On Explainable Features for Translatorship Attribution: Unveiling the Translator’s Style With Causality

Abstract

Translatorship attribution deals with accurately attributing a translation to its translator. The task is challenging because several factors can confound the attribution such as the original author’s style, genre, and topic of the text. The attribution and the identification of the translator’s style could contribute to fields including translation studies and forensic linguistics. In this paper, we pose translatorship attribution as a multiclass classification problem and employ machine learning algorithms. To address the problem of confounding, we use corpora of English translations of the same source material (parallel corpora) to identify the translators’ personal style. We propose two novel feature sets in this task: i) a list of cohesive markers with and without their surrounding punctuation and ii) syntactic n-grams to capture real syntactic information. We employ $\chi ^{2}$ feature selection and, using 10-fold cross-validation, assess the accuracy of several classifiers trained with our proposed features and with word, punctuation, POS, and POS-punctuation n-grams. The results show that the proposed features yield comparable and even higher accuracy results than the reported in the literature on the same corpora and prove that POS-punctuation n-grams are an effective feature set for this task. We also recover the most distinctive features and provide examples of stylistic interpretations of them for each translator. Finally, using insights from causal inference, where confounding is well-defined and studied, we provide a novel explanation for the accepted need of using parallel and contemporaneous corpora on this task and for the different results among types of features.

Keywords

• Computational linguistics
• translator style
• stylometry
• machine learning
• causal inference