Personalized Machine Translation: Preserving Original Author Traits

• Gender classification is strong on original texts but becomes obfuscated in both manual and statistical machine translation.Gender is even harder to determine from manual translations than from SMT, possibly because of the translation process or human translators.Source: §1 · p. 1
• The study evaluates gender preservation across original, manually translated, and automatically translated texts, extending analysis across domains and language pairs.It uses a different classification methodology motivated by the stronger signal of the translation process itself than the author’s gender.Source: §1 · p. 1
• Gender-aware SMT models use domain adaptation to retain authorial gender markers without harming general translation quality.The approach incorporates gender-specific model components or tuning sets and compares them with a gender-agnostic baseline.Source: §1 · p. 1
• Gender markers differ greatly across languages, and source-language gender traits overshadow target-language traits in manual and automatic translations.The source language significantly affects translated-text features and classification accuracy.Source: §1 · p. 2
• Personalized SMT produces translations more faithful to the original texts’ style by preserving gender traits while maintaining translation quality.The paper frames this as a path toward more personalized machine translation.Source: §1 · p. 2

• Most MT systems disregard personal traits and train a single model, while conflicting writing styles can hinder performance and produce unfaithful translations.Existing customization work mainly addresses text domains rather than writer traits.Source: §2 · p. 2
• Reader-preference personalization predicts users’ preferred MT systems, whereas this work targets traits of the writer.Mirkin and Meunier (2015) modeled overall system choice rather than authorial characteristics.Source: §2 · p. 2
• Translation research shows that authorial gender signals can vanish in both manual and automatic translation, motivating personalized MT models.Mirkin et al. (2015) suggested domain adaptation but did not provide experimental results for personalized MT models.Source: §2 · pp. 2–3
• Previous research largely studied gender and authorial traits within single languages, while little work examined their variation across languages or translation.Gender differences have been identified using content-independent function words and POS-tag ngrams, and some shallow syntactic patterns persist across languages.Source: §2 · p. 2
• The paper claims to be the first to automatically identify speaker gender in manual and automatic translations across multiple domains and language pairs.It also examines how gender-marker distributions differ between source and target languages.Source: §2 · p. 3

• The released resource contains bilingual English–French and English–German corpora enriched with speaker MEP identity, gender, and age metadata.It includes metadata for 3,586 MEPs and restricts sentence pairs to utterances originally produced in English, French, or German.Source: §3 · p. 3, §3.2 · p. 4

3.1 Identification of MEP gender

• Wikidata provides gender information for 2,618 MEPs, covering 73% of the 3,586-member population.Among these records, 1,882 are male and 736 female.Source: §3.1 · p. 3
• Genderize predicts gender for 2,785 MEPs, with 2,730 retained after filtering low-confidence entries.The retained set covers 76% of all MEPs and is 73% male and 27% female.Source: §3.1 · p. 3
• The annotation procedure prioritizes Wikidata, then uses agreed or single predictions from Genderize and AlchemyVision, with manual annotation for the remaining 108 MEPs.The final resource assigns gender to 3,586 MEPs overall.Source: §3.2 · p. 4

3.2 Resource evaluation and statistics

• The final resource contains 947 female and 2,639 male MEPs, with estimated overall gender-information accuracy of 99.88%.The estimate assumes that manual annotation is correct and relies on Wikidata as the gold standard.Source: §3.2 · p. 4

Evaluation methodology

• Gender preservation is measured by comparing author-gender classification accuracy on original, human-translated, and machine-translated texts.Higher classification accuracy is treated as evidence that gender markers remain prominent in the text.Source: §4 · p. 4, §4.1 · p. 4

Datasets

• The study uses Europarl data annotated for demographic traits and a separate TED-talks corpus to test robustness across genres.The TED corpus provides a substantially different genre in which demographic traits may manifest differently.Source: §4.2 · pp. 4–5

Experimental design

• Experiments cover English–French and English–German translation in both directions, with 2K sentence-pairs for tuning and 15K per gender for testing.Additional bilingual training data comes from unannotated Europarl or combined TED male and female data.Source: §4.2 · p. 5

