Tell Me More! Using Multiple Features for Binary Text Classification with a Zero-Shot Model

Unsupervised zero-shot text classifiers have large application potential in science and in the private sector. However, especially in settings where these zero-shot classifiers encounter a “domain shift”, the performance of text classifiers is not yet sufficiently accurate for most applications. In this study, we propose an algorithm that tackles this issue by combining the zero-shot prediction information from various combinations of class features. While facing a trade-off in performance and the ratio of predicted text instances, we show that this multi-feature approach can drastically improve the performance of zero-shot classifiers (MCC =. 488 compared to MCC =. 328). Finally, we devise a variety of baselines which can be used to compare the performance of multi-feature zero-shot algorithms.