Whereas we previously saw that the NLU sentiment scores were distributed throughout the -1 to +1 scale, in the last several months I’ve seen that there are virtually no neutral scores assigned. In the past it was very common for NLU to assign scores in this range.
Of the last 10,000 articles we've submitted to the API, the service has returned:
positive (>0.2) 5,274
negative (<-0.2) 4,692
I think it is unlikely that in 10,000 pieces of content, just 34 are truly neutral. The team should be aware that the performance of the model does not seem appropriate.
Why is it useful?
|Who would benefit from this IDEA?||As a customer, I want to be able to compare sentiment scores for content across time. The exaggerated change in the NLU model means that this is no longer feasible.|
How should it work?