关于 CNN_BILSTM_CRF model的一些问题

[hit-joseph]

先膜拜大佬： 我想把这个模型用在一个中文的序列标注问题上：这里面有POS的标记：这个和CNN_character的特征冲突吗，你的项目里面是手动标记特征和CNN_character的特征可以共存吗？另外看了一下数据的预处理的格式：Friday [Cap]1 [POS]NNP O ，我只用到了POS的特征数据是不是应该写成Friday [POS]NNP O ，[POS]是必须要的吗，还是你只是作为一个标记？ 烦请指教

[jiesutd]

1. Character CNN/LSTM是可以和自定义feature一起用的，不冲突。 2.你是对的，如果不用某个feature，你的输入文件里面就不需要那一列。并不是所有的都需要的。

[hit-joseph]

one more question，in the file named demo.train.config ,there are some Hyperparameters: cnn_layer=4 char_hidden_dim=30 I can't really uderstand what those 2 parameters means, do you mind explain it? Thanks

[jiesutd]

You may refer the configuration explanation here: https://github.com/jiesutd/NCRFpp/blob/master/readme/Configuration.md

[hit-joseph]

thank you .this file I have read already , so I JUST want to make it sure, if I use lstm as word-sequence layer, cnn_layer is useless and can be annotated, and the parameter of char_hidden_dim=50 means the feature i extract from word and dim is 50 ,and than joint it after word2vec and pos_vec?

[jiesutd]

Do you mean "ignored" rather than "annotated"? Yes, if you choose the LSTM to encode the word sequence then the settings of CNN can be ignored.

About the char_hidden_dim=50: "the feature i extract from word and dim is 50", this is right. "joint it after word2vec and pos_vec", it is concatenated with word embeddings(not word2vec) and feature embeddings.

[hit-joseph]

if I SET char_emb_dim=100 (this parameter in I/O part) char_hidden_dim=50(this parameter in Hyperparameters part) is that means i input the pretrained char_embedding which in 100 dim ,and after cnn_char_layer I GET char_features_embdding in 50dim?and it can concatenate with word embedding?

[jiesutd]

exactly.

[hit-joseph]

thank you very much!