Transformer networks have revolutionized NLP representation learning since they were introduced. Though a great effort has been made to explain the representation in transformers, it is widely recognized that our understanding is not sufficient. One important reason is that there lack enough visualization tools for detailed analysis. In this paper, we propose to use dictionary learning to open up these "black boxes" as linear superpositions of transformer factors. Through visualization, we demonstrate the hierarchical semantic structures captured by the transformer factors, e.g., word-level polysemy disambiguation, sentence-level pattern formation, and long-range dependency. While some of these patterns confirm the conventional prior linguistic knowledge, the rest are relatively unexpected, which may provide new insights. We hope this visualization tool can bring further knowledge and a better understanding of how transformer networks work. The code is available at https://github.com/zeyuyun1/TransformerVis

It has been more than a decade since President George W. Bush set out to get electronic health records for every American. In the 15 years since his pronouncement, there's been significant implementation of EHRs across the country, propagating an incomprehensible amount of data. In many cases, that data sits dormant and untapped of its potential. Some healthcare organizations contend that it's financial constraints that provide limitations. Lack of dollars makes it difficult for all but the most advanced and lucrative healthcare organizations to put machine learning or artificial intelligence in place to make the most of the data.