The Sentiment analysis tool within Engage helps you to manage your reputation and identify new opportunities by labelling comments as positive, negative, or neutral. Doing so helps filter out the noise so you can quickly identify and act on relevant content. You can even filter for sentiment specific messages.
In this article:
How is sentiment calculated?As content is indexed, Engage will determine whether the text in each message is positive, negative, or neutral.
Sentiment analysis is based on cutting edge AI research in the fields of Deep Learning and Natural Language Processing (NLP). Our data scientists are pioneering the use of Transformer Architecture Language Models (AI models famous for being able to suggest how any text should be completed) to determine sentiment in social data.
Models are pre-trained on billions of words, to have a deep general knowledge of over 100 languages before being applied to sentiment analysis, offering a more sophisticated understanding of context, slang, and dialects. Some examples of how sentiment can be indicated are listed below:
- Words (including misspellings)
- Sentence structure
- Multi-word hashtags
The sentiment accuracy for most queries in a supported language is estimated to be between 60-75%. While studies have shown that two humans will agree on the sentiment of posts (tweets) just 80% of the time, Engage continues to test models on the widest range of use cases possible, rather than just those which may achieve the highest level of accuracy (for example, product reviews).
Learn more about the data science behind Brandwatch’s sentiment analysis by reading this interview with one of our data scientists.
What languages are supported for sentiment?
Brandwatch officially supports sentiment in over 40 evaluated languages but can classify content in any language.
The accuracy for most queries in supported languages can be expected to range between 60-75%.
|Mixed Hindi English
Unsupported languagesOur sentiment model is language-agnostic and will attempt to assign sentiment to posts in any language (or posts with no language identified such as emoji-only posts). This includes unsupported languages not yet evaluated and even languages not included in pre-training models. The nature of our multilingual learning model allows it to use what it knows about languages it has been trained with to classify posts in languages, or dialects, it doesn’t know.
Unsupported languages may not see as high a recall rate as the model will only classify sentiment where it has confidence in the prediction. And Brandwatch does not guarantee the precision for these languages.
To manually override a sentiment score, simply click on the face icon and choose the appropriate sentiment — positive, neutral, or negative.
Filtering by sentiment
You can filter your Engage feeds by sentiment by clicking the edit icon, select Advance Filters and choose the sentiment you want to filter by.
To filter conversation comments by sentiment, click on the conversation, press the filter icon, then choose the sentiment you want to filter comments by.