Argument Mining for Film Reviews

The paper presents the creation of a new dataset for Argument Mining (AM) in Russian, focusing on film reviews. The dataset, a reworked version of the blinoff/kinopoisk collection, consists of 3,280 reviews annotated for stance and argumentative premises. A novel rule-based system is introduced for determining the final argument label, which accounts for the nuanced balance of pros and cons statements in a review. The study evaluates two transformer-based models fine-tuned on the dataset: DeepPavlov/rubert-base-cased and ai-forever/ruRoberta-large. The performance of the models is analyzed in detail, testing hypotheses related to a film’s release period (20th vs. 21st century), its rating position (top-250 vs. bottom-100), and the congruence between stance and premises. The results demonstrate that ruRoberta outperforms ruBERT for stance and premise detection. The analysis reveals only one statistically significant finding: both models perform substantially better on congruent texts, where stance and premise align. The apparent differences in film century and rating do not reach statistical significance given the current test sample size. The newly created dataset and the fine-tuning of encoder models establish a foundation for future research in argument mining for Russian-language texts.