Towards Effective Authorship Attribution: Integrating Class-Incremental Learning

TL;DR

This work reframes Authorship Attribution as Class-Incremental Learning to address open-world, time-evolving authorships. It defines the CIL-AA paradigm, surveys existing CIL approaches, and presents a BERT-based incremental learner with multiple strategy variants (finetuning, decoupled, distillation, replay, and regularization). Extensive experiments on diverse AA datasets quantify forgetting dynamics (PD) and overall accuracy (AvgA), showing that distillation and replay methods yield the strongest resilience to forgetting, with dataset characteristics strongly influencing which variant performs best. The study introduces BuildCILData for creating sequential session splits, provides a public implementation, and discusses practical implications and future directions, including FSCIL and stylometric extensions. Overall, the paper demonstrates the potential of CIL to make AA systems robust to continual learning scenarios while highlighting avenues for further enhancement and realistic deployment in streaming settings.

Abstract

AA is the process of attributing an unidentified document to its true author from a predefined group of known candidates, each possessing multiple samples. The nature of AA necessitates accommodating emerging new authors, as each individual must be considered unique. This uniqueness can be attributed to various factors, including their stylistic preferences, areas of expertise, gender, cultural background, and other personal characteristics that influence their writing. These diverse attributes contribute to the distinctiveness of each author, making it essential for AA systems to recognize and account for these variations. However, current AA benchmarks commonly overlook this uniqueness and frame the problem as a closed-world classification, assuming a fixed number of authors throughout the system's lifespan and neglecting the inclusion of emerging new authors. This oversight renders the majority of existing approaches ineffective for real-world applications of AA, where continuous learning is essential. These inefficiencies manifest as current models either resist learning new authors or experience catastrophic forgetting, where the introduction of new data causes the models to lose previously acquired knowledge. To address these inefficiencies, we propose redefining AA as CIL, where new authors are introduced incrementally after the initial training phase, allowing the system to adapt and learn continuously. To achieve this, we briefly examine subsequent CIL approaches introduced in other domains. Moreover, we have adopted several well-known CIL methods, along with an examination of their strengths and weaknesses in the context of AA. Additionally, we outline potential future directions for advancing CIL AA systems. As a result, our paper can serve as a starting point for evolving AA systems from closed-world models to continual learning through CIL paradigms.

Figures (5)

• Figure 1: Illustration of CIL for AA. The model first receives a set of authors along with multiple samples for training and evaluates its performance based on these authors. In subsequent sessions, the model updates itself only with new authors, different from those in previous sessions. Evaluation is performed using a combined set of all authors encountered thus far, including those introduced in the current session. Different colors correspond to different authors and their documents.
• Figure 2: Empirical validation set categories for 5 datasets.
• Figure 3: Averaged Accuracy of Models on Datasets for 10-session setup.
• Figure 4: Confusion matrix of various models on Blog50 (upper row) and CCAT50 (lower row) after last incremental stage (6-session setup).
• Figure 5: Results of PD rate using 2, 5, 10, and 20 examples for FT-EK and LWF-EK models using Blog50 and IMDb62 datasets at 6-session setup.