Extremely Low Resource Text simplification with Pre-trained Transformer Language Model

Recent text simplification approaches regard the task as a monolingual text-to-text generation inspired by machine translation. In particular, the transformer-based translation model outperform previous methods. Although machine translation approaches need a large-scale parallel corpus, parallel corpora for text simplification are very small compared to machine translation tasks. Therefore, we attempt a simple approach which fine-tunes the pre-trained language model for text simplification with a small parallel corpus. Specifically, we conduct experiments with the following two models: transformer-based encoder-decoder model and a language model that receives a joint input of original and simplified sentences, called TransformerLM. Thus, we show that TransformerLM, which is a simple text generation model, substantially outperforms a strong baseline. In addition, we show that fine-tuned TransformerLM with only 3,000 supervised examples can achieve performance comparable to a strong baseline trained by all supervised data.