forward计算中只用了cls，所以bert后的cnn对齐还有必要吗？另外请教是否尝试过融入词性特征呢？

[jkkl]

first_token_tensor = sequence_output[:, 0]

[cooelf]

The alignment would be useful for improving the local interactions between the adjacent tokens. I simply concatenated them with the token embeddings, but did not see obvious improvements.

[jkkl]

@cooelf Thanks for your Reply。In my scenario，compared to bert-base-chinese, it has improved by abount 5 points.