Lattice @MultiGEC-2025: A Spitful Multilingual Language Error Correction System Using LLaMA

Abstract

This paper reports on our submission to the NLP4CALL shared task on Multilingual Grammatical Error Correction (MultiGEC-2025) (Masciolini et al., 2025). We developed two approaches: fine-tuning a large language model, LLaMA 3.0 (8B), for each MultiGEC corpus, and a pipeline based on the encoderbased language model XLM-RoBERTa. During development, the first method significantly outperformed the second, except for languages that are poorly supported by LLaMA 3.0 and have limited MultiGEC training data. Therefore, our official results for the shared task were produced using the neural network system for Slovenian, while fine-tuned LLaMA models were used for the eleven other languages. In this paper, we first introduce the shared task and its data. Next, we present our two approaches, as well as a method to detect cycles in the LLaMA output. We also discuss a number of hurdles encountered while working on the shared task.