Adapting text simplification decisions to different text genres and target users

Abstract

Hemos analizado las alineaciones a nivel de oración de dos corpus paralelos de textos originales y sus simplificaciones creados con diferentes objetivos. Hemos clasificado las alineaciones que se observan y diseñado un algoritmo de clasificación capaz de predecir si las oraciones de un texto serán eliminadas, segmentadas, o transformadas durante el proceso de simplificación. Hemos realizado una evaluación cruzada en cada uno de los corpus así como una evaluación en la cual se entrena en algoritmo en un corpus y se lo evalúa en el otro.

We investigate sentence deletion and split decisions in Spanish text simplification for two different corpora aimed at different groups of users. We analyse sentence transformations in two parallel corpora of original and manually simplified texts for two different types of users and then conduct two classification experiments: classifying between those sentences to be deleted and those to be kept; and classifying between sentences to be split and those to be left unsplit. Both experiments were first run on each of the two corpora separately and then run by using one corpus for the training and the other for testing. The results indicated that both sentence decision systems could be successfully trained on one corpus and then used for a different text genre in a text simplification system aimed at a different target population.