Discriminative and Generative Transformer-based Models For Situation Entity Classification

TL;DR

This paper explores Transformer-based variational autoencoders for situation entity classification, demonstrating improved performance with ample data and insights into low-data regimes, guiding future semantic prediction research.

Contribution

Introduces a Transformer-based variational autoencoder approach for SE classification, highlighting its effectiveness across different data sizes and providing guidance for low-label scenarios.

Findings

Transformer autoencoder improves over state-of-the-art with ample data.

Generative RNNs outperform transformers with extremely small datasets.

Model performance varies significantly with training data size.

Abstract

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.