First version of Zero-Shot Sequence Labeler

[alanakbik]

First version of TARS few-shot sequence tagger. Train like this:

# init corpus and map label names to descriptions
corpus = WNUT_17(label_name_map={
    "location": "location name",
    "corporation": "corporation name",
    "person": "person name",
    "creative-work": "name of song, movie, book or other creative work",
    "product": "name of product or consumer good",
    "group": "name of music band, sports team or non-corporate organization",
}
)

dictionary = corpus.make_label_dictionary('ner')
print(dictionary)

# init the TARS sequence tagger
tars_tagger = TARSTagger(
    'ner_wnut',
    dictionary,
    tag_type='ner',
    embeddings='bert-base-uncased',
    num_negative_labels_to_sample=1,
    prefix=True,
)

# train the model
trainer = ModelTrainer(tars_tagger, corpus)

trainer.train('resources/taggers/few-shot-sequence-tagger',
              learning_rate=0.02,
              mini_batch_size=16,
              mini_batch_chunk_size=1,
              max_epochs=20,
              monitor_test=True,
              embeddings_storage_mode="none",
              )

This PR also makes a number of smaller changes:

• Change the best model logic so that again the best-model.pt file is saved. The information which epoch was the best model is no longer encoded directly in the filename but rather will be added explicitly to the logs.
• Changes in loss averaging so that we get consistency between training / testing and in particular TARS-like approaches. This means that the sequence tagger no longer returns an averaged loss over all words in the mini-batch, but rather a summed loss plus the information over how many words the loss was summed.