transformers-for-lawyers
AI apps/benchmark for legaltech
Searching better search options for lawyers
DEMO with huggingface, Jina and European Court of Justice judgmentsThis site contains 113 judgments of the European Court of Justice
from 2019 and 2020 concerning tax issues.
All sentences from judgments have been encoded via BERT model
(bert-base-uncased provided by
Huggingface's
transformers library), an example
of a very powerful NLP model that has conquered AI applications.
The infrastructure of the search experience is based on
Jina - a wonderful scalable library to design neural search engines,
based on the newest Deep Learning strategies.
The entire concept - as well as Jina and Huggingface - has a great future in legal tech, because lawyers
need to use a lot of documents, and searching among them is highly challenging...
...That's why law is so compelling and hard
How does it work?
- Write a phrase / sentence
- Click Enter
- You get the most similar sentence (the lower the score, the better)
Enjoy!...
... and be aware that this is a playground. Sometimes BERT doesn't give proper hits,
but sometimes analogies are pretty impressive, like:
QUERY: that complaint was rejected
RESULTS:
- That request was rejected.
- Its application was rejected, as was the objection that it subsequently lodged.
- That request was rejected.
- That is unfair and unlawful.
- That argument cannot be accepted.
Remarks
I am aiming to test other approaches (like other transformer architectures), and fine-tune it, in order to prepare an ultimate benchmark of AI solutions for legaltech, so stay tuned and follow me at LinkedIn, Twitter and on my blog at inteliLex.
If you would like to test it on more documents and play with the code,
please clone this
git repository and contribute to it.
Feel free to contact me: [email protected].
Works on Ubuntu 18.04 and Docker
Below please find how to launch it on Ubuntu. If you are working on (for example) Windows 10 you need to have a Docker engine running and you can skip to the last part "Run on Docker"
1. Upload
Upload documents in *.txt format to search_engine/data and frontendApp/src/assets.
2. Set global environment variables
export PARALLEL=1
export SHARDS=6
export CLIENT_PORT=80
export TMP_WORKSPACE=test_index
export JINA_PORT=567983. Run locally
In the search_engine directory:
python3.7 app.py -t index
gunicorn -w 1 --bind 0.0.0.0:6500 main:appIn the frontendApp directory:
npm install
ng serveAnd you can open the website on http://localhost:4200
Run on Docker
You can easily create Docker apps, but you need to set up the proper variables in DockerFiles for each app.