Literador: a comprehensive tutoring system for Spanish writing

Abstract: In a professional setting and in adult education, a well-written text needs to convey meaning by presenting ideas in a coherent and cohesive form that facilitates readability. This involves factors such as the balance in the use of coreferences, the abstraction of the language used, and the lexical diversity of the text. In this work, we proposed Literador, an Intelligent Tutoring System for Spanish writing. By incorporating different Natural Language Processing (NLP) tools, Literador can analyze and provide feedback on the different aspects of a text, beyond just content. The three main NLP tools that have been incorporated into Literador are BETO, TRUNAJOD, and regular expressions. BETO is a Spanish-trained model of the state-of-the-art BERT language model. Because BETO can capture a highly accurate representation of the Spanish language, it can be used to analyze the meaning of a text, such as determine the accuracy of its main ideas. On the other hand, we use the text complexity analyzer TRUNAJOD to analyze the readability of a text. By integrating over 50 different indicators that measure lexical and syntactical elements, TRUNAJOD can provide insight into key writing factors that can affect the text’s internal connectivity, narrativity, coherence, concretion, and lexical diversity. Finally, we use regular expressions to identify the use of key lexical elements in a text. Regular expressions are used to identify structural sequences that indicate logical entailment, dependency, or adequate closing words for a specific type of text. By combining all of these sources of information, Literador can provide fine-grained feedback on what aspects of a text need improvement. This information is integrated following a feedback strategy that prompts different types of messages depending on the student’s level and tries, which also intends to avoid information overflow.

Keywords: Natural Language Processing, Intelligent Tutor, Feedback

DOI: 10.54941/ahfe100875

Cite this paper:

​