Our tonality tool is built around a data model, where the model has learned to interpret tonality by having processed a larger amount of data. We have different models for different languages. The model is a multi-class classification (with the classes positive, negative and neutral) where a binary classification has been used for each class.
Another way we collect Sentiment is by getting answers about different tonalities, such as;
- positive
- love
- desire
- negative
- hate
- fear
- violence
For these different types, we get different points for how positive or negative it is. Then, we interpret it - if the positive outweighs the negative, it becomes positive, and vice versa.
Please note, if we have an item that has a lot of emojis instead of text, it becomes much harder to detect Sentiment, as a text with a lot of words becomes easier to translate to positive/negative. If the Sentiment can't be decided, it usually ends up as Neutral. If you don't want Neutral items, you can change this in the platform to positive/negative on posts you don't see any sentiments on.
All media types get Sentiment, but it works best on items with more text.