Sentiment or Emotion?

The MORES project has been exploring emotions in political communication. We’ve employed several methods to research this topic, including sentiment and emotion analyses. Both are techniques in natural language processing (NLP) used to understand human emotions and opinions expressed in text. As a result, businesses and researchers have embraced them. But young researchers and laities often mistake them for each other. Each one has its own scope and methodology, making them non-interchangeable. What is the working principle behind their operation? In this blog post, I summarise their commonalities and differences and the type of results they achieve.

Sentiment analysis, also called opinion mining, assesses the sentiment or attitude of a speaker or writer towards a topic or the overall polarity of a document. Usually, this method categorises sentence sentiments as positive, negative, or neutral. The goal is to categorise the sentiment of a text based on its overall orientation, whether at the document, sentence, or aspect level, without pinpointing specific emotions. The techniques employed range from lexicon-based approaches to machine learning approaches, or a combination of both.

While sentiment analysis and emotion analysis are related and often overlap, they serve different purposes and require different methodologies.

In contrast, emotion analysis can identify distinct emotions, such as joy, anger, sadness, surprise, and others, which are expressed through written communication. It goes beyond polarity categorisation to capture nuanced emotional states. Emotion analysis requires complex models that can distinguish between nuanced emotions. It is more suited for applications that require an understanding of specific emotional conditions, such as during election campaigns, policy debates, or group-based conflicts.

Both are valuable tools in NLP. But while sentiment analysis and emotion analysis are related and often overlap, they serve different purposes and require different methodologies. Sentiment analysis typically needs labelled datasets with sentiment annotations and can be performed with smaller datasets or simpler features. Emotion analysis usually requires more detailed and specific emotion-labelled datasets, often necessitating larger and more diverse datasets to capture a wide range of emotional expressions.

In sentiment analysis, the main challenge is that words can have different sentiments in different contexts, with sarcasm and irony being particularly difficult to detect, leading to misclassification. In emotion analysis, the biggest issues are the complexity of manifestation and context dependency, with emotions often better captured using multimodal data (e.g., text, voice, facial expressions, images). For example, if a politician says that “the idea behind this policy was great, but the implementation is tragicomic…,” sentiment analysis might classify this as neutral or mixed; emotion analysis might classify ‘great’ as joy, but it would have a hard time detecting the speaker’s potential irony about how the policy was handled.

The MORES project is developing various computational models to identify primary emotions (e.g., fear, anger, joy) and complex moral emotions (e.g., guilt, shame, pride) using both sentiment and emotion in multilingual settings.

Sentiment analysis has advanced significantly with the rise of deep learning and large-scale pre-trained language models, remaining a critical component in business intelligence and customer experience management. Emotion analysis is gaining traction with the development of more sophisticated NLP models and the growing interest in understanding human emotions using AI, with emerging applications in politics, social psychology, mental health, and social robotics.

The MORES project is developing various computational models to identify primary emotions (e.g., fear, anger, joy) and complex moral emotions (e.g., guilt, shame, pride) using both sentiment and emotion in multilingual settings. This integrated approach is based on a sequential investigation. To filter the general approach to the subject of the discussion, sentiment analysis is an excellent tool. Then, AI-assisted emotion analysis will be used to delve deeper into specific emotions within each sentiment category. Over the next few months, we will be sharing preliminary results derived from an enhanced understanding of political text data. Make sure you subscribe to our newsletter so you don't miss anything.

Publication: What kinds of emotions are mobilised by different policy fields? A text mining analysis of parliamentary speeches (2024). By Zsolt Boda, Orsolya Ring, and Gabriella Szabó. Download here.

Meet the Expert

Learn more about Gabriella Szabó on the Database of Experts.

Newsletter

Stay updated and gain exclusive scientific insights from MORES by subscribing to our newsletter. Join today.

Get Involved

Questions or comments about the blog post? Contact us at [email protected].

Further Reading

Chatterjee, A., Das, B., & Das, A. (2023). Towards the analyzing of social-media data to assess the impact of long lockdowns on human psychology due to the Covid-19 pandemic. In Computational Intelligence Applications for Text and Sentiment Data Analysis (pp. 225-238). Academic Press.

Liu, B. (2020). Sentiment analysis: Mining opinions, sentiments, and emotions. Cambridge University Press.

Plaza-del-Arco, F. M., Curry, A., Curry, A. C., & Hovy, D. (2024). Emotion Analysis in NLP: Trends, Gaps and Roadmap for Future Directions. arXiv preprint arXiv:2403.01222.