CHiPS: Character Histograms and Positional Signals for Lightweight Authorship Attribution in Romanian Texts CHiPS：用于罗马尼亚文本轻量级作者归属的字符直方图和位置信号

TL;DR

• CHiPS is a lightweight, character-level authorship attribution method for Romanian texts that avoids tokenization, syntactic analysis, pretrained language models, and transformer fine-tuning.
• It uses two complementary style fingerprints: CH-SVM (character-histogram classifier based on one-character marginal distributions) and FFT12-LR (positional-signal classifier using Fourier/Welch spectral descriptors of impulse trains).
• The leakage-safe fusion variant CHiPS-F achieves 0.9310 accuracy and 0.9341 macro-F1 on a locked grouped ROST split with strict source-text-level evaluation and grouped five-fold model selection.
• The method intentionally restricts features to transparent, interpretable character evidence under strict leakage control, not aiming for maximum possible accuracy but demonstrating robust performance with minimal computational overhead.
• On the secondary ROSTories-cleaned corpus, CHiPS-R achieves 0.8919 accuracy and 0.8708 macro-F1, showing consistent performance across different dataset configurations.

Why It Matters

This work demonstrates that high-performance authorship attribution can be achieved with simple, transparent, and computationally efficient methods without relying on complex deep learning architectures or large pretrained models. For AI practitioners, it provides an important benchmark for lightweight NLP solutions in forensic text analysis, particularly for languages like Romanian where resources may be limited. The strict leakage control and focus on interpretable features make this approach valuable for real-world applications where model transparency and reproducibility are critical.

Technical Details

• CH-SVM Component: Uses character-histogram features based solely on one-character marginal distributions, avoiding n-grams with length ≥ 2 to maintain simplicity and prevent data leakage through contextual patterns.
• FFT12-LP Component: Represents selected characters and punctuation classes as binary indicator sequences (impulse trains) over character positions, then extracts Fourier and Welch spectral descriptors to capture positional writing style patterns.
• Fusion Strategy: Implements CHiPS-F as a decision-level fusion of the two components designed to be leakage-safe, ensuring no information from test documents contaminates training.
• Evaluation Protocol: Uses locked grouped ROST splits with source-text-level evaluation and grouped five-fold model selection to rigorously prevent data leakage between training and testing sets.
• Optional Enhancement: Includes a top-5 listwise reranker trained exclusively on out-of-fold predictions to further improve performance while maintaining the integrity of the evaluation protocol.

Industry Insight

The success of CHiPS suggests that for many authorship attribution tasks, especially in resource-constrained environments or for specific languages, simpler statistical approaches can achieve competitive results compared to heavy transformer-based models. This has significant implications for developing deployable forensic tools that require low latency, minimal computational resources, and explainable decisions. Researchers should consider incorporating strict leakage control protocols when evaluating new methods to ensure their reported performance reflects genuine predictive capability rather than data contamination. The approach also highlights the value of combining multiple complementary feature types (statistical histograms + spectral positional signals) to capture different aspects of writing style within a single lightweight framework.

TL;DR

• 提出CHiPS方法，用于罗马尼亚语文本的轻量级作者归属识别。
• CHiPS包含两种风格指纹：基于字符直方图的CH-SVM和基于位置信号的FFT12-LR。
• 实验在ROST数据集上达到0.9310准确率和0.9341宏F1分数，展示了受限透明特征的有效性。
• 方法无需分词、句法分析或预训练语言模型，避免了高阶n-gram特征。
• 研究重点在于严格泄漏控制下受限字符证据的性能极限。

为什么值得看

该论文为资源受限场景下的作者归属问题提供了一种高效且透明的解决方案，其轻量化设计对低资源语言（如罗马尼亚语）的NLP应用具有参考价值。同时，其对“泄漏安全”评估框架的探索，为学术界提供了区分模型泛化能力与数据泄露风险的重要基准范式。

技术解析

• 双模态特征提取：系统采用互补策略，CH-SVM利用单字符边缘分布构建直方图特征，而FFT12-LR将特定字符和标点符号表示为脉冲序列，通过傅里叶变换提取频谱描述符作为位置信号。
• 架构简化：整个流程完全基于字符级别，摒弃了传统的分词器、句法解析器以及大型预训练语言模型（PLM），显著降低了计算开销和部署门槛。
• 防泄漏机制：引入了CHiPS-F融合变体，通过决策层融合确保测试数据不泄露到训练过程中；此外还设计了仅基于交叉验证外预测进行训练的Top-5列表重新排序器，进一步增强了模型的鲁棒性。
• 实验设置：在包含392个源文本组、10位作者的锁定分组ROST分割集上进行评估，采用源文本级别的评估标准和分组五折模型选择协议，确保了结果的统计可靠性。

行业启示

• 可解释性与效率的平衡：在需要高透明度或低延迟的应用场景中（如法律取证或实时内容审核），此类基于简单统计特征的轻量级模型比复杂的深度学习模型更具实用价值。
• 数据隐私保护的重要性：严格的泄漏控制评估标准提醒从业者，在开发作者归属等敏感任务时，必须建立严谨的数据隔离机制，以防止模型过拟合于训练集中的非风格化信息。
• 低资源语言的处理思路：对于缺乏大规模标注数据和预训练模型的语言，利用语言学先验知识（如字符频率和位置模式）结合传统机器学习方法，往往能取得令人满意的性能表现。