Interactive Summarization of Large Document Collections

Abstract

We present a new system for custom summarizations of large text corpora at interactive speed. The task of producing textual summaries is an important step to understand large collections of topicrelated documents and has many real-world applications in journalism, medicine, and many more. Key to our system is that the summarization model is refined by user feedback and called multiple times to improve the quality of the summaries iteratively. To that end, the human is brought into the loop to gather feedback in every iteration about which aspects of the intermediate summaries satisfy their individual information needs. Our system consists of a sampling component and a learned model to produce a textual summary. As we show in our evaluation, our system can provide a similar quality level as existing summarization models that are working on the full corpus and hence cannot provide interactive speeds.

An interactive query-based approach for summarizing scientific documents

“…Furthermore, in recent years some studies have proposed novel methods to involve the user in query-based interactive summarization (Avinesh et al , 2018; Hättasch et al , 2019). Avinesh et al (2018) have proposed an iterative summarization approach interactively summarizing large text collections.…”

Report on the International Conference on Design of Experimental Search & Information Retrieval Systems (DESIRES 2018)

“…DESIRES short papers are all single-authored and are presented in a gong-show fashion: 10 minutes for the presentation plus up to 5 minutes for questions. Short papers touched upon several hot and new research topics in IR at large comprising fact checking [18], keyword search over structured data [11], moving beyond single query optimization [7], off-line web search [3], user profiling for searching digital libraries [2], interactive document summarization [12], medical case-based retrieval [22], privacy-safe on-device personal search [24], cost-aware total recall e-discovery [10], topic modeling [25], machine-assisted annotation [16], IR and NLP advances from under an industrial lens [19] and searching for arguments [28].…”