3 ms·
"Because the calculation is automated and devoid of human involvement, criteria are consistently applied across articles and sources." How is the training data
by tqi 7y ago
"Because the calculation is automated and devoid of human involvement, criteria are consistently applied across articles and sources."
How is the training data generated?
- ajoy 7y agoFor the tone of voice analysis, our training data was a large set of articles from Reuters and Associated Press - both wire services that are used by most news outlets. We supplemented this with a dictionary of emotional weights for words and some other heuristics. For author expertise, we classify every article into one of 1000 different topics based on the IPTC taxonomy. We’ve now evaluated about 7M articles creating an expertise database on 50,000 journalists, i.e. what topics have they written on in the past and how much do they focus on that.