Publicação
Large language models (LLMs) for legal analysis: RAG and beyond for optimizing domain adaptation in Portuguese legal domain
| Resumo: | This study explores RAG systems tailored to the Portuguese legal domain, highlighting challenges in underrepresented languages. Fixed-size chunking strategies, particularly Token Text Splitter, were found to be most effective, while more advanced techniques like Recursive and Semantic splitting showed little benefits. Larger chunk sizes improved retrieval accuracy and answer quality, though the impact of chunk overlap remains inconclusive. Self-reflection techniques show promising results, particularly for weaker LLMs. Techniques such as adding a pre-post translation proved to be an efficient technique for mitigating language bias. |
|---|---|
| Autores principais: | Barros, Tiago Mendonça Alencar |
| Assunto: | Retrieval-augmented generation RAG Large language models LLM Artificial intelligence AI Hallucination Question answering RAG evaluation Vector store Chunking Legal AI Knowledge graph GraphRAG RDF Graph-based reasoning Self-assessment Self-reflection Multi-agent systems MAS Document reranking Relevance ranking Legal information retrieval Portuguese legal retrieval Machine translation Natural language processing LLM bias Prompt engineering Hierarchical indexing Hierarchical retrieving Chain-of-thought |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | This study explores RAG systems tailored to the Portuguese legal domain, highlighting challenges in underrepresented languages. Fixed-size chunking strategies, particularly Token Text Splitter, were found to be most effective, while more advanced techniques like Recursive and Semantic splitting showed little benefits. Larger chunk sizes improved retrieval accuracy and answer quality, though the impact of chunk overlap remains inconclusive. Self-reflection techniques show promising results, particularly for weaker LLMs. Techniques such as adding a pre-post translation proved to be an efficient technique for mitigating language bias. |
|---|