train on bert base

[L-hongbin]

Hello, I'd to know how about the result of this model training on Bert_base? I have trianed on bert base with c2f , python run.py train_bert_base_ml0_d2, but only get a result about 67 F1

[lxucs]

There might be sth off in your configuration. I have trained spanbert_base weight with F1 score 77+.

[sushantakpani]

Hi,
What is the difference train_bert_base_ml0_d1 and train_bert_base_ml0_d2?
Which configuration is for c2f 2019?

[lxucs]

c2f 2019 is train_bert_xxx_ml0_d2. d2 uses Attended Antecedent as higher-order inference while d1 only uses local decisions.

[sushantakpani]

c2f 2019 is train_bert_xxx_ml0_d2. d2 uses Attended Antecedent as higher-order inference while d1 only uses local decisions.

@lxucs Thank you.

[L-hongbin]

There might be sth off in your configuration. I have trained spanbert_base weight with F1 score 77+.

I have trained on bert_base with the following configuration，but only get the F1 score 73.2+, and the c2f-coref-bert_base model in the paper https://arxiv.org/pdf/1908.09091.pdf is about 73.9.

max_top_antecedents = 50
max_training_sentences = 11
top_span_ratio = 0.4
max_num_extracted_spans = 3900
max_num_speakers = 20
max_segment_len = 128
bert_learning_rate = 1e-05
task_learning_rate = 0.0002
loss_type = "marginalized"
mention_loss_coef = 0
false_new_delta = 1.5
adam_eps = 1e-06
adam_weight_decay = 0.01
warmup_ratio = 0.1
max_grad_norm = 1
gradient_accumulation_steps = 1
coref_depth = 2
higher_order = "attended_antecedent"
fine_grained = true
dropout_rate = 0.3
ffnn_size = 3000
ffnn_depth = 1
cluster_ffnn_size = 3900
cluster_reduce = "mean"
easy_cluster_first = false
cluster_dloss = false
num_epochs = 20
feature_emb_size = 20
max_span_width = 30
use_metadata = true
use_features = true
use_segment_distance = true
model_heads = true
use_width_prior = true
use_distance_prior = true
conll_eval_path = "./dev.english.v4_gold_conll"
conll_test_path = "./test.english.v4_gold_conll"
genres = [
"bc"
"bn"
"mz"
"nw"
"pt"
"tc"
"wb"
]
eval_frequency = 1000
report_frequency = 100
log_root = "./"
num_docs = 2802
bert_tokenizer_name = "bert-base-cased"
bert_pretrained_name_or_path = "bert-base-cased"

[lxucs]

@L-hongbin I was talking about spanbert_base, not bert_base

[L-hongbin]

@L-hongbin I was talking about spanbert_base, not bert_base

Thanks for your reply, so you don't have the results on bert_base?

[lxucs]

@L-hongbin The results are similar to the reported numbers but I don't have the exact numbers on my hand now.

[L-hongbin]

@L-hongbin The results are similar to the reported numbers but I don't have the exact numbers on my hand now.

Thanks~