Repositorio Digital de la UPF
Publicación
Artículo científico (JournalArticle)
Corpus-based sentence deletion and split decisions for Spanish text simplification, Eliminación de frases y decisiones de división basadas en corpus para simplificación de textos en español
This study addresses the automatic
simplification of texts in Spanish in order to make them
more accessible to people with cognitive disabilities.
A corpus analysis of original and manually simplified
news articles was undertaken in order to identify
and quantify relevant operations to be implemented
in a text simplification system. The articles were
further compared at sentence and text level by
means of automatic feature extraction and various
machine learning classification algorithms, using three
different groups of features (POS frequencies, syntactic
information, and text complexity measures) with the
aim of identifying features that help separate original
documents from their simple equivalents. Finally, it
was investigated whether these features can be used
to decide upon simplification operations to be carried
out at the sentence level (split, delete, and reduce).
Automatic classification of original sentences into those
to be kept and those to be eliminated outperformed the
classification that was previously conducted on the same
corpus. Kept sentences were further classified into those
to be split or significantly reduced in length and those
to be left largely unchanged, with the overall F-measure
up to 0.92. Both experiments were conducted and
compared on two different sets of features: all features
and the best subset returned by an attribute selection
algorithm., Este estudio aborda el problema de
simplificación automática de textos en español con
el fin de hacerlos más accesible a las personas
con discapacidades cognitivas. Análisis de corpus
de artículos originales y artículos simplificados
manualmente se ha realizado para identificar y
calificar relevantes operaciones que tienen que ser
implementadas en el sistema de simplificación de
textos. Luego los artículos se han comparado al nivel
de frase y texto mediante extracción automática de
características y diversos algoritmos de aprendizaje
de máquina para clasificación usando tres distintos
grupos de características (frecuencias de partes de
oración (POS), información sintáctica y medidas de la
complejidad de textos) con el propósito de identificar las
características que ayuden a distinguir los documentos
originales de sus simples equivalentes. Finalmente, se
ha investigado la posibilidad de usar esas características
en operaciones de simplificación a nivel de frase (dividir,
eliminar y reducir). Clasificación automática de frases
originales en las que deben preservarse y las que
deben eliminarse ha superado la clasificación anterior
sobre el mismo corpus. Las frases guardadas luego se
clasificaron en las que se dividen o reducen de manera
significativa en su longitud y las que se quedan sin
cambios mayores con la F-medida de 0.92. Ambos
experimentos se realizaron y compararon sobre dos
distintos conjuntos de características: el de todas
características y el mejor subconjunto recuperado por el
algoritmo de selección de atributos., The research described in this paper was partially
funded by the European Commission under
the Seventh (FP7 - 2007-2013) Framework
Programme for Research and Technological
Development (FIRST 287607). This publication
[communication] reflects the views only of
the authors, and the Commission cannot be
held responsible for any use which may be
made of the information contained therein.
We acknowledge partial support from the
following grants: Avanza Competitiveness grant
number TSI-020302-2010-84 from the Ministry
of Industry, Tourism and Trade, Spain and grant
number TIN2012-38584-C06-03 and fellowship
RYC-2009-04291 (Programa Ramón y Cajal
2009) from the Spanish Ministry of Economy and
Competitiveness.