Structuring Legal Documents with Deep Learning
blog.doctrine.fr
blog.doctrine.fr
As far as I can tell, it's intended to classify each paragraph as one of "facts", "pleas", "grounds", etc. On its own, this doesn't seem particularly useful. I assume it feeds into the broader Doctrine platform (litigation search and case summaries perhaps?). I jumped to their website but couldn't say for sure as I don't speak French.
We classify each paragraph as one of "facts", "pleas", "grounds", etc. in order to generate a table of contents and help users read a court decision more easily. For example, one may want to directly jump to the operative part of the judgement: he just has to click on the "Operative part" of the table of contents, instead of having to read all the decision and guess when that part actually starts.
It's indeed intended to improve our users experience on Doctrine's website https://www.doctrine.fr (legal search engine).
What is the tool's intended output? Is it a discrete classification of documents (or their problem spaces) into predefined categories? What's the number of categories (and some examples)?
What's the cost of simplifying/ignoring most or all of the contextual details in each case? Wouldn't that make a small number of output classes dysfunctionally procrustean?
98.5% accuracy sounds pretty good, but how does system performance compare to the state-of-the-art alternatives that use more traditional probabilistic tag-based methods? Does it tend to fail consistently, perhaps on specific topics or complications?
A legal document from France seems to have come common topics, structured, but as lawyers are free to write in any fashion, the simple fact of finding where in the document is a specific part is time-consuming enougth
In our case using Doc2Vec achieve better results, will try the wang2vec to see if it improves something.
I just had a trial which turned on a couple of sentences in a contract. I could have a paralegal pull every relevant legal in a couple of hours, through “dumb” terms and connecters searching. More than a year of litigation boiled down to applying a rather small volume of law to a large volume of human facts. Even in appeals, which are case law heavy, the process of finding and categorizing the law takes vanishingly little time compared to synthesizing it into a legal framework, and applying it to a large volume of facts.