What's the difference between multilabel and multinomial text classification?
Include an example of each.
Instructor solution
Multilabel and multinomial classifiers are both multiclass classifiers, meaning there are more than two labels that exist as potential outputs (as opposed to binary classifiers that determine if something is or is not a label, such as if an email is or is not spam).
Multilabel classifiers ("any of") mean that one input can have multiple labels as outputs. One example of such a classifier is one that assigns topics to a news article. A news article could be tagged as "sports," "basketball," and "Cleveland Cavaliers" all at once. With multilabel classifiers, each topic would essentially have its own binary classification model––such as "is this sports or not"––and each decision is made independently.
Conversely, multinomial classifiers ("one of") assign just one label to each input. As an example, sentiment analysis commonly has three labels––positive, neutral, and negative––and one input document would never be labeled as both positive and neutral (or positive and negative, etc.). Multinomial classifiers also build out binary classifiers for each topic, but only the label that returns the highest score or the highest probability is selected as the label. For instance, a multinomial sentiment analysis classifier may output a 97% chance the document is positive, 2% it is neutral, and 1% it is negative, and thus, we would predict that the document is positive.
Think you've got it?
You may exit out of this review and return later without penalty.