Now onto Zach Arundel from the DfT Labs (Department for Transport) looking at their experiments with using AI to categorise Parliamentary Questions and other correspondence, around 3000 items per year #ArchivesAI
Identified pain points, assigning questions to the correct person (civil servants tend to move around), cutting out the need to retype data into different systems, getting data out of hard copy correspondence #ArchivesAI
Use Google Cloud Vision API to read image of hard copy, identify addressee, sender, addresses etc #ArchivesAI
Much of AI is pattern matching, began with Support Vector Machines #ArchivesAI. Simple divisions can start to categorise, eg tell difference between muffins and cup cakes based on sugar and fat content #ArchivesAI (getting #ArchiveCake in too)
For PQs, this can be enough to help divide between relevant teams #ArchivesAI
As a lab, they tried various different AI approaches, moving on from the Support Vector Machines, then neural nets, then Word2Vec, Convoluted Neural Net, could then compare effectiveness of different approaches #ArchivesAI
The fanciest method doesn't always give the best/most accurate results. Match appropriate tool to your task. #ArchviesAI
Though you can buy in expertise, important to also have enough internal knowledge to understand to some level what machine learning is and if claims for it are reasonable #ArchivesAI
Some examples, comparisons of the results from different approaches #ArchivesAI 
Working in four week sprints, demonstrating initial value of an approach, then helping the department go out to market for full-scale solutions #ArchviesAI
Questions: @johnlsheridan - how do we preserve these models as Public Records in their own right?
@archivista13 asks about training data, how much work did they have to do to generate the training data? In this case effectively had a large reservoir of already classified PQs (though of course this recorded who had answered a given question in the past - the best?) #ArchviesAI
