Automatic Analysis of Semantic Coherence in Academic Abstracts Written in Portuguese

SciPo is a system whose ultimate goal is to support novice writers in producing academic texts in Brazilian Portuguese through presentation of critiques and suggestions. Currently, it focuses on the rhetorical structure of texts, being capable of automatically detecting and criticizing the rhetorical structure of Abstract sections. We describe a system that enhances SciPo’s functionality by evaluating aspects of semantic coherence in academic abstracts. This system identifies features of sentences based on semantic similarity measures and rhetorical structure. Different machine learning algorithms were trained and evaluated with these features, resulting in three classifiers capable of detecting specific coherence issues on sentences with regard to a rhetorical structure model for abstracts. Results indicate that the system yields higher performance than the baseline for all classifiers.