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~