Whodunit? Learning to Contrast for Authorship Attribution

Anonymous

Abstract: Authorship attribution is the task of identifying the author of a given text. Most existing approaches use manually designed features that capture a dataset's content and style. However, this dataset-dependent approach yields inconsistent performance. Thus, we propose to fine-tune pretrained language representations using a combination of contrastive learning and supervised learning (Contra-X). We show that Contra-X advances the state-of-the-art on multiple human and machine authorship attribution benchmarks, enabling improvements of up to 6.8\%. We also show Contra-X to be consistently superior to cross-entropy fine-tuning across different data regimes. Crucially, we present qualitative and quantitative analyses of these improvements. Our learned representations form highly separable clusters for different authors. However, we find that contrastive learning improves overall accuracy 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 analyze the effect of combining contrastive learning with cross-entropy fine-tuning for authorship attribution.

Paper Type: long