Classification setup

• Classification uses function words and frequent POS trigrams with linear-kernel SVMs and 10-fold cross-validation to emphasize stylistic and syntactic characteristics.Texts are divided into approximately 1,000-token chunks, and lexical-feature values are normalized by token count.Source: §4.3 · p. 5

MT setup

• Phrase-based SMT systems are trained with Moses, 5-gram Kneser–Ney language models, and MERT tuning after sentence cleaning, tokenization, and punctuation normalization.The preprocessing removes empty, long, and misaligned sentences before modeling.Source: §4.4 · p. 5

Model objective

• Gender-aware SMT models improve preservation of gender traits while maintaining general translation quality.The models are designed as simple domain-adaptation techniques for personalized translation.Source: §5 · p. 5

Personalization strategy

• Gender personalization is implemented with gender-specific phrase tables, language models, or tuning sets, compared against a baseline that disregards gender.All configurations use a single reordering table trained on the complete training set.Source: §5 · p. 6

Baseline

• The baseline MT-B system trains one translation model and one language model on mixed gender-specific and unannotated data without distinguishing gender.Its tuning set is a random 2K-sentence sample preserving the underlying gender distribution.Source: §5 · p. 6

Personalized models

• MT-P1 builds three phrase tables and three language models for male, female, and additional in-domain bilingual data.Each genderized model variant is tuned with its corresponding gender-specific tuning set.Source: §5 · p. 6
• MT-P2 combines gender-specific phrase tables and language models with general phrase-table and language-model components.Both genderized variants are tuned using gender-specific data, while translation quality is evaluated separately.Source: §5 · p. 6

MT evaluation

• Baseline and personalized systems had comparable general translation quality, with BLEU differences generally small across language pairs and datasets.For example, en-fr and fr-en baselines scored 38.65 and 37.35, while personalized models scored 38.42, 38.34 and 37.16, 37.16.Source: §6 · p. 6

Classification accuracy

• English, French, and German originals yielded 77.3%, 81.4%, and 76.1% gender-classification accuracy, respectively.These results establish a strong gender signal in original texts before translation.Source: §6 · p. 6, Table 4 · p. 6, Table 5 · p. 7
• Human and non-personalized machine translation significantly obfuscated gender traits, partly reflecting editing and translator fingerprints in Europarl proceedings.The paper notes that TED’s subtitled and edited format also affects interpretation of human-translation results.Source: §6 · pp. 6–7
• Gender-aware MT improved Europarl gender-classification accuracy by 3.8 to 7.0 percentage points over baseline, from fr-en to de-en.Both personalized models preserved gender traits better than manual and gender-agnostic automatic translations.Source: Table 4 · p. 6, §6 · p. 7, Table 5 · p. 7
• 77.2% and 77.7% accuracy were achieved by TED MT-P1 and MT-P2, significantly outperforming baseline and manual translation.The TED originals reached 80.4%, while human translation and baseline MT generally weakened the gender signal.Source: §6 · p. 7, Table 6 · p. 7

Analysis of gender markers

• Gender markers are carried into manual and machine translations strongly enough to overshadow the target language’s original markers.InfoGain weights reveal source-language marker persistence in translated English, including French je, dire, doit, and doivent reflected in English choices.Source: §7 · p. 7, Figure 2 · p. 8
• Translated texts show a hybrid gender outcome dominated by source-language tendencies, with target-language markers contributing to a lesser extent.The analysis finds that translated English markers can mirror corresponding French or German gender markers rather than the original English tendencies.Source: §7 · pp. 7–8
• Gender traits remain weaker in translations than in originals, despite the persistence of source-language tendencies.The paper characterizes translation as combining weakened original signals with source- and target-language marker effects.Source: §7 · pp. 7–8

Capturing the “personalization” effect

• Personalized SMT can select gender-associated alternatives, such as English exactly over really or think over consider and believe.These examples come from sentences where personalized translations were correctly classified while baseline translations were misclassified.Source: §7 · p. 8

8 Conclusions

• Personalized SMT models show preliminary promise for preserving gender traits in automatic translation.The paper frames these results as an initial step toward personalized machine translation.Source: §8 · p. 8
• This approach could facilitate user modeling in other languages and augment English data with translated content.Source: §8 · pp. 8–9
• Future work will test more sophisticated MT models, additional demographic traits, domains, and language pairs.Source: §8 · p. 9