I am an attorney that has spent the last 15 years drafting and negotiating contracts, primarily in the tech and IP space.
Reviewing and drafting contracts is a time consuming and laborious process, requiring line-by-line analysis of a given contract. Although reviewing and redlining a contract often takes an inordinate amount of time, the dirty little secret (that lawyers don’t like to share) is that lawyers often reuse, recycle, or paraphrase certain provisions based on past work. Reviewing and revising a contract is often a “cut-and-paste” exercise, pulling in past used provisions from a variety of sources. I’m convinced that this process can be automated via GPT3 and NLP in general.
I have a fairly robust document data set of various contracts and accompanying metadata (about 15 years of legal work). I’m interested in partnering up with a developer to take my data set, incorporate with GPT3, and create a MVP that can analyze a given contract and automatically insert preferred contract provisions/ language based on past work/ drafting tendencies.
Please contact me if you are a developer with any interest in the above. Thanks,
This sounds like a perfect use-case for finetuning, and also a prime example of how AI will help automate even highly intellectualized jobs. Let’s sync up and discuss, as I’m already brainstorming how this can be done.
I dont think it would work since the algorithm complexity is of quadratic nature. We basically would need to train a GPT model on your data and run it together with GPT-3. GPT-3 can be used for data distillation but still would end up with the training part.
Short-cuts or a cheaper implementation like the ppl at Jarvis AI did but in legal field could cause bad backlash.
It is not impossible, just require a lot of capital in research for a reliable implementation.
I’m currently studying the contract review mechanism with a lawyer in France. Not sure if you’re still looking for a partner but maybe worth having a conversation.
Let me know.
Nice to hear from you, and appreciate you reaching out.
Yes, I’ve been working with a number of programmers the last few months on leveraging GPT-3 for legal analysis and the like.
I’m curious as to what you and your partner have been working on, and would be interested to compare notes. I’m based in Amsterdam, NL. Where are you based?
Thanks and looking forward to hearing back from you.
I’m in France (Hendaye), so we basically have the same time zone. Sending you a PM with my phone number.
The goal is to save the lawyer’s time reviewing the contracts and detecting issues or misalignments with contract policies.
We are at the beginning of the thing (2 weeks).
Combine classic text treatment algorithms with GPT-3 classifications/completions to take raw text (often badly formatted) and translate it into structured entity objects (with the help of GPT-3) to simplify data extraction, analysis, and checks against the existing policies (with classic programming).
Solved so far:
- getting badly formatted text string of a contract and formatting it into standard contract form:
Name of the contract text \n text parties article 1 article one text \n text etc.
Name of the contract
Article 1 - Article Title
Article 2 - Article Title
- label the text based on the function in the contract:
eg. output as object with properties :
Name of the contract - header
Parties - title
Parties definitions - text
Article 1 - Article Title - title
Article body - text
Article 2 - Article Title - title
Signature bloc - meta text
Working on the hierarchy of the objects within a contract and their classification.
And you guys, where are you at?
Serge, thanks for the quick response.
In combining email responses:
I don’t have any availability left today to chat, but will tomorrow late morning/ early afternoon. Let me know what times best work for you, and I’ll accommodate.
I think the use case for our idea and your idea are slightly different (We are focusing more on identifying non-standard provisions, listing a recommendation, and then actually implementing that recommendation automatically in the text via redline.) However, the issues remain the same - i.e. structuring unstructured legal data so it can be properly fed into a LLM like GPT-3, Bert, whatever.
We are about six months in, so have gotten a bit further in this process, but still have a lot of work to do. It’s definitely taking us longer than expected, but slowly but surely getting there. Hugging Face has a number of smart document ai transformers we’ve been playing with recently that have shown initial promise. I’d be happy to share where we are (but please note, I’m the lawyer and not one of our tech team).
Hi, have you looked at the existing companies doing legal contract analysis? This is a very advanced field with many big players.
My lawyer had several demos from what is available here in France and said exactly that: "Damn, they call it advanced?”.
So I think there is some space for me to play a bit.
In about three weeks we could get some decent results out of nowhere. So more time will probably make big players move a bit to free the space for us.
I bet you guys need someone from linguistics in your team. That one helps me all the time, especially in this domain.
Point taken. I am not familiar with the European market. Regarding North America, I’d say contract analysis is saturated and a start-up would have a very hard time breaking in.
Hey Dennis, how’s the progress? Good results in non-standard provisions identification/classification?