HYBRID ARCHITECTURE FOR AUTOMATIC TRANSLATION FROM BRAZILIAN PORTUGUESE TO LIBRAS IN AN EDUCATIONAL CONTEXT: AN APPROACH BASED ON SEQ2SEQ, MAPPING DICTIONARY, AND WORD EMBEDDINGS

Abstract

This paper presents the hybrid architecture developed for an automatic translation system from Brazilian Portuguese to Brazilian Sign Language (Libras) gloss, designed to support deaf students in mainstream classrooms. The approach combines three complementary translation strategies: (i) a Sequence-to-Sequence (Seq2Seq) neural model with Long Short-Term Memory (LSTM) networks, trained on a proprietary corpus of 800 sentence pairs; (ii) a high-precision direct mapping dictionary with approximately 500 entries; and (iii) word embeddings trained with the Gensim library for out-of-vocabulary (OOV) token handling. Module integration is managed by an orchestration layer based on the Strategy design pattern, enabling dynamic selection of the best strategy per sentence. The system also incorporates offline automatic speech recognition (Vosk engine) and asynchronous WebSocket communication, achieving total latency of 150–350 ms and supporting 10 simultaneous sessions. Results indicate 95% accuracy for training corpus sentences and 33% for unseen sentences, with translation confidence between 80 and 95%. The work validates the technical feasibility of the hybrid approach and identifies corpus expansion and Transformer-based architectures as priority future directions.