Show HN: Using AI to Summarize Terms and Conditions
88% of people never read the terms and conditions of websites or services they use. However, most people want to know what they are agreeing to in those terms. That is why we created Legal Leaf. We strongly believe that everyone should have easy access to those agreements, in language they can understand.
Legal Leaf works behind the scenes, in your browser, to read and summarize these terms using powerful AI. We're constantly working to improve the accuracy of these summaries. The results are displayed in the top right corner without affecting web speeds.
Legal Leaf is a beta product still going through development, but it's improving rapidly and we would love a group of willing beta testers.
41 comments
[ 2.3 ms ] story [ 82.8 ms ] threadOur students compared and benchmarked Sumy against a bunch of other popular summarization techniques, https://rare-technologies.com/text-summarization-in-python-e...
You're about to upload a photo. The T&C of this website gives them irrevocable legal rights to use this photo in [social media settings, advertising campaigns]. Click No to stop the upload, Yes to continue, or Always to stop warning for this website.
What I really hate is when I get T&C at a credit card terminal.
I'm also concerned by the fact this tool could miss some important pieces of information or subtleties. How can its reliability be improved ?
1) A quick side-by-side of a sample Terms/Conditions versus Leaf's summarized version would be helpful. It would help me understand the product more before I install it.
2) What ML/NLP tools did you use for this? It looks like Sumy for Python summarization, along with a specific list of clauses (will, agree, must, etc). When you get a chance, I'd be curious to know more about the technical process.
Also, I noticed that you are stemming words - you may also be interested in lemmatization, which is a slightly more complicated way of converting words into their base forms (like running -> run or ran -> run). Lemmatization also takes into account part of speech context. Given that legal documents are fairly grammatical (I'm assuming?), lemmatization should work well here. I've been fairly happy with Spacy's lemmatization results (https://spacy.io/)
[1] https://en.wordpress.com/tos/
Not being a lawyer, this raises some questions. Some companies go the extra mile to make their ToS quite short and readable for their users, but that text is still reviewed by a lawyer (presumably). But if the summary is auto generated, that review isn't necessarily in place unless leaf.legal is just a summary tool subject to lawyer review before approval and publication to my website.
Also, how do we keep track of which version of the automated summary that is seen by which user? This seems like it could have legal ramifications. For example, the user who violates the ToS because something isn't in the summary. Tomorrow legal.leaf updates its algorithm and regenerates all the summaries for its clients and now the missing ToS article is in the summary.
I imagine there are some pretty solid answers on how to handle these situations. Would love to hear how you are approaching them.
For now, we have a blanket disclaimer that our product is merely for summary and should not represent the will of the company whose terms you are reading.
However, you're absolutely right, we want to make sure that the companies are also well represented. We are building out an avenue for companies to "contest" the summary on their page to be more accurate, or write a personalized summary.
The question can be looked at for any one, what happens if they don't read the ToS at all, and then do something that violates those? We think Legal Leaf is a step in the right direction towards education, but there is still lots of work left to do.
But to be fair, they were citing an award they were given. "Best use of AI/Blockchain"
Also if there was a sample on the home pageit would be great since I'm on a mobile and can't open Chrome store right now
But you're right. There is good room for thought there.
We're working on an example to put on the homepage. It's a good suggestion.
I am currently (and for some time actually) interested into the same problem. Summarization of text. It is a hard one to master. Not a lot of work has been done. I will be happy to be of any help.
(Bonus - a recently published paper about extractive summarization - https://arxiv.org/abs/1708.04439)
Either way, thanks for the feedback!
If the TOS can be summarized into a shortened version that is understandable and readable, then was the original TOS too long and complicated to begin with?
I wonder if summarizing can really distill that which the TOS covers. I further wonder if someone reads a summarized TOS and then violates a part not covered in the summarized version, then who do they blame?