Field Extraction from Forms with Unlabeled Data
This addresses the problem of automating form processing without labeled data, which is incremental as it builds on pseudo-labeling and noise-handling techniques.
The paper tackles field extraction from forms using only unlabeled data by developing a rule-based method to generate noisy pseudo-labels and a transformer-based model with a refinement module to handle label noise, achieving effective results as demonstrated experimentally.
We propose a novel framework to conduct field extraction from forms with unlabeled data. To bootstrap the training process, we develop a rule-based method for mining noisy pseudo-labels from unlabeled forms. Using the supervisory signal from the pseudo-labels, we extract a discriminative token representation from a transformer-based model by modeling the interaction between text in the form. To prevent the model from overfitting to label noise, we introduce a refinement module based on a progressive pseudo-label ensemble. Experimental results demonstrate the effectiveness of our framework.