adaxry/gcdt

Code for the paper: GCDT: A Global Context Enhanced Deep Transition Architecture for Sequence Labeling

gcdt's Issues

Let's beat the state of the art on NER once again!

Hi Adaxry!

You have used BERT large which at the time was the state of the art pre trained transformer.
But nowadays is had been succeeded by XLnet which is a major improvement.
Very often switching from BERT large to XLnet large give multiple percents accuracy gains "for free"!

https://paperswithcode.com/paper/xlnet-generalized-autoregressive-pretraining
You could either use the official implementation (not actively maintained) or the one maintained by tensorflow.
But I believe the most actively maintained is
https://github.com/huggingface/transformers which allow you through the same API to switch to another transformer easily.

I and Humanity really needs more accurate named entity recognition! :m

您好 想问一下在使用bert embedding时候，维度怎么和其他词向量拼接的呀

比如bert embedding 是768 维

glove 是300维

然后想问一下你是怎么对齐的呀

谢谢~

您好。请问Conll03的testb测试结果里，某些句子会出现Predict标签数和Gold标签数不一致，会是什么原因呢？

我试用了GCDT代码，训练了5个模型，有3个模型会出现预测标签缺失的情况(bad data生成，样例如下)。
打印了预测值的输出，发现模型的预测值提前生成了标签，导致预测值的输出被截断。
这是什么原因呢？可以修复吗？

---

原文： ['Pace', 'outdistanced', 'three', 'senior', 'finalists', '--', 'Virginia', 'Tech', 'defensive', 'end', 'Cornell', 'Brown', ',', 'Arizona', 'State', 'offensive', 'tackle', 'Juan', 'Roque', 'and', 'defensive', 'end', 'Jared', 'Tomich', 'of', 'Nebraska', '.']

Gold： ['S-PER', 'O', 'O', 'O', 'O', 'O', 'B-ORG', 'E-ORG', 'O', 'O', 'B-PER', 'E-PER', 'O', 'B-ORG', 'E-ORG', 'O', 'O', 'B-PER', 'E-PER', 'O', 'O', 'O', 'B-PER', 'E-PER', 'O', 'S-ORG', 'O']
Gold token数：27

Predicted：['S-PER', 'O', 'O', 'O', 'O', 'O', 'B-ORG', 'E-ORG', 'O']
Predicted token数：9

How to feed data to the model in placeholder mode instead?

Why do you choose this kind of GRU variants, instead of LSTM?

Hi~

I tried a similar architecture, with all your GRU variants replaced with standard LSTMs/BiLSTM, and it reached 91.96 (compared with your model in Figure 1, w/o BERT) on CoNLL 2003 NER task.

Yes, I know you mentioned this kind of thing in Section 1,

While in BiLSTMs, even stacked BiLSTMs, the transition depth between consecutive hidden states are inherently shallow

But I don't know what do you mean by transition depth and how do you measure it? Could you give me more details about this? And have you tried LSTM before? If yes, could you please show me your experimental results? If no, could you explain why you need to design this kind of GRU variants?

Thanks~