A machine learning system that reads Spanish medical text and extracts structured knowledge: the medical concepts mentioned and how they relate. Built as the UH-MMM team’s entry in the eHealth-KD Challenge at IberLEF 2021, an international research competition, where it ranked 5th overall in the main evaluation scenario.
The challenge has two subtasks. First, Named Entity Recognition finds and classifies the entities in each sentence (concepts, actions, predicates, references). Then Relation Extraction detects the semantic relations between them, turning free text into a small knowledge graph.
Both tasks share one neural architecture: BiLSTM layers encode context, and Dense layers decode the output tags. The models join word-level embeddings (FastText, trained on Spanish medical text) with character-level representations, which helps with the rare and unseen words common in medical text. The entity model also uses multi-task learning, predicting part-of-speech tags alongside entity tags to improve its grasp of sentence structure.
The pipeline is built with TensorFlow and Keras, with spaCy and NLTK for text preprocessing, and covers the full path from raw challenge data to the submission format. The team’s workshop paper linked above describes the approach.