from typing import List, Tuple
import pickle
import tensorflow as tf
import numpy as np
from ..stemmer_morph_analyzer import StemmerAnalyzer
from ..part_of_speech_tagger import PoSTagger
from ..tokenizer import TreebankWordTokenize
from ..utils import check_and_download, load_keras_tokenizer
from .utils import dp_pos_to_displacy_format, decode_arc_label_vector
from ._treestack_utils import (create_dependency_parser_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 + "DP_TreeStack_prod.weights"
EVAL_WEIGHTS_LOC = RESOURCES_PATH + "DP_TreeStack_eval.weights"
WORD_EMBEDDING_MATRIX_LOC = os.path.abspath(os.path.join(os.path.dirname( __file__ ), '..', 'resources/TBWTokenized_word_embedding.matrix'))
PROD_WEIGHTS_LINK = "https://vnlp-model-weights.s3.eu-west-1.amazonaws.com/DP_TreeStack_prod.weights"
EVAL_WEIGHTS_LINK = "https://vnlp-model-weights.s3.eu-west-1.amazonaws.com/DP_TreeStack_eval.weights"
WORD_EMBEDDING_MATRIX_LINK = "https://vnlp-model-weights.s3.eu-west-1.amazonaws.com/TBWTokenized_word_embedding.matrix"
TOKENIZER_WORD_LOC = os.path.abspath(os.path.join(os.path.dirname( __file__ ), '..', 'resources/TB_word_tokenizer.json'))
TOKENIZER_POS_LOC = os.path.abspath(os.path.join(os.path.dirname( __file__ ), '..', 'part_of_speech_tagger/resources/PoS_label_tokenizer.json')) # using the tokenizer of part_of_speech_tagger
TOKENIZER_TAG_LOC = os.path.abspath(os.path.join(os.path.dirname( __file__ ), '..', 'stemmer_morph_analyzer/resources/Stemmer_morph_tag_tokenizer.json')) # using the tokenizer of stemmer_morph_analyzer
TOKENIZER_LABEL_LOC = RESOURCES_PATH + "DP_label_tokenizer.json"
# Data Preprocessing Config
SENTENCE_MAX_LEN = 40
TAG_MAX_LEN = 15
WORD_OOV_TOKEN = '<OOV>'
# Loading Tokenizers
# Have to load tokenizers here because model config depends on them
tokenizer_word = load_keras_tokenizer(TOKENIZER_WORD_LOC)
tokenizer_pos = load_keras_tokenizer(TOKENIZER_POS_LOC)
tokenizer_tag = load_keras_tokenizer(TOKENIZER_TAG_LOC)
tokenizer_label = load_keras_tokenizer(TOKENIZER_LABEL_LOC)
LABEL_VOCAB_SIZE = len(tokenizer_label.word_index)
POS_VOCAB_SIZE = len(tokenizer_pos.word_index)
# Model Config
WORD_EMBEDDING_VECTOR_SIZE = 128 # Word2Vec_medium.model
WORD_EMBEDDING_VOCAB_SIZE = 63_992 # Word2Vec_medium.model
# WORD_EMBEDDING_MATRIX and TAG_EMBEDDING MATRIX are initialized as Zeros, will be overwritten when model is loaded.
WORD_EMBEDDING_MATRIX = np.zeros((WORD_EMBEDDING_VOCAB_SIZE, WORD_EMBEDDING_VECTOR_SIZE))
TAG_EMBEDDING_MATRIX = np.zeros((127, 32))
POS_EMBEDDING_VECTOR_SIZE = 8
NUM_RNN_STACKS = 2
RNN_UNITS_MULTIPLIER = 3
TAG_NUM_RNN_UNITS = WORD_EMBEDDING_VECTOR_SIZE
LC_NUM_RNN_UNITS = TAG_NUM_RNN_UNITS * RNN_UNITS_MULTIPLIER
LC_ARC_LABEL_NUM_RNN_UNITS = TAG_NUM_RNN_UNITS * RNN_UNITS_MULTIPLIER
RC_NUM_RNN_UNITS = TAG_NUM_RNN_UNITS * RNN_UNITS_MULTIPLIER
ARC_LABEL_VECTOR_LEN = SENTENCE_MAX_LEN + 1 + LABEL_VOCAB_SIZE + 1
FC_UNITS_MULTIPLIERS = (8, 4)
WORD_FORM = 'whole'
DROPOUT = 0.2
[docs]class TreeStackDP:
"""
Tree-stack Dependency Parser class.
- This dependency parser is *inspired* by `Tree-stack LSTM in Transition Based Dependency Parsing <https://aclanthology.org/K18-2012/>`_.
