From Confusion to Clarity: Confusion-Aware Retrieval and Knowledge Injection for Text Classification
The paper targets the failure point after retrieval: labels that look close enough that the model cannot reliably separate them.
The proposed framework finds label pairs the model confuses, adds those confusable labels into the candidate set, and injects generated rules for telling them apart. It does not require fine-tuning. On WOS, Flipkart, and LEDGAR, the authors report up to a 10.0-point Macro F1 gain over retrieval baselines. They also say the rules transfer to smaller 2B-20B models, with gains up to 11.5 points. ArXiv · AI/CL/LG's note
The proposed framework finds label pairs the model confuses, adds those confusable labels into the candidate set, and injects generated rules for telling them apart. It does not require fine-tuning. On WOS, Flipkart, and LEDGAR, the authors report up to a 10.0-point Macro F1 gain over retrieval baselines. They also say the rules transfer to smaller 2B-20B models, with gains up to 11.5 points. ArXiv · AI/CL/LG's note
score 4