Data Augmentation through Paraphrase Generation for Real World Few Shot Entity Tagging

John Chen, Mahnoosh Mehrabani, Srinivas Bangalore

Abstract: Few shot entity tagging is important because different applications of natural language processing typically have different semantics, necessitating custom models. Here, we study few shot entity tagging in a real world scenario insofar that the training data consists of small number of examples per entity type, every entity type has the same number of examples, and there is not any development set. We perform paraphrase generation for many different domains using a T5 model trained on generic paraphrase data. We find that this method produces gains in tagging accuracy across many different domains, and gains are accentuated with an ensemble voting approach.