Lightweight Passage Retrieval for Open Domain Multi-Document Summarization

View a PDF of the paper titled LightPAL: Light-weight Passage Retrieval for Open Area Multi-Doc Summarization, by Masafumi Enomoto and 4 different authors

View PDF
HTML (experimental)

Summary:Open-Area Multi-Doc Summarization (ODMDS) is the duty of producing summaries from massive doc collections in response to person queries. This job is essential for effectively addressing numerous data wants from customers. Conventional retrieve-then-summarize approaches fall quick for open-ended queries in ODMDS duties. These queries typically require broader context than initially retrieved passages present, making it difficult to retrieve all related data in a single search. Whereas iterative retrieval strategies has been explored for multi-hop query answering (MQA), it is impractical for ODMDS because of excessive latency from repeated LLM inference. Accordingly, we suggest LightPAL, a light-weight passage retrieval methodology for ODMDS. LightPAL leverages an LLM to pre-construct a graph representing passage relationships, then employs random stroll throughout retrieval, avoiding iterative LLM inference. Experiments display that LightPAL outperforms naive sparse and pre-trained dense retrievers in each retrieval and summarization metrics, whereas reaching greater effectivity in comparison with iterative MQA approaches.