GITT 2026

Students who received training (even very short) on gender-fair language made fewer misgendering mistakes for German and French. Fluency is impacted a bit, but there is an obvious learning curve, with fluency improving as the conversation goes on #GITT2026

Switching gears to a project with Laura Robaey on gender fair language in interpreted interactions with service providers - how do student interpreters handle this and how can we prepare them for this kind of language? #GITT2026

People are increasingly familiar with nonbinary pronouns and attitudes are somewhat more positive through time as well, but usage remains stable/low. More accepted for generic use than for nonbinary reference. Introduction seems slower in Dutch and French (frequent masculine generics?) #GITT2026

Compared to the naming task, when asked to describe an employee at company X, gender is mentioned less frequently and gender representations are more mixed / diverse. Compared to men, women and nonbinary people are more likely to mention gender #GITT2026

What happens to mental representation? Green=male, red=female, blue=ambiguous descriptions depending on pronouns used in descriptions. Gender inclusive pronouns lead to more diverse representation for women, but did not eliminate male bias in men #GITT2026

Similar results for inclusive pronouns for reference to unknown referent in employee guidelines: no impact on content question accuracy - somewhat more obvious differences for French but still need to be interpreted with caution: what is the practical implication? #GITT2026

Nonbinary pronouns for nonbinary referent had no impact on accuracy compared to other strategies, but some other differences detected #GITT2026

Pronouns are intriguing as linguistic phenomenon: do they really impact comprehension? Do they actually reduce male bias? To what extent have they been successfully introduced in language? Which factors influence this? Big WEAVE project on small words tries to answer these questions!

Gender neutral pronouns are introduced into languages in different ways and serve different purposes (specific nonbinary reference, specific reference to unknown referent, generic reference to people in general) #GITT2026

Banger line to wrap up with: "although speech technologies themselves are not inclusive, you can use them to make a more inclusive society" #GITT2026

Crowdfunding now live for ASR for live captioning for people with a speech impairment at www.supporttudelft.nl/project/veel...

Dysarthric speech remains challenging for humans and systems, unless we add data augmentation - fine tuned systems go down to 20-30 WER #GITT2026

But what about human listeners? Brand new work (ongoing) shows that word error rate is really high for human listeners as well 👀👀 Google Telephony outperformed humans on diverse speech #GITT2026

Different types of data augmentation in practice #GITT2026 massive improvements (word error rate reductions) are possible for both native and non-native speech.

Different types of diversity need different types of data augmentation - these findings didn't make reviewers happy but have been verified repeatedly! #GITT2026

Possible solutions: data augmentation (SpecAugment, speed, volume, pitch changes, voice conversion) or different training strategies (fine-tuning, domain-adversarial training, multi-task learning)

Architectures and language matter - can lead to very different bias patterns #GITT2026 some similarities and differences across languages "bias is there, but we don't really know what the origin is"

Can we locate origin of bias so we can mitigate it? Study with Dutch and Flemish speech and with Mandarin. Trained on CGN, tested on Jasmin, some differences between Dutch and Flemish. taalmaterialen.ivdnt.org/download/tst...

We need actually inclusive speech technology. Biggest issues = data, bias against different speaker groups, lack of definition of fairness / bias / inclusiveness #GITT2026

Speech technology only works for some of the 7000+ languages in the world and only for standard speech, whereas a lot of speech is actually diverse (-18, +60, speech impediments, different accents, etc.) - can impact anything from healthcare to education to accessibility #GITT2026

And last but definitely not least, #GITT2026 's very own @bsavoldi.bsky.social

@amelija166mp.bsky.social on gender bias in translated product reviews and Deniz Albayrak on WinoTR #GITT2026

Sophie Hennuy presenting the Inclupedie (see www.inclupedie.eu) and Marie Dewulf on a real world corpus for gender bias in the webcare domain #GITT2026

Complexity of the data: not a lot of variety in the emotions to be detected. Focus on 'love' ❤️ - humans more likely to translate 'love' reviews as feminine. LLMs still default to masculine, with the exception of Llama #GITT2026

Does source text emotion function as an implicit gender cue, both in human translation and LLMs (focus on binary gender given the Russian language)? Text = English Amazon reviews, translators and LLM did not receive gender-specific instructions #GITT2026

Consistency and automatic labeling are similar across models. Failure happens e.g. across unit boundaries or for compound nouns. Masculine bias, inefficient strategies and incoherence issues remain #GITT2026

Reasoning models seem to perform better than non-reasoning ones (or at least "fail less often" 😉) #GITT2026

Reasoning helps when going from binary input to nonbinary output, nonbinary input leads to more chaos (but reasoning can mitigate some unwanted effects)

LegISTyr test set focuses on legal terminology and tests the impact of strategy used in the source (binary or non-binary inclusive) on target rendering in LLM workflow (with and without reasoning) #GITT2026

Another call for human intervention: queer misrepresentation happens by different kinds of technology, albeit in different ways #GITT2026

LLM mostly leads to misgendering (only into feminine!) and some intriguing intralingual and interlingual calques or hybrid morpholgical hallucinations

NMT usually leads to pro-drop or misgendering by binary translations (also this presentation has a lot of brilliant memes but typing and taking pictures at the same time is hard - you'll have to join us in the room next edition for the full experience 😉) #GITT2026

Queer novel translation (Schwindel) with 3 different strategies, followed by annotation of bias and misgendering, translation failures, and non-binary translation strategies

LLM and WinoMT have similar performance with different weaknesses, aka human annotation and validation remains essential #GITT2026

Results look okay, until you look at the 'unknown' categories: alignment shift, tagging errors, translation variability for WinoMT, forced binary annotation for LLM in linguistic indeterminacy contexts

Process of gender annotation relies on automatic steps in the pipeline which can lead to errors #GITT2026 - to what extent do labels correlate with human annotations and can LLM be used instead?

As well as the perception of the linguistic strategies used for non-binary referent translation by Dutch audiences:

@jdaems.bsky.social and team investigated non-binary pronoun translation in Netflix's series 'Sex Education' #GITT2026