Discriminative and Generative Transformer-based Models For Situation Entity Classification
This work addresses situation entity classification for NLP researchers, providing guidance for low-label training regimes, but it is incremental as it builds on existing Transformer and generative methods.
The authors tackled situation entity classification by developing a Transformer-based variational autoencoder that encodes sentences into a latent space for text generation and classification, showing it improves over previous discriminative models with plentiful data and performs better with smaller data, though generative RNNs outperform in extremely low-data scenarios (e.g., 4 instances per label).
We re-examine the situation entity (SE) classification task with varying amounts of available training data. We exploit a Transformer-based variational autoencoder to encode sentences into a lower dimensional latent space, which is used to generate the text and learn a SE classifier. Test set and cross-genre evaluations show that when training data is plentiful, the proposed model can improve over the previous discriminative state-of-the-art models. Our approach performs disproportionately better with smaller amounts of training data, but when faced with extremely small sets (4 instances per label), generative RNN methods outperform transformers. Our work provides guidance for future efforts on SE and semantic prediction tasks, and low-label training regimes.