Abstract
Unsupervised extractive summarization typically relies on sentence centrality, which often leads to "facet bias"—the over-representation of a single dominant topic at the expense of others. This paper introduces Facet-Aware Rank (FAR), a novel unsupervised model that mitigates this bias by integrating document-level relevance and noise-filtering mechanisms. Evaluated on eight benchmark datasets, FAR consistently outperforms existing unsupervised methods and even rivals supervised models on specific tasks.
Introduction
Unsupervised extractive summarization aims to identify the most salient sentences in a document without the need for labeled training data. Traditional graph-based methods rank sentences by centrality, assuming that sentences most similar to others are the most important. However, this approach often fails when documents contain multiple distinct topics or "facets."
The Facet Bias Problem
Facet bias occurs when a cluster of highly similar sentences dominates the centrality ranking, causing the summarizer to select redundant information from one facet while ignoring others. For instance, in a news report about a football manager's dismissal, a centrality-based model might select multiple sentences about the firing itself, missing out on fan reactions or historical context. This redundancy isn't just a stylistic flaw; it is a failure to capture the document's full scope.
Methodology: Facet-Aware Rank (FAR)
The FAR model addresses facet bias through two primary mechanisms:
- Modified Directed Graph Ranking: The model introduces a similarity threshold ($\epsilon$). Relationships between sentences that fall below this threshold are treated as noise and ignored. This prevents irrelevant sentences from inflating each other's importance in the graph.
- Facet-Aware Centrality Scoring: Instead of relying solely on inter-sentence similarity, FAR measures how well a candidate summary represents the entire document vector. By representing the whole document as a single vector (using max-pooling over BERT embeddings), the model can rank sentences based on their contribution to a summary that "covers" the entire document's semantic space.
Additionally, the researchers improved sentence embeddings by post-training BERT on a local context task, enhancing the model's ability to understand domain-specific sentence relationships.
Experimental Results
FAR was tested on eight datasets across single-document (SDS), multi-document (MDS), and long-document (LDS) scenarios, including CNN/Daily Mail, NYT, arXiv, and MultiNews.
- State-of-the-Art Performance: FAR consistently outperformed PacSum, the previous leading unsupervised model, across almost all datasets.
- Long Document Gains: On the arXiv dataset, FAR improved the ROUGE-1 score by 1.89 points over PacSum, demonstrating its effectiveness in complex, multi-faceted texts.
- Competitive with Supervised Models: On the NYT dataset, FAR's performance surpassed some supervised baselines (like REFRESH), a significant achievement for an unsupervised approach.
Conclusion
By shifting the focus from local sentence popularity to global document coverage, FAR effectively solves the facet bias problem. The model's ability to filter noise and prioritize multi-faceted relevance makes it a robust solution for summarizing complex, long-form, and multi-document content without the need for expensive labeled data.
Summary updated