NLP, bert-based model, neural networks, machine learning, text classification, intent classification, text extraction, voice-to-command, embedded systems
Abstract
This study introduces an innovative implementation of text processing for a voice command interaction system designed for embedded systems, which operates locally, thereby eliminating the need for external resources. The system comprises three modules: a BERT-based TextClassifier for intent classification, a StringParser for extracting pertinent text information, and a TimeParser for temporal information extraction. We developed a unique dataset to train these models, consisting of over 12,000 commands across 66 classes. The TextClassifier model was then benchmarked against other methods from the existing literature. We employed quantization techniques to adapt the model for local operation on embedded devices, achieving an application context inference time of less than 5 milliseconds. The system demonstrated an accuracy exceeding 99% for text classification and over 98% for string parsing.
