In this lab, we'll use everything we've learned so far to build a model that can classify a text document as one of many possible classes!

You will be able to:

• Perform classification using a text dataset, using sensible preprocessing, tokenization, and feature engineering scheme
• Use scikit-learn text vectorizers to fit and transform text data into a format to be used in a ML model

For this lab, we'll be working with the classic Newsgroups Dataset, which is available as a training data set in sklearn.datasets. This dataset contains many different articles that fall into 1 of 20 possible classes. Our goal will be to build a classifier that can accurately predict the class of an article based on the features we create from the article itself!

Let's get started. Run the cell below to import everything we'll need for this lab.

import nltk
from nltk.corpus import stopwords
import string
from nltk import word_tokenize, FreqDist
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.metrics import accuracy_score
from sklearn.datasets import fetch_20newsgroups
from sklearn.ensemble import RandomForestClassifier
from sklearn.naive_bayes import MultinomialNB
import pandas as pd
import numpy as np
np.random.seed(0)

Now, we need to fetch our dataset. Run the cell below to download all the newsgroups articles and their corresponding labels. If this is the first time working with this dataset, scikit-learn will need to download all of the articles from an external repository -- the cell below may take a little while to run.

The actual dataset is quite large. To save us from extremely long runtimes, we'll work with only a subset of the classes. Here is a list of all the possible classes:

For this lab, we'll only work with the following five:

• 'alt.atheism'
• 'comp.windows.x'
• 'rec.sport.hockey'
• 'sci.crypt'
• 'talk.politics.guns'

In the cell below:

• Create a list called categories that contains the five newsgroups classes listed above, as strings
• Get the training set by calling fetch_20newsgroups() and passing in the following parameters:
  • subset='train'
  • categories=categories
  • remove=('headers', 'footers', 'quotes') -- this is so that the model can't overfit to metadata included in the articles that sometimes acts as a dead-giveaway as to what class the article belongs to
• Get the testing set as well by passing in the same parameters, with the exception of subset='test

categories = None
newsgroups_train = None
newsgroups_test = None

Great! Let's break apart the data and the labels, and then inspect the class names to see what the actual newsgroups are.

In the cell below:

• Grab the data from newsgroups_train.data and store it in the appropriate variable
• Grab the labels from newsgroups_train.target and store it in the appropriate variable
• Grab the label names from newsgroups_train.target_names and store it in the appropriate variable
• Display the label_names so that we can see the different classes of articles that we're working with, and confirm that we grabbed the right ones

data = None
target = None
label_names = None
label_names

Finally, let's check the shape of data to see what our data looks like. We can do this by checking the .shape attribute of newsgroups_train.filenames.

Do this now in the cell below.

# Your code here

(2814,)

Our dataset contains 2,814 different articles spread across the five classes we chose.

Now that we have our data, the fun part begins. We'll need to begin by preprocessing and cleaning our text data. As you've seen throughout this section, preprocessing text data is a bit more challenging that working with more traditional data types because there's no clear-cut answer for exactly what sort of preprocessing and cleaning we need to do. Before we can begin cleaning and preprocessing our text data, we need to make some decisions about things such as:

• Do we remove stop words or not?
• Do we stem or lemmatize our text data, or leave the words as is?
• Is basic tokenization enough, or do we need to support special edge cases through the use of regex?
• Do we use the entire vocabulary, or just limit the model to a subset of the most frequently used words? If so, how many?
• Do we engineer other features, such as bigrams, or POS tags, or Mutual Information Scores?
• What sort of vectorization should we use in our model? Boolean Vectorization? Count Vectorization? TF-IDF? More advanced vectorization strategies such as Word2Vec?

These are all questions that we'll need to think about pretty much anytime we begin working with text data.

Let's get right into it. We'll start by getting a list of all of the english stopwords, and concatenating them with a list of all the punctuation.

In the cell below:

• Get all the english stopwords from nltk
• Get all of the punctuation from string.punctuation, and convert it to a list
• Add the two lists together. Name the result stopwords_list
• Create another list containing various types of empty strings and ellipses, such as ["''", '""', '...', '``']. Add this to our stopwords_list, so that we won't have tokens that are only empty quotes and such

stopwords_list = None

Great! We'll leave these alone for now, until we're ready to remove stop words after the tokenization step.

Next, let's try tokenizing our dataset. In order to save ourselves some time, we'll write a function to clean our dataset, and then use Python's built-in map() function to clean every article in the dataset at the same time.

In the cell below, complete the process_article() function. This function should:

• Take in one parameter, article
• Tokenize the article using the appropriate function from nltk
• Lowercase every token, remove any stopwords found in stopwords_list from the tokenized article, and return the results

def process_article(article):
    pass  

Now that we have this function, let's go ahead and preprocess our data, and then move into exploring our dataset.

In the cell below:

• Use Python's map() function and pass in two parameters: the process_article function and the data. Make sure to wrap the whole map statement in a list().

Note: Running this cell may take a minute or two!

processed_data = None

Great. Now, let's inspect the first article in processed_data to see how it looks.

