Harnessing conversational AI for legal knowledge

The aim of this project is to design an AI model that can read primary legal documents and assist users with finding information. If we can make legal documents more accessible, this could have an impact on law reform, and contribute to better informed public discourse on legal issues.

Legal documents are often lengthy and difficult to navigate. Recent advances in natural language processing (NLP) has brought interest in applying AI models in the legal domain. The project will involve building a prototype model using open-source large language models, for test usage by students at Melbourne Law School. Depending upon the prototype’s performance, the model could be further developed for public benefit use.

The tool will have two main components:

1. A ChatGPT style conversational, plain-language interface, where the user can ask questions e.g. ‘why did the court side with the employer in this case?’
2. An information retrieval component.

Our project addresses two key barriers to application of AI in the legal domain:

1. Acceptance in legal domain – There has been reluctance within the legal community to rely upon AI models, given the consequences of incorrect legal information. We will address this issue by building a model which answers the user’s query, and extracts the relevant part of the legal document and presents this to the user, so users can verify the answer for themselves.
2. Data labelling – The requirement for specialised legal knowledge to label data has posed a practical barrier to building training datasets of adequate size.This project will utilise a crowdsourcing approach to train the model, with a tiered approach of using law students and legal experts to maximise the gains from domain expertise.

MDAP will provide expertise in NLP and its applications in the legal domain design and build, test and refinement of the prototype model, crowdsourcing methods for data labelling. The datasets involved will be publicly available legal documents.