Multi-document question answering (MD-QA) is the process of responding to questions that call for the synthesis of data from various documents.
For information retrieval systems in the actual world, this is a critical capability.
However, the majority of question-answering methods currently in use are either intended for single document QA or open-domain QA with a huge corpus.
To solve the difficulties of MD-QA, researchers from Vanderbilt University and Adobe Research offer a novel strategy called Knowledge Graph Prompting (KGP).
Challenges in Multi-Document QA
Compared to typical QA activities, MD-QA presents the following particular difficulties:
Reasoning across many passages from various publications is referred to as “multi-hop reasoning” and is sometimes necessary to provide an answer. such as connecting items between papers.
Retrieval latency: It takes time to find relevant portions among a huge number of documents.
Multiple modalities: Documents may have different types of structures, including paragraphs, lists, tables, graphics, etc. It is challenging to reason across various modes.
Through the use of built-in knowledge graphs and language model guidance, KGP seeks to address these issues.
Constructing Knowledge Graphs over Documents
The creation of a knowledge graph (KG) to depict the connections between material across documents is the first fundamental principle of KGP. A node in the network is created for each passage or structural component, such as a table or graphic. Edges between nodes are added based on:
Lexical similarity: If two nodes share rare or important tokens, they are related. Associating similar passages is aided by this.
Finding nodes with comparable semantic properties can be done by comparing passage embeddings.
Structural relationships: Relationships like “Page 1 contains Passage A” are added to the document to represent its structure.
