This query necessitates retrieving the two most relevant documents to provide accurate answers. Static-relevant documents are relatively easy to retrieve due to their direct relevance to the query, such as ‘Peter Andreas Heiberg’ and ‘child/son’.
However, retrieving dynamic-relevant documents poses challenges as they are only tangentially related to the query, such as spouse/wife.
Additionally, the vast amount of information on spouse in the knowledge base may cause dynamic-relevant documents to be ranked lower in the retrieval process.
Notably, there is a high relevance between static and dynamic relevant documents, such as Johan Ludvig Heiberg and wife. Considering ‘spouse/wife’ along with the query can facilitate the retrieval of dynamic-relevant documents, thus enabling the extraction of accurate answers.
The study identifies the need to create synergies between multiple documents and establish contextual relevance not only from one document, but from all relevant and applicable documents.
DR-RAG is described as multi-hop question answering framework. This framework does remind much of previous research done on this approach.
The differentiating factor of DR-RAG might be the classifier which the researches designed to determines whether the retrieved documents contribute to the current query by setting a predefined threshold.
The mechanism is aimed at reducing redundant documents and ensures that the retrieved documents are concise and efficient.
Considering the image below, which is an overview of DR-RAG:
Step 1: Retrieve static-relevant documents (SR-Documents) based on their high relevance with the query.
Step 2: Concatenate SR-Documents with the query to retrieve multiple dynamic-relevant documents (DR-Documents).
Step 3: Select each DR-Document individually and combine it with the query and SR-Documents. Feed these combinations into a classifier to determine the most relevant DR-Document.
