Research
What we work on
Knowledge Modeling & Semantic Reasoning
Knowledge graphs, ontologies and their embeddings — and how to reason over them to answer complex, multi-hop questions.
We study how structured knowledge — knowledge graphs, ontologies and their embeddings — can be combined with neural models. Recent work includes knowledge-graph embedding models for factoid entity questions, generating SPARQL queries from natural-language questions with triple-augmented language models, and adversarial multi-hop QA benchmarks grounded in knowledge graphs.
Text Processing & Retrieval
Neural ranking, learning to rank, query expansion and performance prediction for ad-hoc and entity retrieval.
Our retrieval work covers learning to rank (listwise and self-distilled ranking models), contextual representations for entity retrieval, query expansion, and semantic query performance prediction — bridging classic IR with modern neural encoders.
Biomedical Scholarly Text Retrieval
Automating systematic literature reviews: Boolean query formalization, study screening and structured search with language models.
Systematic literature reviews are slow and expensive. We develop language-model-based tools that formalize Boolean queries, perform structured search, screen studies and benchmark fully automated scholarly search, with a focus on making open, medium-sized models practical for biomedical evidence synthesis.
Responsible Text Neural Models
Measuring and mitigating bias in large language models, and quantifying uncertainty and ambiguity in LM-based question answering.
We investigate how large language models can be made more trustworthy: noise-driven bias mitigation (Schema-Tune), measuring and reducing gender bias in generative models, detecting ambiguous questions via answer diversity, and characterizing epistemic uncertainty and unanswerability in machine reading comprehension.