from typing import List, Tuple
import pickle
import numpy as np
import sentencepiece as spm
from ..tokenizer import TreebankWordTokenize
from ..utils import check_and_download, load_keras_tokenizer
from .utils import ner_to_displacy_format
from ._spu_context_utils import (
create_spucontext_ner_model,
process_single_word_input,
)
# Resolving parent dependencies
from inspect import getsourcefile
import os
import sys
current_path = os.path.abspath(getsourcefile(lambda: 0))
current_dir = os.path.dirname(current_path)
parent_dir = current_dir[: current_dir.rfind(os.path.sep)]
sys.path.insert(0, parent_dir)
RESOURCES_PATH = os.path.join(os.path.dirname(__file__), "resources/")
PROD_WEIGHTS_LOC = RESOURCES_PATH + "NER_SPUContext_prod.weights"
EVAL_WEIGHTS_LOC = RESOURCES_PATH + "NER_SPUContext_eval.weights"
WORD_EMBEDDING_MATRIX_LOC = os.path.abspath(
os.path.join(
os.path.dirname(__file__),
"..",
"resources/SPUTokenized_word_embedding_16k.matrix",
)
)
PROD_WEIGHTS_LINK = "https://vnlp-model-weights.s3.eu-west-1.amazonaws.com/NER_SPUContext_prod.weights"
EVAL_WEIGHTS_LINK = "https://vnlp-model-weights.s3.eu-west-1.amazonaws.com/NER_SPUContext_eval.weights"
WORD_EMBEDDING_MATRIX_LINK = "https://vnlp-model-weights.s3.eu-west-1.amazonaws.com/SPUTokenized_word_embedding_16k.matrix"
SPU_TOKENIZER_WORD_LOC = os.path.abspath(
os.path.join(
os.path.dirname(__file__),
"..",
"resources/SPU_word_tokenizer_16k.model",
)
)
TOKENIZER_LABEL_LOC = RESOURCES_PATH + "NER_label_tokenizer.json"
# Data Preprocessing Config
TOKEN_PIECE_MAX_LEN = 8
SENTENCE_MAX_LEN = 40
# Loading Tokenizers
spu_tokenizer_word = spm.SentencePieceProcessor(SPU_TOKENIZER_WORD_LOC)
tokenizer_label = load_keras_tokenizer(TOKENIZER_LABEL_LOC)
sp_key_to_index = {
spu_tokenizer_word.id_to_piece(id): id
for id in range(spu_tokenizer_word.get_piece_size())
}
sp_index_to_key = {
id: spu_tokenizer_word.id_to_piece(id)
for id in range(spu_tokenizer_word.get_piece_size())
}
LABEL_VOCAB_SIZE = len(tokenizer_label.word_index)
WORD_EMBEDDING_VOCAB_SIZE = len(sp_key_to_index)
WORD_EMBEDDING_VECTOR_SIZE = 128
WORD_EMBEDDING_MATRIX = np.zeros(
(WORD_EMBEDDING_VOCAB_SIZE, WORD_EMBEDDING_VECTOR_SIZE)
)
NUM_RNN_STACKS = 2
RNN_UNITS_MULTIPLIER = 2
NUM_RNN_UNITS = WORD_EMBEDDING_VECTOR_SIZE * RNN_UNITS_MULTIPLIER
FC_UNITS_MULTIPLIER = (2, 1)
DROPOUT = 0.2
[docs]class SPUContextNER:
"""
SentencePiece Unigram Context Named Entity Recognizer class.
- This is a context aware Deep GRU based Named Entity Recognizer that uses `SentencePiece Unigram <https://arxiv.org/abs/1804.10959>`_ tokenizer and pre-trained Word2Vec embeddings.
- It achieves 0.9928 Accuracy and 0.9833 F1 score on test sets of "wikiann", "gungor.ner" and "teghub-TurkishNER-BERT" datasets.
- It achieves 0.9766 F1 score for "ORG", 0.9852 F1 score for "PER" and 0.9742 F1 score for "LOC" entities. (Treating entity of interest as positive, all others as negative class.)
- For more details about the training procedure, dataset and evaluation metrics, see `ReadMe <https://github.com/vngrs-ai/VNLP/blob/main/vnlp/named_entity_recognizer/ReadMe.md>`_.
"""
def __init__(self, evaluate):
self.model = create_spucontext_ner_model(
TOKEN_PIECE_MAX_LEN,
SENTENCE_MAX_LEN,
WORD_EMBEDDING_VOCAB_SIZE,
LABEL_VOCAB_SIZE,
WORD_EMBEDDING_VECTOR_SIZE,
WORD_EMBEDDING_MATRIX,
NUM_RNN_UNITS,
NUM_RNN_STACKS,
FC_UNITS_MULTIPLIER,
DROPOUT,
)
# Check and download word embedding matrix and model weights
check_and_download(
WORD_EMBEDDING_MATRIX_LOC, WORD_EMBEDDING_MATRIX_LINK
)
if evaluate:
MODEL_WEIGHTS_LOC = EVAL_WEIGHTS_LOC
MODEL_WEIGHTS_LINK = EVAL_WEIGHTS_LINK
else:
MODEL_WEIGHTS_LOC = PROD_WEIGHTS_LOC
MODEL_WEIGHTS_LINK = PROD_WEIGHTS_LINK
check_and_download(MODEL_WEIGHTS_LOC, MODEL_WEIGHTS_LINK)
# Load Word embedding matrix
word_embedding_matrix = np.load(WORD_EMBEDDING_MATRIX_LOC)
# Load Model weights
with open(MODEL_WEIGHTS_LOC, "rb") as fp:
model_weights = pickle.load(fp)
# Insert word embedding weights to correct position (0 for SPUContextNER model)
model_weights.insert(0, word_embedding_matrix)
# Set model weights
self.model.set_weights(model_weights)
self.spu_tokenizer_word = spu_tokenizer_word
self.tokenizer_label = tokenizer_label
[docs] def predict(
self, sentence: str, displacy_format: bool = False
) -> List[Tuple[int, str, int, str]]:
"""
Args:
sentence:
Input sentence/text.
displacy_format:
When set True, returns the result in spacy.displacy format to allow visualization.
Returns:
NER result as pairs of (token, entity).
"""
tokenized_sentence = TreebankWordTokenize(sentence)
num_tokens_in_sentence = len(tokenized_sentence)
int_preds = []
for t in range(num_tokens_in_sentence):
# t is the index of the token/word
X = process_single_word_input(
t,
tokenized_sentence,
self.spu_tokenizer_word,
self.tokenizer_label,
int_preds,
)
# Predicting
logits = self.model(X).numpy()[0]
int_pred = np.argmax(logits, axis=-1)
int_preds.append(int_pred)
preds = [
self.tokenizer_label.sequences_to_texts([[int_pred]])[0]
for int_pred in int_preds
]
ner_result = [(t, e) for t, e in zip(tokenized_sentence, preds)]
if not displacy_format:
return ner_result
else:
return ner_to_displacy_format(sentence, ner_result)