A Multi-label Few-Shot Learning with Combinations of Layers

Multi-class and multi-label classification on noisy call transcript data generated by speech-to-text (STT) systems is challenging due to the different human accents and transcription errors. The multi-labeling task is even more complicated if the data points have only single or no labels. This study has three main contributions to solving these problems: (1) To overcome the labeling problem, we train a multi-class classification model and use a minimal set of manually annotated data to determine a threshold. We obtain a multi-label classifier by utilizing a multi-class classifier with this threshold. (2) To overcome the noise issue, we propose concatenating well-known feature extraction techniques such as word2vec, tf-idf, transformers, and fuzzy embeddings. This combined feature extraction method is more resilient to noise with proper configurations than stand-alone techniques. (3) This is an industry task; we must protect our client’s data. Hence to carry out our success on French private client data to benchmark data, we propose a noising pipeline that artificially mimics the observed STT transcription errors. We combined these solutions in an NLP framework, enabling us to achieve state-of-the-art results with fewer resources, such as manually annotated data or multiple GPU utilization.