A clear example of this is our use of word embeddings. Word2Vec debuted in 2013, upending the world of language processing: perhaps the most profound discovery since that of the Rosetta Stone in 1799. Representing words as vectors in a space of meaning, this has unlocked all manner of NLP advances: sentiment classification, text generation, remarkably good machine translation. Today’s latest tech (think Google’s BERT or OpenAI’s GPT-3) all exploit word vectors to deliver their astonishingly (and dangerously) good results. However, at their core, it’s the word2vec innovation that underpins their success.

We imported that specific idea for use in our firm’s products; we don’t (yet) train our own models, since the vectors available freely from the open source community far surpass the quality of any we could train from our small datasets. Our products work remarkably well; not because we run huge models with many layers, but because we use the foundational word2vec technology.

Another, even more basic example is that of page classification. Image classifiers were the original use case for CNN’s, and continue to be an area of hot research: GANS producing “Van Gogh” paintings rely on them to discern fact from fiction, and virtual personal “shoppers” use them to comb online retails for items matching your taste.

For our use cases, where documents are regularly structured, we discovered that sophisticated (and sometimes finicky) CNN’s are overmatched to our tasks at hand: a combination of image rasterization and cutting hyperplanes deliver exceptional performance, for a fraction of the complexity.