TobiasLee / MTA-LSTM-TensorFlow

TensorFlow reimplementation of Topic-to-Essay Generation with Neural Networks.
https://tobiaslee.top/2018/11/02/customized-RNN-cell/
70 stars 24 forks source link

关于生成的UNk #8

Closed nlpofwhat closed 4 years ago

nlpofwhat commented 4 years ago

请问一下你生成的文本为什么没有UNK,我生成的文本又一堆的UNK,END,PAD字符

TobiasLee commented 4 years ago

README 中的例子是去掉特殊符号的

nlpofwhat commented 4 years ago

那你的效果达到了论文的效果吗,感觉原论文给出的测试太少了,我做的实验发现很多句子只是语法还可以,但是语义上不通顺,指代不明一而且一堆的UNK,END,PAD。可能与原文唯一的区别就是他用了大量的预料进行词向量的预训练,但是我不认为这个对模型有很大的改善

------------------ 原始邮件 ------------------ 发件人: "Tobias Lee"notifications@github.com; 发送时间: 2019年8月14日(星期三) 晚上6:40 收件人: "TobiasLee/MTA-LSTM-TensorFlow"MTA-LSTM-TensorFlow@noreply.github.com; 抄送: "@My heart will go on."952167295@qq.com; "Author"author@noreply.github.com; 主题: Re: [TobiasLee/MTA-LSTM-TensorFlow] 关于生成的UNk (#8)

README 中的例子是去掉特殊符号的

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

TobiasLee commented 4 years ago

我没有做评测,肉眼观察还是质量可以的;pre-train word embedding 对于生成的效果还是有提升的

nlpofwhat commented 4 years ago

请问您最后的损失函数降到多少了,我的损失函数从400多到200后面就很难降下去了

nbdwddbf commented 4 years ago

请问您Data文件夹里面的数据有嘛,我是一个新手,求分享。

nlpofwhat commented 4 years ago

原论文有下载连接的

nbdwddbf commented 4 years ago

原论文有下载连接的

论文里面的下载链接就是“https://github.com/hit-computer/MTA-LSTM“ 这个仓库啊,没有那个vec.text,求分享vec.text,以及文章的TrainingData用的是那个composition.txt还是zhihu.txt啊,求教!!!

nlpofwhat commented 4 years ago

下面有数据的下载链接,vec.txt是可以自己训练的词向量

------------------ 原始邮件 ------------------ 发件人: "nbdwddbf"notifications@github.com; 发送时间: 2019年8月23日(星期五) 下午2:06 收件人: "TobiasLee/MTA-LSTM-TensorFlow"MTA-LSTM-TensorFlow@noreply.github.com; 抄送: "@My heart will go on."952167295@qq.com; "Author"author@noreply.github.com; 主题: Re: [TobiasLee/MTA-LSTM-TensorFlow] 关于生成的UNk (#8)

原论文有下载连接的

论文里面的下载链接就是“https://github.com/hit-computer/MTA-LSTM“这个仓库啊,没有那个vec.text,求分享vec.text,以及文章的TrainingData用的是那个composition.txt还是zhihu.txt啊,求教!!!

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

nbdwddbf commented 4 years ago

请问您这个项目中,vec是通过哪个txt来训练的啊

nbdwddbf commented 4 years ago

可以分享一下您的data这一部分数据呢?谢谢啦

nbdwddbf commented 4 years ago

如您能分享,非常的感谢!!!(附邮箱:2927572356@qq.com)

nlpofwhat commented 4 years ago

这是下载链接,我用的是这个https://pan.baidu.com/s/1_JPh5-g2rry2QmbjQ3pZ6w

------------------ 原始邮件 ------------------ 发件人: "nbdwddbf"notifications@github.com; 发送时间: 2019年8月23日(星期五) 下午2:14 收件人: "TobiasLee/MTA-LSTM-TensorFlow"MTA-LSTM-TensorFlow@noreply.github.com; 抄送: "@My heart will go on."952167295@qq.com; "Author"author@noreply.github.com; 主题: Re: [TobiasLee/MTA-LSTM-TensorFlow] 关于生成的UNk (#8)

可以分享一下您的data这一部分数据呢?谢谢啦

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

nbdwddbf commented 4 years ago

可以加您qq嘛?我有不懂的想咨询您

nlpofwhat commented 4 years ago

欢迎讨论

------------------ 原始邮件 ------------------ 发件人: "nbdwddbf"notifications@github.com; 发送时间: 2019年8月23日(星期五) 下午2:39 收件人: "TobiasLee/MTA-LSTM-TensorFlow"MTA-LSTM-TensorFlow@noreply.github.com; 抄送: "@My heart will go on."952167295@qq.com; "Author"author@noreply.github.com; 主题: Re: [TobiasLee/MTA-LSTM-TensorFlow] 关于生成的UNk (#8)

可以加您qq嘛?我有不懂的想咨询您

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or mute the thread.

nbdwddbf commented 4 years ago

已加,求通过