Manually dragging and dropping tabs on a DocuSign agreement can get repetitive, especially when faced with a field-heavy document. In this episode, we’re sitting down with Raphael Alabi, Senior Machine Learning Engineer, to talk about our auto-tagging feature. In collaboration with Google last year, we created the fastest document tagging solution and successfully deployed it to our production environment.
The solution involves using computer vision (CV) in addition to state-of-the-art natural language processing (NLP) algorithms. The CV algorithm enabled us to identify potential tag locations within a document as well as the words contained within the document; while LSTM/BERT based NLP algorithm enabled the correct differentiation of the tags into signature, date and text tags. The architecture is very flexible as it can accommodate more tag differentiations via tuning of the LSTM/BERT models. The architecture is also scalable since inference can be run on multiple GPUs thus enabling a faster millisecond response from the NLP/CV models.