Understanding Convolutional Neural Networks for NLP

Dive into the world of Convolution Neural Networks (CNN), learn how they work, how to apply them for NLP, and how to tune CNN hyperparameters for best performance.

Pooling Layers

A key aspect of Convolutional Neural Networks are pooling layers, typically applied after the convolutional layers. Pooling layers subsample their input. The most common way to do pooling it to apply a max operation to the result of each filter. You don’t necessarily need to pool over the complete matrix, you could also pool over a window. For example, the following shows max pooling for a 2×2 window (in NLP we typically are apply pooling over the complete output, yielding just a single number for each filter):

Max pooling in CNN. Source: http://cs231n.github.io/convolutional-networks/#pool
Fig. 5 Max pooling in CNN. Source: http://cs231n.github.io/convolutional-networks/#pool

Why pooling? There are a couple of reasons. One property of pooling is that it provides a fixed size output matrix, which typically is required for classification. For example, if you have 1,000 filters and you apply max pooling to each, you will get a 1000-dimensional output, regardless of the size of your filters, or the size of your input. This allows you to use variable size sentences, and variable size filters, but always get the same output dimensions to feed into a classifier.

Pooling also reduces the output dimensionality but (hopefully) keeps the most salient information. You can think of each filter as detecting a specific feature, such as detecting if the sentence contains a negation like “not amazing” for example. If this phrase occurs somewhere in the sentence, the result of applying the filter to that region will yield a large value, but a small value in other regions. By performing the max operation you  are keeping information about whether or not the feature appeared in the sentence, but you are losing information about where exactly it appeared. But isn’t this information about locality really useful? Yes, it  is and it’s a bit similar to what a bag of n-grams model is doing. You are losing global information about locality (where in a sentence something happens), but you are keeping local information captured by your filters, like “not amazing” being very different from “amazing not”.

In imagine recognition, pooling also provides basic invariance to translating (shifting) and rotation. When you are pooling over a region, the output will stay approximately the same even if you shift/rotate the image by a few pixels, because the max operations will pick out the same value regardless.


The last concept we need to understand are channels. Channels are different “views” of your input data. For example, in image recognition you typically have RGB (red, green, blue) channels. You can apply convolutions across channels, either with different or equal weights. In NLP you could imagine having various channels as well: You could have a separate channels for different word embeddings (word2vec and GloVe for example), or you could have a channel for the same sentence represented in different languages, or phrased in different ways.

Convoltuonal Neural Networks applied to NLP

Let’s now look at some of the applications of CNNs to Natural Language Processing. I’ll try it summarize some of the research results. Invariably I’ll miss many interesting applications (do let me know in the comments), but I hope to cover at least some of the more popular results.

The most natural fit for CNNs seem to be classifications tasks, such as Sentiment Analysis, Spam Detection or Topic Categorization. Convolutions and pooling operations lose information about the local order of words, so that sequence tagging as in PoS Tagging or Entity Extraction is a bit harder to fit into a pure CNN architecture (though not impossible, you can add positional features to the input).

[1] Evaluates a CNN architecture on various classification datasets, mostly comprised of Sentiment Analysis and Topic Categorization tasks. The CNN architecture achieves very good performance across datasets, and new state-of-the-art on a few. Surprisingly, the network used in this paper is quite simple, and that’s what makes it powerful.The input layer is a sentence comprised of concatenated word2vec word embeddings. That’s followed by a convolutional layer with multiple filters, then a max-pooling layer, and finally a softmax classifier.  The paper also experiments with two different channels in the form of static and dynamic word embeddings, where one channel is adjusted during training and the other isn’t. A similar, but somewhat more complex, architecture was previously proposed in [2]. [6] Adds an additional layer that performs “semantic clustering” to this network architecture.

Kim, Y. (2014). Convolutional Neural Networks for Sentence Classification
Fig. 6Kim, Y. (2014). Convolutional Neural Networks for Sentence Classification

[4] Trains a CNN from scratch, without the need for for pre-trained word vectors like word2vec or GloVe. It applies convolutions directly to one-hot vectors. The author also proposes a space-efficient bag-of-words-like representation for the input data, reducing the number of parameters the network needs to learn. In [5] the author  extends the model with an additional unsupervised “region embedding” that is learned using a CNN  predicting the context of text regions. The approach in these papers seems to work well for long-form texts (like movie reviews), but their performance on short texts (like tweets) isn’t clear. Intuitively, it makes sense that using pre-trained word embeddings for short texts would yield larger gains than using them for long texts.

Building a CNN architecture means that there are many hyperparameters to choose from, some of which I presented above: Input represenations (word2vec, GloVe, one-hot), number and sizes of convolution filters, pooling strategies (max, average), and activation functions (ReLU, tanh). [7] performs an empirical evaluation on the effect of varying hyperparameters in CNN architectures, investigating their impact on performance and variance over multiple runs. If you are looking to implement your own CNN for text classification, using the results of this paper as a starting point would be an excellent idea. A few results that stand out are that max-pooling always beat average pooling, that the ideal filter sizes are important but task-dependent, and that regularization doesn’t seem to make a big different in the NLP tasks that were considered. A caveat of this research is that all the datasets were quite similar in terms of their document length, so the same guidelines may not apply to data that looks considerably different.

[8] explores CNNs for  Relation Extraction and Relation Classification tasks. In addition to the word vectors, the authors use the relative positions of words to the entities of interest as an input to the convolutional layer. This models assumes that the positions of the entities are given, and that each example input contains one relation. [9] and [10] have explored similar models.

Another interesting use case of CNNs in NLP can be found in [11] and [12], coming out of Microsoft Research. These papers describe how to learn semantically meaningful representations of sentences that can be used for Information Retrieval. The example given in the papers includes recommending potentially interesting documents to users based on what they are currently reading. The sentence representations are trained based on search engine log data.

Most CNN architectures learn embeddings (low-dimensional representations) for words and sentences in one way or another as part of their training procedure. Not all papers though focus on this aspect of training or investigate how meaningful the learned embeddings are. [13] presents a CNN architecture to predict hashtags for Facebook posts, while at the same time generating meaningful embeddings for words and sentences. These learned embeddings are then successfully applied to another task – recommending potentially interesting documents to users, trained based on clickstream data.

Character-Level CNNs

So far, all of the models presented were based on words. But there has also been research in applying CNNs directly to characters. [14] learns character-level embeddings, joins them with pre-trained word embeddings, and uses a CNN for Part of Speech tagging. [15][16] explores the use of CNNs to learn directly from characters, without the need for any pre-trained embeddings. Notably, the authors use a relatively deep network with a total of 9 layers, and apply it to Sentiment Analysis and Text Categorization tasks. Results show that learning directly from character-level input works very well on large datasets (millions of examples), but underperforms simpler models on smaller datasets (hundreds of thousands of examples). [17] explores to application of character-level convolutions to Language Modeling, using the output of the character-level CNN as the input to an LSTM at each time step. The same model is applied to various languages.

What’s amazing is that essentially all of the papers above were published  in the past 1-2 years. Obviously there has been excellent work with CNNs on NLP before, as in Natural Language Processing (almost) from Scratch, but the pace of new results and state of the art systems being published is clearly accelerating.

Questions or Feedback? Let me know in the comments. Thanks for reading!

Paper References