💬 NLP

HALvest-Contrastive: Retrieval-Like Authorship Attribution with Patch-Level Late Interaction

This paper introduces HALvest, a large multilingual corpus of scholarly papers, and a novel Patch-Level Late Interaction (PLI) retrieval framework that significantly improves authorship attribution accuracy by comparing sequences of token patches rather than compressing documents into single vectors, thereby effectively mitigating topical confounds.

Original authors: Francis Kulumba, Wissam Antoun, Guillaume Vimont, Laurent Romary, Florian Cafiero

Published 2026-05-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Francis Kulumba, Wissam Antoun, Guillaume Vimont, Laurent Romary, Florian Cafiero

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 figure out if two different books were written by the same person. Usually, this is easy if the books are about totally different things (like one is about cooking and the other about space). But what if both books are about cooking?

If you ask a computer, "Do these two recipes sound like they came from the same chef?" the computer might get tricked. It might say "Yes!" just because both recipes use words like salt, oven, and bake. It's confusing the topic (cooking) with the style (the chef's unique handwriting).

This paper, HALvest-Contrastive, is like a new, super-smart training camp for computers to learn how to spot a writer's true "handwriting" without getting distracted by the topic.

Here is how they did it, broken down into simple parts:

1. The Problem: The "Topic Trap"

In the past, computers trying to identify authors often cheated. If they saw two texts about "AI," they assumed they were written by the same person because they both used the word "algorithm."

The Analogy: Imagine trying to recognize your friend's voice at a loud party. If everyone is shouting about "football," you might think two people shouting about football are your friend just because they are talking about the same thing. You need to hear how they say it, not what they say.

2. The Solution: A Special Library (HALvest)

The authors built a massive library called HALvest containing 17 billion words from open-access academic papers.

The Trick: They created a special version called HALvest-Contrastive. In this version, they forced the computer to compare two papers written by the same author, but about different topics.
- Example: Paper A is about "Quantum Physics." Paper B is about "History of Art." Both are written by Dr. Smith.
- Because the topics are totally different, the computer cannot cheat by looking for shared words. It has to learn Dr. Smith's unique style (how they use commas, sentence length, or transition words) to realize, "Hey, these two very different papers were written by the same person!"

3. The New Way of Comparing: "Late Interaction"

Traditionally, computers would take a whole document, crush it down into a single "summary number" (a vector), and compare those numbers.

The Old Way (Single Vector): Imagine taking a whole novel, squishing it into a single drop of ink, and comparing that drop to another drop. You lose all the details.
The New Way (Late Interaction): Instead of squishing the book, the computer keeps every single word's "fingerprint" separate. It looks at the first word of Book A and finds the best match in Book B, then the second word, and so on. It sums up all these little matches.
- The Result: This is like comparing two books page-by-page, word-by-word, rather than just comparing their covers. It worked much better.

4. The "Patch" Discovery: Finding the Sweet Spot

The researchers wondered: "Do we need to compare every single word, or can we group them?"

They tried grouping words into little chunks called patches (like grouping words into small phrases).
The Finding: Comparing every single word is great, but it's slow and uses a lot of memory. Grouping them into small patches (about 2 to 4 words at a time) gave almost the same accuracy but was much faster.
The Magic Formula: They found a simple rule of thumb for how big these patches should be: Take the square root of the total number of words, multiply by 0.18, and that's your patch size.
- Analogy: It's like packing a suitcase. If you have a small trip (few words), you pack small items individually. If you have a huge trip (many words), you bundle things into boxes. There is a perfect "box size" that saves space without losing anything important.

5. What They Proved

Topic Decoupling Works: When they removed the "topic shortcuts," simple word-counting methods (like looking for common keywords) failed miserably. But the smart neural networks kept working, proving they actually learned the style, not just the topic.
Style vs. Meaning: They showed that their new system is very good at spotting who wrote it, while standard search engines (which look for meaning) are terrible at it. A search engine might think two papers are similar because they are both about "math," but this system knows they were written by different people because the "math voice" is different.
Better than the Old Way: Using this new "patch-level" method improved accuracy significantly compared to the old "squish-it-all-into-one-number" method.

Summary

The paper introduces a new dataset and a new method to teach computers how to recognize a writer's unique style, even when they are writing about completely different subjects. They discovered that looking at small groups of words (patches) rather than single words or whole documents is the "Goldilocks" zone—just right for speed and accuracy.

Important Note: The paper strictly focuses on academic papers and fan-fiction (for testing). It does not claim this technology is ready for spying on people, catching spies, or medical uses. It is purely a tool for understanding how computers can learn to recognize writing styles.

Technical Summary: HALvest-Contrastive

Problem Statement

Authorship Attribution (AA) faces a persistent challenge known as topical confounding. Traditional statistical and neural approaches often struggle to distinguish between stylistic similarity and topical overlap; two writers covering the same subject frequently appear more similar to a model than a single writer covering two different subjects. While neural AA systems have adopted contrastive learning (e.g., triplet loss), they often rely on compressing entire documents into single vectors via mean pooling, a strategy inherited from early statistical methods but potentially suboptimal for capturing nuanced stylistic signals. Furthermore, existing benchmarks often fail to rigorously decouple topic from authorship, allowing models to rely on "topical shortcuts" (vocabulary overlap) rather than learning true stylistic distributions.

