Transformer Network Application: Named-Entity Recognition¶

import pandas as pd
import numpy as np
import tensorflow as tf
import json
import random
import logging
import re

tf.get_logger().setLevel('ERROR')

df_data = pd.read_json("ner.json", lines=True)
df_data = df_data.drop(['extras'], axis=1)
df_data['content'] = df_data['content'].str.replace("\n", " ")

def mergeIntervals(intervals):
    sorted_by_lower_bound = sorted(intervals, key=lambda tup: tup[0])
    merged = []

    for higher in sorted_by_lower_bound:
        if not merged:
            merged.append(higher)
        else:
            lower = merged[-1]
            if higher[0] <= lower[1]:
                if lower[2] is higher[2]:
                    upper_bound = max(lower[1], higher[1])
                    merged[-1] = (lower[0], upper_bound, lower[2])
                else:
                    if lower[1] > higher[1]:
                        merged[-1] = lower
                    else:
                        merged[-1] = (lower[0], higher[1], higher[2])
            else:
                merged.append(higher)
    return merged

def get_entities(df):
    
    entities = []
    
    for i in range(len(df)):
        entity = []
    
        for annot in df['annotation'][i]:
            try:
                ent = annot['label'][0]
                start = annot['points'][0]['start']
                end = annot['points'][0]['end'] + 1
                entity.append((start, end, ent))
            except:
                pass
    
        entity = mergeIntervals(entity)
        entities.append(entity)
    
    return entities

def convert_dataturks_to_spacy(dataturks_JSON_FilePath):
    try:
        training_data = []
        lines=[]
        with open(dataturks_JSON_FilePath, 'r') as f:
            lines = f.readlines()

        for line in lines:
            data = json.loads(line)
            text = data['content'].replace("\n", " ")
            entities = []
            data_annotations = data['annotation']
            if data_annotations is not None:
                for annotation in data_annotations:
                    #only a single point in text annotation.
                    point = annotation['points'][0]
                    labels = annotation['label']
                    # handle both list of labels or a single label.
                    if not isinstance(labels, list):
                        labels = [labels]

                    for label in labels:
                        point_start = point['start']
                        point_end = point['end']
                        point_text = point['text']
                        
                        lstrip_diff = len(point_text) - len(point_text.lstrip())
                        rstrip_diff = len(point_text) - len(point_text.rstrip())
                        if lstrip_diff != 0:
                            point_start = point_start + lstrip_diff
                        if rstrip_diff != 0:
                            point_end = point_end - rstrip_diff
                        entities.append((point_start, point_end + 1 , label))
            training_data.append((text, {"entities" : entities}))
        return training_data
    except Exception as e:
        logging.exception("Unable to process " + dataturks_JSON_FilePath + "\n" + "error = " + str(e))
        return None

def trim_entity_spans(data: list) -> list:
    """Removes leading and trailing white spaces from entity spans.

    Args:
        data (list): The data to be cleaned in spaCy JSON format.

    Returns:
        list: The cleaned data.
    """
    invalid_span_tokens = re.compile(r'\s')

    cleaned_data = []
    for text, annotations in data:
        entities = annotations['entities']
        valid_entities = []
        for start, end, label in entities:
            valid_start = start
            valid_end = end
            while valid_start < len(text) and invalid_span_tokens.match(
                    text[valid_start]):
                valid_start += 1
            while valid_end > 1 and invalid_span_tokens.match(
                    text[valid_end - 1]):
                valid_end -= 1
            valid_entities.append([valid_start, valid_end, label])
        cleaned_data.append([text, {'entities': valid_entities}])
    return cleaned_data

data = trim_entity_spans(convert_dataturks_to_spacy("ner.json"))

