Sentiment analysis
We use a language model to read thousands of reactions and determine sentiment by topic over time. That’s how the Respect Foundation monitored the Week van Respect online.
A social media monitoring tool and our LLM helped the Respect Foundation understand sentiment around the Week van Respect
Through our Data For Good initiative, we helped the Respect Foundation measure the campaign’s online impact: sentiment in the coverage and its actual reach.
How are people talking about us, and is that changing?
After a campaign, a change or an outage, comments pour in through social media, reviews and open-ended responses. No one reads them all. A language model does, and shows you where the tone shifts.
Every response is assigned a tone and a topic, so you can count them
Sentiment by message
Positive, negative, neutral or mixed, as our LLM assessed coverage of the Week van Respect.
What it’s about
Which topic drives the tone: delivery, price, staff or a campaign. For WUA, we categorised open-ended responses by topic.
Tracked over time
How sentiment shifts after a campaign, a change or a service disruption, by channel and by week.
Sentiment alongside reach
How many people saw a message alongside how people talked about it, as in the impact analysis for the Respect Foundation.
A few hundred comments scored by hand, then the model
-
Sources and objectives
Which comments you want to monitor, where they come from and what you will do with the results.
-
Check against human assessments
The model scores a sample that you assess yourselves, so you know how often it gets it right.
-
Monitor and alert
A regular overview by channel and topic, with an alert if the tone suddenly shifts.
Are there enough comments to measure sentiment meaningfully?
A good fit if
Comments come in continuously, and no one reads them all. You want to know how the tone is changing and what is driving it.
Not a good fit if
You receive only a handful of comments each month, or you want to route emails and tickets to different teams. Reading them or using classification is a better fit.
What determines the scope
The number of sources and languages, whether you are allowed to retrieve the data yourself, how granular the topics need to be and whether it needs to run continuously.
Can a model understand sarcasm?
Not always. That is why we test it on your own comments and flag uncertain cases instead of silently including them in the count.
A monitoring tool, a standard feature or a custom language model?
A monitoring tool
Choose this if you mainly want to monitor social media using the channels and categories the tool provides.
A standard feature
Choose this if a positive or negative rating for each post is enough, such as the sentiment feature in Databricks or Azure, where your data already sits.
A custom language model
Choose this if you want sentiment by topic, using your own categories and covering all your sources. This is what we recommend for campaigns and services.
We used language models to analyse campaign responses and open-ended answers
Data Science Lab has been operating since 2016 and is a Microsoft and Databricks partner. We analysed sentiment for the Respect Foundation under our Data For Good label and categorised open-ended answers for WUA.

Our conversation starts with the responses you’re currently missing
Stef Stoelwinder
Business DevelopmentStef is your point of contact. In our first conversation, we’ll discuss which responses you want to track and what you’ll do with the results.
E-mail info@datasciencelab.nl
Phone +31 (0)20 3636 724
Which campaign or service would you like to know people’s opinions on?
Tell us what comments you’re receiving and where they’re posted. We’ll reply by email or arrange an initial conversation.
Message sent
Thanks for your message. We’ll be in touch as soon as possible.
Frequently asked questions
Sentiment analysis uses a model to determine whether the tone of a text is positive, negative, neutral or mixed. Aspect-based sentiment analysis also identifies what the opinion is about, such as price or delivery.
You can find other applications under AI applications.
Text analysis is broader: it includes classifying texts by topic, summarising them and linking them together. Sentiment analysis focuses on tone and how it changes.
Read more about the broader field under NLP and text analysis.
Yes, if the posts can be retrieved. For the Respect Foundation, a social media monitor tracked posts about the Week van Respect, and our LLM assessed their sentiment.
Which channels you can retrieve posts from depends on each platform’s terms.
Where possible, we work where your data already lives. Databricks has an AI function for sentiment analysis in SQL, and Azure AI Language can assess sentiment by aspect. For a custom classification or difficult texts, we use a language model.
We make the choice based on your own feedback.
By comparing it with people. You and the model each rate a sample, and we show you where the ratings differ. That tells you how much you can rely on the results.
We repeat this test if the topics change.
Yes, language models can read Dutch, English and informal language. We test abbreviations, emoji and specialist terminology separately because they vary by audience.
We show you ambiguous cases rather than discarding them.
We store only what is needed: the text, date, channel and result, without names where possible. If customer data is involved, we discuss the approach with your privacy officer beforehand.
See also DPIA for AI.
When there is too little feedback to detect a shift, or when nobody will use the results to do anything differently. In those cases, simply reading the feedback is quicker.
If you want to sort emails or tickets, email classification is a better fit.