This project was part of research published as a chapter in Richard Toye and David Thackeray's edited collection, Electoral Pledges in Britain Since 1918 available here.
One of the challenges of presenting this work is the scale of the data and related data visualisations. The image below is a heat map showing each of the 'promise lines' identified in the database. Each horizontal line represents a single speech and these are stacked in chronological order from c. 2000 through to 2013. Each horizontal line (or speech) is subdivided into blocks representing each percentile of the speech. So the left-most block corresponds to the first 1% of the speech. Each of those blocks is then colour-coded to indicate how many predicted commitment lines were identified in that block (from zero to four).
You may need to zoom in.
A more user friendly explanation of the technical elements of the Neural Network used will be provided in due course. However, the basic details of the model required to reproduce it in Keras are available here.