Summary
In the previous chapters, we looked at the performance of a collection of classifiers on a range of datasets, with datasets with varying numbers of emotions, varying sizes, and varying kinds of text and, most importantly, with some datasets that assigned exactly one label to each tweet and some that allowed zero or more labels per tweet. The conclusion at the end of Chapter 10, Multiclassifier was that “different tasks require
different classifiers.” This holds even more strongly now that we have tried our classifiers on data that does not match the data they were trained on, with the DNN and SVM classifiers that performed well on some of the previous datasets doing extremely poorly on the case study data.
These two classifiers seem to have assigned neutral to almost all the tweets in this dataset. This seems likely because the clues that these classifiers are sensitive to are missing from, or at any rate rare in, the data, and hence they are not assigning...