When you need to tackle an NLP task — say, text classification or sentiment analysis — the sheer number of available software options can be overwhelming. Task-specific packages, generic libraries and cloud APIs all claim to offer the best solution to your problem, and it can be hard to decide which one to use. In this blog post we’ll take a look at some of the available options for a text classification task, and discover their main advantages and disadvantages.
The NLP task in this blog post is a classic instance of text classification: we’ll teach the computer to detect the topic in a news article. The task itself is pretty straightforward: the topics (sports, finance, entertainment, etc.) are very general, and their number is limited to four or five. We’ll use two freely available data sets to train and test our classifiers: the AG News dataset and the Sogou News dataset. Our focus is on three available solutions: FastText, a command-line tool, Scikit-learn, a machine learning library for Python, and MonkeyLearn, a cloud service. We’ll keep parameter optimization to a minimum and evaluate the off-the-shelf models that the various solutions offer.
In July, Facebook’s AI lab released FastText, a command-line tool for building word embeddings and classifying text. The open-source software package has been available on Github since early August. Reactions to the release were mixed. One news source claimed “Facebook’s FastText AI is the Tesla of Natural Language Processing”, and “researchers just supercharged autonomous artificial intelligence”. Other people were less enthusiastic, and pointed out FastText is merely a re-implementation of existing techniques.
Revolutionary or not, FastText is extremely easy to use. When you’ve prepared your training data correctly (one piece of text per line, with the class of the text prefixed by “__label__”), you can train a classifier with a single command:
Testing your trained classifier on a set of held-out test data is also a breeze:
FastText surely lives up to its name. On my laptop, it takes 3 to 4 seconds to train a classifier on the single words of the 120,000 AG training texts. The 450,000 training samples in the Sogou corpus keep it busy for about 2 minutes. With one optional parameter setting, the model takes both single words and bigrams (2-word sequences) as features. This doubles the training time to 7 seconds for the AG corpus and 4 minutes for Sogou. That’s impressive.
The trained classifiers are also pretty accurate. In my experiments, the unigram model with the standard settings achieved an accuracy of 91.5% on the AG test set, and 93.2% on the Sogou test set. Adding bigrams as features led to an increase in accuracy of 0.2% for AG (91.7%), and 2.3% for Sogou (95.5%). These figures are slightly lower than those reported in the paper that accompanied the release of FastText, possibly due to slight differences in tokenization or parameter settings. The paper further demonstrates that this accuracy is state-of-the-art, not only on the AG and Sogou corpora, but also on other “sentiment” datasets (although, contrary to what the authors suggest, AG and Sogou are topical rather than sentiment data).
This state-of-the-art performance may be due to FastText’s reliance on neural networks rather than more traditional, linear models, where parameters cannot be shared among features. Its networks learn low-dimensional representations for all features in a text, and then average these to a low-dimensional representation of the full text. The resulting models are therefore able to represent similarities between different features. For example, the model might learn that big and large are near-synonyms, and should be treated in a similar manner. That can be really useful, particularly when you’re classifying short texts.
It’s clear FastText is a neat little software package that deals with large volumes of data easily and produces high-quality classifiers. Let’s find out how it compares to the competition.
While FastText only has neural networks to learn a classifier, it can often be worthwhile to explore some alternative approaches. That’s where more generic machine learning software libraries come in. One great example of such a library is Scikit-learn for Python. Scikit-learn offers a wealth of machine learning approaches and makes it really easy to experiment with various models and parameter settings. When you’re doing text classification, its Multinomial Naive Bayes classifier is a simple baseline to try out, while its Support Vector Machines can help you achieve state-of-the-art accuracy.
Training and testing a model with Scikit-learn is more involved than with FastText, but not very
much so. The tutorial Working
with Text Data summarizes the most important steps. Basically, what we need is a pipeline with
three components: a vectorizer that extracts the features from our texts, a transformer that
weights these features correctly, and finally, our classifier. In my code below, the
CountVectorizer tokenizes our texts and models each text as a vector with the
frequencies of its tokens. The
TfidfTransformer converts these frequencies to the more
informative tf-idf weights, before the classifier builds a classification model. Training an SVM
instead of a Multinomial Naive Bayes model is as simple as replacing
LinearSVC on line three; extending the features with bigrams can be done by setting
(1, 2). The
fit command on line four trains a model by sending the training data through the
predict command uses the trained model to classify the test data. Finally, we measure
the accuracy by checking how often the predicted label is the same as the test label.
Compared to FastText, Scikit-learn is painfully slow. For example, training an SVM on the unigrams and bigrams of the Sogou corpus takes about 17 minutes on my laptop, compared to 4 minutes for FastText. It also requires several times more memory. The models can be made significantly smaller and faster by setting a minimum frequency for their features, but when you’re dealing with lots of data, speed and memory usage can become a concern.
In terms of accuracy, however, there is much less difference between the two solutions. In fact, I obtained a slightly higher accuracy with Scikit-learn than with FastText: when it is trained on unigrams and bigrams, its SVM classifies 92.8% of the AG test examples correctly (FastText: 91.7%, 92.5% in the paper), and 97.1% of the Sogou examples (FastText: 95.5%, 96.8% in the paper). As expected, the Naive Bayes classifier does less well, with an accuracy of 90.8% for AG, and 90.2% for Sogou.
