Tracking entities in technical procedures -- a new dataset and baselines

We present a new dataset, TechTrack, to advance research in the understanding of procedural text - i.e., text that describes a sequence of actions geared towards achieving an end goal. We focus on procedures from technical "Howto"s. Such text is typically seen in FAQs and manuals, which consist of step-by-step answers to questions such as "How to print a test page on a printer?" or "How to troubleshoot a network connection?". The answers consist of step-by-step instructions for completing the task or troubleshooting the problem in the question. This kind of procedural text has specific entities of interest (for example, printer, printer driver, ethernet card, etc.), and these entities have attributes whose values change over the course of actions described in the text. An example procedure from the WikiHow website is shown in Figure 1.