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

Summary

The paper describes MorphoFiltered-Gemini, a system entered in subtask 1 (MWE identification) of the PARSEME 2.0 shared task. It covers all 17 languages, including Ukrainian, and all syntactic MWE types rather than only verbal ones. The authors address a known weakness of LLMs: they tag too many word combinations as multiword expressions.

Their pipeline works as follows:

• CUPT annotations are converted to BIO tags.
• Gemini 2.0 Flash-Lite is prompted in batches, with separate prompts for span detection and for category labelling.
• Few-shot examples from training data are used only for some languages.
• A lightweight post-filter based on UD part-of-speech patterns discards implausible spans, such as DET+NOUN or ADJ+NOUN, and removes single-token predictions.

Filters are enabled per language, depending on whether they improved development results.

The system did not win on F1. However, it ranked third in token-based precision and had the most even performance across languages and categories of all submissions. This suggests that minimal, language-agnostic morphological constraints can make LLM-based MWE tagging more stable, which is useful for low-resource or typologically diverse languages.

Key points

• Official scores: MWE-based P/R/F1 was 20.95/14.50/17.14, and token-based P/R/F1 was 34.14/24.20/28.32. The system ranked third in token-based precision and had the top Shannon evenness score (0.97; entropy 3.71, richness 56.53).
• Filter coverage: The POS-pattern filter was used for 11 languages, including UK. It was disabled for Hebrew, Japanese and Slovene, where it hurt results. The single-token filter was kept for EL, FA, GRC, KA and LV.
• Baselines tried: A purely rule-based morphological baseline reached only 0.152 token F1. Using Gemini merely to validate those rule-based candidates brought almost no gain, because recall was too low to begin with.
• Seen vs. unseen MWEs: Expressions seen in train/dev data scored F1 26.25 (P 75.06), while unseen ones scored only F1 8.89. This points to strong reliance on caching and memorization.
• Main limitation: The BIO representation cannot encode discontinuous or overlapping MWEs, so discontinuous expressions scored zero. Remaining errors include missed split light-verb constructions and false positives on compositional patterns, such as Portuguese contractions.