Questions about Token-Token Matrix Decoding

[Jingkustc]

Hi there. Thanks for the awesome work.

I am curious about the implementation of Token-Token Matrix Decoding. But I find no implementation of the TT decoding in your code now. Could you kindly provide the code for implementing the TT decoding?

Furthermore, according to your paper, "If a token plays different roles in the same/different events, it is regarded a new token". However, when we are predicting on a new document, it seems hard to know if a token would play different roles. Could you please tell me how to achieve this in the evaluation scenario.

Thanks for your kindness and looking forward to your reply.

[ylxwan]

他的方法没有采用传统的LSTM+CRF的BIO标注法，而是直接分词。但是从代码中我如果没有读错的话，他在无论是在训练还是评估中，分词的时候都直接使用了真实值，这一点是我不能理解的。所以他在对对新文档进行预测时，直接通过真实值就知道了一个 token 是否会扮演不同的角色。