We offer automatic sentiment detection for Dutch, English, French, German, Greek, Italian, Spanish, Portuguese, and Turkish. In this article, we provide more information about the following subjects:
When you wish to get a general overview of how people feel about your topic, you can go to Insights > Overview > Sentiment. The statistics you'll find here are entirely based on the number of mentions in the Inbox, grouped per sentiment value.
The pie chart on the left gives you the general sentiment break up for your selected period, whereas the chart on the right will break it up into the chosen date grouping. For this last chart, you can also leave out selected values by hovering over the right top of the chart, clicking the striped button and then (de)selecting one or multiple values by clicking them.
Hovering over bars on the right chart will cause a pop-up listing the amounts for the selected sentiment values and situating the selected bar within the chosen date range and date grouping. Hovering over the pie chart will give you the breakdown per sentiment value, both in numbers and percentage. For both charts, clicking on a specific part will present you with a pop-up showing the underlaying mentions.
As you might have noticed, only mentions with determined sentiment are included in the metrics you can find in Insights. The mentions for which the sentiment could not be determined automatically are labeled as unknown and not included here. If you wish, you can change the sentiment manually for the unknown ones, so all mentions end up being included.
We use state of the art machine learning and classification algorithms, specifically, maximum entropy classification of texts represented as “bags of words” enriched with additional meta- and language specific features such as the source and length of texts. The algorithms then compute the probability that a certain bag of words occurs in a certain polarity class (positive, negative, or neutral). The algorithms were trained on a combination of manually annotated and semi-automatically mined training examples. Our users themselves also contribute to the accuracy of the algorithms because they extensively curate the automatic sentiment annotation. This allows us to build large amounts of training data and significantly increases the coverage of our service compared to other services which are typically trained on much less data.
We dare say our sentiment analysis has greatly improved over the last couple of and performs at least as good as other services available (but with increased coverage). More precisely, we achieve accuracy and precision of around 75% which, on average, is almost as good as humans. If you have interest in helping or collaborating to improve sentiment analysis, we welcome and encourage our users to do so!
When new mentions come into the Inbox, sentiment is automatically assigned to them. However, if you notice that the automatically determined sentiment isn't accurate, you can change this manually.
Are you interested in seeing:
The sentiment of all mentions on a specific, connected profile? Filter on 'Profile is ….' and select the profile you wish to view.
The sentiment of all mentions from, for example, Twitter (not only the mentions from a Twitter profile you connected, but also the ones gathered by a keyword search). Filter on 'Source is …' and select the source you would like to view.
The sentiment overview of all mentions containing your specific hashtag. Filter on 'Hashtag is …' and type the name of the hashtag.