Imagine reading a long article or a thick legal contract and knowing, with confidence, exactly which lines prove that an event happened—or did not happen. That is now possible thanks to a research team at Soochow University. They have built a new neural network that not only determines if an event described in a document is real but also highlights the exact sentences that led it to that conclusion. In head-to-head comparisons with earlier approaches, this new model improved overall fact-checking accuracy by 2.5 points and exact-match accuracy by almost 5 points on a standard benchmark.
“We aimed to open the black box of AI decision-making,” says Prof. Zhong Qian, the lead researcher. “By showing exactly which sentences support our model’s verdict, we make its reasoning as clear as stepping through a well-explained proof.”
Why This Makes a Difference
In our fast-paced digital world, false or misleading claims can spread rapidly. Journalists racing to cover breaking stories need tools that do not simply raise a flag but also explain their reasoning. Legal teams reviewing lengthy contracts cannot afford to miss a single misleading clause. This model’s ability to pinpoint the precise text that supports—or contradicts—an event’s truth helps professionals across fields see exactly why a claim stands or falls. It is a step toward AI systems that feel less like inscrutable black boxes and more like transparent partners.
A Tool for Many Fields
This innovation goes far beyond the newsroom. In the media, it could slash the time needed to verify eyewitness accounts or viral online claims by highlighting the most telling sentences. In law, it could breeze through pages of dense text to mark where a statement is grounded in fact or where it may be speculative. Even scientists and developers of future AI systems will benefit from a clear example of how to strike a balance between accuracy and interpretability. In every case, the result is greater trust and faster, more reliable decision-making.
What the Tests Showed
When the team applied their model to an English corpus used by researchers worldwide, it achieved a factual-accuracy score of 66.9%—up from 64.4%—while exact matches rose to 42.9%, nearly five points higher than the previous best pipeline approach. The gains were most dramatic in cases involving speculative language or outright negatives—areas where earlier models often struggled. The model even maintained its edge when tested on a Chinese version of the same dataset, demonstrating its ability to adapt across languages.
How It Comes Together
At the heart of the new approach is a method of examining a document from multiple angles simultaneously. The system builds a web of connections among words, sentences, and special cues such as “not” or “perhaps.” It then homes in on the exact stretch of text that carries the weight of the truth decision. Finally, it blends those pinpointed clues with the overall story of the document to arrive at a final verdict. The result is both precise and coherent, with no loose ends.
This work appears in the June 2025 issue of Frontiers of Computer Science. The authors plan to share their code and detailed annotations, allowing others to build upon their outcomes. As AI becomes a crucial part of our daily lives, innovations like this one promise to keep machines honest, transparent, and valuable—whether we are checking the latest news, reviewing a contract, or simply reading for pleasure.
DOI: 10.1007/s11704-024-3809-6