Research map / Graph learning / Knowledge and temporal graphs
MultiSum: A Multi-Facet Approach for Extractive Social Summarization Utilizing Semantic and Sociological Relationships
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MultiSum, an unsupervised extractive social summarization model that fuses social and semantic post graphs with a two-stream GCN and selects posts via a centrality score combined with summary-document similarity to cover multiple facets.
Problem
Social summarization of many short, sparse posts on a topic suffers from weak post representations and from centrality-based selection that over-focuses on one facet, causing facet bias and redundancy.
Approach
Model multiple post relations (social consistency/contagion and semantic similarity) jointly with a multi-stream GCN encoder trained with content, graph, and summary-similarity reconstruction losses, and select summaries by multiplying summary-document similarity with directed centrality so the summary covers multiple facets.
Key results
- On TWEETSUM, MultiSum reaches R-1 45.953 and R-2 14.236, above the best baseline SCMGR (45.829 / 14.081).
- On Weibo, MultiSum reaches R-1 36.501 and R-SU* 12.935, above SCMGR (36.405 / 12.722).
- Human and LLM evaluation: MFS covers more facets than SRE (human 8 vs 7 out of 10 facets) and has lower redundancy in most evaluators (human 2 vs 3).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the social relation encoder (ablation) | MultiSum · TWEETSUM | R-1: 44.981 -> 45.953 | Helped |
| add the social relation encoder (ablation) | MultiSum · TWEETSUM | R-2: 13.126 -> 14.236 | Helped |
| add the semantic relation encoder (ablation) | MultiSum · TWEETSUM | R-1: 45.122 -> 45.953 | Helped |
| add the semantic relation encoder (ablation) | MultiSum · TWEETSUM | R-L: 43.213 -> 43.802 | Helped |
| combine social and semantic relations in the multi-stream encoder (ablation) | MultiSum · TWEETSUM | R-1: 44.916 -> 45.953 | Helped |
| combine social and semantic relations in the multi-stream encoder (ablation) | MultiSum · TWEETSUM | R-2: 13.017 -> 14.236 | Helped |
| combine social and semantic relations in the multi-stream encoder (ablation) | MultiSum · TWEETSUM | R-L: 42.915 -> 43.802 | Helped |
| combine social and semantic relations in the multi-stream encoder (ablation) | MultiSum · TWEETSUM | R-SU*: 19.792 -> 20.618 | Helped |
Surprising findings
- The multi-facet summarizer performs about the same as the sparse reconstruction extractor on ROUGE (TWEETSUM R-1 44.916 vs 44.886), with the gain mainly in facet coverage rather than ROUGE.
- GPT-4 rated MFS as more redundant than SRE (4 vs 4 tie), unlike the other evaluators.
Limitations
- Gains on Weibo are small and the model underperforms on abstractive references, since it is extractive.
- Facet evaluation relies on a single sub-topic with three annotators and LLM judges, a limited evaluation scope.
- Several hyperparameters (alpha, beta, lambda) are tuned by grid search on ROUGE-1.
Experimental setup
Datasets: TWEETSUM Weibo
Benchmarks: TWEETSUM Weibo
Metrics: ROUGE-1 ROUGE-2 ROUGE-L ROUGE-SU* facet count redundancy score
Baselines: Centroid LSA LexRank DSDR MDS-Sparse PacSum Spectral MTGNN SNSR SCMGR
Builds on: PacSum SCMGR BERTweet Kipf and Welling GCN Liang et al. 2021 facet-aware summarization
Closest papers by meaning
- PersonalSum: A User-Subjective Guided Personalized Summarization Dataset for Large Language Models (NeurIPS 2024 · similarity 0.84)
- Language Models of Code Are Few-Shot Planners and Reasoners for Multi-Document Summarization with Attribution (AAAI 2025 · similarity 0.83)
- Unsupervised Extractive Summarization with Learnable Length Control Strategies (AAAI 2024 · similarity 0.82)
- EventSum: A Large-Scale Event-Centric Summarization Dataset for Chinese Multi-News Documents (AAAI 2025 · similarity 0.82)
- Distilling Structured Rationale from Large Language Models to Small Language Models for Abstractive Summarization (AAAI 2025 · similarity 0.82)
- Learning from Failure: Improving Meeting Summarization without Good Samples (AAAI 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.