Emotion detection/analysis

Emotion detection, or emotion analysis, is an NLP task that identifies the specific emotions expressed in text, such as joy, sadness, anger, fear or disgust. Standard sentiment analysis tells you whether text is positive or negative; emotion detection tells you which feeling is behind it. Two negative reviews can need very different responses depending on whether the customer is angry or disappointed.

How it works

An emotion model is a classifier trained on text labelled with emotions. For each document, sentence or target phrase, it returns a score per emotion. IBM Watson NLU has an emotion feature covering joy, sadness, fear, disgust and anger, which is the set used in MLforSEO’s customer review project. Open-source transformer models on Hugging Face offer alternatives; one widely used English model predicts seven classes (anger, disgust, fear, joy, neutral, sadness and surprise). Large language models can also label emotions from a prompt, though their outputs are less consistent between runs.

How it’s used in SEO and marketing

• Customer reviews. Turn thousands of reviews into emotion scores, then chart which products or locations attract anger versus joy.
• Social listening. Track the emotional response to a campaign or announcement over time.
• Query research. Spot frustration in queries such as “why does my … keep” to plan troubleshooting content with the right tone.
• Prioritising responses. Route angry complaints ahead of mildly disappointed ones.

Example

In MLforSEO’s review analysis, reviews are scraped, run through IBM Watson NLU’s emotion feature in a Colab notebook, and the five emotion scores are added as columns. In Data Studio, a calculated field picks the strongest emotion per review as a “Category Label (Emotion)”, so you can filter to every review where anger dominates and read them together.

Limitations and things to watch

• Sarcasm, mixed feelings and cultural differences confuse models.
• Label sets differ between tools, so scores from two providers are not directly comparable.
• Scores reflect the model’s confidence, which is different from how intensely the person felt.
• Short texts give the model little to work with, and language support varies by provider.
• Check that your chosen API and its emotion feature are still available for your region and language before building a workflow on it.

Learn it on MLforSEO

• Semantic Analysis of Customer Reviews with Machine Learning
• Sentiment Analysis: A Deep Dive Into the Theory, Methods, and Applications
• IBM Watson NLU Sentiment and Emotion Analysis Template for Google Sheets
• Customer Reviews Semantic Analysis: Data Studio Template
• Introduction to Machine Learning for SEO course