In summary, Scikit-learn is a really versatile library that allows you to quickly experiment with many different models and parameter settings. The more advanced of these often obtain a very good performance. However, out of the box it’s less suitable for modelling large data sets than FastText.
A third simple approach to text classification is to use an online API. While most available services have pre-trained models for traditional NLP tasks such as named entity recognition, sentiment analysis, and general text classification, some of them allow people to train their own models and hopefully achieve a better accuracy on domain-specific data. One of these is MonkeyLearn.
One of the main selling points of MonkeyLearn is its user-friendliness. Three dialogue windows help us set up a text classification model. During this process, we tell the system we would like to train a classifier, that this classifier should categorize texts by topic, and that our texts are news articles in English (or Chinese for Sogou). These settings help Monkeylearn choose the best pre-processing steps (filtering, emoticon normalization, etc.) and select the best combination of parameters for its models. When this is done, it creates an environment where we can manipulate the parameters of our model, train it and explore its performance.
Before we can train our model, we need to upload the training data, either as a csv or Excel file, or as text data, through a simple API call. Unfortunately, there’s a pretty strict limit on the number of training data MonkeyLearn accepts. The free plan allows for 3,000 training examples, the Ultra Gorilla for 40,000, and the Enterprise plan is made to measure. These limits may not be an issue when there is little tagged data for your problem, but in this age of big data they do feel a bit restrictive. Luckily the folks behind MonkeyLearn were friendly enough to give me access to the Ultra Gorilla plan for a few weeks. As a result, I chose to work with 40,000 training examples for the AG corpus, and 20,000 for Sogou.
While the MonkeyLearn user interface is particularly attractive for newcomers to NLP, more experienced users still have the possibility to tweak the main parameters of their classifier. These include the type of model (Naive Bayes or SVM), the type of features (unigrams, bigrams, trigrams, or a combination of these), the maximum number of features (the default is 10,000), the stopwords, etc. There’s also a reference page that explains what these parameters mean.
Building a classifier takes a considerable amount of time. For example, MonkeyLearn needs around 20 minutes to train an SVM on the unigrams and bigrams in 20,000 Sogou news texts. Scikit-learn does that in just 41 seconds on my laptop. However, our patience is rewarded with a nice overview of our model and its performance: its accuracy, the keywords in our text data, and a really useful confusion matrix that helps us analyze what mistakes the classifier tends to make. These statistics are based on 4-fold cross-validation on the training data, which explains the longer training times at least partially.
In order to test the trained model on our test set, we can make use of another of MonkeyLearn’s distinguishing features: all trained classifiers are immediately available for production, through an API. As with Scikit-learn, we experimented with a Naive Bayes classifier and an SVM, both with and without bigrams. On the AG corpus, there is very little difference between these settings: all models achieve an accuracy between 87.7% and 88.0%. With the same training examples, our Scikit-learn models gave an accuracy of 89.3% (SVM, unigrams only), while FastText achieved 88.6% (unigrams only). The Sogou corpus shows bigger differences. The best-performing MonkeyLearn model is the SVM with unigrams and bigrams, with an accuracy of 94.2%. This is about the same as a Scikit-learn SVM trained on the same examples (94.8%), and considerably better than FastText, which scores 89.1% with unigrams, and 86.8% with ungirams and bigrams.
MonkeyLearn takes away some of the pains of developing a classifier: it assists users in the training process, helps them evaluate their models, and immediately exposes their classifiers through an API. In terms of accuracy, it’s in the same ballpark as the other approaches we’ve tested here, but training takes a while and using large data sets can become expensive.
As machine learning and AI grow in popularity, tackling basic NLP tasks is getting easier and easier. The wealth of available options means NLP practitioners are now often spoilt for choice. While comparing the accuracy of the models may be a logical first step, it did not reveal a clear winner among the three approaches to text classification I tested here. This is true in particular because I didn’t do any parameter optimization, and some approaches may have performed better with other parameter settings.
This means other factors will be decisive when you weigh up the different software options. If you’re working with big data sets, and speed and low memory usage are crucial, FastText looks like a good choice. The figures I obtained with 20,000 examples do suggest it works less well with small data sets, however. If you need a flexible framework that allows you to compare a wide range of machine learning models quickly, Scikit-learn will help you out. Its best classifiers can give great performance on both small and large data sets, but training large models can take some time. If you’re relatively new to NLP and looking for a text classifier that’s easy to build and integrate, MonkeyLearn is worth checking out. Keep in mind it’s less suitable for big-data problems and you give up some control.
There’s no doubt there are many more possibilities than the three frameworks I’ve explored here. As new options come along frequently, we’ll be faced with the agony of choice for quite some time to come. Still, a few simple experiments can go a long way to help you take your pick.
Yves is the Natural Language Processing expert at NLP Town. He holds an MSc in Speech & Language Processing from the University of Edinburgh, and a PhD in Computational Linguistics from the University of Leuven. After spending some time as a post-doctoral researcher at Stanford University, he traded academia for industry. In 2014 he founded NLP Town to help companies and organizations implement NLP solutions successfully.