Skip to main content

Named Entity Recognition Package

Project description

TakeBlipNer Package

Data & Analytics Research

Overview

NER (Named Entity Recognition) is an NLP problem that aims to locate and classify entities in a text. This implementation uses BiLSTM-CRF for solving the NER task utilizing PyTorch framework for training a supervised model and predicting in CPU. For training, it receives a pre-trained FastText Gensim embedding, a PosTagging Model and a .csv file. It outputs three pickle files: model, word vocabulary and label vocabulary. Example of classes that can be trained:

  • Financial [FIN]
  • Generic [GEN]
  • Company [COMP]
  • Number [NUMBER]
  • Document [DOC]
  • Location [LOC]
  • Person [PERS]
  • Phone [PHONE]
  • Address [ADDR]
  • Email [EMAIL]
  • Date [DATE]
  • Week Day [WD]
  • Money [MONEY]
  • Relatives [REL]
  • Vocatives [VOC]

Some additional information is used to identify where the recognized entity begins and ends.

  • The letter B indicates the beginning of the CLASS class entity;
  • The letter I indicates that the respective token is a continuation of the class with the name CLASS started;
  • The letter O indicates that no entity related to the token was found;

For example, the sentence "ligar internet a cabo!" would be classified as "O O B-GEN I-GEN I-GEN", where B-GEN represents the beginning of the GEN entity (token "internet") and the next two tokens are the continuation of the entity (tokens "a cabo"). The entity found in the sentence would be "internet a cabo" of the GEN class.

Here are presented these content:

Training NER Model

To train you own NER model using this package, some steps should be made before:

  1. Import main packages
  2. Initialize file variables: embedding, postagging, train and validation .csv files;
  3. Initialize NER parameters;
  4. Instantiate vocabulary and label vocabulary objects;
  5. Save vocabulary and label models;
  6. Read PosTagging pickle file, embedding and initializing PosTagging object;
  7. Initialize BiLSTM-CRF model and set embedding;
  8. Initialize LSTMCRFTrainer object;
  9. Train the model.

An example of the above steps could be found in the python code below:

  1. Import main packages:
import torch

from TakeBlipNer import utils, vocab, nermodel
from TakeBlipNer.train import LSTMCRFTrainer
from TakeBlipPosTagger.predict import PosTaggerPredict
  1. Initialize file variables:
wordembed_path = '*.kv'
save_dir = '*'
input_path = '*.csv'
val_path = '*.csv'
postag_model_path = '*.pkl'
postag_label_path = '*.pkl'
  1. Initialize NER parameters

In order to train a model, the following variables should be created:

  • sentence_column: String with sentence column name in train file
  • unknown_string: String which represents unknown token;
  • padding_string: String which represents the pad token;
  • batch_size: Number of sentences Number of samples that will be propagated through the network;
  • shuffle: Boolean representing whether the dataset is shuffled.
  • use_pre_processing: Boolean indicating whether the sentence will be preprocessed
  • separator: String with file separator (for batch prediction);
  • encoding: String with the encoding used in sentence;
  • save_dir: String with directory to save outputs (checkpoints, vocabs, etc.);
  • device: String where train will occur (cpu or gpu);
  • word_dim: Integer with dimensions of word embeddings;
  • lstm_dim: Integer with dimensions of lstm cells. This determines the hidden state and cell state sizes;
  • lstm_layers: Integer with layers of lstm cells;
  • dropout_prob: Float with probability in dropout layers;
  • bidirectional: Boolean whether lstm cells are bidirectional;
  • alpha: Float representing L2 penalization parameter;
  • epochs: Integer with number of training epochs;
  • ckpt_period: Period to wait until a model checkpoint is saved to the disk. Periods are specified by an integer and a unit ("e": 'epoch, "i": iteration, "s": global step);
  • val: Integer whether to perform validation;
  • val_period: Period to wait until a validation is performed. Periods are specified by an integer and a unit ("e": epoch, "i": iteration, "s": global step);
  • samples: Integer with number of output samples to display at each validation;
  • learning_rate: Float representing learning rate parameter;
  • learning_rate_decay: Float representing learning rate decay parameter;
  • max_patience: Integer with max patience parameter;
  • max_decay_num: Integer with max decay parameter;
  • patience_threshold: Float representing threshold of loss for patience count.

