NLP@U period-2 meeting

The 2025-2026 period-2 meeting of NLP@U will be hosted by the Faculty of Science.

There will be presentations by Jelle Zuidema and Hosein Mohebbi (titles and abstracts below). After the talks, there will be a round of short pitches (see information below), and lunch. In addition, there will be a coffee break and plenty of time to network and chat informally.

Invited talks

Jelle Zuidema - Opening the blackbox of state-of-the-art language and speech models

I will review approaches for opening the blackbox of large language models, self-supervised speech models and related deep learning technology. I will illustrate techniques such as probing, lensing, interchange interventions and sparse autoencoders, mostly drawing on our own work on the Wav2Vec2 en Hubert models (De Heer Kloots & Zuidema, 2024; Pouw et al. 2025; de Heer Kloots et al., 2025). I will end with some reflections on the limits of posthoc interpretability and the need for moving beyond posthoc interpretability to making interpretability part of the model design before training (as with concept bottleneck and other bottleneck models).

* Heer Kloots, M. de, Zuidema, W. Human-like Linguistic Biases in Neural Speech Models: Phonetic Categorization and Phonotactic Constraints in Wav2Vec2.0. Interspeech, 2024, pp. 4593-4597.
* Pouw, C., de Heer Kloots, M., Alishahi, A., Zuidema, W. (2025), Perception of phonological assimilation by neural speech recognition models. Computational Linguistics, 50(4):1557–1585.
* de Heer Kloots, M., Mohebbi, H., Pouw, C., Shen, G., Zuidema, W., Bentum, M. (2025) What do self-supervised speech models know about Dutch? Analyzing advantages of language-specific pre-training. In Proceedings Interspeech 2025, 256-260.

Hosein Mohebbi - Actionable Interpretability for Speech Models

Actionable interpretability refers to a recently growing focus within interpretability research on translating our insights and understanding from models into something useful and practical. In this talk, I’ll explore this idea in the domain of speech through two case studies, one in depth and the other more briefly. First, I’ll discuss if we can disentangle linguistic content from task-relevant acoustic features in foundation speech models to achieve improved controllability. Second, I’ll discuss if audio frame representations in automatic speech recognition (ASR) models can be adaptively merged to improve model efficiency.

Pitches and discussion

We are inviting submissions for the pitches-and-discussion section. Participants may submit a tentative title and a description. Each selected participant will have two (2) minutes to either:

1. Describe an idea you are working on, and ask if anyone is interested in discussing this idea with you, or
2. Propose a topic for debate immediately afterwards.

Each pitch will be followed by thirteen (13) minutes of discussion.
Please fill out this form to submit your pitch.

Programme