Research

Artificial intelligence, from the lab to the people who use it

We are curious about how machines find, read and reason with information, and we want what we build to help people. That curiosity runs from search engines and language models to logic and evolutionary computation, and it takes us out of the lab: spotting early signs of depression online, measuring hate speech on social networks, answering health questions with sources you can check. Below, each line introduces its people, the questions keeping them busy and their latest papers.

“early signs of depression in social media”

Early risk detection and digital mental health

  • Early risk detection
  • Mental health
  • Depression
  • NLP for health
  • LLM personas
  • eRisk

We build models and test collections to detect early signs of depression, ADHD and other risks from what people write online, and we study how language models can reason about symptoms in a clinically reliable way. Since 2017 the group co-organises eRisk, the CLEF lab on early risk prediction on the Internet.

Questions we are working on

  • Can a system rank evidence for each DSM-5 symptom, sentence by sentence?
  • Can LLM-based personas stand in for patients when evaluating conversational screening?
  • How do we make LLM reasoning about depression consistent enough for clinicians?

Recent papers

“hate speech detection in Galician and Spanish”

Social network analysis: hate speech, bias and online harms

  • Hate speech detection
  • Online harms
  • Bias in LLMs
  • Content moderation
  • Conspiracy narratives
  • Computational social science

We study what circulates on social networks and how models read it. We assemble large meta-collections of hateful and non-hateful posts, build datasets of conspiracy and hyperpartisan narratives, study how language models annotate and judge this content, and measure the demographic and geographic biases they carry. Part of this work runs inside the HYBRIDS doctoral network on democracy and online discourse.

Questions we are working on

  • Can an LLM evaluate hate speech it cannot reliably annotate?
  • Does personalising a classifier to a region make it fairer or more prejudiced?
  • How do we cover low-resource languages such as Galician?

Recent papers

“can LLMs replace human relevance judgments”

Search evaluation and LLMs as assessors

  • IR evaluation
  • Test collections
  • LLM relevance judgments
  • Pooling
  • Relevance feedback
  • Reproducibility

Test collections are expensive to build. We design pooling strategies, statistical tests and LLM-assisted judging that keep evaluation reliable while cutting the cost of human assessment, and we check where automatic judgments break down.

Questions we are working on

  • Which documents should humans judge, and which can an LLM judge?
  • How do we compare many systems at once without false discoveries?

Recent papers

“trustworthy answers to health questions”

Search experience, RAG and health information access

  • Retrieval-augmented generation
  • Health information access
  • Consumer health search
  • Query suggestion
  • User simulation
  • Faithfulness

From query suggestions and keyphrase search to retrieval-augmented generation, we study how people search and how to give them answers that are faithful to the evidence. In the XMED and C3HS projects we look at health searches by ordinary people, where a wrong answer can do real damage.

Questions we are working on

  • Does the position of evidence in the context change what a RAG system answers?
  • How do we evaluate query suggestions over a whole session, not just the top results?

Recent papers

“answer set programming with temporal logic”

Knowledge representation, explainable reasoning and clinical AI

  • Explainable AI
  • Answer set programming
  • Logic programming
  • Temporal reasoning
  • Clinical decision support
  • Precision dosing

We develop logical foundations for AI systems that can explain their conclusions: Answer Set Programming, temporal and deontic logic, and their use to generate explanations in hybrid systems that combine logic and learning, the focus of the GEISER project. The same concern for explanations takes us to the clinic: with the pharmacy service of the A Coruña University Hospital we build interpretable models to personalise drug dosing after transplantation, where a recommendation is only useful if the pharmacist can see why it was made.

Questions we are working on

  • When are two logic programs with constraints strongly equivalent?
  • How do we represent obligations and permissions that may conflict?
  • Can interpretable AI personalise tacrolimus doses across hospitals?

Recent papers

“genetic programming interpretable prediction models”

Bio-inspired computing and evolutionary learning

  • Genetic programming
  • Evolutionary computation
  • Memetic algorithms
  • Interpretable machine learning
  • Protein structure prediction
  • Differential evolution

Nature solves hard search problems by variation and selection, and we use the same idea to build models. With genetic programming, differential evolution and memetic algorithms we search spaces far too large to explore one by one: the shapes a protein can fold into, the combinations of financial ratios that anticipate a company failure, or the latent space of a generative model. What genetic programming evolves is a readable formula rather than a black box, so anyone can check why it predicts what it predicts, the same question we ask of logic-based systems in the GEISER project.

Questions we are working on

  • Can genetic programming select the variables and build a readable classifier at the same time?
  • Can evolutionary search refine protein structures predicted by deep learning?
  • Does clustering the latent space of a variational autoencoder give better generated images?

Recent papers