2.4.3. Working with text data¶
The goal of this section is to explore some of the main scikit-learn tools on a single practical task: analysing a collection of text documents (newsgroups posts) on twenty different topics.
In this section we will see how to:
- load the file contents and the categories
- extract feature vectors suitable for machine learning
- train a linear model to perform categorization
- use a grid search strategy to find a good configuration of both the feature extraction components and the classifier
2.4.3.1. Downloading the data and loading it from Python¶
The dataset is called “Twenty Newsgroups”. Here is the official description, quoted from the website:
The 20 Newsgroups data set is a collection of approximately 20,000 newsgroup documents, partitioned (nearly) evenly across 20 different newsgroups. To the best of our knowledge, it was originally collected by Ken Lang, probably for his paper “Newsweeder: Learning to filter netnews,” though he does not explicitly mention this collection. The 20 newsgroups collection has become a popular data set for experiments in text applications of machine learning techniques, such as text classification and text clustering.
To download the dataset, go to $TUTORIAL_HOME/data/twenty_newsgroups and run the fetch_data.py script.
In the following we will use the built-in dataset loader for 20 newsgroups from scikit-learn. Alternatively it is possible to download the dataset manually from the web-site and use the sklearn.datasets.load_files function by pointing it to the 20news-bydate-train subfolder of the uncompressed archive folder.
In order to get faster execution times for this first example we will work on a partial dataset with only 4 categories out of the 20 available in the dataset:
>>> categories = ['alt.atheism', 'soc.religion.christian',
... 'comp.graphics', 'sci.med']
We can now load the list of files matching those categories as follows:
>>> from sklearn.datasets import fetch_20newsgroups
>>> twenty_train = fetch_20newsgroups(
... subset='train', categories=categories,
... shuffle=True, random_state=42)
The returned dataset is a scikit-learn “bunch”: a simple holder object with fields that can be both accessed as python dict keys or object attributes for convenience, for instance the target_names holds the list of the requested category names:
>>> twenty_train.target_names
['alt.atheism', 'comp.graphics', 'sci.med', 'soc.religion.christian']
The files themselves are loaded in memory in the data attribute. For reference the filenames are also available:
>>> len(twenty_train.data)
2257
>>> len(twenty_train.filenames)
2257
Let’s print the first lines of the first loaded file:
>>> print "\n".join(twenty_train.data[0].split("\n")[:2])
From: sd345@city.ac.uk (Michael Collier)
Subject: Converting images to HP LaserJet III?
>>> print twenty_train.target_names[twenty_train.target[0]]
comp.graphics
Supervised learning algorithms will require a category label for each document in the training set. In this case the category is the name of the newsgroup which also happens to be the name of the folder holding the individual documents.
For speed and space efficiency reasons scikit-learn loads the target attribute as an array of integers that corresponds to the index of the category name in the target_names list. The category integer id of each sample is stored in the target attribute:
>>> twenty_train.target[:10]
array([1, 1, 3, 3, 3, 3, 3, 2, 2, 2])
It is possible to get back the category names as follows:
>>> for t in twenty_train.target[:10]:
... print twenty_train.target_names[t]
...
comp.graphics
comp.graphics
soc.religion.christian
soc.religion.christian
soc.religion.christian
soc.religion.christian
soc.religion.christian
sci.med
sci.med
sci.med
You can notice that the samples have been shuffled randomly (with a fixed RNG seed): this is useful if you select only the first samples to quickly train a model and get a first idea of the results before re-training on the complete dataset later.
2.4.3.2. Extracting features from text files¶
In order to perform machine learning on text documents, we first need to turn the text content into numerical feature vectors.
2.4.3.2.1. Bags of words¶
The most intuitive way to do so is the bags of words representation:
- assign a fixed integer id to each word occurring in any document of the training set (for instance by building a dictionary from words to integer indices).
- for each document #i, count the number of occurrences of each word w and store it in X[i, j] as the value of feature #j where j is the index of word w in the dictionary
The bags of words representation implies that n_features is the number of distinct words in the corpus: this number is typically larger that 100,000.
If n_samples == 10000, storing X as a numpy array of type float32 would require 10000 x 100000 x 4 bytes = 4GB in RAM which is barely manageable on today’s computers.
Fortunately, most values in X will be zeros since for a given document less than a couple thousands of distinct words will be used. For this reason we say that bags of words are typically high-dimensional sparse datasets. We can save a lot of memory by only storing the non-zero parts of the feature vectors in memory.
scipy.sparse matrices are data structures that do exactly this, and scikit-learn has built-in support for these structures.
