ctb_bert_graph.yaml

# @package _global_
d_model: 2048

encoder: bert-base-chinese
use_tags: true
use_words: false
d_kqv: 64
d_ff: 2048
word_emb_dropout: 0
tag_emb_dropout: 0.2
relu_dropout: 0.4
residual_dropout: 0.2
attention_dropout: 0
num_attn_layers: 4
num_attn_heads: 8

decoder: graph
num_gcn_layers: 5

learning_rate: 4e-5
weight_decay: 1e-6
subbatch_max_tokens: 1000