A Supervised Learning Approach Towards Profiling the Preservation of Authorial Style in Literary Translations

Gerard LynchCOLING 2014

TL;DR

This paper explores whether a literary translator's unique style is preserved across their translations of different authors, using supervised learning to analyze Constance Garnett's translations of Russian literature. The study demonstrates that machine learning can effectively distinguish between authors based on their translated texts, achieving over 90% accuracy, suggesting that authorial style is indeed discernible even after translation.

What Problem Does This Paper Address?

The field of translation studies grapples with the concept of "translationese" – the potential for translations to exhibit linguistic features distinct from original texts. A related, yet less explored, question is how a translator's individual style manifests across their translations of different authors. While some argue that the best translators are "invisible," their style indistinguishable, this paper investigates whether authorial fingerprints remain detectable. This is crucial for understanding the interplay between original authorial voice, translator's choices, and the resulting translated text.

Methods

The researchers employed supervised learning techniques, specifically Support Vector Machines (SVM), Naive Bayes, and Simple Logistic Regression classifiers. They extracted two main types of features from the translated texts:

  • Document-level metrics: These are holistic measures of the text, such as type-token ratio (lexical diversity), readability scores (Coleman-Liau Index, Automated Readability Index), average sentence length, and ratios of different word classes (nouns, pronouns, conjunctions, numerals).
  • N-gram features: These capture sequential patterns of words (unigrams) and part-of-speech tags (POS bigrams). Noun unigrams were excluded to mitigate topic-based classification.

The methodology involved training classifiers on segments of translated text to predict the original author. They used ten-fold cross-validation to evaluate performance and information gain to rank feature importance.

Datasets

The study utilized a corpus of translated literary works by Constance Garnett. Specifically, it included:

  • Eight works by Ivan Turgenev
  • Seven works by Fyodor Dostoyevsky
  • Eleven collections of short stories by Anton Chekhov

These texts were sourced from Project Gutenberg. The corpus was divided into 10-kilobyte segments, resulting in 942 segments for analysis. No new benchmark datasets were introduced; the focus was on analyzing an existing body of translated work.

Key Insights and Ideas

  • Authorial Style Persists: The most significant finding is that authorial style is remarkably preserved in translation, to the extent that machine learning models can reliably distinguish between authors based on their translated works. This challenges the notion of complete translator invisibility.
  • Feature Importance: Both document-level metrics and n-gram features are valuable for distinguishing authorial style. However, specific features like the ratio of nouns to total words, pronouns to total words, and occurrences of certain adverbs and reporting verbs emerged as particularly strong discriminators.
  • Combinatorial Power: Combining different feature sets (document-level metrics and n-grams) significantly boosted classification accuracy, highlighting the complementary nature of these feature types.
  • Translator's Role: While authorial style is preserved, the study implicitly acknowledges the translator's role in shaping how that style is presented. The question remains whether the translator is merely a conduit or actively constructs a specific stylistic rendition for each author.

Key Results

  • Document-level metrics alone: Achieved an accuracy of 87% using an SVM classifier.
  • Word unigram features (100 words): Achieved 89.5% accuracy with SVM and 91.5% with Simple Logistic Regression.
  • POS bigram features (1021 unique): Achieved 83% accuracy with SVM.
  • Combined feature sets (1153 features): Achieved an impressive 95% accuracy with SVM and Simple Logistic Regression.

Qualitatively, features like the ratio of nouns to total words, pronouns to total words, and specific adverbs (e.g., "almost," "perhaps") and verbs (e.g., "observed," "replied," "repeated") were identified as key discriminators.

Limitations and Future Work

  • Single Translator: The study focuses on only one translator, Constance Garnett. It's acknowledged that different translators might exhibit varying degrees of stylistic preservation or influence.
  • Source Text Relationship: The paper doesn't deeply explore how the translator's choices relate to the original source text. Future work could investigate whether stylistic shifts reflect the original or represent the translator's own constructed idiolect for a given author.
  • Temporal Evolution: The authors note that Garnett's translation career spanned decades, and her style might have evolved. This temporal aspect could influence the results and warrants further investigation.
  • Broader Scope: Future research is suggested to encompass a wider range of translators and languages, and to employ deeper linguistic processing (e.g., semantic role labeling, deep syntactic parsing) to uncover more nuanced stylistic patterns.

Summary updated

BibTeX
@inproceedings{lynch2014supervised,
  author = {Gerard Lynch},
  title = {{A Supervised Learning Approach Towards Profiling the Preservation of Authorial Style in Literary Translations}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2014},
}
Figure