VietLegalQA: Unsupervised Legal Question Answering for Vietnamese Using Cloze Translation Approach

The acquisition of training data for AI modeling, especially in Natural Language Processing (NLP) and Question Answering (QA), is well-known as laborious and time-consuming process. However, the accessibility of such datasets is typically limited to specific domains and languages. In this paper, we introduce a framework designed for generating training data for such scenario, especially in the Vietnamese language and legal domain. Our approach only requires a source of document as input data, then it can generate a high quality dataset that can be used for model training from that data. We first extract named entities (NE) and part-of-speech (POS) components from these contexts, serves them as the basis for the QA pairs dataset. Subsequently, we transform these contexts into fill-in-the-blank (cloze) question formats and serve them as natural questions. Combining the answer and question pairs, we then construct a synthetic dataset. Finally, a simple filter was applied to eliminate the noise data from the synthetic dataset. In our experiments, the model using our method achieved a state-of-the-art performance on ALQAC dataset, a 68.52 in F1 score and 39.80 on exact match (EM) score.