Open HyunkuKwon opened 4 years ago
The question I have when using sentiment analysis in the projects is that one sentiment classification (or normalized score) doesn't tell the whole story of the people who wrote the content. For example, when leaving reviews on Amazon, a consumer might write half of the paragraph on good things and the rest on complaining. The sentiment score might be neutral (0), but actually the consumer has mixed opinions about the product <0.8, -0.8>. So I am wondering how can we customize the sentiment analysis algorithms into this specific application?
I have a question about the evaluation metrics used for sentiment analysis. We see all of the same metrics e.g. accuracy, f1, precision, recall, etc. However, all of these assume there is a "true" sentiment. Given the incredible subjectivity of sentiment, what are the best practices for getting "ground truth" sentiment in order to then compare to the predictions? Is this something that is often done by crowd sourcing?