Methodology

1. The HALvest Corpus and HALvest-Contrastive Benchmark

The authors introduce HALvest, a 17-billion-token multilingual corpus harvested from 778,000 open-access scholarly papers via the HAL repository. From this, they derive HALvest-Contrastive, an English-only benchmark designed to enforce topic decoupling:

Positive Pairs: Anchor and positive passages are drawn from different papers written by the same author-set, ensuring minimal topical overlap while maintaining stylistic consistency.
Negative Pairs: Negatives are mined from the same disciplinary field but by different author-sets.
Labeling: The ground truth is the author-set (co-authors) rather than a single individual, reflecting the metadata structure of scholarly repositories and avoiding arbitrary single-author heuristics.
Span Lengths: The dataset supports variable span lengths ( $k \in \{2, 4, 6, 8, 10\}$ sentences) to evaluate how stylistic signals scale with text length.

2. Retrieval-Like Architecture for AA

The paper applies the "retrieval playbook" to authorship attribution, moving away from standard triplet loss and single-vector pooling:

Loss Function: The authors employ InfoNCE (Noise Contrastive Estimation) instead of triplet loss. InfoNCE leverages large in-batch negative pools, which naturally concentrates gradients on hard negatives (topically similar but different-author documents), aiding in the decorrelation of style from topic.
Interaction Granularity: The paper investigates three scoring modes:
1. Single-Vector Pooling: Standard mean pooling of token embeddings followed by cosine similarity.
2. Late Interaction (LI): Documents are represented as full sequences of token vectors. Scoring uses MaxSim (sum of maximum cosine similarities between query and document tokens), a technique common in dense retrieval (e.g., ColBERT).
3. Patch-Level Late Interaction (PLI): Token vectors are grouped into contiguous patches (either fixed $n$ -grams or whole words), pooled into single patch vectors, and then matched via MaxSim at the patch level. This reduces storage and computational costs while maintaining interaction granularity.

Key Contributions

HALvest and HALvest-Contrastive: The release of a large-scale, multilingual scholarly corpus and a rigorously constructed English benchmark that minimizes topical overlap between same-author pairs, providing a stricter evaluation protocol for stylometry.
Empirical Validation of Topic Decoupling: The authors demonstrate that lexical baselines (e.g., BM25) suffer significant performance drops when topical shortcuts are removed, whereas neural contrastive models retain accuracy, confirming that the benchmark successfully isolates stylistic signals.
Granularity Study: A systematic comparison of pooling strategies reveals that Late Interaction significantly outperforms single-vector pooling. Crucially, the paper introduces Patch-Level Late Interaction (PLI), showing that grouping tokens into patches yields comparable or slightly better accuracy than full token-level interaction while reducing the vector count.
Empirical Patch-Size Fit: The authors identify a practical heuristic for the optimal patch size $n^*$ relative to the token count $S$ : $n^* \approx 0.18 \cdot \sqrt{S}$ . This suggests that the optimal interaction granularity scales sub-linearly with document length.

Results

Topic Decoupling: On the "Base" (restricted) configuration, BM25 accuracy drops by ~11 percentage points compared to the "Unrestricted" configuration, while neural models (RoBERTa) show robustness, indicating they learn signals beyond lexical overlap.
Performance Gains: Late Interaction (LI) and PLI variants consistently outperform mean pooling across all metrics (Recall@k, nDCG@k) and span lengths. The gain is approximately 20% over mean pooling when keeping the encoder fixed and swapping the scorer.
Granularity Trade-offs: While token-level LI is competitive, fixed-policy PLI (specifically $n$ -gram variants) often matches or slightly exceeds token-level performance with lower computational overhead. The optimal $n$ varies by span length but follows the $\sqrt{S}$ trend.
Cross-Domain Generalization: Models trained on HALvest-Contrastive achieve strong zero-shot triplet accuracy (0.740) on the PAN19 fan-fiction dataset, suggesting the learned representations capture authorship style rather than just academic register or topic.
Scaling Negatives: Increasing the in-batch negative pool size (from 256 to 512) yields consistent, albeit small, improvements, aligning with findings in information retrieval literature.

Significance and Claims

The paper argues that scholarly prose is an ideal testbed for stylometry due to its low-entropy, standardized nature, which makes stylistic residuals easier to isolate from topical noise. By applying retrieval techniques (InfoNCE and Late Interaction) to AA, the authors demonstrate that:

AA is distinct from semantic retrieval: Fine-tuned stylistic models separate same-author pairs effectively, whereas semantic encoders (like E5) fail to distinguish authorship despite high semantic similarity.
Granularity matters: The "retrieval playbook" largely transfers to AA, but the optimal interaction granularity is subtle. Token-level matching is not uniformly superior; patching offers a computationally efficient alternative that preserves performance.
Practical Utility: The derived formula $n^* \approx 0.18 \cdot \sqrt{S}$ provides a practical starting point for practitioners designing AA systems, balancing accuracy with storage and inference costs.

The authors conclude that while the retrieval paradigm improves AA performance, the specific choice of interaction granularity (patches vs. tokens) is a critical design choice that differs from standard semantic retrieval tasks. The work provides open-source code, corpus construction scripts, and the dataset itself to facilitate future research.