Explainable AI for Practitioners

Chapter 5. Explainability for Text Data

Language models play a central role in modern-day deep learning use cases and the field of natural language processing (NLP) has advanced rapidly, especially over the last few years. NLP is focused on understanding how human language works and is at the heart of applications such as machine translation, information retrieval, sentiment analysis, text summarization, and question answering. The models built for these applications rely on text data to understand how human language works, and many of the deep learning architectures commonly used today, like LSTMs (long short-term memory), attention, and transformer networks, were developed specifically to handle the nuances and difficulties that arise when working with text.

Perhaps the most significant of these advances is the transformer architecture, introduced in the paper “Attention Is All You Need.”¹ Transformers rely on the attention mechanism and are particularly well-equipped for handling sequential text data. This is partly because of their computational efficiencies and because they are better able to maintain context since text is processed as a whole rather than sequentially. Soon after transformers hit the scene, BERT, which stands for Bidirectional Encoder Representations from Transformers, was introduced and it beat all the GLUE² (General Language Understanding Evaluation) benchmarks for NLU (natural language understanding) tasks ranging from sentiment classification, textual ...