Structured Language Generation Model for Robust Structure Prediction

We propose Structured Language Generation Model (SLGM), a mixture of new loss function and inference method for better generalization of structured outputs. Previous studies on structure prediction (e.g. NER, RE) make use of explicit dataset information, which would boost performance, yet it might pose challenges to robust generalization in real-world situations. Instead, our model gives generalized format information about data indirectly. With format information, we could reduce sequence-to-sequence problem into classification problem via loss calibration and formatted decoding. Our experimental results showed SLGM successfully maintain performance without dataset information, and showed much less format errors. We also showed our model can work like adapters on individual dataset, with no additional training.