Error in MultilabelNaiveBayes about skweak HOT 5 CLOSED

This seems to be the same (or similar) kind of error I ran into (#43). We didn't really manage to pinpoint the exact source of the problem at that time.
One workaround that worked for me was replacing -np.inf by a large negative number. But if you can help trace down why this sometimes happens and sometimes doesn't, it would be great, too!

from skweak.

Could you send a minimal data example, so I can have a look at it?
Typically, when this happens, it means that the HMM has reached a position where no state is possible (all values have a probability of zero), given the current transition and emission probabilities.

from skweak.

Could you send a minimal data example, so I can have a look at it? Typically, when this happens, it means that the HMM has reached a position where no state is possible (all values have a probability of zero), given the current transition and emission probabilities.

example:
Sentence: Eastern Equine Encephalitis Virus (EEEV) is a dangerous disease

we have predicted labels before from 2 models.
model1: ["negative", "not safe"]
model2: ["positive", "not safe"]

how to generate a labelling function for this kind of multi labelling.

from skweak.

same error occured to me while working on an NER Task with the train set from the CoNLL dataset.

from skweak.

Ok, I found what was going on (and sorry for the answering delay): even though the aggregation model allows for multi-label predictions, skweak still assumes that each labelling source only provides a single label for a given text. In other words, the aggregator may predict multiple labels, but each labelling source must stick to one single label at a time.

Fortunately, there is an easy way to change your code: just create a separate labelling source for each possible label outputted by a given model, like model1_negative, model1_positive, model1_not_safe, etc.

from skweak.