We set out to create a resource for veterans suffering from PTSD and ended up building open-source legal research tools (LA-MPS), that can work across any legal domain. LA-MPS stands for Legal Apprentice, Marker, Pad, Search. The suite of tools makes it possible to train the AI for your personal legal apprentice on your desktop computer. The process, that works with as few as 20 labeled cases, can produce highly accurate case retrieval. A detailed discussion uses real examples taken from the Bureau of Veterans affairs.
A lawyer annotates the training set with Legal Marker. Marker generates annotated sentences and paragraphs from the original text in the form of json. The resulting enriched cases (now referred to as Legal Semantic json or ls-json). The ls-json data is used to train a TensorFlow classification model that is 80 % accurate with sentence classification in this domain. A pipeline is used to automatically enrich the remaining cases; in our example we did over 10,000. The ls-json cases are loaded into Legal Search for use by the team. An ASP.Net core solution provides the UI for Search which lets the team share notes in a Legal Pad application driven by SignalR. The results can be used to argue a specific case or return new annotations for the next round of training.
The process results in a very efficient way to extract findings, evidence and legal reasoning from a huge corpus of un-researched but relevant cases. Comparisons will be made with general tools like google search and CHAT-GPT4