This month, we’re introducing some significant enhancements to our existing Text AI Models - Sentiment and Topics, and expanding our coverage to include more languages! We’re also now fully releasing our Entities analysis model, which was previously available in Beta to some clients, meaning it will now be available for everyone to use, at no additional cost.

First, let’s talk about the changes we’ve made to Sentiment analysis on the platform. The ability to gauge the overall nature of a conversation is important for any social listening use case. As such, we have doubled down on our language coverage from 12 to a whopping 24 new languages. Many of these languages are a part of our expansion into the Asian and Pacific Island regions, so if you've been trying to understand what people thought about Dogecoin in Singlish, then look no further, because you can do that on TRAC!

If you’re using the sentiment scores available in the Results exports or via the TRAC API, you will probably need to understand the scoring approach we now use. Sentiment scoring is on a scale from -1 to 1, where the negative side is negative and the positive side is positive (intuitive, right?).

Previously, a 1/-1 would represent a measurement of certainty regarding sentiment. So for example, a post with a rating of -0.95 meant we're 95% certain that this post is negative. But with our latest enhancements, these decimal values are now reflective of intensity rather than accuracy. So now, if a post gets a -0.95, it means it's really negative, whereas a post with -0.15 is just a little bit negative.

You can try out our expanded Sentiment coverage in any of your TRAC searches, and if there are any languages you think we should add, then do let us know!

We've also enhanced our existing Topic analysis model, making significant improvements in comparability and language coverage. This means moving from a bottom-up method, which involves the open discovery of new topics (by looking at the most relevant keywords in a document), to a top-down taxonomy led classification, which now identifies the main themes in a document, even if they aren’t explicitly mentioned. Whilst the previous method allowed us to quickly understand new topics as they emerged, our new taxonomy led approach means that data across searches is now much more comparable, plus we can cover more languages - from 11 to 17 languages in total.

The enhanced Topic analysis will be available automatically across all TRAC searches from now onwards and any data collected prior to this update will not be re-analysed.

Finally, let Pulsar give structure to the data you're collecting and discover the content that matters using our improved Entity extraction model. The new Entity model can extract and identify specific data from social posts and news and classify that data according to predefined named entities. For now, we’re exposing two types of entities, People e.g. Tim Cook, Elon Musk and Organizations: e.g. Apple, Twitter.

Entity recognition is available in the following languages:

Arabic, English, Chinese, Danish, Dutch, Finnish, French, German, Italian, Japanese, Korean, Norwegian, Polish, Portuguese, Russian, Spanish, and Swedish.

To enable Entity analysis in a search, you’ll need to switch this on in your search, during search set-up, when you’re in the ‘Analysis’ section. Click on ‘Text’, as shown below, and you’ll see an option for Entities. For existing searches you’ll need to head over to Search Settings -> Summary -> Click on Analysis -> Text -> Entities.

And that's it for now! As mentioned previously, all our Text AI models are available at no additional cost for all TRAC clients. We are constantly fine tuning our models, so your feedback is welcome - whether it's on performance, accuracy or coverage, we'd love to hear from you!