A semantic network approach to measuring sentiment.

Area of Science:

• Computational Linguistics
• Social Media Analysis
• Network Science

Background:

• Traditional sentiment analysis often uses bag-of-words, neglecting word context.
• Existing methods struggle to capture nuanced sentiment expression in text.

Purpose of the Study:

• To introduce and validate an aspect-based network analysis model for sentiment computation.
• To evaluate the model's internal and external validity against human annotations.
• To assess the model's predictive validity in analyzing news sentiment.

Main Methods:

• Developed an aspect-based network analysis model using shortest paths between sentiment words and target entities.
• Utilized two ground-truth datasets with human-annotated tweet sentiment (positive/negative).
• Applied the model to analyze sentiment in television news coverage of coronavirus (Jan-Mar 2020).

Main Results:

• The network model showed strong correlation with human annotations, with distinct negativity/positivity scores for each class.
• Validated hypotheses regarding sentiment in broadcast vs. non-broadcast news, negative bias, sentiment volume, and uncertainty.
• Found increased sentiment for coronavirus, panic, and social distancing across different news channel types.

Conclusions:

• The aspect-based network analysis model offers a more robust approach to sentiment analysis than traditional methods.
• The model demonstrates utility in analyzing real-world sentiment dynamics, such as in news media.
• Context-aware sentiment analysis is crucial for accurate interpretation of textual data.