Evidence-grounded NLP and LLMs

NLP and LLM systems that connect answers, claims and decisions to explicit evidence.

DTALab develops NLP and LLM-based systems whose outputs are grounded in explicit evidence. The goal is to move beyond answer generation and black-box prediction by making language-based AI systems traceable, inspectable and auditable.

This research line focuses on methods for claim verification, question answering, retrieval, report analysis and evidence-grounded decision support. In fact-checking and verification tasks, the aim is not only to classify a claim as supported or refuted, but also to identify which sources support the decision, which sources contradict it, and which information is irrelevant or insufficient.

A specific focus is on the use of Large Language Models for the integrated analysis of textual and tabular data. This includes Table Question Answering, Multi-table Question Answering, extraction of information from complex reports, and evaluation of how models use evidence distributed across documents, tables and semi-structured sources.

The research includes methods and tools for:

  • claim verification over textual, tabular and structured evidence;
  • evidence retrieval and source attribution;
  • explainable fact-checking and evidence-level explanations;
  • RAG and LLM pipelines whose outputs can be inspected and audited;
  • Table QA and Multi-table QA over complex reports;
  • evaluation of how models use external evidence;
  • analysis of unsupported, contradictory or insufficient evidence.

This theme supports applications in information integrity, compliance, education, document intelligence, ESG reporting, financial reporting and decision-support workflows where AI outputs must be justified and validated by domain experts.

Selected publications