Vectors and Muppets, part II

In the last post, I mentioned two NLP embeddings models: word2vec and BERT. I haven't talked in much detail of the latter yet. In a nutshell, BERT vectors are to word2vec vectors pretty much what the Starkiller is to the Deathstar, or what Borgs are to Klingons. It's bigger, the producers had a bigger budget, it's more efficient at what it does. In a word: supersized.

That is not to say that BERT models are just a scaled up version of word2vec. From a purely engineering standpoint, BERT models are what we call "Transformers". A Transformer is a type of neural network, in the same sense that a Ferrari is a type of car. In our jargon, we call types of networks "architectures". Transformers, in particular, have been found to perform impressively well on a wide range of tasks, from machine translation to language modeling—the task of predicting the next word in a sentence.

On the other hand, word2vec is based on a log-linear classifier architecture—that's a very simple, straightforward class of neural networks. The two main differences between a Transformer and a log-linear classifier is that a) Transformers are a lot more complex and b) Transformers can take as input sequences of arbitrary length, whereas a log-linear classifier can only process input of a fixed size.

The different characteristics of the two architectures entail that the computation of word vectors in a BERT model will be quite different than what we saw with word2vec. In particular, in word2vec, we went about our business one word at a time, trying to reconstruct the context around one word at each step. On the other hand, Transformers like BERT can handle inputs of arbitrary length: hence I can feed the entire sentence at once in my model. This means that the vectors we get from a BERT model are word vectors that take into account which sentence this word occurs in. We have unique vector representations for a word in a sentence. Instead of representing the word "tie" with one vector regardless of all the contexts it occurs in, we'll represent the word "tie" differently in different sentences such as "I tie my shoes" or "My dad wears a silly tie".

The way we train BERT models is technically different from how we train word2vec models, but it does closely resemble the CBOW task I mentionned last time. Basically, we train the model to solve a fill-in-the-blanks exam. We start by selecting a few words at random, we then replace them with a blank. So from a sentence such as "The most fearsome animal is the underfed mallard", we would create a blanked input such as "The most ____ animal is ____ underfed mallard". We then pass that blanked-out input to the BERT model we train. BERT's task is to fill those blanks and guess what words correspond to the blanks ____ in the original sentence.

The last point worth mentioning is the hype. BERT models and their variants took the top of the charts very, very quickly. Any benchmark we made, BERT embeddings could seemingly solve it to a high degree of accuracy. The key point is that, unlike word2vec vectors which were useful as inputs in more complex models, BERT models can often be plugged in whatever task you fancy with very minimal change. Want to guess which customer reviews express a negative feedback? Get a pre-trained BERT model, retrain it for a smidge longer to discriminate between postive and negative comments, et voilĂ ! The YELP, she is done.

Another consequence of the hype is the popularity of it. Transformer models, as I pointed out earlier, are complex. They are in fact "black boxes", like most other neural network architectures. We don't have a satisfactory explanation of how they process the input data we feed them. We're only able to observe what goes in and what goes out (and that what goes out can be generally impressive). BERT is a very new model (it was proposed in 2018), so we haven't had the time to build a thorough understanding of its workings and its quirks. Nonetheless, research is being conducted: there's a whole cottage industry built around understanding what BERT models do, how they do it, why they do it. That field of study is called BERTology (I kid you not). Some of the fun findings they suggest is that BERT models would be implicitly modeling the syntax of the data we give it as input.

Another side effect of the hype is the very impressive number of models inspired from BERT. We have BERT and BART and alBERT and RoBERTA and ELECTRA and PEGASUS and mBERT and mBART and distilBERT and CamemBERT and BERTje and herBERT and BARThez and FlauBERT and BETO and UmBERTo and BioBERT and PhoBERT and FinBERT and RuBERT and GilBERTo. As a fun exercise, you can try to guess which language or literary genre each of those correspond to. Full disclosure: I would have to look some of these up.

Anyways, my point is that BERT is popular. Obviously, a plug-and-play high-performance NLP model is bound to bring a lot of hype. To be fair, there's a lot to unpack about the downside of BERT models, and it's a rather depressing topic. So I'll just tackle that topic another time, and leave it at that for today.