Cross-Domain Affective Analysis of Large-Scale Textual Data Using Transformer-Based NLP Models
Bhadra, Nivedita (2025)
2025
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2025052817311
https://urn.fi/URN:NBN:fi:amk-2025052817311
Tiivistelmä
In this thesis, we present a comprehensive, cross-domain analysis of large-scale textual data using advanced Natural Language Processing (NLP) techniques. Our study began with the Common Crawl News (CC-News) dataset—an extensive, time-indexed corpus of news articles collected from global sources. To evaluate the generalizability of our approach, we extended the analysis to three additional datasets: the Financial Phrase-Bank, a Social Media Popularity corpus, and a Kaggle Sentiment Review dataset, each representing distinct textual genres and affective tones.
We implemented a scalable pipeline incorporating transformer-based models for sentiment and emotion classification, alongside BERTopic-based topic modeling. Leveraging big data frameworks and GPU-accelerated inference, we explored patterns in sentiment distribution, emotional tone, and topic-specific trends across domains. Our findings reveal clear domain-specific affective signatures and highlight both the robustness and limitations of pretrained models in multilingual, cross-domain settings. Beyond academic contributions, the results hold practical relevance for media bias detection, audience-sensitive content filtering, and real-time sentiment tracking. This work underscores the value of combining affective computing and large-scale NLP in understanding how emotional tone and topical focus vary across diverse communication platforms.
We implemented a scalable pipeline incorporating transformer-based models for sentiment and emotion classification, alongside BERTopic-based topic modeling. Leveraging big data frameworks and GPU-accelerated inference, we explored patterns in sentiment distribution, emotional tone, and topic-specific trends across domains. Our findings reveal clear domain-specific affective signatures and highlight both the robustness and limitations of pretrained models in multilingual, cross-domain settings. Beyond academic contributions, the results hold practical relevance for media bias detection, audience-sensitive content filtering, and real-time sentiment tracking. This work underscores the value of combining affective computing and large-scale NLP in understanding how emotional tone and topical focus vary across diverse communication platforms.