TextBrew: Automated Model Selection and Hyperparameter Optimization for Text Classification

Abstract

In building a machine learning solution, algorithm selection and hyperparameter tuning is the most time-consuming task. Automated Machine Learning is a solution to fully automate the process of finding the best model for a given task without actually having to try various models. This paper introduces a new AutoML system, TextBrew, explicitly built for the NLP task of text classification. Our system provides an automated method for selecting transformer models, tuning hyperparameters, and combining the best models into one by ensembling. Keeping in mind that new state-of-the-art models are being constantly introduced, TextBrew has been designed to be highly flexible and thus can support additional models easily. In our work, we experiment with multiple transformer models, each with numerous different hyperparameter settings, and select the most robust models. These models are then trained on multiple datasets to obtain accuracy scores, which are then used to build the metadataset to train the meta-model. Since text classification datasets are not as abundant, our system generates synthetic data to augment the meta-dataset using CopulaGAN, a deep generative model. The meta-model is an ensemble of five models, which predicts the best candidate model with an accuracy of 78.75%. The final model returned to the user is an ensemble of all the best models that can be trained under the given time constraint. Experiments on various datasets and comparisons with existing systems demonstrate the effectiveness of our system.

Beyond BERT: Exploring the Efficacy of RoBERTa and ALBERT in Supervised Multiclass Text Classification

“…BERT pioneered bidirectional training, considering both left and right contexts in all layers [52], [53]. In refining this approach, RoBERTa removed the next sentence prediction objective and integrated dynamic masking during training [54], [55]. ALBERT, addressing computational challenges, implemented cross-layer parameter sharing and a factorized embedding parameterization, enhancing efficiency [56], [57].…”