Can someone tell me what characterizes a positive or negative sentiment in Azure Text Analytics API? For example "he died today" Depending on the context the death can be good or bad ... what will determine the return to be positive or negative?
0
votes
1 Answers
4
votes
Under the hood, we have two classifiers.
First, an objectivity model which tells us the probability that a given record is objective.
For instance if someone were to say “This Christmas party was held at the Hyatt”, that is an objective sentence, but a sentence like “I enjoyed the party at the Hyatt, the venue was beautiful.” is subjective.
Then, for those records that are subjective, we analyze the polarity of the records, telling us whether the sentence expresses a positive sentiment or a negative sentiment.
Those objective sentences are mapped to 0.5. The non-objective sentences take into consideration the polarity result to map between 0 and 1 (0== most negative, 1= most positive, around 0.4 slightly negative, around 0.6 slightly positive).
For all practical purposes, you could say that scores between 0.45 and 0.65 are neutral, 0.45 and lower are negative, and 0.65 and higher are positive.
Now, how does the model classify something as objective or not? It's all about the training data, and the labels provided to that data.
