[3C3-GS-6] Language media processing: generation

[3C3-GS-6-03] Selective training with generated samples from generative language models for generalized zero-shot text classification

〇Sohta Kannan¹, Taro Yano², Kunihiro Takeoka², Masafumi Oyamada², Takeshi Okadome¹ (1. Kwansei Gakuin University, 2. NEC Corporation)

Keywords:text classification, zero-shot classification, data augmentation, text generation

Generalized zero-shot text classification (GZSTC) is a task to classify text into a set of classes including unseen classes, which has no teacher data.
GZSTC is widely applied to such as news and product classification.
One of the existing approaches to zero-shot text classification is to use a language model to generate pseudo samples of the unseen class and use them as teacher data.
However, these existing approaches use language models that have been pre-trained by data from a wide range of domains, and thus cannot generate only samples that correspond to the target domain, which adversely affects the training of the classifier.
In this paper, we propose a GZSTC method that improves classification performance.
This method removes the out-of-domain samples generated by the language model and creates a data set with only the samples corresponding to the target domain.
Experiments on real data show the improvement of the classification performance of the proposed method against the baseline.