YleSent

YleSent is a sentiment analysis project that processes Yle news articles. It predicts sentiment value using a logistic regression model and visualizes average sentiment trends by day and month. The goal is to explore how public sentiment evolves over time and influential keywords.

Motivation

Motivation for this project was that we felt that the news often feels too negative, which can affect mood and perception. People who follow news regularly might be interested if news coverage is too negative or balanced. We want to show trends clearly and help people put news negativity in perspective. This will help users to consume news more critically and makes it easier to spot trends in the news sentiment. We want to increase transparency about how media present news.

Challenges

The sentiment analysis model presented some challenges because the training data (ScandiSent) differs from actual news data in style and content. To address this, we re-trained the model using 3,000 Yle articles, labeled via supervised annotation with ChatGPT. This improved prediction accuracy and made visualizations more meaningful by enabling the model to detect subtler sentiment cues.

Network graph of news topics and their sentiments

Average rate of the news sentiments by day and month

Average monthly sentiments and counts of some interesting news categories

You can select the categories to view by clicking the topic list on the right.

Words with most positive effect on sentiment

Words with most negative effect on sentiment