Prompt-Tuning for Targeted Sentiment Analysis in Russian

Yuliana Solomatina, Natalia V Loukachevitch

Abstract: This paper describes prompt-based methods for targeted sentiment classification of the Russian news texts, proposed during the RuSentNE-2023 competition. This task is challenging in two following aspects: first, it is required to extract sentiment towards specific entities, not of the overall text; second, the sentiments in the news texts are often implicit, which means they are harder to detect. In the present study, several strategies of prompt-tuning BERT-like models were explored. The best result was achieved when incorporating external knowledge into the prompt verbalizer. We demonstrate that prompt-tuning methods help achieve high results while being less computationally intensive than other fine-tuning and ensembling strategies.