Example of parameters creation:

sentence_column = 'Message'
label_column = 'Tags'
unknown_string = '<unk>'
padding_string = '<pad>'
batch_size = 64
shuffle = False
use_pre_processing = True
separator = '|'
encoding = 'utf-8'
device = 'cpu'
word_dim = 300
lstm_dim = 300
lstm_layers = 1
dropout_prob = 0.05
bidirectional = False
alpha = 0.5
epochs = 10
ckpt_period = '1e'
val = True
val_period = '1e'
samples = 10
learning_rate = 0.001
learning_rate_decay = 0.01
max_patience = 5
max_decay_num = 5
patience_threshold = 0.98
device = 'cpu'
  1. Instantiate vocabulary and label vocabulary objects:
input_vocab = vocab.create_vocabulary(
    input_path=input_path,
    column_name=sentence_column,
    pad_string=padding_string,
    unk_string=unknown_string,
    encoding=encoding,
    separator=separator,
    use_pre_processing=use_pre_processing)

label_vocab = vocab.create_vocabulary(
    input_path = input_path,
    column_name =label_column,
    pad_string = padding_string,
    unk_string = unknown_string,
    encoding = encoding,
    separator = separator,
    is_label = True)
  1. Save vocabulary and label models:
vocab.save_vocabs(save_dir, input_vocab, 'ner-vocab-input.pkl')
vocab.save_vocabs(save_dir, label_vocab, 'ner-vocab-label.pkl')
vocab.save_vocabs(save_dir, postag_label_path, 'vocab-postag.pkl')
  1. Read PosTagging pickle file, embedding and initializing PosTagging object:
postag_bilstmcrf = torch.load(postag_model_path)
embedding = utils.load_fasttext_embeddings(wordembed_path, '<pad>')

postag_model = PosTaggerPredict(
    model=postag_bilstmcrf,
    label_path=postag_label_path,
    embedding=embedding)
  1. Initialize BiLSTM-CRF model and set embedding:
crf = nermodel.CRF(
    vocab_size=len(label_vocab),
    pad_idx=input_vocab.f2i[padding_string],
    unk_idx=input_vocab.f2i[unknown_string],
    device=device).to(device)

bilstmcr_model = nermodel.LSTMCRF_NER(
    device=device,
    crf=crf,
    vocab_sizes=[len(input_vocab), len(postag_model.label_vocab)],
    word_dims=[word_dim, len(postag_model.label_vocab)],
    hidden_dim=lstm_dim,
    layers=lstm_layers,
    dropout_prob=dropout_prob,
    bidirectional=bidirectional,
    alpha=alpha
).to(device)

bilstmcr_model.reset_parameters()

fasttext = postag_model.fasttext

bilstmcr_model.embeddings[0].weight.data = torch.from_numpy(fasttext[input_vocab.i2f.values()])
bilstmcr_model.embeddings[0].weight.requires_grad = False

utils.load_postag_representation(bilstmcr_model.embeddings[1], postag_model.label_vocab)
bilstmcr_model.embeddings[1].weight.requires_grad = False
  1. Initialize LSTMCRFTrainer object:
trainer = LSTMCRFTrainer(
    bilstmcrf_model=bilstmcr_model,
    epochs=epochs,
    input_vocab=input_vocab,
    input_path=input_path,
    postag_model=postag_model,
    postag_label_vocab=postag_model.label_vocab,
    label_vocab=label_vocab,
    save_dir=save_dir,
    ckpt_period=utils.PeriodChecker(ckpt_period),
    val=val,
    val_period=utils.PeriodChecker(val_period),
    samples=samples,
    pad_string=padding_string,
    unk_string=unknown_string,
    batch_size=batch_size,
    shuffle=shuffle,
    label_column=label_column,
    encoding=encoding,
    separator=separator,
    use_pre_processing=use_pre_processing,
    learning_rate=learning_rate,
    learning_rate_decay=learning_rate_decay,
    max_patience=max_patience,
    max_decay_num=max_decay_num,
    patience_threshold=patience_threshold,
    val_path=val_path,
    tensorboard=None)
  1. Train the model:
trainer.train()

Prediction

The prediction could be done in two ways, with a single sentence or a batch of sentences.

