Spelling Check with Sparse Distributed Representations Learning
Thasayu Soisoonthorn, Maleerat Maliyaem, Herwig Unger · 2023
This study focused on enhancing learning sequences using a method inspired by the brain, following Hawkins's approach. Capable of not only recognizing existing sequences but also learning new ones and ensuring fault-tolerant operations, the learning method was evaluated through a spelling check. The evaluation utilized the standard TREC-5 Confusion Track dataset to automatically correct incorrect words. The new method was compared with other techniques, such as Levenshtein Distance, pyspellchecker, LSTM, and Elmosclstm (Semantically Conditioned LSTM and Elmo Transformer), which is the state-of-the-art. The results demonstrated that the highest accuracy at the word level was 79.35%%, surpassing Elmosclstm's 74.41%. Additionally, at the sentence level, the brain-inspired method achieved 90.75% accuracy, outperforming Elmosclstm's 72.18%.