TL;DR
This paper investigates the explainability of GraphSum, a state-of-the-art graph-based transformer for multi-document summarization (MDS). The authors demonstrate that the model's internal attention weights—specifically in the later decoding layers—correlate strongly with the actual source of information, revealing a significant "positional bias" where the model prioritizes the opening paragraphs of input documents.
What Problem Does This Paper Address?
Multi-document summarization (MDS) is the challenging task of condensing information from several overlapping sources into a single, coherent summary. While Transformer-based models have pushed the state-of-the-art, they remain "black boxes." We know they work, but we don't always know why they choose specific pieces of information or where that information came from in the source documents.
Graph-based models like GraphSum attempt to add structure by representing text units (like paragraphs) as nodes in a similarity graph. However, two major questions remained:
- Granularity: Is it better to represent the graph using sentences or paragraphs?
- Explainability: Can we use the model's internal attention mechanisms to track the "source origin" of generated summary sentences?
Methods
The researchers extended the existing GraphSum architecture to compare different textual granularities and analyze its internal mechanics:
- Granularity Comparison: They modified GraphSum to support sentence-level representations, building TF-IDF similarity graphs where nodes are individual sentences rather than paragraphs.
- Attention Extraction: They focused on the Global Graph Attention mechanism within the decoder. As shown in Figure 1, this mechanism uses the graph structure to guide which input units the model attends to during generation.
- Source Origin Analysis: To see if attention weights actually reflect information flow, they created a processing pipeline (Figure 2). They aggregated token-level attention weights to a sentence level ($A'$) and compared them against a reference metric ($R$)—calculated using ROUGE scores between the generated summary and the original input paragraphs.
Datasets
The study utilized two standard MDS benchmarks:
- MultiNews: A collection of news articles from newser.com with professional human-written summaries.
- WikiSum: A massive dataset where the goal is to generate a Wikipedia lead section based on search engine results and cited references.
Key Insights and Ideas
- Paragraphs Win on Efficiency: Surprisingly, paragraph-level representations slightly outperformed sentence-level ones. The authors attribute this to the "long-tail" nature of news: many paragraphs in MultiNews are only 1-2 sentences long anyway. Moving to sentences increases graph complexity and memory consumption without providing a significant information gain.
- Attention as a Map: The "Global Attention" weights aren't just random numbers; they are highly correlated with the ROUGE-based similarity of the source text. This means we can treat these weights as a built-in "explanation" for which source paragraphs influenced a specific summary sentence.
- Layer Specialization: Not all layers are equal. The researchers found that decoding layers cluster into two groups. The later layers (5–8) show a much higher correlation with the source origin than the early layers, suggesting that the model refines its "focus" on specific source content as it moves deeper into the decoder.
- The Lead Bias: The model exhibits a strong positional bias. It learns that in news, the most important information is usually at the beginning.
Key Results
- Performance: On MultiNews, paragraph-level units achieved a ROUGE-1 score of 44.82, compared to 43.82 for sentence-level units (at the same batch size).
- Correlation: In the MultiNews dataset, the correlation between attention weights and the source origin metric reached as high as 0.69 in decoding layer 6.
- Positional Bias Visualization: Heatmaps (Figure 6) clearly show that for almost every sentence in a generated summary, the model is most heavily attending to the first few paragraphs of the input documents.
Limitations and Future Work
The authors identify several areas for further exploration:
- Resource Constraints: Sentence-level models require significantly more VRAM due to the larger graph structure, which limited the batch sizes the authors could test.
- ROUGE as a Proxy: Using ROUGE to determine "source origin" is an approximation. The authors suggest that human-annotated source origins would provide a more robust ground truth for evaluating explainability.
- Content vs. Position: It remains unclear if the model picks the first paragraphs because of their positional encoding or their actual content. Future work involving shuffling input documents could help disentangle these factors.
- Dataset Metadata: The WikiSum analysis was limited because the dataset lacks metadata linking paragraphs back to specific source documents, preventing a full positional bias analysis across documents.
Summary updated