What are “features” and “representations”?
A representation is a collection of features. A feature is an aspect of a dataset that can be quantified.
For example, if you’re given a collection of documents, whether or not a given document contains the word “cat” could count as a feature, and that feature could count as evidence for the document being about pets.
Metonymi representations are “distributed,” which means there’s no single on/off, cat/no cat switch to flip that makes up any single feature.
Will the Metonymi representations work for what I’m doing?
We’ve tested the representations on a variety of data domains and tasks, and the features have predictive power on all of them. Our experience has been that the features are better suited for complex semantic tasks like this, rather than simple vocabulary tasks where sparse representations might be better.
We invite you to test the representations for yourself by following the instructions here on our Github page.
How does your technology work?
We’ve created a deep learning suite by examining a wide array of texts from Wikipedia and popular news sites and training it against an unsupervised objective function. We can’t really tell you what that objective function is, because then everyone would want to do it!
What languages does the technology support?
Metonymi knows English and has a vocabulary of over 100K words.
Our AI is still learning as we train it to analyze other languages and expand its knowledge of English.
How big can the documents it reads be?
It will read documents of up to 130 words; so it’s perfect for social media posts, reviews, and short articles.