Jia-Chen Gu

@andrew-begain 1.正例是正常连续的对话，将最后一句当作正例回复，之前的当作对话历史；负例是从整个语料库中随机采样。具体可参考https://arxiv.org/pdf/1612.01627.pdf 2. 我们没有重新进行负采样，正负样例follow了原始数据集，只不过加了id信息。NA表示没有正确回复，这是Douban测试集的特性 3. 我们仅仅是对原始的数据集进行了加id操作，没有筛选，你可以联系原作者 https://github.com/MarkWuNLP/MultiTurnResponseSelection

@andrew-begain 关于如何制作一个好的数据集，你可以参考现有常用的数据集的论文，例如Ubuntu (https://aclanthology.org/W15-4640.pdf), Douban(https://arxiv.org/pdf/1612.01627.pdf), E-commerce(https://aclanthology.org/C18-1317.pdf) 也可参考我们维护的一个leaderboard(https://github.com/JasonForJoy/Leaderboards-for-Multi-Turn-Response-Selection) 如何召回一些好的回复，可以用一些轻量级的IR算法，例如TF-IDF和BM25，一些计算量比较大但更准确的，例如DPR

@hozayfarifai Please refer to the section https://github.com/JasonForJoy/SA-BERT#dependencies

@KJGithub2021 All experiments were running on a single NVIDIA GeForce 1080 (12G) GPU card. The default training parameters is 10 epochs and 96 batch_size with evaluation every 1000 steps. You...

@KJGithub2021 It took about 90h (including evaluation on the dev set every 1000 steps) under the default setting, i.e., 10 epochs and 96 batch_size on a single NVIDIA GeForce 1080...

@KJGithub2021 About 50h on the Ubuntu V2 dataset. You might try: 1) Enlarge batch_size with more advanced GPU cards 2) Evaluate with less frequent steps, e.g. evaluate every 2k steps

@KJGithub2021 Sorry, we do not have any experience of resuming model training from a saved checkpoint using google colab. No suggestion can be provided.

Hi @jasonwu0731, Thanks for your interest in our work. Sorry about that we do not have a plan to release a pytorch version of code or checkpoint recently. You could...