Análise automática de coerência usando o modelo grade de entidades para o português

Abstract

A well-written text should be coherent, it means that each new piece of information can be interpreted as the preceding context, a key factor for the readability and interpretability the text. Coherence is usually reflected by the structure of discourse and its relations, which can be analyzed by theories such as RST (Rhetorical Structure Theory), Centering, and models like LSA (Latent Semantic Analysis), which are the main sources of studies for developing tools that can distinguish coherent texts of incoherent in the generation, automatic summarization and evaluation of texts. For the Portuguese language, SciPo (Scientific Portuguese) is an example of a writing tool which has a coherence analysis module that detects potential problems of semantic coherence in abstracts. Based on LSA, this module analyzes different semantic relationships among sentences, according to a pre-established set of dimensions. For one of these dimensions, named Linearity-break, evaluation results were not satisfactory, suggesting the use of other coherence models. In this context, our purpose was to investigate the applicability of the Entity-Grid model, a statistical model based on the Centering theory, in the evaluation of coherence in scientific abstracts written in Portuguese. More specifically, we aimed at assessing whether such model could be employed in the implementation of a classifier capable of detecting local coherence problems, similar to those assessed by the dimension Linearity-break, aiming at future inclusion of such classifier as part of coherence analysis module. Our experimental results are close to those of the original entity-grid model for English language and very similar to the results reported by related works for other languages. Results are also close to those obtained by human judges in an annotation experiment, showing that the entity-grid model can be applied in the context of SciPo system.