Multi-author writing style change detection, a core challenge in PAN@CLEF evaluations, requires precise localization of author transition points in collaboratively authored documents. This paper presents SCL-DeBERTa, novel framework that integrates supervised contrastive learning (SCL) with DeBERTa for fine-grained style boundary detection. in the PAN 2025 evaluation, Our team, xxsu-team, proposed a novel method that achieved F1 scores of 0.955 (ranking 3rd), 0.825 (ranking 1st), and 0.829 (ranking 2nd) on the three subtasks, respectively, establishing a new benchmark for multi-author writing style analysis. This performance demonstrates significant advancements in feature extraction robustness and cross-task adaptability.
The identification of writing styles in multi-author documents has emerged as a core task in digital text forensics, with numerous solutions based on pre-trained language models proposed in recent years. Early approaches primarily relied on traditional feature engineering, such as lexical statistical features and syntactic pattern analysis, but these methods exhibited limited generalization ability in cross-topic scenarios.
With the rise of Transformer models, researchers began exploring the adaptability of diferent
pretrained architectures: Team Gladiators[
In contrast, we adopted the lightweight DeBERTa as the encoder to embed data samples. Its
disentangled attention mechanism[
The PAN25 writing style analysis evaluation task focuses on paragraph-level stylistic change detection in multi-author documents, requiring participating systems to precisely identify all writing style transition boundaries within documents under strictly constrained conditions of author identity and topic variation.
The dataset used in this study for stylistic change detection consists of multiple document pairs, with each document corresponding to two files: • problem-<id>.txt: Contains the original text content.
• truth-problem-<id>.json: Contains the annotations for stylistic changes.
The annotation information for each document is stored as a binary array, indicating whether a stylistic change occurs between adjacent sentences: changes = [1, 2, . . . , −1 ] where ∈ {0, 1} (1) Here, represents the total number of sentences in the document. The overall distribution of the dataset is shown in Table 1.
Each generated training sample contains the following elements: sample = {input_ids, attention_mask, label , doc_id , pair_idx } (2) ⏟Token ⏞IDs ⏟ Attention⏞Mask
Styl⏟istic ⏞Label Docum⏟ent I⏞dentifier Sent⏟ence Pa⏞ir Index
To support the supervised contrastive learning framework, additional document-level metadata is included in the samples: • doc_id: A unique document identifier used to group sentence pairs from the same document. • pair_idx: The position index of the sentence pair within the document.
This metadata enables the model to learn representations of stylistic consistency within a document, enhancing its ability to detect stylistic changes in long documents.
Success Successful
Error Exception ❌ ❌
⚠ ⚠
• File I/O exception handling.
• Encoding error handling.
To ensure the robustness of data processing, the following error handling mechanisms are implemented: • Data consistency checks.
• Automatic filtering of invalid samples.
These mechanisms ensure that most usable data can still be processed efectively, even in cases of partial data corruption or non-standard formatting.
DeBERTa Encoder
CLS Token CLSFeatures
Token Features Hierarchical Features
CClalsassisfiicfiactaiotinon
Header Header
StyleCChahnagnege Style
Probability
Probability Projection Head
Style Vector
SupervisedContrastive Loss The multi-author writing style detection task, established by PAN Lab in CLEF conferences, requires precise localization of author transition points in collaboratively authored documents. Formally, given a document = {1, 2, . . . , } composed of paragraphs, the goal is to identify the boundary set = { | author() ̸= author(+1)}.
We propose SCL-DeBERTa, a novel framework integrating supervised contrastive learning with
DeBERTa[
2 · ReLU(W 1hCLS))
Projection head: z = W4 · ReLU(W 3hCLS) where W1 ∈ R256×768 , W2 ∈ R1×256 , W3 ∈ R256×768 , W4 ∈ R128×256 are learnable parameters.
