This project will be a collaboration between Language Technology Group and "Ønskevenn" startup registered with Akershus universitetssykehus (Ahus). It can be a group project with up to 2 students working jointly.
It is aimed at building a constrained, developmentally adapted conversational support system for Norwegian speakers. Its target audience are children with neurodevelopmental disorders and related developmental and learning difficulties, including autism spectrum conditions, ADHD, developmental language disorder (DLD), intellectual disability, learning difficulties, and combinations of these conditions. A tentative name for this system is "Vinni". It usually initiates interaction with a child by asking a question or giving a brief request.
In NLP terms, Vinni is a large language model (and some harness around it) which is intentionally constrained to generate linguistically appropriate dialogue calibrated to specific developmental language levels. Target users often struggle with syntactic complexity, abstraction level, and pragmatic inference, hence the emphasis on short, bounded and professionally controlled micro-dialogues. Examples of developmentally appropriate language can include:
- short sentences
- familiar and concrete vocabulary
- limited syntactic complexity
- limited abstract language
- avoidance of irony, sarcasm and unnecessary ambiguity
- one main idea or instruction at a time
- concrete examples when needed
Logistics
The initial work will be carried out using simulated child profiles and simulated dialogue scenarios. Real child data may become available from autumn 2027, depending on the progress of the project and the necessary privacy, ethical and institutional approvals.
Generation of simulated dialogues is a part of this Master project. Ønskevenn/Ahus will provide guidelines on the variation in these dialogues, and (importantly) manual annotation of the LLM responses into "acceptable/not acceptable", based on clinical and developmental criteria.
Because of sensitive nature of the target domain, only open-weights (and preferably open training data) LLMs will be considered. Another constraint is that the models must be well-performing in Norwegian. A specific set of models to experiment with will be determined in consultation with the LTG and Ønskevenn supervisors. It is expected that the project will include experiments with:
- baseline models
- models with persistent system prompts or guardrails on top, constraining dialogues so that they are predictable, developmentally supportive, and safely bounded
- adapted or fine-tuned baseline models
References
Constrained generation is a well-developed NLP field. Below we are providing a glimpse into relevant publications:
- NeuroLogic A*esque Decoding: Constrained Text Generation with Lookahead Heuristics (Lu et al., NAACL 2022)
- CoDa: Constrained Generation based Data Augmentation for Low-Resource NLP (Evuru et al., Findings 2024)
- Small LLMs for Medical NLP: A Systematic Analysis of Few-Shot, Constraint Decoding, Fine-Tuning and Continual Pre-Training in Italian (Ferrazzi et al., LREC 2026)