Narrative Consolidation: Formulating a New Task for Unifying Multi-Perspective Accounts
This paper introduces "Narrative Consolidation" as a distinct NLP task for unifying overlapping documents into a coherent, chronologically sound text, establishing a new benchmark based on the Biblical Gospels and demonstrating that explicit temporal alignment is the critical factor for success while highlighting the need for principled mechanisms to surpass simple length-based heuristics.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine trying to tell a single, coherent story from four different witnesses who were all at the same scene but saw it from different angles. One might describe the weather in detail, another the specific words spoken, and a third the sequence of actions, yet none of them have the full picture alone. In the world of computer science, this is a problem known as multi-document summarization. For decades, researchers have taught computers to read many documents and squeeze them into a shorter, punchier version, stripping away repetition to get to the "main points." This works well for news briefs or meeting notes where brevity is the goal. But it fails miserably when the goal is to reconstruct a complex, unfolding story where the order of events matters more than the length of the text. If a computer simply picks the most important sentences from four different accounts of a historical event, it might end up with a jumbled mess where the ending appears before the beginning, or where crucial details from one witness are lost because they seemed redundant compared to another.
A team of researchers from Brazil has identified this gap and proposed a new way to think about the problem, calling it "narrative consolidation." Instead of trying to make a story shorter, their goal is to make it whole. They wanted to build a system that takes multiple overlapping accounts and weaves them into a single, fluid narrative that respects the strict timeline of what actually happened. To test if this was possible, they turned to an ancient challenge that theologians have wrestled with for centuries: harmonizing the four Gospels of the Bible. These four texts tell the story of the final week of Jesus's life, but they do so with different details, different emphases, and sometimes different sequences of events. The researchers treated these texts as a dataset, creating a new benchmark to see if a computer could stitch them together into one perfect, chronological story without losing any of the unique details each source offered.
The researchers built a new tool called the Temporal Alignment Event Graph, which acts like a scaffold for the story. Instead of letting the computer guess which sentences are important based on how often words repeat, this tool forces the computer to follow a pre-established timeline of events. It organizes the text by first identifying the 169 specific events that make up the story of that final week, such as the triumphal entry into the city or the last supper. For each of these events, the computer looks at how all four Gospels describe it and must choose the best version to include in the final story. The team tested several different strategies, ranging from a simple method that picks the longest description of each event to a complex system that tries to analyze the relationships between the different accounts.
What they found was surprising and humbling. The most powerful factor in creating a good story was not the sophistication of the computer's selection algorithm, but simply having the correct timeline to follow. When the researchers gave their systems a clear, ordered list of events to follow, the quality of the resulting story jumped dramatically. A system that ignored the timeline and just picked the most "important" sentences produced a narrative that was chronologically confused and scored very poorly. However, once the timeline was locked in, even a very simple strategy—picking the longest available description for each event—performed remarkably well, often outperforming the more complex, graph-based systems. The simple rule of "choose the longest version" worked because, in this specific dataset, the longer accounts tended to contain the most details, and since the sources were complementary rather than contradictory, more detail meant a better story.
The study also revealed what does not work. The researchers tested whether the computer could learn to pick the best version by looking at how similar the different accounts were to each other. They found that this approach provided almost no useful information. The different descriptions of the same event were so similar in their wording that the computer could not use surface-level similarities to decide which one was better. The only thing that truly helped the computer make a smart choice was the structure of the story itself—the temporal connections between events. This suggests that for this type of task, the hard part is not choosing the right words, but understanding the correct order of things. The researchers conclude that while their new tool provides a solid foundation, the biggest challenge remaining is teaching computers to figure out that timeline in the first place without being given one. Until that is solved, the best way to unify multiple stories remains to first establish the sequence of events, and then let the details fill in the gaps.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.