您好，数据集缺少Phonetic和Visual，请问可以补充一下吗

[wingwwww]

下载的数据集里，只有pinyin、abb、char、hybrid四种，缺少Phonetic和Visual对应的数据集，请问可以麻烦您提供一下吗。
而且在使用提供的pretrain模型时，对pinyin、abb、char这三种的识别较好，但对谐音字的识别效果较差，例如在输入论文中提到的征服的防疫措施让人失望这句话进去，无法正确输出政府的防疫策略让人失望，请问是否需要在pre-trained modle上进一步微调？或者能否提供以及训练好的识别五种变体的模型

[fenffef]

1. Phonetic and Visual data are included within the hybrid dataset. If you wish to generate separate datasets for them, please refer to the script here: https://github.com/sww9370/RoCBert/blob/main/attacker/building.py
2. Yes, you need to further fine-tune the pre-trained model to achieve better performance on specific tasks (such as correcting homophone errors).

[wingwwww]

thank you very much

[fenffef]

😄 If you find our work helpful, please consider citing our paper:
Feng X, Gu T, Chang L, et al. PROTECT: Parameter-Efficient Tuning for Few-Shot Robust Chinese Text Correction. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2024.