Comparative Analyses of Multilingual Sentiment Analysis Systems for News and Social Media

In this paper, we present an evaluation of three in-house sentiment analysis (SA) systems originally designed for three distinct SA tasks, in a highly multilingual setting. For the evaluation, we collected a large number of publicly available gold standard datasets, in different languages and varied text types. The aim of using different domain datasets was to achieve a clear snapshot of the level of overall performance of the systems and thus obtain a better quality of an evaluation. We compare the results obtained with the best performing systems evaluated on their basis and performed an in-depth error analysis. Based on the results, we can see that some systems perform better for different datasets and task than they were designed, showing that we could replace one system with another and gain an improvement in performance. Our results are hardly comparable with the original dataset results because the datasets often contain a different number of polarity classes than we used, and for some datasets, there are even no basic results. In some cases that we were able to compare, we achieved slightly worse results due to our focus on multilinguality.

PRIBAN Pavel;  BALAHUR-DOBRESCU Alexandra; 

2024-06-05

Springer Verlag

JRC115701

0302-9743 (online),   

https://link.springer.com/chapter/10.1007/978-3-031-24340-0_20,    https://publications.jrc.ec.europa.eu/repository/handle/JRC115701,   

10.1007/978-3-031-24340-0_20 (online),