Deep learning toolset for legal domain

History -- we joined the following competitions:

Global Legal Hackathon 2018, Decoding Law -- make law accessible for all
Access to Justice Hackathon 2018, LegGo -- 21st-century legal information system

Quick start

Typically you'd want to install Tensorflow under Anaconda.

For example, download the Anachonda install script and run it:

 bash Anaconda3-5.1.0-Linux-x86_64.sh

Then create a Tensorflow environment called "tf":

conda create -n tf pip python=3.6

Then activate the environment:

source activate tf

You will see the command prompt changed.

Then install Tensorflow with:

pip install tensorflow==1.5

(As of 2018 March, 1.6 seems to have a bug)

Then install NLTK toolkit with:

conda install -c anaconda nltk

Then install NLTK corpus with:

python
    import nltk
    nltk.download()

and choose "corpus" to download.

Lastly, download the word vectors file from either:

fastText: https://fasttext.cc/docs/en/english-vectors.html
(file name = wiki-news-300d-1M.vec.zip)

or GloVe from Stanford:  https://nlp.stanford.edu/projects/glove/
(file name = glove.840B.300d.zip)

Then unzip the file and put it in the directory \data. (Or you can modify the path in my code, it's possible to run with the zipped version directly.)

Then you can run our code:

python main-RNN.py

Try out some queries after the ? prompt.

Architecture

Currently the code does stage 1, "broad classifications".

In the future we will try to do stage 2.

Demo

The demo classifies queries into 3 categories:

nuisance
dangerous driving
work injuries

This is an example run:

New demo

The new demo isn't very successful (more wrong than correct answers):

Now increased to 10 categories:

The above file is part of the decision tree used by a telephone inquiry system (TelLaw, provided by HK government). Each sub-topic is associated with a pre-recorded message of ~500 words' length. Since the message text is not long, and the sub-categories are all very similar (under FAMILY LAW), our AI does not perform very well, but there may be ways to improve it.

Notes on AI

The AI / deep learning aspects are discussed further in these notes:

AI notes