Text Technology/Digital Linguistics Colloquium HS 2026

Time & Location: every 2 weeks on Tuesdays 10:15-12:00 in room BIM 4-033.

Online participation via the MS Teams Team CL Colloquium is also possible.

Responsible: Marius Huber

15 Sep 2026

Mark Fišel: Modelling Extremely Low-resource Finno-Ugric Languages and Dialects

I will describe an on-going effort on developing NLP tools for the Finno-Ugric family, the members of which range from languages with 1M+ speakers and "an army and fleet" to critically endangered dialects and languages with few resources and even fewer speakers. The latest development direction in this work is on multi-tasking models and general-purpose LMs, which is, however, hindered by the availability of data, low domain coverage, diverse quality in data from different sources and variety in orthography. I will present the already developed models and performed experiments, and will be happy to discuss open questions.

Jana Massoud & Sina Ahmadi: Towards a quantitative language resource index

The term "low-resource" in NLP is too vague to be useful. This talk presents our project that aims to move towards a quantitative resource index that scores 8,168 languages across six key dimensions (from corpora and models to dictionaries). Our analysis reveals a heavily skewed landscape where 75% of the world's languages are entirely absent from contemporary NLP. We validate this index against downstream tasks, offering a new, fine-grained tool to accurately situate languages in the tech ecosystem and identify where resources are needed most.

29 Sep 2026

Yiyang Chen: Towards Automatic, Label-Free Expressive TTS

 How can a text-to-speech (TTS) system infer an appropriate emotional expression from text and use it to guide speech synthesis? This talk explores automatic emotion prediction and control in zero-shot TTS. I will introduce our progression from selecting emotional speech references using predicted valence, arousal, and dominance (VAD) to EmoPilot, which predicts emotion cues directly in the TTS backend’s learned emotion space. These cues support reference-audio retrieval or direct conditioning, bypassing explicit emotion-category and VAD predictions. I will then discuss our next steps towards context-aware emotion prediction: the same sentence can convey different emotions depending on its surrounding context. Our upcoming Romansh long-form audio generation project provides a motivating setting, with training speech limited in both quantity and emotional expressiveness. Can emotional expression learned from speech in other languages transfer to Romansh synthesis? Reference and synthesized audio examples will illustrate the research questions and provide a starting point for discussing future directions and applications.

Tanja Samardžić: Controlled token allocation in optimising multilingual vocabularies

Multilingual (pre-)training of large language models is necessary but it is prone to various biases due to the uneven data availability and structural differences between languages. This problem becomes apparent already at the level of tokenisation: if no care is taken, a multilingual vocabulary (constructed by means of BPE) ends up dominated by items representing high-resource languages. To prevent this, various controlling techniques have been proposed, from upsampling tokens from low-resource languages to minimising token sharing across languages and setting cross-lingual parity as one of the BPE's objectives. In this talk, I will present experiments testing potential advantages of several new techniques aimed at designing better multilingual vocabularies. Our preliminary findings suggest that controlled token allocation is more beneficial with larger vocabularies allowing systematic improvements in model-intrinsic metrics (bits-per-byte, information parity) and also in some downstream tasks (belebele).

13 Oct 2026

Anna Bondar: Mixture-of-Readers (MoRe): Enabling Scanpath Generation for Different Reader Groups

Eye movements during reading constitute a rich behavioural signal of moment-to-moment cognitive processing. As such, they have long served as a central dependent variable in psycholinguistic research, and, recently, have been leveraged for cognitively enhancing language models. However, eye-tracking data are scarce, which has motivated the development of deep learning models of eye movements during reading. Previous work has shown that these models benefit from training on multiple reader groups, partly because this increases the amount of training data. A central limitation of previous approaches, however, is that, despite being trained on heterogeneous data, the models generate overly homogeneous scanpaths and fail to capture group-specific reading patterns. Preserving such group-level differences is important for both psycholinguistic and NLP research. In this paper, we introduce Mixture-of-Readers (MoRe), a model based on a mixture-of-experts architecture for generating group-specific scanpaths. We experiment with different routing strategies and evaluate our approach by predicting scanpaths for dyslexic and neurotypical L1, as well as L2 readers across English and Danish languages. We evaluate the models in terms of predictive performance and their ability to preserve scanpath diversity between and within reader groups. We further analyse the expert selection patterns across the reader groups.

