[Paper Review] Personalized Machine Translation: Preserving Original Author Traits

Ella Rabinovich, Shachar Mirkin+3 more|arXiv (Cornell University)|Oct 18, 2016

Authorship Attribution and Profiling4 citations

TL;DR

This paper proposes gender-aware statistical machine translation (SMT) systems that preserve authorial gender traits in translations by applying domain-adaptation techniques using gender-annotated parallel data. It demonstrates that personalized SMT models retain gender markers significantly better than standard models while maintaining translation quality, showing that gender signals from original texts are otherwise lost in both human and machine translation.

ABSTRACT

The language that we produce reflects our personality, and various personal and demographic characteristics can be detected in natural language texts. We focus on one particular personal trait of the author, gender, and study how it is manifested in original texts and in translations. We show that author's gender has a powerful, clear signal in originals texts, but this signal is obfuscated in human and machine translation. We then propose simple domain-adaptation techniques that help retain the original gender traits in the translation, without harming the quality of the translation, thereby creating more personalized machine translation systems.

Motivation & Objective

• To investigate whether gender traits in original texts are preserved during human and machine translation.
• To assess the extent to which gender markers are obfuscated in both manual and automatic translations.
• To develop and evaluate domain-adaptation techniques that incorporate gender information into SMT systems.
• To create a new parallel corpus annotated with gender and age for multilingual evaluation.
• To demonstrate that personalized SMT can maintain translation quality while enhancing stylistic fidelity to the original author’s traits.

Proposed method

• Annotate a new parallel corpus (Europarl) with gender and age information for English, French, and German.
• Train standard SMT systems and apply domain-adaptation techniques to create gender-aware SMT models using gender-annotated parallel data.
• Use InfoGain feature selection to identify gender-discriminative words in source and target languages.
• Compare gender classification accuracy on original texts, manual translations, and automatic translations (SMT and personalized SMT).
• Apply personalized SMT models by fine-tuning on gender-specific subsets of the parallel data to preserve authorial traits.
• Analyze translation outputs for gender markers (e.g., ‘really’ vs. ‘exactly’, ‘think’ vs. ‘believe’) to evaluate personalization effects.

Experimental results

Research questions

• RQ1 To what extent are gender traits preserved in human and machine translations compared to original texts?
• RQ2 Why is gender classification more difficult on manual translations than on SMT outputs?
• RQ3 Can domain-adaptation techniques be used to create gender-aware SMT systems that retain authorial gender markers?
• RQ4 How do gender markers in the source language influence gender traits in the target language translation?
• RQ5 What specific lexical and syntactic features reflect gender differences in original and translated texts?

Key findings

• Gender has a strong signal in original texts, but this signal is significantly weakened in both manual and automatic translations.
• Manual translation results in lower gender classification accuracy than SMT, suggesting that human translators may obscure gender markers more than SMT systems.
• Gender-aware SMT models achieve higher gender classification accuracy on translated texts compared to standard SMT, without degrading overall translation quality.
• The translation process results in a hybrid gender signal, where gender markers from the source language dominate those of the target language.
• Specific lexical markers such as ‘really’ (female) and ‘exactly’ (male) in English are preserved or reintroduced in personalized SMT outputs, demonstrating effective personalization.
• The choice of source language significantly affects the detectability of gender markers in translations, with German and French sources showing stronger persistence of gender signals in English translations.