Automatic computation of poetic creativity in parallel corpora

Abstract

Text representation has been broadly studied in information retrieval and natural language processing. The process of building a good representation is an aspect to consider because it captures the most important aspects of the text in order to support other tasks such as classification, sentiment analysis, and automatic translation. The most common techniques used to represent text are based on the hand-crafted features paradigm: the bag of words model, n-grams analysis, and recently, the learned-features paradigm: the word-to-vec model and other deep learning techniques. However, both of them need of a large corpus to extract general structure to represent each document. In this paper we present a method to represent small poetry parallel corpora, in which a set of features are extracted from the poem to build a vector representation of the poem. We use this representation to compute poetic creativity. We evaluated the proposed method in a bilingual corpus English → Spanish, being the source language English and the target one Spanish. Up to best of our knowledge, this is the first attempt to automatically measure poetic creativity in parallel corpora.