2.4.3.2.2. Tokenizing text with scikit-learn¶
scikit-learn offers a provides basic tools to process text using the Bag of Words representation.
To build such a representation we will proceed as follows:
- tokenize strings and give an integer id for each possible token, for instance by using whitespaces and punctuation as token separators.
- count the occurrences of tokens in each document.
- normalize and weighting with diminishing importance tokens that occur in the majority of samples / documents.
In order to do the first two steps, scikit-learn provides the :class:sklearn.feature_extraction.text.CountVectorizer class:
>>> from sklearn.feature_extraction.text import CountVectorizer
This class exposes many utility functions, in particular the analyzer function used for tokenizing the text:
>>> analyze = CountVectorizer().build_analyzer()
>>> text = "A WONDERFUL test phrase!"
>>> analyze(text)
[u'wonderful', u'test', u'phrase']
Note that punctuation and single letter words have automatically been removed.
It is further possible to configure CountVectorizer to extract n-grams instead of single words:
>>> CountVectorizer(min_n=1, max_n=2).build_analyzer()(text)
[u'wonderful', u'test', u'phrase', u'wonderful test', u'test phrase']
The analyzer is used internally by CountVectorizer to build a dictionary of features and transform documents to feature vectors:
>>> count_vect = CountVectorizer(stop_words='english')
>>> X_train_counts = count_vect.fit_transform(twenty_train.data)
>>> X_train_counts.shape
(2257, 35481)
Once fitted, the vectorizer has built a dictionary of feature indices:
>>> count_vect.vocabulary_.get(u'algorithm')
4683
The index value of a word in the vocabulary is linked to its frequency in the whole training corpus.
2.4.3.2.3. From occurrence counts to normalized frequencies¶
Occurrence count is a good start but there is an issue: longer documents will have higher average count values than shorter documents, even though they might talk about the same topics.
To avoid these potential discrepancies it suffices to divide the number of occurrences of each word in a document by the total number of words in the document: these new features are called “tf” for Term Frequencies.
Another refinement on top of tf is to downscale weights for words that occur in many documents in the corpus and are therefore less informative than those that occur only in a smaller portion of the corpus.
This downscaling is called tf–idf for “Term Frequency times Inverse Document Frequency”.
Both tf and tf–idf can be computed as follows:
>>> from sklearn.feature_extraction.text import TfidfTransformer
>>> tf_transformer = TfidfTransformer(use_idf=False).fit(X_train_counts)
>>> X_train_tf = tf_transformer.transform(X_train_counts)
>>> X_train_tf.shape
(2257, 35481)
>>> tfidf_transformer = TfidfTransformer()
>>> X_train_tfidf = tfidf_transformer.fit_transform(X_train_counts)
>>> X_train_tfidf.shape
(2257, 35481)
2.4.3.3. Training a classifier¶
Now that we have our feature, we can train a classifier to try to predict the category of a post. Let’s start with a naïve Bayes classifier, which provides a nice baseline for this task. scikit-learn includes several variants of this classifier; the one most suitable for word counts is the multinomial variant:
>>> from sklearn.naive_bayes import MultinomialNB
>>> clf = MultinomialNB().fit(X_train_tfidf, twenty_train.target)
To try to predict the outcome on a new document we need to extract the features using almost the same feature extracting chain as before. The difference is that we call transform instead of fit_transform on the transformers, since they have already been fit to the training set:
>>> docs_new = ['God is love', 'OpenGL on the GPU is fast']
>>> X_new_counts = count_vect.transform(docs_new)
>>> X_new_tfidf = tfidf_transformer.transform(X_new_counts)
>>> predicted = clf.predict(X_new_tfidf)
>>> for doc, category in zip(docs_new, predicted):
... print '%r => %s' % (doc, twenty_train.target_names[category])
...
