Our Legal Text Summarizer project, developed during the Airbus Hackathon, is a practical application of existing advancements in text summarization to address the challenges posed by extensive legal documents.
The model is a combination of classical transformers for text summarization and RAG (Retrieval Augmented Generation). The classical transformers are the main components of the model as the RAG is quite energy hungry and we wanted to make an efficient model regarding energy consumption.
In order to use the text summarizer you must provide a config file in the yaml format :
parameters_textsum_archi:
n_brut_force: 5
models:
number: 2
models_names: ["google-t5/t5-base", "facebook/bart-base"]
models_params: [
["./weight/google-t5/tokenizer", "./weight/google-t5/t5-base"],
["./weight/facebook/tokenizer", "./weight/facebook/bart-base"]
]
modelsRAG:
chunk_size: 1000
chunk_overlap: 100
api_key: "11111111111111111111111111111111111111"
repo_id: "mistralai/Mistral-7B-Instruct-v0.2"
dbRAG:
datapath: "./data.json"There are two ways to use the text summarizer:
- if you want to use the model to evaluate it on the dataset you should have a dataset to test it on :
python main.py --config_file_path config.yaml --data_test_path data.json --evaluation_mode- if you want to use the model just to summarize one text, you can use the following command and put the text in the input window
python main.py --config_file_path config.yaml -
Evaluation
-
One shot summary
The model is composed of two part:
- in the first part we use classical transformers trained on text summarization task like T5 and BART in order to produce summaries using bootstrapping strategies in order to choose the summary from the model that has the highest similarity with the original text. Then, we use an ensemble strategy to choose the summary with again the highest similarity between the outputs of T5 and BART
- the second part of the model is used only if the summary produced by T5 or BART does not have a similarity higher than 0.85. It's a RAG where the vector database is a database composed of the data provided in the config file. The vector database is built using FAISS. Then the prompt is enriched and passed through the LLM that has been chosen in the config file.
This project is under the licence MIT.