Abstract

This paper presents a new task in educational NLP, recommending professional interventions for Special Needs Education (SNE) students using NLP techniques. The task is formulated as a multi-label classification problem in which each training example is formed by the student’s diagnosis along with various free text observations made by teachers and professionals, and the target classes correspond to a set of interventions recommended based on that information. Using the previously mentioned structure, we build the Special Needs Education Corpus (SNEC), a new corpus of over 3,000 Chilean special needs students. We also train several machine learning models using different settings and feature representations of our data. Our results indicate that textual features are the most useful in terms of classification performance and that other non-textual features, such as diagnosis and other chosen interventions, are also beneficial. We also observed a positive effect of representing text inputs with a dense BERT-based representation over using sparse n-grams and non-contextual word embeddings. Our corpus and source code are available at https://github.com/dccuchile/SNEC.