The Sounds of Silence: Reducing Noise One Iteration at a Time with Transfer Learning and Synthetic Data
Much research is dedicated to novel deep learning architectures and complex feature engineering approaches to improve performance in deep learning models for NER (named entity recognition) tasks. However, the research devotes less attention to the everyday challenges when your data is just plain “meh.” The majority of text data generated in real-world applications is quite messy. Transforming the mess into large amounts of high-quality, labeled training data may not be feasible given budget or time constraints. In this presentation, we’ll demonstrate the results of addressing noise by varying the training data utilized for a NER machine learning pipeline effectively side-stepping these constraints.