Abstract
We present a method to automatically generate a concise summary by identifying and synthesizing similar elements across related text from a set of multiple documents. Our approach is unique in its usage of language generation to reformulate the wording of the summary.
Introduction
The proliferation of online news sources has led to significant information redundancy. When a major event occurs, numerous agencies report on it, often sharing a core set of facts while differing in peripheral details. Traditional summarization techniques, primarily based on sentence extraction, struggle in this multi-document context. Extracting all similar sentences leads to verbosity and repetition, while selecting a single representative sentence may introduce bias or irrelevant specifics unique to one source.
This paper addresses the challenge of Information Fusion: identifying the common information across multiple documents and synthesizing it into a single, coherent sentence.
Methodology
The proposed system employs a multi-stage pipeline that treats summarization as a natural language generation task rather than a simple selection task.
1. Theme Identification and Representation
Initially, similar sentences are grouped into "themes." To identify the intersection of information within a theme, the system converts sentences into a dependency-based representation called DSYNT. This representation captures the predicate-argument structure (e.g., "who did what to whom") while abstracting away from surface-level variations like word order or active/passive voice.
2. Paraphrasing and Intersection
The system utilizes a set of paraphrasing rules to match phrases across documents. Analysis of news corpora revealed that approximately 85% of paraphrases are syntactic or lexical (e.g., "regime" vs. "government"). By finding the intersection of these DSYNT structures, the system isolates the facts common to all sources.
3. Temporal Ordering and Generation
To maintain chronological flow, the system assigns timestamps based on the earliest publication date of a phrase. Finally, the FUF/SURGE generation engine is used to realize the fused fragments into fluent English sentences. This allows for complex revisions, such as repositioning temporal markers (e.g., "On Friday") to ensure grammatical flow in the merged output.
Evaluation
The system was evaluated using 200 sentence pairs from the Topic Detection and Tracking (TDT) corpus and a manually constructed test set of 26 themes. The algorithm successfully identified 74% of the predicate-argument structures. It demonstrated high accuracy in matching core components, correctly identifying 69% of subjects and 74% of main verbs.
Discussion and Key Insights
A significant finding of this research is that "surface-level" paraphrasing (grammar and synonyms) accounts for 82% of rephrasing in news, suggesting that deep world knowledge is not always necessary for effective fusion. The authors characterize content planning for summaries as "straddling the border between interpretation and generation."
However, challenges remain. The "negation problem"—where one source reports an event and another denies it—requires deeper contextual understanding to avoid factual errors during fusion. Additionally, while individual fused sentences are fluent, improving overall document coherence through better connectives and anaphora resolution remains a goal for future work.
Conclusion
By moving beyond simple extraction to a generative "fusion" model, this approach produces summaries that are more balanced and less redundant than traditional methods. It leverages redundancy as a signal of importance, transforming the challenge of multi-document overlap into a mechanism for synthesis.
Summary updated