This application continuously monitors many news sites and collects the messages related to COVID-19. Next, the sentiment in these sentences is detected and classified as positive or negative. The results are displayed in a graphical interface, which quickly gives an impression of the current sentiment on this topic.
In the application you can select news sites in different languages. You can also filter specific sources and click through to the sentences and messages in which the sentiment was found.
Start the demo, and click on the different news sources, messages or aspects of the COVID-19 news to filter and see how this affects the sentiment.
The news items are retrieved from the various daily newspapers via RSS feeds. Then the messages are filtered based on corona/covid related terms that are tracked in an ontology. The sentiment is assigned by a specially trained roBERTa classifier. In addition to a general sentiment, a distinction is made between four different aspects of the pandemic: infections, measures, tests and vaccinations.
To make this distinction, the classifier has been trained for Natural Language Inference. This means that the so-called “entailment” between two texts can be detected. This is important because several sentiments can be expressed simultaneously in a single sentence.
For example the sentence: “Despite the reduced willingness to vaccinate, the number of infections has decreased“: This sentence expresses a negative sentiment for the aspect of vaccinations, but the sentiment is positive for the aspect of infections.
In order to be able to process news items in other languages as well, this system uses machine translation to translate all messages to English. As a result, this system can suffice with only one sentiment classifier for many different languages.
This media sentiment detection system is completely generic in design and is ready to use for any conceivable topic without major changes to the system. For example, we have used the same set-up to monitor the sentiment surrounding all makes and types of automobiles available in the Netherlands. Unlike traditional sentiment detection systems, we can simultaneously discern sentiment on different aspects of the subject. For cars, this was sentiment related to appearance, handling, price, fuel consumption, equipment, comfort and safety.
With the Semlab media sentiment detection system, sentiment in many languages for any topic of choice can be detected in real time. Moreover, the sentiment of different aspects of this topic can be distinguished simultaneously. Unlike other suppliers, we do not supply a “black box”. Our system is completely transparent. The reported sentiment can always be related to the message containing the passage in which this sentiment was expressed. And these messages can be displayed directly for review.