Recognizing Selected Features in the News Text

Agata Kaczmarek, Filip Tobolewski, Maria Ganzha, Marcin Paprzycki, Anna Wróblewska

Abstract: Today, huge amounts of text news are generated, making it challenging to keep up with them. At the same time, customers realize that companies may or may not match their values and/or preferences. To address this issue, Actaware Inc. is developing a solution to automatically capture news events related to specified companies and assess news sentiment. In this context, natural language processing has been applied to real-world data collected from multiple sources, e.g. Washington Post, Reuters, or BBC. To categorize text news and analyze sentiment, state-of-the-art approaches were used, including BERT- and GPT-based, and specially developed ones. The experimental evaluation of the proposed solutions considered not only standard metrics but also their sustainability, cost, and explainability. Moreover, the resulting pipeline, accepted by the company, is presented.