AUEB-Archimedes at RIRAG-2025: Is obligation concatenation really all you need?

TL;DR

This paper describes systems developed for RIRAG-2025 that retrieve and refine regulatory answers using multiple models and a novel metric, achieving high scores but highlighting issues with current evaluation methods.

Contribution

The paper introduces a retrieval and reranking approach combined with a neural component for extracting obligations, and demonstrates the limitations of current metrics in evaluating generated answers.

Findings

High RePASs score (0.947) achieved with obligation extraction from passages.

Refined answers with iterative improvements score lower (0.639) but are more coherent.

Current metrics may overestimate answer quality due to extraction-based scoring.

Abstract

This paper presents the systems we developed for RIRAG-2025, a shared task that requires answering regulatory questions by retrieving relevant passages. The generated answers are evaluated using RePASs, a reference-free and model-based metric. Our systems use a combination of three retrieval models and a reranker. We show that by exploiting a neural component of RePASs that extracts important sentences ('obligations') from the retrieved passages, we achieve a dubiously high score (0.947), even though the answers are directly extracted from the retrieved passages and are not actually generated answers. We then show that by selecting the answer with the best RePASs among a few generated alternatives and then iteratively refining this answer by reducing contradictions and covering more obligations, we can generate readable, coherent answers that achieve a more plausible and relatively high score (0.639).