Exploring Multi-Document Information Consolidation for Scientific Sentiment Summarization

Modern natural language generation systems with LLMs exhibit the capability to generate a plausible summary of multiple documents; however, it is uncertain if models truly possess the ability of information consolidation to generate summaries, especially on those source documents with opinionated information. To make scientific sentiment summarization more grounded, we hypothesize that in peer review human meta-reviewers follow a three-layer framework of sentiment consolidation to write meta-reviews and it represents the logic of summarizing scientific sentiments in meta-review generation. The framework is validated via human annotation. Based on the framework, we propose evaluation metrics to assess the quality of generated meta-reviews, and we find that the hypothesis of the sentiment consolidation framework works out empirically when we incorporate it as prompts for LLMs to generate meta-reviews in extensive experiments.