Merge pull request #3432 from flairNLP/span_classifier_ner_type

add prediction label type for span classifier

docs/tutorial/tutorial-training/how-to-train-span-classifier.md

@@ -154,7 +154,7 @@ from flair.nn.multitask import make_multitask_model_and_corpus
 
 # 1. get the corpus
 ner_corpus = NER_MULTI_WIKINER()
-nel_corpus = ZELDA(column_format={0: "text", 2: "ner"})  # need to set the label type to be the same as the ner one
+nel_corpus = ZELDA(column_format={0: "text", 2: "nel"})  # need to set the label type to be the same as the ner one
 
 # --- Embeddings that are shared by both models --- #
 shared_embeddings = TransformerWordEmbeddings("distilbert-base-uncased", fine_tune=True)
@@ -171,12 +171,13 @@ ner_model = SequenceTagger(
 )
 
 
-nel_label_dict = nel_corpus.make_label_dictionary("ner", add_unk=True)
+nel_label_dict = nel_corpus.make_label_dictionary("nel", add_unk=True)
 
 nel_model = SpanClassifier(
     embeddings=shared_embeddings,
     label_dictionary=nel_label_dict,
-    label_type="ner",
+    label_type="nel",
+    span_label_type="ner",
     decoder=PrototypicalDecoder(
         num_prototypes=len(nel_label_dict),
         embeddings_size=shared_embeddings.embedding_length * 2, # we use "first_last" encoding for spans

flair/data.py

@@ -1363,7 +1363,7 @@ def __init__(
             test, train = randomly_split_into_two_datasets(train, test_size, random_seed)
             log.warning(
                 "No test split found. Using %.0f%% (i.e. %d samples) of the train split as test data",
-                test_portion,
+                test_portion * 100,
                 test_size,
             )
 
@@ -1375,7 +1375,7 @@ def __init__(
             dev, train = randomly_split_into_two_datasets(train, dev_size, random_seed)
             log.warning(
                 "No dev split found. Using %.0f%% (i.e. %d samples) of the train split as dev data",
-                dev_portion,
+                dev_portion * 100,
                 dev_size,
             )
 

flair/models/entity_linker_model.py

@@ -94,6 +94,7 @@ def __init__(
         label_dictionary: Dictionary,
         pooling_operation: str = "first_last",
         label_type: str = "nel",
+        span_label_type: Optional[str] = None,
         candidates: Optional[CandidateGenerator] = None,
         **classifierargs,
     ) -> None:
@@ -107,6 +108,7 @@ def __init__(
                 text representation we take the average of the embeddings of the token in the mention.
                 `first_last` concatenates the embedding of the first and the embedding of the last token.
             label_type: name of the label you use.
+            span_label_type: name of the label you use for inputs of predictions.
             candidates: If provided, use a :class:`CandidateGenerator` for prediction candidates.
             **classifierargs: The arguments propagated to :meth:`flair.nn.DefaultClassifier.__init__`
         """
@@ -121,6 +123,7 @@ def __init__(
 
         self.pooling_operation = pooling_operation
         self._label_type = label_type
+        self._span_label_type = span_label_type
 
         cases: Dict[str, Callable[[Span, List[str]], torch.Tensor]] = {
             "average": self.emb_mean,
@@ -153,9 +156,16 @@ def emb_mean(self, span, embedding_names):
         return torch.mean(torch.stack([token.get_embedding(embedding_names) for token in span], 0), 0)
 
     def _get_data_points_from_sentence(self, sentence: Sentence) -> List[Span]:
+        if self._span_label_type is not None:
+            spans = sentence.get_spans(self._span_label_type)
+            # only use span label type if there are predictions, otherwise search for output label type (training labels)
+            if spans:
+                return spans
         return sentence.get_spans(self.label_type)
 
     def _filter_data_point(self, data_point: Sentence) -> bool:
+        if self._span_label_type is not None and bool(data_point.get_labels(self._span_label_type)):
+            return True
         return bool(data_point.get_labels(self.label_type))
 
     def _get_embedding_for_data_point(self, prediction_data_point: Span) -> torch.Tensor:
@@ -170,6 +180,7 @@ def _get_state_dict(self):
             "pooling_operation": self.pooling_operation,
             "loss_weights": self.weight_dict,
             "candidates": self.candidates,
+            "span_label_type": self._span_label_type,
         }
         return model_state