Handle Numeric features

This feature is a experimental feature

https://github.com/BrikerMan/Kashgari/issues/90

Some time, except the text, we have some additional features like text formatting (italic, bold, centered), position in text and more. Kashgari provides NumericFeaturesEmbedding and StackedEmbedding for this kine data. Here is the details.

If you have a dataset like this.

  1. token=NLP       start_of_p=True    bold=True     center=True     B-Category
  2. token=Projects  start_of_p=False   bold=True     center=True     I-Category
  3. token=Project   start_of_p=True    bold=True     center=False    B-Project-name
  4. token=Name      start_of_p=False   bold=True     center=False    I-Project-name
  5. token=:         start_of_p=False   bold=False    center=False    I-Project-name

First, numerize your additional features. Convert your data to this. Remember to leave 0 for padding.

  1. text = ['NLP', 'Projects', 'Project', 'Name', ':']
  2. start_of_p = [1, 2, 1, 2, 2]
  3. bold = [1, 1, 1, 1, 2]
  4. center = [1, 1, 2, 2, 2]
  5. label = ['B-Category', 'I-Category', 'B-Project-name', 'I-Project-name', 'I-Project-name']

Then you have four input sequence and one output sequence. Prepare your embedding layers.

  1. import kashgari
  2. from kashgari.embeddings import NumericFeaturesEmbedding, BareEmbedding, StackedEmbedding
  4. import logging
  5. logging.basicConfig(level='DEBUG')
  7. text = ['NLP', 'Projects', 'Project', 'Name', ':']
  8. start_of_p = [1, 2, 1, 2, 2]
  9. bold = [1, 1, 1, 1, 2]
  10. center = [1, 1, 2, 2, 2]
  11. label = ['B-Category', 'I-Category', 'B-ProjectName', 'I-ProjectName', 'I-ProjectName']
  13. text_list = [text] * 100
  14. start_of_p_list = [start_of_p] * 100
  15. bold_list = [bold] * 100
  16. center_list = [center] * 100
  17. label_list = [label] * 100
  20. SEQUENCE_LEN = 100
  22. # You can use WordEmbedding or BERTEmbedding for your text embedding
  23. text_embedding = BareEmbedding(task=kashgari.LABELING, sequence_length=SEQUENCE_LEN)
  24. start_of_p_embedding = NumericFeaturesEmbedding(feature_count=2,
  25.                                                 feature_name='start_of_p',
  26.                                                 sequence_length=SEQUENCE_LEN)
  28. bold_embedding = NumericFeaturesEmbedding(feature_count=2,
  29.                                                 feature_name='bold',
  30.                                                 sequence_length=SEQUENCE_LEN)
  32. center_embedding = NumericFeaturesEmbedding(feature_count=2,
  33.                                                 feature_name='center',
  34.                                                 sequence_length=SEQUENCE_LEN)
  36. # first one must be the text embedding
  37. stack_embedding = StackedEmbedding([
  38.     text_embedding,
  39.     start_of_p_embedding,
  40.     bold_embedding,
  41.     center_embedding
  42. ])
  44. x = (text_list, start_of_p_list, bold_list, center_list)
  45. y = label_list
  46. stack_embedding.analyze_corpus(x, y)
  48. # Now we can embed with this stacked embedding layer
  49. print(stack_embedding.embed(x))

Once embedding layer prepared, you could use all of the classification and labeling models.

  1. # We can build any labeling model with this embedding
  3. from kashgari.tasks.labeling import BLSTMModel
  4. model = BLSTMModel(embedding=stack_embedding)
  5. model.fit(x, y)
  7. print(model.predict(x))
  8. print(model.predict_entities(x))

This is the struct of this model.

Handle Numeric features - 图1