Vision & Research Focus
My research interest and efforts are focused into the area of Representation Learning:
A crucial requirement for any mission-critical model (in the mathematical sense) is that its governing rules, assumptions and constraints strictly map to the underlying reality it is supposed to represent. In order for AI to reach the maturity level required to enable truly informative discovery in cancer research, it needs to transition from utilitarian, best-effort approaches (i.e., produces the desired results), to formally grounded ones (i.e., works as specified).
I have been involved with Artificial Intelligence (AI) and Natural Language Processing work for over a decade, in which this field has seen several big advancements in technology and practical applications, from the generation and organization of massive textual corpora to multimodal generative models and agentic systems. Now, the application of AI in many areas hangs in the ability to explain the answers it provides, with the analysis of information from healthcare, law and other critical sectors, and of the propagation of misinformation on online social media posing as challenging, but necessary testing grounds for the design of explainable AI, wherein lies my current efforts.
Profile
Danilo Carvalho is the AI Lead for the National Biomarker Centre - Cancer Research UK, University of Manchester, as part of the Digital Cancer Research team, working on Safe and Explainable Artificial Intelligence (AI) for cancer research.
[About]
Current Research
◈ Theoretical research on verification and interpretability of neural AI architectures.
◈ Application of explainable AI systems for translational cancer research.
◈ Applied research in the fields of Natural Language Processing, Knowledge Representation and Computer Vision on the analysis of biomedical data.
Areas of Interest
◈ General
- Artificial Intelligence
- Bioinformatics
- Computational Linguistics / Natural Language Processing
- Data Science
- Software Engineering
◈ Specific (summary)
- Explainable AI
- Representation Learning
- Language Models
- Open Information Extraction
- Patent / Bibliographical Databases
Events
Neuro-Symbolic Natural Language Processing [Tutorial at EMNLP 2025]
Montague semantics and modifier consistency measurement in neural language models [Presentation at COLING 2025]
Formal Semantic Controls over Language Models [Tutorial at LREC-COLING 2024]
Latest Publications
- Where Do LLMs Compose Meaning? A Layerwise Analysis of Compositional Robustness
- Learning to Disentangle Latent Reasoning Rules with Language VAEs: A Systematic Study
- LangVAE and LangSpace: Building and Probing for Language Model VAEs
Featured
Reducing resource demands of control for large language models by over 90%
University of Manchester
Testing AI logic in biomedical research
University of Manchester
Projects
LangVAE & LangSpace: Building and Probing for Language Model VAEs [more]
- Train and run Language Model Variational Autoencoders (LM-VAEs)
- Probing of sentence representations
SylloBio-NLI & SylloBio-NLI: Evaluating large language models on biomedical syllogistic reasoning [more]
- Testing LLMs ability to think logically in biomedical research
SAF-Datasets: Dataset loading and annotation facilities for the Simple Annotation Framework [more]
- Easy access to Annotated NLP Datasets
- Multi-level annotations: document, sentence, token