For the last year or so, our Lexum Lab team has been playing around with machine learning algorithms. ”Telling the fortune” of users based on their search history was one option, but this example showed us that it may not turned out according to plan. Instead our team came up with two applications promising to considerably enhance legal information retrieval. And we are currently looking for partner organizations who are interested in trying them out.
The first one is called Facts2Law. Using the latest deep learning techniques, it predicts the most relevant Canadian case law (and eventually legislation) when presented with a statement of facts. The presence of legal citations in the text body can help Facts2Law provide accurate predictions, but they are not necessarily required. The algorithm learns from the wording of millions of decisions available on the CanLII website and leverages the citation network to cluster them by concept. It has a very large range of potential uses. For example, CanLII users will see early next year an initial implementation of this approach. In addition to cases cited in decisions available on CanLII, Facts2Law will also suggest a selection of other cases that could have been cited but were not. Other potential uses beyond enhancing the search experience on CanLII can also be envisioned, such as identifying missing citations in opinions or briefs, providing guidance over the initial steps of legal research, or creating associations between legal documents based on common characteristics.
The second project is called Learning-to-Rank. It uses machine learning to leverage user metrics, behaviors, and interactions with legal information to improve search results ranking. CanLII has a quite substantial user interaction log and this project puts this collective wisdom to use. Learning-to-Rank does what classic keyword-based algorithms cannot do by taking into account the relative importance of some documents for the legal community. It uses a wide range of new signals indicating user engagement (e.g. number of clicks, time spent reading, number of print requests, etc.) to weigh hits and impact ranking. Ultimately, documents receiving more attention are bumped up in search results lists.
Lexum recently announced the launch of pilot projects aimed at identifying use cases and testing the above AI applications on third party datasets. A pilot can be setup with your organization if a substantial volume of legal information can be provided in a supported format and processing can be completed online.
In this situation, Lexum commits to providing a secure environment and entering into a Non-Disclosure Agreement (NDA) if required. In return, we expect your organization to help us validate the results to assess the success of the pilot. Validation is a straightforward process that consists in clicking on emoticon: are you happy, neutral, or sad about the result provided.
If your organization hosts a set of legal documents and is interested in exploring the potential of AI with us, then let’s get in touch!