waltbai/mcpredictor

Integrating Deep Event-Level and Script-Level Information for Script Event Prediction

• Hi, I'm Long Bai.
• I'm a research assistant at Institue of Computing Technology, Chinese Academy of Sciences.
• I'm interested in knowledge graph and natural language processing.
• Email: [email protected]

mcpredictor's Issues

数据集与评测问题

您好，很抱歉再次打扰您，然而论文中还有一些模糊的地方想和您确认一下。我重新生成了数据集并成功复现了论文中结果，该数据集的训练集大小为1440285，和论文中的一致，

但是table 2中的复现结果是使用这个较大的数据集复现出来的吗，还是用的其他数据集？

events: size(batch_size, choice_num, seq_len + 1, 5)

network.py----> forward 中param events: size(batch_size, choice_num, seq_len + 1, 5)，这里5代表什么意思，choice_num又是什么意思？

MCNC dataset Construction Problems

Hi, I am following your great work, and MCNC codebase has too many bugs today, after following your modification, I have tried to fix a lot, but there are still too much, could you share your modified eventchains-aaai folder, I think this may help a lot. I would appreciate it if you could help me !
And what is the purpose of bin/event_pipeline/rich_docs/gigaword.txt, if I need origin text then I need to install it?

SCPredictor-s模型复现问题

1.https://mark.granroth-wilding.co.uk/papers/what_happens_next/
2.https://github.com/waltbai/MCPredictor/blob/main/experiments/preprocess.py
3.https://github.com/waltbai/MCPredictor/blob/main/config/scpredictor.json
你好，我在复现论文SCPredictor-s模型时遇到了一些问题，如果方便的话希望能够得到您的帮助。我使用what_happens_next中的代码处理语料库得到了每个文档对应的实体以及事件，然后使用preprocess.py处理得到了包括single_train,single_dev,single_test在内的数据集，以及pretrain_embedding.npy和word_dict.json，然后使用scpredictor.json中的配置训练模型，然而准确率只有55.82，远低于论文中的结果58.28，之后我将batchsize从1000改为了200，准确率提高到了57.72。想请问config文件中的配置是复现论文所需要的配置吗，或者我有什么遗漏导致无法直接复现结果吗？
感谢您的回复！