A new paper by our researchers has been published on the Springer website. It explores the sentiment of Wikipedia articles about companies using a range of artificial-intelligence models. The study set out to determine how well each model copes with assessing the sentiment of Wikipedia’s characteristically long articles, and to examine how that sentiment varies across industries and article-quality classes.
Wikipedia is one of the most popular sources of information on a wide array of topics. Because anyone can edit it, ensuring that each article maintains a neutral point of view (NPOV) can be a significant challenge. One method of evaluating neutrality is sentiment analysis – identifying the emotional tone embedded in the text.
The publication “Sentiment Analysis of Wikipedia Articles About Companies: A Comparison of Different Models” presents a comparative study of several sentiment-analysis methods, ranging from lexicon-based tools (TextBlob, VADER) to more advanced transformer-based models (RoBERTa, DistilBERT, and PySentimiento). Authors of the publication: Dr. Milena Stróżyna, Dr. Włodzimierz Lewoniewski, Izabela Czumałowska, Aleksandra Wojewoda.
The researchers carried out a detailed process of article selection and preprocessing, and they proposed three distinct approaches for aggregating sentence-level sentiment into an overall article score. Their findings reveal that sentiment assessments vary considerably depending on both the model applied and the industry of the company being described. These results offer practical insights for scholars and practitioners working with sentiment analysis – particularly for long-form texts such as Wikipedia articles – and advance our understanding of the strengths and limitations of different sentiment-analysis models.