Michelle Wastl: Fine-grained Text Comparison: Scaling Word Alignment to Documents and Benchmarking Token Representations

Tasks such as the recognition of semantic differences (RSD), hallucination detection, or label projection require systems to compare two texts at the level of individual tokens. For such tasks, we have found simple word alignment based on embedding similarity to outperform LLM prompting at a fraction of the computational cost.
In the first part of this talk, I will present CTFAlign, a coarse-to-fine approach that extends unsupervised embedding-based word alignment from sentences to full document pairs without relying on sentence segmentation. I will also show how it transfers to several fine-grained text comparison tasks.
How well such methods work depends on the choice of encoder, which is usually guided by text embedding benchmarks such as MTEB. These benchmarks, however, evaluate pooled text representations and therefore say little about the quality of individual token representations, which is what fine-grained comparison relies on. Hence, in the second part of the talk, I will present an ongoing project that brings existing fine-grained text comparison datasets together into a single benchmark for token-level representations. 

27 Oct 2026

Cui Ding: From Population-Level Effects to Individual Differences: Testing Lexical Processing in E-Z Reader and SWIFT

Cognitive models of reading explain how lexical processing, attention, and oculomotor control shape eye movement behavior. Two leading models embody different theories of reading: E-Z Reader assumes serial attention allocation and staged lexical processing that controls saccade programming, whereas SWIFT assumes parallel lexical processing within a distributed attentional gradient and an autonomous saccade timer. These differences should affect both aggregate reading measures—fixation durations, skipping, and regressions—and the way linguistic effects of the stimulus unfold over time and across words.

Prior work. Existing comparisons of E-Z Reader and SWIFT are mostly qualitative or focus on aggregate surface-level reading measures for an “average” reader. Recent studies that use both models treat them mainly as baselines for evaluating deep neural generative models of eye movements. It remains unclear whether E-Z Reader and SWIFT capture how linguistic predictors shape reading: local effects, spillover, temporal effect distribution, and individual-level predictor effects.

Current study. We test E-Z Reader and SWIFT using spillover and individual differences in linguistic effects as diagnostics of reading mechanisms. We ask whether the models simulate population-level effects of word length, lexical frequency, and surprisal; how these effects spill over to the subsequent word; and how these effects vary across readers.

Jan Brasser: Predicting Comprehension Accuracy from Eye Movements on Critical Information Spans

Eye movements during reading offer a window into how well a text has been understood.
However, previous approaches to predicting comprehension from eye movements either aggregate them over whole texts or rely on short, curated stimuli.
This leaves open which aspects of reading behaviour matter for which kind of understanding. Using the German MultiplEYE eye-tracking-while-reading data, we focus on the critical text spans that contain the information needed to answer comprehension questions. We ask which eye-movement measures on these spans predict whether a question is answered correctly, both individually and in combination. We also ask whether this depends on the level of comprehension a question targets: explicitly stated information, bridging inferences, or an evaluation of the text as a whole. Our results suggest that eye movements mainly reflect comprehension that requires integrating information beyond what the text explicitly states. They also suggest that measures that predict well on their own are not necessarily the most useful ones in combination.

10 Nov 2026

Bastian Rieck

TBA

Jason Armitage: Multimodal Accessibility Planning with AI Models (MAP)

MAP is a novel benchmark for evaluating multimodal AI systems as assistants for users with accessibility requirements when planning visits to places in the real world. MAP introduces two novel assessments that evaluate systems on the factual accuracy of the information and the ability to present visual evidence that addresses a user request. Claim verification consists of two question types: feature-check prompts query whether a named place has the requested accessibility features; search-by-feature prompts ask systems to find places of a given type in a city area that meet the requested accessibility requirements. Visual evidence retrieval for accessibility planning checks if a multimodal AI system can select visual evidence for the requested place and accessibility feature. The methodology supports comparison of AI systems in a live setting where places change from week to week by refreshing samples and performing assessments within specified windows of time. In this presentation, a review is provided of the benchmark design, construction of the ground truth, and initial results. A human assessment of selected cases highlights challenges in scoring these tasks and indicates where state-of-the-art systems go wrong.

24 Nov 2026

Hanxu Hu

TBA

Andrianos Michail

TBA

8 Dec 2026

Longqian Ming

TBA

Gerard Sant Muniesa

TBA