Creating a Taxonomy for BBC News Articles (Part 6 based on - A Hybrid Approach to Hypernym Discovery)

The difference between Part 5 and Part 6.
In Part 5, we were using Cosine Distance between input text and output label. 
In Part 6, we are first finding the dot product and then getting a probability using the sigmoid function similar to Logistic Regression.

A sigmoid function is a mathematical function having a characteristic "S"-shaped curve or sigmoid curve.






import pandas as pd
import numpy as np
from sentence_transformers import SentenceTransformer

from sklearn.metrics.pairwise import cosine_similarity # Expects 2D arrays as input
from scipy.spatial.distance import cosine # Works with 1D vectors

from sklearn.metrics import classification_report

smodel = SentenceTransformer('distilbert-base-nli-mean-tokens')

df1 = pd.read_csv('bbc_news_train.csv')

df1.head()





def get_sentence_vector(query):
    query_vec = smodel.encode([query])[0]
    return query_vec

df1['textVec'] = df1['Text'].apply(lambda x: get_sentence_vector(x))

def std_category(x):
    if(x == 'tech'):
        return 'technology'
    elif (x == 'sport'):
        return 'sports'
    else:
        return x

df1['Category'] = df1['Category'].apply(std_category)

import math
def sigmoid(x):
    return 1 / (1 + math.exp(-x))

def get_logistic_regression_probability(x, Y):
    y = smodel.encode([Y])[0]
    d = np.dot(x, y)
    
    s = sigmoid(d)
    
    return s

df1['proba_business'] = df1['textVec'].apply(lambda x: get_logistic_regression_probability(x, 'business')) # CPU times: total: 2min 1s. Wall time: 1min 1s

df1['Category'].unique()

# OUTPUT: array(['business', 'technology', 'politics', 'sports', 'entertainment'], dtype=object)

df1['proba_technology'] = df1['textVec'].apply(lambda x: get_logistic_regression_probability(x, 'technology'))

df1['proba_politics'] = df1['textVec'].apply(lambda x: get_logistic_regression_probability(x, 'politics'))

df1['proba_sports'] = df1['textVec'].apply(lambda x: get_logistic_regression_probability(x, 'sports'))

df1['proba_entertainment'] = df1['textVec'].apply(lambda x: get_logistic_regression_probability(x, 'entertainment'))

def get_prediction(in_row):
    max_proba = 0
    label = ""
    for i in ['proba_business', 'proba_technology', 'proba_politics', 'proba_sports', 'proba_entertainment']:
        d = in_row[i]
        if d > max_proba:
            max_proba = d
            label = i.split('_')[1]
    return label

df1['prediction'] = df1.apply(lambda in_row: get_prediction(in_row), axis = 1)

target_names = ['business', 'entertainment', 'politics', 'sports', 'technology']
print(classification_report(df1['Category'], df1['prediction'], target_names=target_names))

Tags: Technology,Natural Language Processing,