Do this now in the cell below.

processed_data[0]

Now, let's move onto exploring the dataset a bit more. Let's start by getting the total vocabulary size of the training dataset. We can do this by creating a set object and then using it's .update() method to iteratively add each article. Since it's a set, it will only contain unique words, with no duplicates.

In the cell below:

• Create a set() object called total_vocab
• Iterate through each tokenized article in processed_data and add it to the set using the set's .update() method
• Once all articles have been added, get the total number of unique words in our training set by taking the length of the set

total_vocab = None

Great -- our processed dataset contains 46,990 unique words!

Next, let's create a frequency distribution to see which words are used the most!

In order to do this, we'll need to concatenate every article into a single list, and then pass this list to FreqDist().

In the cell below:

• Create an empty list called articles_concat
• Iterate through processed_data and add every article it contains to articles_concat
• Pass articles_concat as input to FreqDist()
• Display the top 200 most used words

articles_concat = None

articles_freqdist = None

At first glance, none of these words seem very informative -- for most of the words represented here, it would be tough to guess if a given word is used equally among all five classes, or is disproportionately represented among a single class. This makes sense, because this frequency distribution represents all the classes combined. This tells us that these words are probably the least important, as they are most likely words that are used across multiple classes, thereby providing our model with little actual signal as to what class they belong to. This tells us that we probably want to focus on words that appear heavily in articles from a given class, but rarely appear in articles from other classes. You may recall from previous lessons that this is exactly where TF-IDF Vectorization really shines!

Although NLTK does provide functionality for vectorizing text documents with TF-IDF, we'll make use of scikit-learn's TF-IDF vectorizer, because we already have experience with it, and because it's a bit easier to use, especially when the models we'll be feeding the vectorized data into are from scikit-learn, meaning that we don't have to worry about doing any extra processing to ensure they play nicely together.

Recall that in order to use scikit-learn's TfidfVectorizer(), we need to pass in the data as raw text documents -- the TfidfVectorizer() handles the count vectorization process on it's own, and then fits and transforms the data into TF-IDF format.

This means that we need to:

• Import TfidfVectorizer from sklearn.feature_extraction.text and instantiate TfidfVectorizer()
• Call the vectorizer object's .fit_transform() method and pass in our data as input. Store the results in tf_idf_data_train
• Also create a vectorized version of our testing data, which can be found in newsgroups_test.data. Store the results in tf_idf_data_test.

NOTE: When transforming the test data, use the .transform() method, not the .fit_transform() method, as the vectorizer has already been fit to the training data.

# Import TfidfVectorizer

vectorizer = None

tf_idf_data_train = None

tf_idf_data_test = None

Great! We've now preprocessed and explored our dataset, let's take a second to see what our data looks like in vectorized form.

In the cell below, get the shape of tf_idf_data.

# Your code here

(2814, 36622)

Our vectorized data contains 2,814 articles, with 36,622 unique words in the vocabulary. However, the vast majority of these columns for any given article will be zero, since every article only contains a small subset of the total vocabulary. Recall that vectors mostly filled with zeros are referred to as Sparse Vectors. These are extremely common when working with text data.

Let's check out the average number of non-zero columns in the vectors. Run the cell below to calculate this average.

non_zero_cols = tf_idf_data_train.nnz / float(tf_idf_data_train.shape[0])
print("Average Number of Non-Zero Elements in Vectorized Articles: {}".format(non_zero_cols))

percent_sparse = 1 - (non_zero_cols / float(tf_idf_data_train.shape[1]))
print('Percentage of columns containing 0: {}'.format(percent_sparse))

As we can see from the output above, the average vectorized article contains 107 non-zero columns. This means that 99.7% of each vector is actually zeroes! This is one reason why it's best not to create your own vectorizers, and rely on professional packages such as scikit-learn and NLTK instead -- they contain many speed and memory optimizations specifically for dealing with sparse vectors. This way, we aren't wasting a giant chunk of memory on a vectorized dataset that only has valid information in 0.3% of it.

Now that we've vectorized our dataset, let's create some models and fit them to our vectorized training data.

In the cell below:

• Instantiate MultinomialNB() and RandomForestClassifier(). For random forest, set n_estimators to 100. Don't worry about tweaking any of the other parameters
• Fit each to our vectorized training data
• Create predictions for our training and test sets
• Calculate the accuracy_score() for both the training and test sets (you'll find our training labels stored within the variable target, and the test labels stored within newsgroups_test.target)

nb_classifier = None
rf_classifier = None

nb_train_preds = None
nb_test_preds = None

rf_train_preds = None
rf_test_preds = None

nb_train_score = None
nb_test_score = None
rf_train_score = None
rf_test_score = None

print("Multinomial Naive Bayes")
print("Training Accuracy: {:.4} \t\t Testing Accuracy: {:.4}".format(nb_train_score, nb_test_score))
print("")
print('-'*70)
print("")
print('Random Forest')
print("Training Accuracy: {:.4} \t\t Testing Accuracy: {:.4}".format(rf_train_score, rf_test_score))

Question: Interpret the results seen above. How well did the models do? How do they compare to random guessing? How would you describe the quality of the model fit?

Write your answer below:

# Your answer here

In this lab, we used our NLP skills to clean, preprocess, explore, and fit models to text data for classification. This wasn't easy -- great job!!