NB: This describes work in progress. It reflects expected results, not necessarily things that are already true.
This repository contains experiments on spaCy's new "pretrain" command, which uses a ULMFit/Elmo/BERT/etc-like process for pre-training. We employ a novel trick which we term Language Modelling with Approximate Outputs (LMAO). Normally language models have to softmax (or hierarchical softmax) over the output vocabulary, which requires large hidden layers. We want to train small hidden layers, as we want spaCy to be cost-effective to run.
Instead of outputting numeric IDs, we predict points in a vector space that's been pre-trained using an algorithm like GloVe or FastText. This lets the model learn large target vocabularies, without requiring extra parameters.
I compared accuracy on the development data of the English-EWT portion of the universal dependencies data for four models:
- No GloVe, no LMAO: Only the training data was used, with no pre-training in
the word vectors or the CNN. This is the
smconfiguration in spaCy. - GloVe: Pre-trained word vectors were used as one of the input features. This
is the
lgconfiguration in spaCy. - No GloVe, LMAO: Like 1, but the CNN and hash embeddings were pretrained with LMAO cloze task.
- GloVe, LMAO: Like 2, but the CNN and hash embeddings were pretrained with the LMAO cloze task.
| Model | LAS |
|---|---|
| -GloVe,-LMAO | 79.1 |
| -GloVe,+LMAO | 81.0 |
| +GloVe,-LMAO | 81.0 |
| +GloVe,+LMAO | 82.4 |
| Stanford '17 | 82.3 |
| Stanford '18 | 83.9 |
All the models had the same number of learnable parameters: token vector width 96, hidden width 64, 2000 embedding rows. The model is very small --- like, 3 MB if total. The pre-training was done on the January 2017 portion of the Reddit comments corpus (about 2 billion words). The pre-training objective is like the BERT cloze task, but instead of predicting word IDs, we predict the pre-trained GloVe vector of the word. This means the output layer is very small (only 300 dimensions). A non-linear layer is used in between the CNN and the output layer. This non-linear layer is a layer-normalized Maxout layer.
The two Stanford models are shown for comparison, to show these are overall quite solid scores given the size of the model. We're achieving comparable accuracy to the Stanford system that won the CoNLL 2017 shared task. The Stanford 2018 results were near state-of-the-art in August 2018, but don't utilise BERT-style pre-training, which I guess would push their results into the high 80s.
The spacy pretrain command now uses a BERT-style masked language model, but
instead of predicting the word IDs, we predict the GloVe vector of the target
words. I've used this objective to pre-train spaCy's small CNN model on
2 billion words of text from Reddit (the January 2017 portion of the Reddit
comments corpus). The pre-trained CNN is very small: it's depth 4, width 96,
and has only 2000 rows in the hash embeddings table. Weights for the serialized
model are only 3.2 MB.
Here's how pretraining affects parser unlabelled accuracy on different sized subsets of the training data.
| #Docs | #Words | Baseline | Pretrain |
|---|---|---|---|
| 1 | 69 | 41.8 | 60.5 |
| 10 | 311 | 47.4 | 66.4 |
| 100 | 2947 | 61.3 | 79.0 |
| 1000 | 31097 | 78.5 | 84.7 |
| 10000 | 255000 | 87.0 | 87.8 |
The effect size here is pretty strong. The pretraining is very effective in the low and medium data cases, but the impact fades as more training data becomes available. One caveat with these results is that static vectors weren't used in the input --- the vectors were trained from scratch. This is a bit unfair to the baseline, since pre-trained vectors are a simple and effective existing way to get some pretrained knowledge into the model. As a quick datapoint, the result for baseline+GloVe vectors for the 1000 document case is 82.0%. So the static vectors aren't as good as the pretraining. We can also use static vectors together with pretraining, but I need to rerun the pretraining for that.
I've also run some text classification experiments, using the same pretrained weights as the parsing experiments. The evaluation is the IMDb task, but with only 1000 training samples available. The GloVe vectors greatly improve results on this task, so the current results aren't so interesting. The pre-training improves over the baseline, but the absolute numbers are too low to be informative. We'll see what happens when we have a pretrained weights file also uses the GloVe vectors as input.
The CNN is very sensitive to hyper-parameters on small text classification
tasks. In order to get a valid comparison, we therefore have to do
hyper-parameter search separately for each experiment configuration. To do
this. I'm using Polyaxon. The experiment code is in the
lmao-imdb-1k directory. See gke-terraform-polyaxon for setup instructions.
I ran 100 configurations with and without pre-training. Without pre-training, the best accuracy achieved was 78.4%; with pre-training, the best accuracy achieved was 81.3%. With GloVe vectors, the baseline goes up to 87%. We'll have to wait and see how the pretraining+GloVe model performs.
Low and medium data use-cases are very important, because they help you perform rapid prototyping. When doing practical NLP, there are an enormous number of decisions to make in how you break down the application requirements into statistical models. To make good decisions, you need to try things out and see how well models perform with different approaches. If you have models which take huge quantities of data before they yield reasonable performance, you have to do a lot of annotation just to find out an idea doesn't work.
A reasonable interpretation of the current results is that pretraining lets you use a larger model than your training data could otherwise support. The previous work has shown that pretraining lets us make very large models, and those models will let us hit state-of-the-art accuracies even when the training corpus is quite large. It wasn't clear from the previous work whether pretraining would also improve results for a small model. My results so far suggest it does, but so far only when the training data is small.
I'm now pre-training a larger CNN, to check that my results match up with previous work in finding that this improves accuracy on the full dataset. This could also be a nice option for spaCy users, who want more accuracy even if it means slower runtimes.