💬 NLP

CHiPS: Character Histograms and Positional Signals for Lightweight Authorship Attribution in Romanian Texts

This paper introduces CHiPS, a lightweight and transparent character-level authorship attribution method for Romanian texts that utilizes character histograms and positional spectral signals to achieve high accuracy without relying on tokenization, syntactic analysis, or large language models, while demonstrating the effectiveness of restricted feature sets under strict leakage control.

Original authors: Sanda-Maria Avram, George C. Ţurcaş

Published 2026-07-28

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sanda-Maria Avram, George C. Ţurcaş

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 a detective trying to solve a mystery, but instead of looking for fingerprints or footprints, you are looking at the invisible "handwriting" of a computer. This field, called authorship attribution, is all about figuring out who wrote a specific text when the name is missing or disputed. It's like trying to guess which friend wrote a anonymous note in a school diary just by looking at how they use commas, where they put spaces, and which letters they seem to love or hate. While modern computers are great at reading the meaning of words (like understanding a joke or a sad story), this paper asks a simpler, more fundamental question: Can we identify an author just by the tiny, unconscious habits they have with the letters and symbols themselves? Think of it as listening to the rhythm of a song rather than the lyrics; even if you change the words, the beat might still give the singer away. This is important because sometimes we need to know who wrote something without relying on complex, heavy-duty computer brains that might be too complicated to check or understand.

The researchers behind this study, Sanda-Maria Avram and George C. Turcaș, decided to build a new detective tool called CHiPS (which stands for Character Histograms and Positional Signals) to solve this puzzle for texts written in Romanian. Their goal wasn't to build the most powerful, super-complex AI possible, but rather to create a "lightweight" and transparent method that anyone could inspect. They wanted to see how much of an author's identity is hidden in the simple distribution of characters and the rhythm of where those characters appear, without using fancy word-segmenting tools or massive pre-trained language models.

Here is how their method works, using a fun analogy: Imagine every author has a unique "style fingerprint." CHiPS looks at this fingerprint in two different ways.

First, there is the CH-SVM part, which acts like a frequency counter. It simply counts how often the author uses specific characters, like the letter 'a', a comma, or a space. It doesn't care about the order of the words or even if the letters are next to each other; it just looks at the total mix. It's like a baker who can identify a specific chef just by the exact ratio of flour, sugar, and salt they always use, regardless of what cake they are making.

Second, there is the FFT12-LR part, which acts like a rhythm detector. Instead of just counting, it looks at where specific characters appear in the text. It treats the text like a musical score, turning the appearance of things like punctuation or capital letters into a signal wave. It then uses math (specifically Fourier analysis, which is like breaking a sound wave into its musical notes) to find patterns in the spacing and rhythm. It's like noticing that a specific author always puts a comma exactly every 15 words, or that they tend to use exclamation marks in bursts at the end of paragraphs.

The researchers combined these two detectives into a team called CHiPS-F. They tested this team on a locked dataset of Romanian texts called ROST, which contained 400 files written by 10 different authors. To make sure the test was fair and the computer didn't "cheat" by memorizing parts of the same story, they kept all fragments of a single story together in either the training group or the test group, never splitting them up.

The results were quite interesting. When they let a standard, powerful computer method that looks at short chains of letters (like "th" or "ing") run the test, it got a perfect score, identifying the author of every single test file correctly. However, the CHiPS team, which was restricted to only looking at single-character counts and rhythm signals, still did incredibly well. The combined CHiPS-F team correctly identified the author in 54 out of 58 test files, achieving an accuracy of 0.9310 (or 93.1%).

The paper explicitly states that CHiPS is not the absolute best possible classifier; the standard method that looks at letter chains was stronger. Instead, the paper's main discovery is that even with strict limits—ignoring word meanings and ignoring letter combinations longer than one character—an author's style is still very visible in their character habits and rhythmic signals. The "rhythm" part of the tool helped fix a few mistakes the "counting" part made, proving that where you put your punctuation is just as important as how often you use it.

They also tried a "re-ranker" tool (CHiPS-R) that would look at the top five guesses and try to pick the right one if the main tool was close but wrong. On a second, larger dataset of cleaned stories, this extra step helped improve the score to 0.8919 accuracy. However, on the main locked test, this extra step didn't improve the results, suggesting that while it might help in some situations, it's not a magic fix for everything.

In the end, the authors conclude that CHiPS is a valuable, transparent tool. It proves that you don't need a massive, complex neural network to find an author's digital fingerprint; sometimes, just looking at the simple, unconscious habits of how a person types and punctuates is enough to give them away. It offers a clear, checkable way to understand authorship that doesn't rely on "black box" AI, making it a useful baseline for future studies, especially in languages where complex tools might not be available.

Technical Summary: CHiPS for Romanian Authorship Attribution

Problem Statement
The paper addresses the closed-set authorship attribution problem for Romanian texts: identifying the most likely author from a set of candidates given a text of unknown origin. While recent approaches often rely on transformer-based models and complex linguistic preprocessing, the authors investigate the extent to which authorial signal exists in raw, low-level character evidence. The study is motivated by the need for transparent, reproducible, and language-agnostic methods that do not depend on tokenization, syntactic parsing, or pretrained language models, which can be scarce or undesirable for low-resource languages like Romanian.