'God is love' => soc.religion.christian
'OpenGL on the GPU is fast' => comp.graphics
2.4.3.4. Building a pipeline¶
In order to make the vectorizer => transformer => classifier easier to work with, scikit-learn provides a Pipeline class that behaves like a compound classifier:
>>> from sklearn.pipeline import Pipeline
>>> text_clf = Pipeline([
... ('vect', CountVectorizer()),
... ('tfidf', TfidfTransformer()),
... ('clf', MultinomialNB()),
... ])
The names vect, tfidf and clf (classifier) are arbitrary. We shall see their use in the section on grid search, below. We can now train the model with a single command:
>>> _ = text_clf.fit(twenty_train.data, twenty_train.target)
2.4.3.5. Evaluation of the performance on the test set¶
Evaluating the predictive accuracy of the model is equally easy:
>>> import numpy as np
>>> twenty_test = fetch_20newsgroups(
... subset='test', categories=categories,
... shuffle=True, random_state=42)
>>> docs_test = twenty_test.data
>>> predicted = text_clf.predict(docs_test)
>>> np.mean(predicted == twenty_test.target)
0.834...
I.e., we achieved 89.4% accuracy. Let’s see if we can do better with a linear support vector machine (SVM), which is widely regarded as one of the best text classification algorithms (although it’s also a bit slower than naïve Bayes). We can change the learner by just plugging a different classifier object into our pipeline:
>>> from sklearn.linear_model import SGDClassifier
>>> text_clf = Pipeline([
... ('vect', CountVectorizer()),
... ('tfidf', TfidfTransformer()),
... ('clf', SGDClassifier(loss='hinge', penalty='l2',
... alpha=1e-3, n_iter=5)),
... ])
>>> _ = text_clf.fit(twenty_train.data, twenty_train.target)
>>> predicted = text_clf.predict(docs_test)
>>> np.mean(predicted == twenty_test.target)
0.912...
scikit-learn further provides utilities for more detailed performance analysis of the results:
>>> from sklearn import metrics
>>> print metrics.classification_report(
... twenty_test.target, predicted,
... target_names=twenty_test.target_names)
...
precision recall f1-score support
alt.atheism 0.93 0.82 0.87 319
comp.graphics 0.88 0.98 0.93 389
sci.med 0.95 0.89 0.92 396
soc.religion.christian 0.90 0.95 0.92 398
avg / total 0.92 0.91 0.91 1502
>>> metrics.confusion_matrix(twenty_test.target, predicted)
array([[261, 10, 12, 36],
[ 5, 380, 2, 2],
[ 7, 33, 352, 4],
[ 7, 9, 4, 378]])
As expected the confusion matrix shows that posts from the newsgroups on atheism and christian are more often confused for one another than with computer graphics.
2.4.3.6. Parameter tuning using grid search¶
We’ve already encountered some parameters such as use_idf in the TfidfTransformer. Classifiers tend to have many parameters as well; e.g., MultinomialNB includes a smoothing parameter alpha and SGDClassifier has a penalty parameter alpha and configurable loss and penalty terms in the objective function (see the module documentation, or use the Python help function, to get a description of these).
Instead of tweaking the parameters of the various components of the chain, it is possible to run an exhaustive search of the best parameters on a grid of possible values. We try out all classifiers on either words or bigrams, with or without idf, and with a penalty parameter of either 100 or 1000 for the linear SVM:
>>> from sklearn.grid_search import GridSearchCV
>>> parameters = {
... 'vect__max_n': (1, 2),
... 'tfidf__use_idf': (True, False),
... 'clf__alpha': (1e-2, 1e-3),
... }
Obviously, such an exhaustive search can be expensive. If we have multiple CPU cores at our disposal, we can tell the grid searcher to try these eight parameter combinations in parallel with the n_jobs parameter. If we give this parameter a value of -1, grid search will detect how many cores are installed and uses them all:
>>> gs_clf = GridSearchCV(text_clf, parameters, n_jobs=1)
The grid search instance behaves like a normal scikit-learn model. Let’s perform the search on a smaller subset of the training data to speed up the computation:
>>> gs_clf = gs_clf.fit(twenty_train.data[:400], twenty_train.target[:400])
The result of calling fit on a GridSearchCV object is a classifier that we can use to predict:
>>> twenty_train.target_names[gs_clf.predict(['God is love'])]
'soc.religion.christian'
but otherwise, it’s a pretty large and clumsy object. We can, however, get the optimal parameters out by inspecting the object’s grid_scores_ attribute, which is a list of parameters/score pairs. To get the best scoring attributes, we can do:
>>> best_parameters, score, _ = max(gs_clf.grid_scores_, key=lambda x: x[1])
>>> for param_name in sorted(parameters.keys()):
... print "%s: %r" % (param_name, best_parameters[param_name])
...
clf__alpha: 0.001
tfidf__use_idf: True
vect__max_n: 1
>>> score
0.910...