the cumulative variance explained by the first $K$ components.
Role: Serves as a baseline for detrending rather than a selection criterion, acknowledging that higher $K$ mechanically captures more information but also more noise.

B. Interpretability Diagnostics

After fitting the regression for a specific $K$ , the semantic gradient is back-projected to the original embedding space.
Nearest neighbors are retrieved for positive and negative poles and clustered.
Metrics:
- Cluster Coherence: Internal consistency of the clusters.
- Alignment: Cosine alignment between cluster centroids and the gradient vector.
Processing: These metrics are aggregated into a weighted score, detrended against the log-variance explained, and standardized to $z$ -scores to prevent bias toward higher dimensions.

C. Stability Diagnostics

Measures the change in the semantic gradient as $K$ increases.
Metric: $\Delta_K = 1 - \cos(\hat{\beta}_K, \hat{\beta}_{K-1})$ , where $\hat{\beta}$ is the normalized gradient.
Logic: Low values indicate high stability. High instability suggests the gradient relies on weakly supported components.

D. Selection Rule

Smoothing: To avoid rewarding isolated local maxima caused by volatility at low $K$ , both the interpretability and stability curves are smoothed using a local neighborhood average (Area Under the Curve, AUCK).
Joint Score: A joint score is calculated for each $K$ :
$\text{joint\_score}_K = \frac{1}{2} (\text{interp\_auck}_K + \text{stab\_auck}_K)$
Decision: The selected $K$ is the smallest value that attains the maximum joint score. This prioritizes parsimonious, stable, and interpretable solutions.

3. Case Study: Admiration and Rivalry in AI Discourse

The method was validated on a corpus of 349 short posts about Artificial Intelligence (AI) written by Prolific participants.

Variables:
- Outcome: Narcissistic traits: Admiration (ADM) (assertive self-enhancement) and Rivalry (RIV) (defensive antagonism).
- Data: Texts embedded using 300-dimensional GloVe (Dolma model) with SIF weighting.
Sweep Parameters: $K \in \{1, 3, \dots, 119\}$ .

Results

Admiration (ADM):
- The sweep selected $K=15$ .
- Regression: Significant model ( $R^2_{adj} = .19, p < 10^{-10}$ ).
- Gradient Interpretation:
  - Positive Pole: Optimistic, collaborative, and future-oriented (e.g., "innovation," "partnership," "cultivation").
  - Negative Pole: Distrustful, antagonistic, and derisive (e.g., "deception," "ridicule," "misleading").
- Stability: The gradient was stable and the clusters were coherent.
Rivalry (RIV):
- The model was not statistically significant ( $p = .095$ ).
- No robust semantic gradient emerged, suggesting Rivalry does not strongly structure AI discourse in this dataset.
Counterfactual Analysis:
- Running SSD with an arbitrarily high dimension ( $K=120$ ) resulted in diffuse, weakly structured clusters that were difficult to interpret, validating the necessity of the sweep to constrain dimensionality.

4. Key Contributions

Methodological Innovation: Introduces the PCA Sweep, a systematic heuristic for selecting PCA dimensions in SSD that prioritizes semantic coherence and gradient stability over predictive power.
Reduction of Researcher Bias: Removes arbitrary dimensionality choices, thereby reducing researcher degrees of freedom and increasing the reproducibility of SSD analyses.
Psychological Insight: Demonstrates that individuals high in Admiration view AI through a lens of collaborative progress and status enhancement, whereas Rivalry does not show a similar structured semantic alignment in this context.
Validation of Interpretability: Shows that lower-dimensional, sweep-selected representations yield more psychologically meaningful clusters than high-dimensional, over-parameterized representations.

5. Significance and Limitations

Significance:

The paper bridges the gap between quantitative regression and qualitative interpretation in NLP.
It provides a "lightweight, model-agnostic" criterion that can be integrated into any embedding-based analysis to enhance transparency.
It reinforces the idea that for interpretive methods, stability and coherence are more valuable metrics than maximizing variance explained or $R^2$ .

Limitations:

Data Scope: The case study relies on a small, survey-elicited dataset of short texts, limiting generalizability to other populations or long-form discourse.
Granularity: The analysis treated whole posts as PCVs rather than using concept-specific lexicons, potentially blurring fine-grained semantic distinctions.
Scope of Control: The sweep only addresses dimensionality selection. Other "upstream" choices, such as the base embedding model (e.g., GloVe vs. BERT) or window size, remain open variables that could influence results.
Causality: The findings are correlational; the link between narcissism traits and AI semantics requires further testing in controlled laboratory settings.

Ethical Note: The authors emphasize that SSD is a tool for hypothesis generation and group-level pattern exploration, not for individual profiling or predictive assessment of psychological traits.