Explosion Blog

I'm not the author but can speak to some of this. Some amount of this was answered by the author himself in his blog post "Dead Code Should Be Buried – Why I Didn't Contribute to NLTK"[1].

While there are portions on NLTK that are useful, it suffers from having too many things under the hood and a lack of consistent maintenance. Many of them are superglued external libraries while others have not been maintained or have never been made to work well with other parts of NLTK.

The biggest example of this was that, for some time, it wasn't certain how the NLTK part of speech tagger was trained[2]. While that has been remedied, the remedy uses part of speech tagging code inspired via earlier work by the author of spaCy[3][4].

spaCy has been made from the ground up to support more modern deep learning features (easy integration of Keras is quite cool but I'm actually meaning the default inclusion of word vectors which has existed since early spaCy) while having a consistent and clean API that covers many sane use cases.

The API is highly Pythonic but the underlying performance critical components are written in Cython, meaning very few parsers are actually able to beat spaCy in speed (especially when using CPU and not GPU). Regarding speed, multi-threading is also a relatively trivial process - Cython releases the GIL around the syntactic parser. As a reference to speed, they used it to tag and parse every Reddit comment from 2015 to then generate "sense2vec"[5] at over…

A quick note about using natural language/sentiment APIs: trained machine learning models must be used apples-to-apples on similar datasets; for example, you can’t accurately perform Twitter sentiment analysis on a dataset using a model trained on professional movie reviews since Tweets do not follow AP Style guidelines. (e.g. for some reason, training Python's NLTK on the IMDb movie review dataset to predict the sentiment of Donald Trump's tweets is a oddly common Hello World, even though the results are misleading and may cause confirmation bias)

Reddit comments are very idiosyncratic, and in this particular case, even moreso than usual. As a result, I am skeptical of trusting the output of such APIs as gospel, even one trained on massive datasets. (however, training a model on a Reddit-only dataset might be interesting, and is an idea I have in the pipeline.)

Last year, spaCy trained a model, sense2vec, on the Reddit dataset and got interesting results: https://explosion.ai/blog/sense2vec-with-spacy

I definitely agree that the same weights won't be optimal for different domains. If you need to parse tweets, you should have a tweet-trained model. The tweet model probably shouldn't be thinking about Jane Austen novels. We want to open a model store where you can buy language and domain specific models.

I think 99% of the time there's one best algorithm, and even one best implementation of it. It's the weights, and sometimes the features, that need to vary.

Finally — I love displaCy too. Ines does great work :). Have you seen that we open-sourced this recently? It's now very easy to run locally, and connect up to the model you're developing. You can use this with any other parser, too. https://explosion.ai/blog/displacy-js-nlp-visualizer

The title is a little misleading. They raised $6 million from SignalFire (who also invested in Grammarly) for 5% of the company [1]. As SignalFire retains that shareholding, they are not less independent (nor more) than that have been in the past couple of years. They are just not looking for more investment and so are more "independent-minded".

I used SpaCy quite expensively pre-2020. It still has a lot of great uses, and more predictable than LLM models. But now many NLP tasks that seemed near impossible, or required a lot of expensive annotated training data (which their product Prodigy is used for) can now be coded in very little time with LLMs.

[1] https://explosion.ai/blog/weve-sold-5-percent-of-explosion

I don't think spaCy will be hurting any time soon. When SyntaxNet was first released last year, Matthew Honnibal had a good writeup [0] of how spaCy vastly outperforms with speed while keeping reasonable accuracy:

>On the time-honoured benchmark for this task, Parsey McParseface achieves over 94% accuracy, at around 600 words per second. On the same task, spaCy achieves 92.4%, at around 15,000 words per second. The extra accuracy might not sound like much, but for applications, it's likely to be pretty significant.

If spaCy is able to increment the accuracy and maintain the large performance gap, it'll still be my go-to NLP framework!

[0] https://explosion.ai/blog/syntaxnet-in-context

Everyone has a different way of approaching. I like video walk-throughs of papers. If you are the same, for the theory bit you may find Yannic Kilcher's videos [0,1] helpful.

On implementation side, the best project out there is [2] with good documentation. Also spacy has transformer package [3] (I personally haven't tried it), so maybe it will be easier for you to jump in if you have prior experience with spacy.

[0] https://www.youtube.com/watch?v=iDulhoQ2pro

[1] https://www.youtube.com/watch?v=-9evrZnBorM

[2] https://huggingface.co/transformers/

[3] https://explosion.ai/blog/spacy-transformers