Whodunit? Learning to Contrast for Authorship Attribution

Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), pages 1142–1157 November 20–23, 2022. ©2022 Association for Computational Linguistics 1142 Whodunit? Learning to Contrast for Authorship Attribution Bo Ai1, Yuchen Wang1, Yugin Tan1, Samson Tan2∗ 1School of Computing, National University of Singapore 2AWS AI Research & Education Abstract Authorship attribution is the task of identifying the author of a given text. The key is finding representations that can differentiate between authors. Existing approaches typically use man- ually designed features that capture a dataset’s content and style, but these approaches are dataset-dependent and yield inconsistent perfor- mance across corpora. In this work, we propose learning author-specific representations by fine- tuning pre-trained generic language represen- tations with a contrastive objective (Contra-X). We show that Contra-X learns representations that form highly separable clusters for different authors. It advances the state-of-the-art on mul- tiple human and machine authorship attribution benchmarks, enabling improvements of up to 6.8% over cross-entropy fine-tuning. However, we find that Contra-X improves overall accu- racy at the cost of sacrificing performance for some authors. Resolving this tension will be an important direction for future work. To the best of our knowledge, we are the first to integrate contrastive learning with pre-trained language model fine-tuning for authorship attribution. 1 Introduction Authorship attribution (AA) is the task of identify- ing the author of a given text. AA systems are com- monly used to identify the authors of anonymous email threats (Iqbal et al., 2010) and historical texts (Mendenhall, 1887), and to detect plagiarism (Gol- lub et al., 2013). With the rise of neural text gener- ators that are able to create highly believable fake news (Zellers et al., 2019), AA systems are also increasingly employed in machine-generated-text detection (Jawahar et al., 2020). When performed on texts generated by human and machine writers, AA can also act as a type of Turing Test for Natural Language Generation (Uchendu et al., 2021, 2020). ∗Work done at the National University of Singapore. Implementation and datasets are available at https:// github.com/BoAi01/Contra-X.git (a) BERT (b) Contra-BERT Figure 1: t-SNE visualization of the fine-tuned rep- resentations (a: baseline; b: Contra-X). Each color denotes one author in the Blog10 dataset. Our con- trastive method effectively creates a tighter represen- tation spread for each author and increased separation between authors. Best viewed in color. Traditional AA methods design features that characterize texts based on their content or writ- ing style (Jafariakinabad and Hua, 2019; Zhang et al., 2018; Sapkota et al., 2015b; Sari et al., 2018). However, the features useful for distinguishing au- thors are often dataset-specific, yielding inconsis- tent performance under varying conditions (Sari et al., 2018). In contrast, learning features from large corpora of data aims to produce general pre- trained models (Devlin et al., 2018) that improve performance on many core natural language pro- cessing (NLP) tasks, including AA (Fabien et al., 2020). However, it is unclear if …

Content truncated to 500 words