ConceptNet Numberbatch 19.08

ConceptNet Numberbatch logo, featuring an otter

It’s been a while since we made a release of ConceptNet Numberbatch. Here, have some new word embeddings.

ConceptNet Numberbatch is our whimsical double-dactyl name for pre-computed word embeddings built using ConceptNet and distributional semantics. The things we’ve been doing with it at Luminoso have benefited from some improvements we’ve made in the last two years.

The last release we announced was in 2017. Since then, we made a few improvements for SemEval 2018, when we demonstrated how to distinguish attributes using ConceptNet.

But meanwhile, outside of Luminoso, we’ve also seen some great things being built with what we released:

  • Alex Lew’s Robot Mind Meld uses ConceptNet Numberbatch to play a cooperative improv game.

  • Tsun-Hsien Tang and others showed how Numberbatch can be combined with image recognition to better retrieve images of daily life.

  • Sophie Siebert and Frieder Stolzenburg developed CoRg, a reasoning / story understanding system that combines Numberbatch word embeddings with theorem proving, a combination I wouldn’t have expected to see at all.

I hope we can see more projects like this by releasing our improvements to Numberbatch.

Expanding the vocabulary

We added a step to the build process of ConceptNet Numberbatch called “propagation”. This makes it easier to use Numberbatch to represent a larger vocabulary of words, especially in languages with more inflections than English.

Previously, there were a lot of terms that we didn’t have vectors for, especially word forms that aren’t the most commonly-observed form. We had to rely on the out-of-vocabulary (OOV) strategy to handle these words, by looking up their neighboring terms in ConceptNet that did have vectors. This strategy was hard to implement, because it required having access to the ConceptNet graph at all times.

I know that many projects that attempted to use Numberbatch simply skipped the OOV strategy, so any word that wasn’t directly in the vocabulary just couldn’t be represented, and this led to suboptimal results.

With the “propagation” step, we pre-compute the vectors for more words, especially forms of known words.

This increases the vocabulary size and the space required to use Numberbatch, but leaves us with a simple, fast OOV strategy that doesn’t need to refer to the whole ConceptNet graph. And it should improve the results greatly for users of Numberbatch who aren’t using an OOV strategy at all.

New results on fairness

Word embeddings are a useful tool for a lot of NLP tasks, but by now we’ve seen lots of evidence of a risk they carry: when they capture word meanings from the ways we use words, they also capture harmful biases and stereotypes. A clear, up-to-date paper on this is “What are the biases in my word embedding?”, by Nathaniel Swinger et al.

It’s important to do what we can to mitigate that. Machine learning involves lots of ethical issues, and we can’t solve them all while not knowing how you’re even going to use our word embeddings, but we can at least try not to publish something that makes the ethical problems worse. So one of the steps in building ConceptNet Numberbatch is algorithmic de-biasing that tries to identify and mitigate these biases.

(If you want to point out that algorithmic de-biasing is insufficient to solve the problem, you are very right, but that doesn’t mean we shouldn’t do it.)

Chris Sweeney and Maryam Najafian published a new framework for assessing fairness in word embeddings. Their framework doesn’t assume that biases are necessarily binary (men vs. women, white vs. black) or can be seen in a linear projection of the embeddings, as previous metrics did. This assessment comes out looking pretty good for Numberbatch, which associates nationalities and religions with sentiment words more equitably than other embeddings.

Please note that you can’t be assured that your AI project is ethical just because it has one fairer-than-usual component in it. We have not solved AI ethics. You still need to test for harmful effects in the outputs of your actual system, as well as making sure that its inputs are collected ethically.

You can find download links and documentation for the new version on the conceptnet-numberbatch GitHub page.

The Python code that builds the embeddings is in conceptnet5.vectors, part of the conceptnet5 repository.


Comments powered by Disqus