Prototype system provides 90% accuracy for dealing with parliamentary questions
The Department for Transport (DfT) has revealed that it has begun to test artificial intelligence (AI) for dealing with its correspondence, with the building of a prototype system for extracting data and sorting messages.
Its DfT Lab team has gone public with details in a blogpost that says the project has been aimed at improving the department’s performance in replying to correspondence promptly.
It set out to solve two problems: reducing the time and errors in transposing data from scanned mail to a database; and establishing to which official each piece of correspondence should be allocated.
The team approached the first with the use of a machine learning system, feeding scanned images to Google Cloud Vision API, which recognises the text and converts it to digital form readable by other programs. It was tested with a set of data of publicly available names, addresses and emails.
Identifying names
It then used the Stanford Named Entity Recogniser, which identifies names, and regular expressions – a sequence of characters that define a search pattern – to capture dates, email addresses and reference numbers. This made it possible to automatically fill required fields.
The second part of the solution involved a focus on responding to parliamentary questions (PQs). About 5,000 were fed to the system and its neural network trained to recognise PQs and allocate them correctly. In addition, it pulled information about the sender’s job from Parliament’s API.
The system proved to be about 90% accurate, which is regarded as good for a proof of concept but not good enough for an actual product. As it stands, the system can sort correspondence into teams but not to the right person for drafting replies.
The team is continuing the work with a focus on private office correspondence.
Image by AJC1, CC BY 2.0 through flickr