from tqdm.notebook import tqdm
def clean_dataset(data):
    cleanedDF = pd.DataFrame(columns=["setences_cleaned"])
    sum1 = 0
    for i in tqdm(range(len(data))):
        start = 0
        emptyList = ["Empty"] * len(data[i][0].split())
        numberOfWords = 0
        lenOfString = len(data[i][0])
        strData = data[i][0]
        strDictData = data[i][1]
        lastIndexOfSpace = strData.rfind(' ')
        for i in range(lenOfString):
            if (strData[i]==" " and strData[i+1]!=" "):
                for k,v in strDictData.items():
                    for j in range(len(v)):
                        entList = v[len(v)-j-1]
                        if (start>=int(entList[0]) and i<=int(entList[1])):
                            emptyList[numberOfWords] = entList[2]
                            break
                        else:
                            continue
                start = i + 1  
                numberOfWords += 1
            if (i == lastIndexOfSpace):
                for j in range(len(v)):
                        entList = v[len(v)-j-1]
                        if (lastIndexOfSpace>=int(entList[0]) and lenOfString<=int(entList[1])):
                            emptyList[numberOfWords] = entList[2]
                            numberOfWords += 1
        cleanedDF = cleanedDF.append(pd.Series([emptyList],  index=cleanedDF.columns ), ignore_index=True )
        sum1 = sum1 + numberOfWords
    return cleanedDF

unique_tags = set(cleanedDF['setences_cleaned'].explode().unique())#pd.unique(cleanedDF['setences_cleaned'])#set(tag for doc in cleanedDF['setences_cleaned'].values.tolist() for tag in doc)
tag2id = {tag: id for id, tag in enumerate(unique_tags)}
id2tag = {id: tag for tag, id in tag2id.items()}

from tensorflow.keras.preprocessing.sequence import pad_sequences

MAX_LEN = 512
labels = cleanedDF['setences_cleaned'].values.tolist()

tags = pad_sequences([[tag2id.get(l) for l in lab] for lab in labels],
                     maxlen=MAX_LEN, value=tag2id["Empty"], padding="post",
                     dtype="long", truncating="post")

gpus = tf.config.list_physical_devices('GPU')
if gpus:
    for gpu in gpus:
        tf.config.experimental.set_virtual_device_configuration(gpu,[tf.config.experimental.VirtualDeviceConfiguration(memory_limit=4096)])

from transformers import DistilBertTokenizerFast #, TFDistilBertModel
tokenizer = DistilBertTokenizerFast.from_pretrained('tokenizer/')

label_all_tokens = True
def tokenize_and_align_labels(tokenizer, examples, tags):
    tokenized_inputs = tokenizer(examples, truncation=True, is_split_into_words=False, padding='max_length', max_length=512)
    labels = []
    for i, label in enumerate(tags):
        word_ids = tokenized_inputs.word_ids(batch_index=i)
        previous_word_idx = None
        label_ids = []
        for word_idx in word_ids:
            # Special tokens have a word id that is None. We set the label to -100 so they are automatically
            # ignored in the loss function.
            if word_idx is None:
                label_ids.append(-100)
            # We set the label for the first token of each word.
            elif word_idx != previous_word_idx:
                label_ids.append(label[word_idx])
            # For the other tokens in a word, we set the label to either the current label or -100, depending on
            # the label_all_tokens flag.
            else:
                label_ids.append(label[word_idx] if label_all_tokens else -100)
            previous_word_idx = word_idx

        labels.append(label_ids)

    tokenized_inputs["labels"] = labels
    return tokenized_inputs

test = tokenize_and_align_labels(tokenizer, df_data['content'].values.tolist(), tags)
train_dataset = tf.data.Dataset.from_tensor_slices((
    test['input_ids'],
    test['labels']
))

text = "Manisha Bharti. 3.5 years of professional IT experience in Banking and Finance domain"
inputs = tokenizer(text, return_tensors="tf", truncation=True, is_split_into_words=False, padding="max_length", max_length=512 )
input_ids = inputs["input_ids"]
#inputs["labels"] = tf.reshape(tf.constant([1] * tf.size(input_ids).numpy()), (-1, tf.size(input_ids)))

pred_labels = []

predictions = np.argmax(output['logits'].reshape(220, -1, 12), axis=-1)

from seqeval.metrics import f1_score, classification_report