This paper describes our system used for a shared task on code-mixed, less-resourced sentiment analysis for Indo-Aryan languages. We are using the large language models (LLMs) since they have demonstrated excellent performance on classification tasks. In our participation in all tracks, we use unsloth/mistral-7b-bnb-4bit LLM for the task of code-mixed sentiment analysis. For track 1, we used a simple fine-tuning strategy on PLMs by combining data from multiple phases. Our trained systems secured first place in four phases out of five. In addition, we present the results achieved using several PLMs for each language.

This paper introduces Cro-FiReDa, a sentiment-annotated dataset for Croatian in the domain of movie reviews. The dataset, which contains over 10,000 sentences, has been annotated at the sentence level. In addition to presentingthe overall annotation process, we also present benchmark results based on the transformer-based fine-tuning approach.

The paper presents a procedure for generating prefixed verbs in Croatian comprising combinations of one, two or three prefixes. The result of this generation process is a pool of derivationally valid prefixed verbs, although not necessarily occuring in corpora. The statistics of occurences of generated verbs in Croatian National Corpus has been calculated. Further usage of such language resource with generated potential verbs is also suggested, namely, enrichment of Croatian Morphological Lexicon, Croatian Wordnet and CROVALLEX.