Fairslator is a web-based tool for detecting and correcting bias in machine translation. Started in 2022 as a personal project, Fairslator has recently (2026) received backing from University of Vienna where it is going to be redeveloped into an open-source, community-contributed tool for rewriting and computer-assisted postediting of machine translation. This contribution introduces the plan for that redevelopment.

The increased representation of non-binary characters in audiovisual media is societally crucial, yet non-binary language (e.g., pronouns) offers new challenges for translators. In Dutch, the acceptability of pronoun strategies for non-binary reference is debated and evolving. Machine translation (MT) is increasingly being used in audiovisual translation, yet its potential for translation in gender-sensitive contexts is understudied. In this paper, we compare three pronoun translation strategies (two human-produced, one MT) for rendering the English non-binary pronoun they into Dutch in an audiovisual context, using a fragment from the Netflix show Sex Education. We explore general acceptability of these strategies and compare this to audience perceptions after viewing subtitled fragments employing a specific strategy. The results support findings from earlier work on non-binary pronouns in Dutch and indicate that the MT-generated translations were perceived as the least suitable.

The assessment of gender bias in Machine Translation critically depends on reliable methods for identifying grammatical gender in system outputs. In this paper, we compare the automated WinoMT annotation pipeline, based on word alignment and morphological tagging, with an instruction-tuned LLM (Qwen3-8B) used to annotate grammatical gender in English–Italian translations. Both approaches achieve similar levels of agreement with a human-annotated gold standard, but exhibit distinct systematic weaknesses. The WinoMT pipeline is sensitive to alignment shifts and morphological tagging limitations, often resulting in indeterminate gender labels. The LLM, in contrast, tends to favour binary gender labels even when noun phrases are morphologically gender-invariant. A qualitative analysis of model outputs and reasoning traces further suggests a lack of stable generalization of grammatical gender rules, even when illustrative exemplars are provided for in-context learning. This highlights important methodological limitations in using LLMs as objective annotators for gender evaluation tasks.

This paper evaluates how Neural Machine Translation (NMT) and Large Language Models (LLMs) process non-binary morphology when translating German literary fiction into Italian. We apply an inductive, mixed-methods framework to analyze 12 NMT and 15 LLM translations from a human-in-the-loop experiment. Results reveal a fundamental divergence. NMT defaults to standard binary grammar but applies it inconsistently, often flipping between masculine and feminine forms for the same subject across different sentences, which effectively erases queer visibility. Conversely, LLMs actively attempt gender-fair language via neutralization and neomorphemes (e.g., the schwa). However, LLMs introduce new systematic errors: driven by semantic cues, they exhibit a contextual bias that, in the present study, frequently led to over-feminization, and their attempts to create inclusive word endings result in structurally invalid words. Ultimately, these findings expose current limitations and provide preliminary empirical guidance to assist post-editors in navigating the complex challenges of gender-fair translation.

This paper investigates how gender-fair strategies in source texts influence the production of non-binary translations in the Italian to German combination. We make use of a controlled test set featuring both binary and non-binary approaches to assess their effectiveness for non-binary renderings in the target language. We also introduce an automatic evaluation framework that classifies target sentences into four categories: non-binary, binary-gendered, single-gendered, and incoherent. Relying on human annotation and analysis, we compare Reasoning LLMs against standard inference, examining whether reasoning improves translation quality and automatic evaluation. Our results show that reasoning models are more successful in shifting from binary to non-binary formulations and in handling linguistic challenges such as epicene terms and special characters, although there are no improvements in sentence-level consistency and evaluation accuracy. A qualitative analysis of German translations shows that reasoning encourages the reformulation of source-side strategies through neutralization, visibility strategies, and paraphrasing, resulting in more natural target texts.

This study investigates whether source-text emotion is associated with grammatical gender choices in English-Russian translation when the author’s gender is not specified. Building on Popović and Lapshinova-Koltunski (2024), we analyse first-person grammatical gender in Amazon review translations produced by professional translators, translation students, and three LLMs: GPT-4, Llama, and Mistral. We assign emotion labels to the English source reviews and test whether these labels are associated with masculine or feminine forms in the Russian translations. Reviews labelled as expressing love are more likely to be translated with feminine grammatical gender by both human translator groups, with small-to-moderate effect sizes. This pattern is absent in GPT-4 and Mistral, but appears in Llama. Other frequent emotion labels do not show the same positive association. We therefore treat the findings as exploratory evidence that specific source-text emotions may be associated with gendered translation choices, while emphasizing the need for larger-scale validation across additional emotions, genres, and language pairs.

Language technologies used in everyday settings such as machine translation systems risk perpetuating societal bias. As previous work shows, biases in these systems not only underscore representational harm but also materialize into economic disparities in resources required to correct errors for the disadvantaged social group. Prior work in creating benchmarks for gender bias in machine translation systems 1) focus primarily on high-resourced language pairs or a low-resourced language paired with a high resource language, 2) use template based benchmarks that usually focus on occupational biases and stereotypes, and 3) translate high-resource benchmarks which may lack cultural significance to low-resourced languages. In this paper, we introduce Yeswa-Stories, a three-way parallel dataset comprising 1,300 aligned sentences in Amharic, Afaan Oromo, and Tigrinya. The dataset focuses on narratives about women and is designed to support research on gender representation in translation. We constructed the dataset in two ways: first, we collected English sentences from Wikipedia articles about notable African women and translated them into the three target languages using human translators. To improve cultural representativeness, we further augment the dataset with locally sourced content reflecting the cultural context where the languages are spoken. Our dataset contributes a new resource for studying gender-inclusive translation in low-resourced settings.

Gender bias in automated translation (AT) is well-documented for English-source language pairs, while source languages generally lacking grammatical gender remain largely understudied. This paper addresses the Chinese-Portuguese direction, a language pair that has received little attention in this context. We developed a Chinese challenge set of 495 sentences constructed from 45 occupations across an eleven-sentence template matrix, systematically varying the type and syntactic position of gender cues: no cue, explicit prenominal modifiers, and coreferential pronouns varying in position and syntactic complexity. We tested two commercial neural machine translation (NMT) systems and six large language models (LLMs) with this challenge set. Results show a clear hierarchy of cue effectiveness: explicit prenominal modifiers yield universal 100% accuracy; in the absence of gender cues, models predominantly default to masculine forms; and coreferential pronouns in complex sentences reveal a pronounced masculine-feminine asymmetry. Also, LLMs demonstrate more symmetric gender cue processing than NMT systems. Our data and code are available at https://github.com/xu-xiaolan/chinese-challenge-set-gender-bias.

We present WinoTR, a Turkish adaptation of the WinoMT challenge dataset (Stanovsky et al., 2019). While WinoMT has been widely studied across multiple languages, its adaptation to Turkish — a morphologically rich language with no grammatical gender — and its analysis through a causal lens remain unexplored. Using 4,752 sentences adapted from the original dataset across pro-stereotypical, anti-stereotypical, and neutral conditions, we apply Double Machine Learning (DML) to estimate the causal effect of gender cues on stereotype-consistent translation output. Our results reveal a striking asymmetry: cue direction has a large and statistically significant effect on translation outcomes, while cue presence alone produces virtually no effect. Even without any gender signal, MT systems default to stereotype-consistent translations in 62.9% of cases across three systems (DeepL, Google Translate, OpenAI). By leveraging this typological property, our causal analysis reveals that gender bias in contemporary MT and LLM-based translation systems runs deeper than surface-level cue processing, persisting as an embedded prior independent of any explicit gender signal in the input.