Juan Alonso, known on Twitter as @kokuma, is the creator of Neologisms, an LSTM neural network designed to generate new verbs in Castilian that surprises for its more than credible and credible inventions.
As its developer explains, the model invents both new verbs as well as existing verbs. “A model is only capable of reproducing the patterns that exist in the data that have been used to train it, and this is reflected in two ways in the verbs generated,” he explains in a thread of tweets.
The model generates each of these verbs letter by letter, and its ‘creativity’ can be regulated with more or less common combinations
Neologisms, an interesting Spanish verb generator
Neologisms can be used directly in the browser, from your website, without installing anything. Just access the site and select one temperature. What this parameter does is generate more conservative and predictable verbs the lower the value and the more daring and even incorrect the higher it is.
This is explained because the model generates each of these verbs letter by letter. The lower the temperature, the less complicated it is and you choose to join letters that frequently appear together based on the training you have received through examples. The higher the temperature, the more creative and daring he gets, using less frequent combinations that can end up being wrong.
descangar, stranded, idioscalar, endortarse, counterfillar, overrun, embrace, marillar, ambar, recoveter …
This Spanish verb generator has its code available on GitHub with the list of verbs that has been used to train the model that makes it possible. For everything related to machine learning uses ml5.js. The verb lemmaries used by Neologisms are those of Ismael Olea and Ignacio Mario Morales.