Poetry (almost) From Scratch – Teaching Versification to a Neural Network

This past week, the Peabody Opera House in St Louis Missouri was the host venue to some of the sharpest developers in the industry for the annual Strange Loop conference.

Strange Loop is a multi-disciplinary conference that brings together the developers and thinkers building tomorrow’s technology in fields such as emerging languages, alternative databases, concurrency, distributed systems, security, and the web.

This year, Criteo Lab’s Yann Schwartz presented one the conference’s talked about presentation on Poetry (almost) From Scratch – Teaching Versification to a Natural Network.

Abstract

Poetry can take many forms but could be defined as a delicate balance between constraints and expression. Computers are pretty good at constraints, less so at expression.

The general problem of cracking the structure and meaning of a text has been one of the goals of linguistics, and Natural Language Processing. This had first been addressed in a reductionist way, by decomposing a sentence, counting, matching then, later, by learning from evidence and statistics. Or, as it has been shown, lately to be more and more effective, by training neural networks with a minimal set of assumptions – as was shown by the seminal paper “Natural Language Processing (almost) from Scratch”.

In this talk we’ll see how to follow these techniques to spot poetry in unlikely textual places, or generate it (almost) from scratch, using a minimal set of assumptions, hoping for the meter and rhyme rules to emerge. Along the way we’ll touch on topics such as the impossible definition of poetry and the tension between simple, classical rules, and deep learning models that can defy interpretation.

Get into the presentation below.