Recognition Is Not Grounding in Multilingual Financial AI

Introduction

Cross-lingual financial analysis involves more than mapping a phrase to an accounting concept. A system must also recover every relevant amount, percentage, period, and other numerical fact that expresses that concept. A model may correctly understand that a passage concerns revenue or liabilities, yet extract only one of several figures needed to support a comparison.

The paper Recognition Is Not Grounding: Benchmarking Multilingual Financial Evidence Extraction examines this failure mode through a new benchmark called MULTIFINGROUND. Its central question is whether concept recognition and evidence recovery succeed on the same piece of financial text.

Key points

• A multilingual disclosure set. The benchmark is built from FY2025 reports by 90 issuers, covering English, French, and Greek. Independently authored reports are aligned through a shared accounting taxonomy rather than treated as simple translations.
• Fact-level traceability. Annotations connect accounting concepts to the specific numerical facts that express them. This makes it possible to check not only whether a label is correct, but also whether the supporting evidence is complete.
• Several complementary tasks. MULTIFINGROUND measures financial fact identification, concept recognition, exhaustive fact selection, and concept discrimination without giving the model candidate labels.
• A large gap after recognition. Across 15 evaluated models, eight API and open-weight reference models reached 97.3% to 99.4% concept-recognition accuracy. However, among correctly recognized groups, 26.6% to 41.9% still missed at least one target fact.
• Specialization is not a guarantee. The analysis reports meaningful variation in concept discrimination, while the evaluated finance-specialized models offered limited gains over the broader reference models.

Why it matters

The results expose a weakness in common financial NLP evaluations. If a benchmark checks only whether a model assigns the right concept, it can overestimate performance in investment research, reporting, audit, or regulatory workflows. Omitting one comparative period, amount, or related figure can alter a ratio, distort a year-over-year interpretation, or weaken a risk assessment.

A stronger evaluation should therefore combine semantic recognition with exhaustive, source-traceable evidence recovery. For model builders, this points toward better joint handling of tables, paragraphs, and facts distributed across sentences, as well as explicit omission checks. For users, a plausible number should not be treated as a complete answer unless its source and coverage are clear.

MULTIFINGROUND is useful because it turns a hidden failure into a measurable one. A model may understand what a disclosure is about while failing to collect all the material needed to substantiate that understanding. In financial applications, that distinction is not cosmetic; it is part of whether an answer can be trusted.

Source: Hugging Face Daily Papers