Single Prediction

To predict a single sentence, the method predict_line should be used. Example of initialization e usage:

Important: before label some sentence, it's needed to make some steps:

  1. Import main packages;
  2. Initialize model variables;
  3. Read PosTagging, NER model and embedding model;
  4. Initialize and usage.

An example of the above steps could be found in the python code below:

  1. Import main packages:
  import torch

  import TakeBlipNer.utils as utils
  from TakeBlipNer.predict import NerPredict
  1. Initialize model variables:

In order to predict the sentences tags, the following variables should be created:

  • postag_model_path: string with the path of PosTagging pickle model;
  • postag_label_path: string with the path of PosTagging pickle labels;
  • ner_model_path: string with the path of NER pickle model;
  • ner_label_path: string with the path of NER pickle labels;
  • wordembed_path: string with FastText embedding files
  • save_dir: string with path and file name which will be used to save predicted sentences (for batch prediction);
  • padding_string: string which represents the pad token;
  • encoding: string with the encoding used in sentence;
  • separator: string with file separator (for batch prediction);
  • sentence: string with sentence to be labeled.

Example of variables creation:

postag_model_path = '*.pkl'
postag_label_path = '*.pkl'
ner_label_path = '*.pkl'
ner_model_path = '*.pkl'
wordembed_path = '*.kv'
save_dir = '*.csv'
padding_string = '<pad>'
encoding = 'utf-8'
separator = '|'
sentence = 'SENTENCE EXAMPLE TO PREDICT'

  1. Read PosTagging, NER model and embedding model:
bilstmcrf = torch.load(model_path)
embedding = utils.load_fasttext_embeddings(wordembed_path, padding)
  1. Initialize and usage:
ner_predicter = NerPredict(
        model=bilstmcrf,
        label_path=label_vocab,
        embedding=embedding,
        save_dir=save_dir,
        encoding=encoding,
        separator=separator)

print(ner_predicter.predict_line(sentence))

Batch Prediction

To predict a batch of sentences in a .csv file, another set of variables should be created and passed to predict_batch method. The variables are the following:

  • input_path: a string with path of the .csv file;
  • sentence_column: a string with column name of .csv file;
  • unknown_string: a string which represents unknown token;
  • batch_size: number of sentences which will be predicted at the same time;
  • shuffle: a boolean representing if the dataset is shuffled;
  • use_pre_processing: a boolean indicating if sentence will be preprocessed;
  • output_lstm: a boolean indicating if LSTM prediction will be saved.

Example of initialization e usage of predict_batch method:

input_path = '*.csv'
sentence_column = '*'
unknown = '<unk>'
batch_size = 64
shuffle = False
use_pre_processing = True
output_lstm = True

ner_predicter.predict_batch(
            filepath=input_path,
            sentence_column=sentence_column,
            pad_string=padding_string,
            unk_string=unknown_string,
            batch_size=batch_size,
            shuffle=shuffle,
            use_pre_processing=use_pre_processing,
            output_lstm=output_lstm)

The batch sentences prediction will be saved in the given save_dir path.

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

TakeBlipNer-0.0.7b0.tar.gz (20.6 kB view hashes)

Uploaded source

Built Distribution

TakeBlipNer-0.0.7b0-py3-none-any.whl (19.7 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Huawei Huawei PSF Sponsor Microsoft Microsoft PSF Sponsor NVIDIA NVIDIA PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page