Methodology: CHiPS
The proposed method, CHiPS (Character Histograms and Positional Signals), is a lightweight, non-neural framework comprising two complementary components and optional fusion/reranking layers. The approach strictly avoids character n-grams of length $n \ge 2$ in its histogram component and relies solely on raw character data after normalization.

CH-SVM (Character-Histogram Classifier):
- Mechanism: Models the one-character marginal distribution of a text.
- Features: Uses normalized frequencies of individual characters (preserving case, diacritics, punctuation, and digits) and a small set of scalar statistics (Shannon entropy, ratios of uppercase/digits/punctuation/diacritics, and log text length).
- Classifier: A Linear Support Vector Machine (LinearSVC).
- Constraint: Deliberately excludes contiguous character subsequences (bigrams, trigrams, etc.) to isolate marginal distributional evidence.
FFT12-LR (Positional-Signal Classifier):
- Mechanism: Treats selected character classes as impulse trains (binary indicator sequences) over character positions.
- Channels: Defines 12 normalized channels (e.g., space, newline, specific punctuation marks, digits, and Romanian diacritics).
- Features: Extracts spectral descriptors using Welch power spectral density estimation. For each channel, it computes 10 band energies (periods 2–4 up to 1024–2048 characters), spectral centroid, and spectral entropy, yielding 144 features per text.
- Classifier: Multinomial Logistic Regression.
Fusion and Reranking:
- CHiPS-F: A decision-level fusion strategy combining CH-SVM and FFT12-LR scores. The fusion weight ( $\alpha$ ) is tuned solely on validation data.
- CHiPS-R: An optional top-5 listwise reranker trained on out-of-fold predictions. It selects the true author from a candidate set of five proposed by the base models, using a logistic regression model gated by score margins to avoid overriding high-confidence anchor decisions.

Experimental Protocol and Data
The study emphasizes a leakage-safe evaluation protocol to prevent data contamination, a critical issue in authorship attribution where fragments of the same source text might otherwise appear in both training and test sets.

Grouped Splitting: Splits are performed at the source-text group level (grouping files by author and base title), ensuring related fragments remain in the same partition (train, validation, or test).
Corpora:
- ROST (Primary): A locked benchmark with 400 files from 392 source-text groups by 10 authors.
- ROSTories-cleaned (Secondary): A larger, cleaned corpus of 1,248 files from 1,240 groups by 19 authors, constructed to include ROST material while removing known data quality issues from the public ro-stories release.
Normalization: A conservative pipeline preserving case, diacritics, and punctuation, standardizing Romanian characters (e.g., cedilla to comma-below), and collapsing whitespace.

Key Results
Experiments were conducted on the locked ROST split (58 held-out source-text groups) and the ROSTories-cleaned set.

ROST (Primary Benchmark):
- CHiPS-F achieved 0.9310 accuracy and 0.9341 macro-F1, correctly classifying 54/58 groups.
- CH-SVM alone reached 0.9138 accuracy (53/58), while FFT12-LR reached 0.8448.
- Baseline Comparison: A matched character 2–5-gram TF–IDF SVM (unrestricted n-grams) achieved 1.0000 accuracy on the same split.
- Reranking: CHiPS-R did not improve upon CHiPS-F on the locked ROST split (dropping to 53/58 correct), though it showed potential in the secondary corpus.
ROSTories-cleaned (Secondary Benchmark):
- CHiPS-R achieved the best results with 0.8919 accuracy and 0.8708 macro-F1, outperforming CHiPS-F (0.8811 accuracy).
Ablation and Interpretability:
- Removing punctuation from CH-SVM reduced accuracy, confirming its importance. Removing digits, diacritics, or uppercase features did not significantly degrade performance on this specific split.
- The model is highly interpretable: CH-SVM weights highlight discriminative characters and scalar statistics (e.g., character entropy, specific punctuation ratios), while FFT12-LR reveals rhythmic patterns in the placement of specific character classes.

Significance and Claims
The paper makes modest, diagnostic claims rather than asserting state-of-the-art superiority in raw classification power:

Transparency and Auditability: CHiPS demonstrates that a restricted, transparent character-level representation (marginal frequencies + positional signals) can capture substantial authorial signal without relying on n-grams ( $n \ge 2$ ), embeddings, or neural fine-tuning.
Methodological Rigor: The study establishes a leakage-safe evaluation protocol for Romanian authorship attribution, ensuring that results are not inflated by source-text fragmentation.
Complementary Evidence: The results show that while unrestricted character n-grams are the strongest classifiers (achieving perfect accuracy on ROST), the marginal distributional evidence (CH-SVM) is robust on its own, and positional signals (FFT12-LR) provide complementary value, particularly in decision fusion.
Baseline Utility: CHiPS serves as a strong, inspectable baseline for low-resource languages where complex NLP pipelines are unavailable, offering a trade-off between predictive performance, feature compactness (102–144 features vs. hundreds of thousands in TF-IDF), and interpretability.

The authors explicitly state that CHiPS is not intended to replace transformer-based systems or feature-rich pipelines but to provide a "diagnostic" view of how far restricted character evidence can go under strict leakage control.

Technical Summary: CHiPS for Romanian Authorship Attribution

More like this