Compound task-class noise in task-agnostic inference for multi-task binary classification: an intuitionistic fuzzy approach

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Multi-task learning often trains one shared representation across tasks, where each example is annotated with both a task identity and a class label whose meaning depends on that task. In real annotation pipelines, however, wrong task assignment can make the attached class label answer the wrong question, which makes it unreliable. We formalize this setting as compound task-class noise , where errors in the observed task identity can change the interpretation of the observed class label and can couple with ordinary class-label noise, rather than as two independent errors. This is especially problematic for task-agnostic inference, where no task name is supplied and the model must infer which task should interpret the class label. We propose IFS-Hesitant , which uses a hesitation score derived from Intuitionistic Fuzzy Sets as a structured task-fit uncertainty signal that down-weights suspicious task-class pairs during training. By separating task fit from class evidence, the hesitation score can down-weight suspicious supervision during training and can also route the input at inference without an external task classifier. Experiments on different multi-task classification datasets show that the method is competitive with top baselines when the task is given and substantially more robust when the task must be inferred without external task labels. The same hesitation score supports accurate routing and post-hoc corruption detection, suggesting that structured hesitation is a useful primitive for robust binary multi-task learning under coupled task and class noise.