FACT-SR: Faithfulness-Aware Multi-Task Learning for Aspect-Based Sentiment Reasoning
The paper proposes FACT-SR, a scale-invariant multi-task learning framework that enhances aspect-based sentiment reasoning by enforcing faithfulness constraints on rationale generation and employing a novel automatic weight learning strategy to balance heterogeneous objectives, resulting in superior classification performance and reasoning reliability across various model scales.
Original paper licensed under CC BY 4.0 (https://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
In the world of artificial intelligence, machines have become remarkably good at reading text and guessing how a person feels about a specific thing. This task, known as aspect-based sentiment analysis, asks a computer to look at a sentence like "The food was cold, but the service was excellent" and determine that the sentiment toward the "food" is negative while the sentiment toward the "service" is positive. For years, researchers focused solely on getting the final guess right. However, as these systems move into real-world decisions, simply knowing the answer is no longer enough; we need to know how the machine arrived at that conclusion. The problem is that when computers are taught to explain their thinking, they often invent details that weren't there, mix up which part of the sentence they are talking about, or jump to a conclusion that doesn't logically follow from their own reasoning. These errors, known as hallucinations and logical gaps, make the explanations untrustworthy, even if the final guess happens to be correct.
A team of researchers at the National University of Defense Technology has developed a new method to fix this, creating a system that forces the computer to be honest about its reasoning before it is allowed to make a prediction. They call their approach FACT-SR, a framework designed to teach smaller, more efficient computer models how to generate explanations that are strictly tied to the evidence provided in the text. Instead of just letting a powerful computer write a story about why it feels a certain way, the researchers built a rigorous checking process. They first asked a large, powerful computer to generate a chain of reasoning for a specific aspect, such as the "food" in a restaurant review. Then, they subjected that reasoning to a strict test: does the explanation actually contain enough information to prove the conclusion? If the computer claimed the food was bad but the text only said it was "cold," the system flagged this as a failure because the reasoning did not fully support the claim. If the computer confused the "food" with the "service," that was also rejected. Only the explanations that passed these logical and factual checks were kept to train the smaller student models.
The researchers then taught these student models to perform three tasks simultaneously: generating the explanation, predicting the sentiment, and adhering to three specific rules of honesty. The first rule ensures that the reasoning steps logically lead to the final conclusion. The second rule demands that every word in the explanation must be grounded in the actual text, preventing the model from inventing facts. The third rule requires the model to keep different topics separate, ensuring that praise for the "service" does not accidentally bleed into a complaint about the "food." To make all these tasks work together without one drowning out the others, the team introduced a new balancing mechanism. This system automatically adjusts the importance of each task as the model learns, ensuring that the model pays attention to the quality of its reasoning just as much as it pays attention to getting the final answer right.
When tested on standard datasets of restaurant and laptop reviews, the results showed that this approach significantly improved both the accuracy of the predictions and the reliability of the explanations. On the restaurant dataset, the new method improved the model's ability to correctly identify sentiment by 6.58 percentage points compared to standard training methods, and by 2.33 points on the laptop dataset. More importantly, the explanations generated by the new system were far less likely to contain made-up facts or logical errors. The researchers found that the model learned to create clearer mental boundaries between different topics, keeping the evidence for "food" distinct from the evidence for "service." This was particularly evident when the model analyzed sentences with multiple conflicting opinions, where it successfully avoided the common mistake of mixing up which part of the sentence was positive and which was negative.
The study also demonstrated that these improvements held true across different sizes of computer models, from smaller, faster versions to larger, more powerful ones. Even when the researchers tested the method on a different type of computer architecture, the results remained strong, suggesting that the core idea of enforcing honest reasoning is a universal improvement. By focusing on the quality of the explanation rather than just the final label, the researchers have shown that it is possible to train artificial intelligence to be not only smarter but also more transparent. The work suggests that for machines to be truly useful in sensitive decision-making areas, they must be taught to justify their answers with facts that are actually present in the evidence, rather than relying on patterns that might lead them to invent a story that sounds right but is fundamentally untrue.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.