- "Inspire" is emphasized because this implementation uses the approach of using Morphological Tags, Pre-trained word embeddings and POS tags as input for the model, rather than implementing the exact network proposed in the paper.
- It achieves 0.6914 LAS (Labeled Attachment Score) and 0.8048 UAS (Unlabeled Attachment Score) on all of test sets of Universal Dependencies 2.9.
- Input data is processed by NLTK.tokenize.TreebankWordTokenizer.
- For more details about the training procedure, dataset and evaluation metrics, see `ReadMe <https://github.com/vngrs-ai/VNLP/blob/main/vnlp/dependency_parser/ReadMe.md>`_.
"""
def __init__(self, evaluate):
self.model = create_dependency_parser_model(WORD_EMBEDDING_VOCAB_SIZE, WORD_EMBEDDING_VECTOR_SIZE, WORD_EMBEDDING_MATRIX,
POS_VOCAB_SIZE, POS_EMBEDDING_VECTOR_SIZE,
SENTENCE_MAX_LEN, TAG_MAX_LEN, ARC_LABEL_VECTOR_LEN, NUM_RNN_STACKS,
TAG_NUM_RNN_UNITS, LC_NUM_RNN_UNITS, LC_ARC_LABEL_NUM_RNN_UNITS, RC_NUM_RNN_UNITS,
DROPOUT, TAG_EMBEDDING_MATRIX, FC_UNITS_MULTIPLIERS)
# 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 (1 for TreeStack Dependency Parsing model)
model_weights.insert(1, word_embedding_matrix)
# Set model weights
self.model.set_weights(model_weights)
self.tokenizer_word = tokenizer_word
self.tokenizer_tag = tokenizer_tag
self.tokenizer_pos = tokenizer_pos
self.tokenizer_label = tokenizer_label
# I don't want StemmerAnalyzer and PosTagger to occupy any memory in GPU!
with tf.device('/cpu:0'):
stemmer_analyzer = StemmerAnalyzer()
self.stemmer_analyzer = stemmer_analyzer
# stemmer_analyzer is passed to PoSTagger to prevent chain stemmer_analyzer initializations
pos_tagger = PoSTagger('TreeStackPoS', evaluate, self.stemmer_analyzer)
self.pos_tagger = pos_tagger
[docs] def predict(self, sentence: str, displacy_format: bool = False, *args) -> List[Tuple[int, str, int, str]]:
"""
Args:
sentence:
Input sentence.
displacy_format:
When set True, returns the result in spacy.displacy format to allow visualization.
Returns:
List of (token_index, token, arc, label).
Raises:
ValueError: Sentence is too long. Try again by splitting it into smaller pieces.
"""
sentence_word_punct_tokenized = TreebankWordTokenize(sentence)
sentence_analysis_result = self.stemmer_analyzer.predict(sentence)
sentence_analysis_result = [sentence_analysis.replace('^', '+') for sentence_analysis in sentence_analysis_result]
num_tokens_in_sentence = len(sentence_word_punct_tokenized)
if num_tokens_in_sentence > SENTENCE_MAX_LEN:
raise ValueError('Sentence is too long. Try again by splitting it into smaller pieces.')
# This is for debugging purposes in case a consistency occurs during tokenization.
if not len(sentence_analysis_result) == num_tokens_in_sentence:
raise Exception(sentence, "Length of sentence and sentence_analysis_result don't match")
# *args exist for API compatability. pos_result is overwritten here.
pos_result = self.pos_tagger.predict(sentence)
pos_tags = [pos for (word, pos) in pos_result]
arcs = []
labels = []
for t in range(num_tokens_in_sentence):
# t is the index of token/word
X = process_single_word_input(t, sentence, sentence_analysis_result,
SENTENCE_MAX_LEN, TAG_MAX_LEN, ARC_LABEL_VECTOR_LEN,
self.tokenizer_word, self.tokenizer_tag, self.tokenizer_label, self.tokenizer_pos,
arcs, labels, pos_tags, WORD_FORM)
# Predicting
logits = self.model(X).numpy()[0]
arc, label = decode_arc_label_vector(logits, SENTENCE_MAX_LEN, LABEL_VOCAB_SIZE)
arcs.append(arc)
labels.append(label)
# 0 arc index is reserved for root, therefore arc = 1 means word is dependent on the first word
dp_result = []
for idx, word in enumerate(sentence_word_punct_tokenized):
dp_result.append((idx + 1, word, arcs[idx], self.tokenizer_label.sequences_to_texts([[labels[idx]]])[0]))
if not displacy_format:
return dp_result
else:
dp_result_displacy_format = dp_pos_to_displacy_format(dp_result, pos_result)
return dp_result_displacy_format