We study the problem of building text classifiers with little or no training data, commonly known as zero and few-shot text classification. In recent years, an approach based on neural textual entailment models has been found to give strong results on a diverse range of tasks. In this work, we show that with proper pre-training, Siamese Networks that embed texts and labels offer a competitive alternative. These models allow for a large reduction in inference cost: constant in the number of labels rather than linear. Furthermore, we introduce label tuning, a simple and computationally efficient approach that allows to adapt the models in a few-shot setup by only changing the label embeddings. While giving lower performance than model fine-tuning, this approach has the architectural advantage that a single encoder can be shared by many different tasks.

Emotion Classification is the task of automatically associating a text with a human emotion. State-of-the-art models are usually learned using annotated corpora or rely on hand-crafted affective lexicons. We present an emotion classification model that does not require a large annotated corpus to be competitive. We experiment with pretrained language models in both a zero-shot and few-shot configuration. We build several of such models and consider them as biased, noisy annotators, whose individual performance is poor. We aggregate the predictions of these models using a Bayesian method originally developed for modelling crowdsourced annotations. Next, we show that the resulting system performs better than the strongest individual model. Finally, we show that when trained on few labelled data, our systems outperform fully-supervised models.