This repository contains the experimental code for Text-to-Text Classifier.
Text classification was conventionally handled by supervised classifiers which treat labels as a set of indices. Recently, the NLP community increasingly treats all tasks as a text-2-text generation paradigm. Especially for some generative large language models, such as the GPT series, are highly reliant on this paradigm.
Although this seems a natural choice for text generation task, but if it really fit text classification tasks? Will it decrease or increase the text classification performance?
We tend to systematically study this research question by conducting experiments on various popular classification datasets.
- 06/02/2023: Project stopped due to the scooping :( (https://arxiv.org/pdf/2211.08099.pdf; https://arxiv.org/pdf/2110.08426.pdf).
- 06/01/2023: Upload FewNERD experiments, a challenging NER datasets where there are 66 fine-grained entity types.
- 05/29/2023: Upload FewRelv1.0 experiments to investigate the cross-label generalization setting. See the README for more details.
- 05/14/2023: Add SuperNI experiments to investigate (1) the difference between classifier and generator on the CLS part of the held-out evaluation; (2) whether using generation tasks (training) can help with generalization on classification tasks (testing). See the README for more details.
- 04/24/2023: Upload few-shot training.
- 04/10/2023: Add code for Indent Indentification task. We use banking_data for experiments. See the README for more details.