Work

MorphoFiltered-Gemini at MWE-2026 PARSEME 2.0 Subtask 1: Tackling LLM Overgeneration via Universal POS-based Constraints

Year: 2026

Type: conference-paper

Abstract: This paper describes MorphoFiltered-Gemini, a system submitted to the PARSEME 2.0 Shared Task (Scholivet et al., 2025), subtask 1 on MWE identification, covering all 17 target languages.The system combines LLM-based predictions generated via the Gemini API with a morphological post-filter designed to reduce false positives.Rather than optimizing peak performance on individual languages, our approach prioritizes cross-lingual stability and precision.As a result, the system exhibits a balanced performance across languages and MWE categories, achieving the highest Shannon evenness score among all submissions.

Source: Workshop on Multiword Expressions (MWE)

DOI: 10.18653/v1/2026.mwe-1.26

Authors: Irina Moise, Sergiu Nisioi

Institution: University of Bucharest

Topic: Natural Language Processing Techniques

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Funder: Unitatea Executiva pentru Finantarea Invatamantului Superior, a Cercetarii, Dezvoltarii si Inovarii

OpenAlex ID: W7140114746

FWCI:

Cites:

Cited by:

Related to: