DataSet

We have scraped and complied a corpus of 3k+ Indian legal judgments and their parallel summaries.

from datasets import load_dataset

dataset = load_dataset("d0r1h/ILC")

train_set = pd.DataFrame(dataset['train'])
test_set = pd.DataFrame(dataset['test'])

Code

git clone https://github.com/d0r1h/ILC.git
cd ILC
pip install -r requirement.txt

Summarzing using Extractive approach

!python Code/Models/extractive.py \
        --output_dir dir_name \
        --text_column text \
        --summary_column summary \
        --data_file data.csv \
        --sentence_count 3 

Training LED using Abstractive approach

!python Code/Models/led_summarization.py \
        --model_name  allenai/led-base-16384 \
        --text_column  Case \
        --summary_column Summary    \
        --max_input_length  8192 \
        --max_output_length  600 \
        --batch_size 2 \
        --num_beams 2 \
        --output_dir output_dir_name

Inference on test-set using led-base-ilc model

Notebook	Colab
led-base-ilc

Results:

Following results are obtained on test-set with transformer based models and extractive methods

Algorithm / model	Rouge-1	Rouge-2	Rouge-L
Extractive
SumBasics	15.69	6.02	14.48
LSA	21.20	7.37	19.76
KLSum	21.40	10.19	19.66
LexRank	33.09	16.81	22.99
TextRank	34.54	18.10	31.11
Abstractive
LedBase	4.31	1.08	4.11
Led-ilc	42.24	23.18	39.30