word embeddings for bureaucratese08 Jul 2020
You can find pre-trained word embeddings for hundreds of different languages - FastText alone has pre-trained embeddings for 157 languages (see here). But a single language can come in multiple “flavors”. I’m not talking about dialects, but about the different vocabulary and writing styles you find in news articles vs social media vs academic papers, etc. Most word embeddings come from a limited number of sources, with Wikipedia and news articles being the most common ones. If you have, say, social media texts, using word embeddings trained on Wikipedia entries may not yield good results.
So I decided to train my own Brazilian Portuguese word embeddings on the source that interests me the most: government publications. Decrees, invitations for bids, contracts, appointments, all that mind-numbingly boring stuff that makes up the day-to-day life of the public sector. Those embeddings might help me in future text-related tasks, like classifying government decrees and identifying certain types of contracts. I imagine it could be useful for other folks working with Brazilian government publications, so here’s how I did that.
I started by scraping the official bulleting where all the acts of the Brazilian government are published: the Diário Oficial da União. To give you an idea of how much text that is, the Diário’s 2020-07-06 issue has a total of 344 pages - with a tiny font and single spaces. (The Brazilian state is humongous and the size of the Diário reflects that.) The Diário is available online going as far back as 2002-01-01 and I scraped all of it. That amounted to about 8GB of zip files. Here is how to scrape it yourself (I used Python for everything):
After the scraping is done you can unzip each of those 400+ files manually or you can automate the job:
This won’t give you all the text in the Diário Oficial da União since 2002-01-01. Some zip files are corrupted and most issues are incomplete. For 2016, for instance, only the May issues are available. And for all years except 2019 and 2020 one of the sections (section 3) is missing entirely (the Diário is divided in three sections - 1, 2, and 3). Also, after you unzip the files you find out that in many cases the text is not in XML but in JPEG format. I wasn’t in the mood to do OCR so I just ignored the JPEG files.
If you want to get in touch with the Diário’s publisher to discuss those problems be my guest. Here I don’t care much about those problems because all I need to train my word embeddings is a ton of data, not all of the data. And with the XML files that I got I have over 4 million government acts, which is probably way more than I need here.
After unzipping everything I trained my word embeddings. I chose to go with gensim’s implementation of word2vec. The beauty of gensim’s implementation is that you can stream the texts one by one straight from disk, without having to keep everything in memory. Now, that’s a little tricky to accomplish. Gensim’s documentation says that instead of a list of documents you can use a generator, but I found that not to be the case. I got this error:
TypeError: You can't pass a generator as the sentences argument. Try an iterator. But I googled around and found a nifty workaround that tricks gensim into using a generator by wrapping it inside an iterator. So here I have a generator (
yield_docs) that yields one document at a time and then I wrap it inside an iterator (
SentencesIterator) so that gensim won’t complain.
About the documents, I have some 4.2 million XML files in total. In theory all these XML files should be easily parsable - they have tags with metadata, main content, etc. But in reality many are invalid. They have unclosed quotation marks and other problems that trip BeautifulSoup’s parser. So I ignored all the metadata and just focused on the stuff inside the
<Texto> (text) tags, which is always a collection of
<p> tags. Now, different paragraphs of the same publication can talk about entirely different issues, so instead of treating each publication (i.e., each XML file) as a document I’m treating each
<p> content as a document. That should yield more coherent word associations. So while I have 4.2 million XML files, in the end I have 72 million documents, one corresponding to each
<p> tag. That’s… a lot of text.
Back to word2vec. I don’t really know the ideal number of dimensions here. I found a nice paper that provides a way to estimate the ideal number of dimensions for any dimensionality reduction algorithm. But it’s too computationally expensive: you need to create a graph where each unique token is a node and the co-occurrences are represented by edges. I tried it but the thing got impossibly slow at around 200k nodes - and I have over 1M unique tokens. By my estimates it would take about half a year for the nodes to be created, and then I would need to find the graph’s maximum clique, which is also computationally expensive. So… no. If I had a specific text classification task in mind I would just try different numbers of dimensions and see what works best, but that’s not what I’m doing right now. So instead of relying on any theoretical or empirical approaches I just went with 300 dimensions because that’s a nice round number and I’ve seen it used in other word embeddings.
I’m discarding all words that appear in fewer than 1000 paragraphs (probably too rare to matter) and I’m using a short window of 5 (maximum distance between current and predicted word in a sentence).
Here’s the code:
And voilà, we have our word embeddings. We have a total of 27198 unique tokens (remember, we ignored any tokens that appeared in fewer than 1000 paragraphs) and 300 dimensions, so our word embeddings are a 27198x300 matrix. If you’re not familiar with word2vec Andrew Ng explains it here. The TL;DR is that word2vec’s output is a matrix where each unique token is represented as a vector - in our case, a 300-dimensional vector. That allows us to do a bunch of interesting stuff with that vocabulary - for instance, we can compute the cosine similarity between any two words to see how related they are. In gensim there is a neat method for that. For instance, suppose we want to find the words most related to “fraude” (fraud):
We can see that bid rigging (“conluio”) and overpricing (“superfaturamento”) are the two most fraud-related words in government publications (“fraudes” is just the plural form of “fraude”). Kinda cool to see it. You can also cluster the word embeddings to find groups of inter-related words; use t-SNE to reduce dimensionality to 2 so you can plot the embeddings on an XY plot; and try a bunch of other fun ideas.
Here I trained the word embeddings from scratch but you could also take pre-trained Brazilian Portuguese embeddings and use the Diário to fine-tune them. You could also tweak the parameters, changing the window (10 here) and the number of dimensions (300). Whatever works best for the task you have at hand.
That’s all for today! And remember, bureaucratese is bad writing - don’t spend too long reading those texts, lest you start emulating them. The best antidote to bureaucratese (or to any bad writing) that I know is William Zinsser.