Learning to sense from events via semantic variational autoencoder

The work was accepted in PLoS One magazine and addresses two main topics:

Event representation through a semantic variational autoencoder (SVAE)
Detecting events of interest through One-Class Learning (OCL)

The work proposes the SVAE method, a Variational Semantic Autoencoder that represents events. The method is formed by: (i) a context-dependent language model Bidirectional Encoder From Transformers (BERT) (considered one of the state-of-the-art for tasks involving text) that captures syntactic and semantic characteristics of texts; and (ii) a Variational Autoencoder, considered one of the state-of-the-art for representation learning. Once represented, the events of interest are detected using OCL. In OCL, a class of interest is defined and the user provides a set of examples of this class for the algorithm to train. This way, when a new example is presented to the OCL algorithm, the example can be assigned to the class of interest or not. Therefore, it is possible to minimize the user’s labeling effort, since the user will not have to label events of no interest. In the OCL scenario, SVAE showed better results in 177 out of 183 event collections than the BERT representation method.

Marcos Gôlo, Rafael Rossi and Ricardo Marcacini. 2021. Learning to sense from events via semantic variational autoencoder. PLoS ONE 16(12): e0260701. https://doi.org/10.1371/journal.pone.0260701
Codes, Results: https://github.com/GoloMarcos/SVAE-plos-one
Datasets: https://github.com/GoloMarcos/OCTCMG