Everything you need to know about our recent improvements to sentiment within Media Intelligence
What is Sentiment?
Sentiment was designed to provide helpful insights into how the media coverage an organization receives has been perceived by their audience. These metrics are used to help understand your organizations reputation, your competitors perception in your market space or overall understanding of how well your chosen message has been received. This in turn can help drive an effective business strategy. Sentiment analysis determines whether a text is positive, negative, or neutral by extracting particular words or phrases.
The Issue We Faced
Language is often vague or highly contextual, making it very difficult for a machine to understand without human help. As such, human annotated data is essential when training a machine learning platform to analyze sentiment. Understanding the sentiment of an article is key to gaining insightful, actionable data.
We traditionally used an algorithm that takes a document and applies sentiment for the entire text. The longer the document, the less accurate the sentiment.
Moving away from document based sentiment to sentence based sentiment. This means each sentence of an article or post is evaluated individually and gets its own sentiment (positive, negative, neutral).
For the document based sentiment, the sentiment is then calculated based on the sum of sentence sentiment. Head on to our Engineering Blog to learn more from our data scientists on how sentiment is calculated.
With the new model, we see significant shift of sentiment for every language. 55-65% less sentiment overwrites globally!
How You Can Help
However, the major innovation of the new sentiment comes with the customer feedback loop. Meaning every time a customer overrides a sentiment score in the platform, we feed this data back into our models to help retrain and improve it’s accuracy.
As our model and the technologies we use continue to grow, our world class data science team is continuously seeking new and intuitive ways to help improve our sentiment offering. One example they have already highlighted is Entity Based Sentiment. This seeks to determine the sentiment of certain people, companies, or other entities within a document.