MIT EECS | Philips Undergraduate Research and Innovation Scholar
Adapting PRIMERA for Query-Focused Multi-Document Summarization
Electrical Engineering and Computer Science
- Machine Learning
- Natural Language and Speech Processing
Multi-document summarization (MDS) is the task of extracting information from multiple sources, often about the same topic, into one summary. Query-focused multi-document summarization (qMDS) creates comprehensive yet concise summaries tailored to a particular query. We propose a qMDS model qPRIMERA based on the state-of-the-art MDS model PRIMERA. Our adaptations and points of experimentation are two-fold. First, we use a query-to-sentence scorer to filter unrelated sentences from the document. Second, we combine a query-to-document score with a cluster ROUGE score to identify salient sentences to mask for the Gap Sentence Generation (GSG) objective. qPRIMERA is competitive with other state-of-the-art qMDS models.