Polifonia Mini Textual Corpus

The Polifonia Mini Textual Corpus is made of a balanced selection of sentences representative of the different periods and styles of the corpus. It was created to perform the transformation of sentences into AMR graphs. The reason why the Polifonia Mini Textual Corpus is made of sentences is twofold: (i) the Polifonia Textual Corpus' documents' length prevents them from fitting into GPU memory; (ii) AMR parsers' models are trained to transform no more than a few sentences into AMR graphs; therefore, their performance on large documents parsing is unreliable.

Creating a sub-sample of sentences representative of all the modules of the Polifonia Textual Corpus allows for performing iterative validations of the text2AMR pipeline output. For example, it facilitates testing the algorithms developed to minimise the loss of information that segregating texts into sentences inherently brings (such as the loss of co-reference information). It also favours a more manageable Quality Assurance of the AMR graphs produced by the text2AMR models.

When the results obtained on processing the Polifonia Mini Textual Corpus are considered satisfactory, we will apply the text2AMR pipeline to the Polifonia Textual Corpus in its entirety.

Full description at: https://github.com/polifonia-project/Polifonia-Knowledge-Extractor