This project, I improve model which was trained, so you can download it, and use it to prediction!
- this project just support every sentences with 45 char length
- download model file, pwd: vv1k
- just use like this
-
first
bs = BertSim(gpu_no=0, log_dir='log/', bert_sim_dir='bert_sim_model\\', verbose=True)
-
second
similarity sentences
text_a = '技术侦查措施只能在立案后采取' text_b = '未立案不可以进行技术侦查' bs.predict([[text_a, text_b]])
you will get result like this: [[0.00942544 0.99057454]]
not similarity sentence
text_a = '华为还准备起诉美国政府' text_b = '飞机出现后货舱火警信息' bs.predict([[text_a, text_b]])
you will get result like this: [[0.98687243 0.01312758]]
-
name | type | detail |
---|---|---|
gpu_no | int | which gpu will be use to init bert ner graph |
log_dir | str | log dir |
verbose | bool | whether show tensorflow log |
bert_sim_model | str | bert sim model path |
In this project, I just use bert pre model to fine tuning, so I just use their original code. I try to create new one, but the new one just same as the original code, so I given up.
Because of my domain work, my work is based on judicial examination education, so I didn't use common dataset, my dataset were labeled by manual work, it include 80000+, 50000+ are similar, 30000+ are dissimilar, this is my dataset link pwd ptje
In original code, they just got the model pool output, I think there may be other ways to increase the accuracy, I tried some ways to increase the accuracy, but I found one, just concat the [CLS] embedding of the fourth from bottom to tailender in encoder output list, if you want to use my way, just do like this。
- Delete the following code
output_layer = model.get_pooled_output()
- Use the following code, it can increase the accuracy 1%.
output_layer = tf.concat([tf.squeeze(model.all_encoder_layers[i][:, 0:1, :], axis=1) for i in range(-4, 0, 1)], axis=-1)