On 1st & 2nd July 2017 Freeths and sharedo joined forces to create Freedo//m to take part in the 24 hour online courts hackathon.
The Hackathon was organised by the Society for Computers and Law, Legal Geek, the Judiciary of England and Wales and HM Courts & Tribunals Service. The SCL President, Professor Richard Susskind OBE, and Legal Geek’s Jimmy Vestbirk devised and drove the event forward. The hackathon was focused on moving forwards the idea of “access to justice”, and the goal of providing online courts.
Seven exalted travellers ventured down to London for a weekend of enlightenment. Meeting up with another 200+ folk, at the University of Law, intent on exploring new ideas in the hope of building a better court system for the future (and totally ignore the need for sleep in the process).
The team consisted of Ian Stockley (Analyst – Freeths), Guy Berwick (Partner and Head of Innovation – Freeths), Chris Wilson (Architect/Developer – sharedo), Sean Wimpenny (Architect/Developer – sharedo), Phil Swinburn (Head of Product Management – sharedo), David Thorpe (Business Development Director – sharedo) and Tony Johnson (Technical Director – sharedo).
After the challenges had been announced and the team were free to break out to determine which challenges they would look to address and the general flow and functionality of the application. As a group, they decided to focus in on the “outcome prediction” area, as it would allow for interesting uses of machine learning/artificial intelligence to provide a user with a possible outcome, and also a general indication of compensation values.
In order to focus on that though, the application would need to be more fully formed than just displaying a simple result. The team would also need to identify the type of case, and capture data about it from the user, and so, they also opted to partially cover the issues of “dispute classification” and “form filling”, by focussing on a user experience led design (although basic for a 24 hour build!), and the use of additional AI techniques (text analysis, sentiment analysis, topic extraction) to understand natural English language input to classify a problem against a case type.
The final solution was aimed at potential litigants – members of the public with a dispute. They would arrive at the web application, which presents them simply with a text box (like google) that asks them to “describe your legal problem”. The user might enter “I am being threatened with eviction” or “My flat has severe rising damp” – at which point, the textual analysis engine extracts what it believes to be the key topic of their problem. This is then cross referenced to a list of possible case types to provide the user with a suggestion list of what cases they might have meant (with detailed explanations of course).
The idea of this step in the application is to simplify the classification of the case, and in most cases to understand the context of their problem to present one single case type that would be most relevant. Our POC at this point would present any matching case types inferred from the sentiment and topics of the user’s input, allowing the user to select the type of case most appropriate, at which point the data capture would begin. (For the 24 hour POC we only implemented one case type – housing disrepair).
The data capture forms then ask pertinent and context sensitive questions about the user’s complaint. If they are claiming for damp issues for example, which rooms have been damaged by damp, do they run a tumble dryer, do they open a window, or have it vented and so on. As the user is filling in the various guided forms, it is constantly providing feedback from the artificial intelligence/machine learning system – telling them, based on answers so far, we think they have this % chance of winning and should be seeking compensation of £x. As more answers are provided, the AI of course gives more accurate answers (based on the limited training data we primed the machine learning experiment with).
With refinement (this was 24 hours after all), for a user unfamiliar with law but understanding they have a grievance, a system such as this, linking onto a wider online courts environment would provide any user, with limited legal knowledge, with the ability to self-serve their own disputes – to correctly classify their case, understand the correct type resolution required (eg: court), to provide the right data in the right context, and to ultimately start the process of negotiation and settlement before escalation, and manage their dispute through to litigation (if necessary) and final closure. Ultimately providing everyone with an internet connection with “access to justice”.
Obviously, the application is just a POC as it was created in 24 hours, but you can see how it works by following these instructions:
- On the home page (google type search), use phrases similar to those listed below to auto detect topics (using azure topic and keyword extraction which cross references the primary topic/keywords against a list of case types that might correspond);
- My property has severe rising damp
- There has been damage caused by a flood
- I have been threatened with eviction
- My house has issues
- Once the case had been classified, the next stage is data capture (please note that only “housing disrepair” is currently supported)
- Complete the wizard (which you can jump around at will), the data on the right panel will constantly refresh from the machine learning experiment (and some limited training data) in azure ML, to tell you what case law areas could apply, the chance of winning the case based on your answers, and the possible compensation range other similar claims have resulted in.
- The final screen shows the calculation in more detail – how much you could expect to get as a base, how much you might get penalised for if it is your own fault (like claiming for damp issues when you tumble dry without a window open), to arrive at the estimated net figure. It also does some basic precedent matching to return relevant case information.
Philip Swinburn, Guy Berwick, David Thorpe, Tony Johnson, Chris Wilson (hidden), Sean Wimpenny & Ian Stockley (taking the photo).