We introduce a label-guided contrastive loss[
1 ∑︁
∑︁ log =1 | ()| ∈ ()
exp(z · z / ) ∑︀̸= exp(z · z / ) where () = { | = } denotes positive samples sharing the same style label, and = 0.07 is the temperature hyperparameter.
The joint optimization objective combines classification and contrastive losses: ℒtotal = BCE(^, ) + · ⏟ ⏞
ℒcls with = 0.3 controlling the contrastive regularization strength.
ℒ scl style re⏟gularization ⏞ (3) (4) (5) (6) (7)
The PAN 2025 multi-author writing style analysis evaluation task focuses on paragraph-level stylistic change detection, requiring participants to analysis the stylistic consistency of consecutive sentence pairs, participants must accurately identify all author transition boundaries. The task is designed to strictly control the synchronization of author identity and topic variation, ensuring that stylistic discrimination is not afected by topic shifts. Systems are evaluated across three dificulty levels based on the complexity of document topics.
1. Easy Subtask: The document topics are highly consistent, with frequent and obvious author transitions, and significant stylistic diferences between paragraphs. 2. Medium Subtask: The document topics are moderately complex, with more subtle author transitions and smaller stylistic diferences between paragraphs. 3. Hard Subtask: The document topics are highly mixed, with extremely subtle author transitions and almost imperceptible stylistic diferences between paragraphs.
The PAN 2025 multi-author writing style analysis evaluation task centers on paragraph-level stylistic change detection, challenging participants to conduct pure stylistic change detection at the sentence level. By examining the stylistic consistency between consecutive sentence pairs, participants are required to precisely identify all author transition boundaries. The task is meticulously designed to enforce strict synchronization between author identity and topic variation, ensuring that stylistic discrimination remains unafected by topic shifts. Systems are assessed across three dificulty levels, determined by the complexity of document topics.
In this study, we selected DeBERTa as the pre-trained model and utilized a supervised contrastive learning framework to optimize model parameters. Based on this foundation, we proposed SCLDeBERTa, which enhances the ability of supervised contrastive learning to improve style diferentiation. By combining classification and contrastive losses, the model achieves precise detection of author transition boundaries. The trained model was evaluated on three distinct task datasets, resulting in customized models tailored for each task. The hyperparameter settings used in this experiment are detailed in Table 2.
As shown in Table 3, SCL-DeBERTa achieved F1 scores of 0.955 (ranking 3rd), 0.825 (ranking 1st), and 0.829 (ranking 2nd) on the three subtasks:
This table presents the leaderboard results on the PAN 2025 benchmark. Our team, xxsu-team, with the proposed SCL-DeBERTa method, achieved top-tier performance across all three subtasks, ranking 3rd, 1st, and 2nd on Task1, Task2, and Task3, respectively. Compared to other strong baselines, our approach demonstrates superior robustness and adaptability in multi-author writing style analysis, establishing a new benchmark for this challenging task.
In this study, we proposed SCL-DeBERTa, a novel framework that integrates supervised contrastive learning with the DeBERTa encoder to address the challenging task of multi-author writing style detection. By combining classification and contrastive losses, the model efectively enhances style diferentiation and achieves precise detection of author transition boundaries.
While SCL-DeBERTa has shown promising results, several avenues for future research remain open. First, we plan to explore the integration of additional pre-trained language models, such as GPT or LLAMA, to further enhance the model’s stylistic representation capabilities. Second, extending the framework to handle multilingual datasets could broaden its applicability to diverse linguistic contexts. Third, we aim to investigate the impact of incorporating external knowledge, such as syntactic or semantic features, to improve the model’s performance on the Hard subtask. Finally, optimizing the computational eficiency of the framework, particularly for large-scale datasets, will be a key focus to ensure its scalability in practical scenarios.
This work is supported by the National Social Science Foundation of China (24BYY080).
During the preparation of this work, the authors utilized the DeepSeek language model and GPT-4.1 for grammaiand spelling checks. Following the use of this tool, the authors carefully reviewed and edited the contenas necessary and assume full responsibility for the content of this publication