The primary objective of this project is to retrieve Vietnamese legal documents based on user-provided queries. The retrieval process is divided into three key stages:
-
First Stage Retrieval: Utilizing BM25
- In this initial stage, the BM25 algorithm is employed to swiftly narrow down the search scope and present a collection of potentially relevant documents.
-
Second Stage Retrieval: Fine-tuning BERT with LoRA (Low-Rank Adaptation)
- In this phase, the BERT model is enhanced through the application of the LoRA adapter layer. This innovative approach to fine-tuning BERT is especially valuable for low-resource languages like Vietnamese. It facilitates the fine-tuning process on a small dataset without the necessity for a large corpus.
-
Ensemble of BM25 and BERT: Enhanced Retrieval
- The final results are achieved through an ensemble approach, combining the outcomes from both stages. This synergistic approach capitalizes on the unique strengths of each model, resulting in a robust retrieval system proficient in comprehending both lexical and semantic aspects of the query.
Sample Query:
Trường hợp công dân không trong độ tuổi nhập ngũ, nếu đi du học, xuất khẩu lao động không cần phải khai báo tạm vắng, đúng hay sai?
(English Translation: "In the case of citizens who are not of military age, if they study abroad or export labor, they do not need to declare temporary absence. Is it true or false?")
Output:
"relevant_articles": [
{
"law_id": "Luật Cư trú",
"article_id": "31"
}
]
