Two weeks ago, I did a presentation for my colleagues of the paper from Yurochkin et al. (2019), from NeurIPS 2019. It contains an interesting approach to document classification leading to strong performance, and, most importantly, excellent interpretability.

This paper seems interesting to me because of it uses two methods with strong theoretical guarantees: optimal transport and topic modelling. Optimal transport looks very promising to me in NLP, and has seen a lot of interest in recent years due to advances in approximation algorithms, such as entropy regularisation. It is also quite refreshing to see approaches using solid results in optimisation, compared to purely experimental deep learning methods.

Introduction and motivation

The problem of the paper is to measure similarity (i.e. a distance) between pairs of documents, by incorporating semantic similarities (and not only syntactic artefacts), without encountering scalability issues.

They propose a “meta-distance” between documents, called the hierarchical optimal topic transport (HOTT), providing a scalable metric incorporating topic information between documents. As such, they try to combine two different levels of analysis:

• word embeddings data, to embed language knowledge (via pre-trained embeddings for instance),
• topic modelling methods (e.g. Latent Dirichlet Allocation), to represent semantically-meaningful groups of words.

Background: optimal transport

The essential backbone of the method is the Wasserstein distance, derived from optimal transport theory. Optimal transport is a fascinating and deep subject, so I won’t enter into the details here. For an introduction to the theory and its applications, check out the excellent book from Peyré and Cuturi (2019), (available on ArXiv as well). There are also very nice posts by Gabriel Peyré on the CNRS maths blog (in French). Many more resources (including slides for presentations) are available at https://optimaltransport.github.io. For a more complete theoretical treatment of the subject, check out Santambrogio (2015), or, if you’re feeling particularly adventurous, Villani (2009).

For this paper, only a superficial understanding of how the Wasserstein distance works is necessary. Optimal transport is an optimisation technique to lift a distance between points in a given metric space, to a distance between probability distributions over this metric space. The historical example is to move piles of dirt around: you know the distance between any two points, and you have piles of dirt lying around¹. Now, if you want to move these piles to another configuration (fewer piles, say, or a different repartition of dirt a few metres away), you need to find the most efficient way to move them. The total cost you obtain will define a distance between the two configurations of dirt, and is usually called the earth mover’s distance, which is just an instance of the general Wasserstein metric.

More formally, if we have to sets of points \(x = (x_1, x_2, \ldots, x_n)\), and \(y = (y_1, y_2, \ldots, y_n)\), along with probability distributions \(p \in \Delta^n\), \(q \in \Delta^m\) over \(x\) and \(y\) (\(\Delta^n\) is the probability simplex of dimension \(n\), i.e. the set of vectors of size \(n\) summing to 1), we can define the Wasserstein distance as \[ W_1(p, q) = \min_{P \in \mathbb{R}_+^{n\times m}} \sum_{i,j} C_{i,j} P_{i,j}\\ \text{\small subject to } \sum_j P_{i,j} = p_i \text{ \small and } \sum_i P_{i,j} = q_j, \] where \(C_{i,j} = d(x_i, x_j)\) are the costs computed from the original distance between points, and \(P_{i,j}\) represent the amount we are moving from pile \(i\) to pile \(j\).

Now, how can this be applied to a natural language setting? Once we have word embeddings, we can consider that the vocabulary forms a metric space (we can compute a distance, for instance the euclidean or the cosine distance, between two word embeddings). The key is to define documents as distributions over words.

Given a vocabulary \(V \subset \mathbb{R}^n\) and a corpus \(D = (d^1, d^2, \ldots, d^{\lvert D \rvert})\), we represent a document as \(d^i \in \Delta^{l_i}\) where \(l_i\) is the number of unique words in \(d^i\), and \(d^i_j\) is the proportion of word \(v_j\) in the document \(d^i\). The word mover’s distance (WMD) is then defined simply as \[ \operatorname{WMD}(d^1, d^2) = W_1(d^1, d^2). \]

If you didn’t follow all of this, don’t worry! The gist is: if you have a distance between points, you can solve an optimisation problem to obtain a distance between distributions over these points! This is especially useful when you consider that each word embedding is a point, and a document is just a set of words, along with the number of times they appear.

References