def __init__(self,
n_words,
n_feats,
n_labels,
feat='char',
n_embed=100,
n_feat_embed=100,
n_char_embed=50,
bert=None,
n_bert_layers=4,
mix_dropout=.0,
embed_dropout=.33,
n_lstm_hidden=400,
n_lstm_layers=3,
lstm_dropout=.33,
n_mlp_span=500,
n_mlp_label=100,
mlp_dropout=.33,
feat_pad_index=0,
pad_index=0,
unk_index=1,
**kwargs):
super().__init__()
self.args = Config().update(locals())
# the embedding layer
self.word_embed = nn.Embedding(num_embeddings=n_words,
embedding_dim=n_embed)
if feat == 'char':
self.feat_embed = CharLSTM(n_chars=n_feats,
n_embed=n_char_embed,
n_out=n_feat_embed,
pad_index=feat_pad_index)
elif feat == 'bert':
self.feat_embed = BertEmbedding(model=bert,
n_layers=n_bert_layers,
n_out=n_feat_embed,
pad_index=feat_pad_index,
dropout=mix_dropout)
self.n_feat_embed = self.feat_embed.n_out
elif feat == 'tag':
self.feat_embed = nn.Embedding(num_embeddings=n_feats,
embedding_dim=n_feat_embed)
else:
raise RuntimeError("The feat type should be in ['char', 'bert', 'tag'].")
self.embed_dropout = IndependentDropout(p=embed_dropout)
# the lstm layer
self.lstm = LSTM(input_size=n_embed+n_feat_embed,
hidden_size=n_lstm_hidden,
num_layers=n_lstm_layers,
bidirectional=True,
dropout=lstm_dropout)
self.lstm_dropout = SharedDropout(p=lstm_dropout)
# the MLP layers
self.mlp_span_l = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_span, dropout=mlp_dropout)
self.mlp_span_r = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_span, dropout=mlp_dropout)
self.mlp_label_l = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_label, dropout=mlp_dropout)
self.mlp_label_r = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_label, dropout=mlp_dropout)
# the Biaffine layers
self.span_attn = Biaffine(n_in=n_mlp_span, bias_x=True, bias_y=False)
self.label_attn = Biaffine(n_in=n_mlp_label, n_out=n_labels, bias_x=True, bias_y=True)
self.crf = CRFConstituency()
self.criterion = nn.CrossEntropyLoss()
self.pad_index = pad_index
self.unk_index = unk_index
def __init__(self,
n_words,
n_labels,
n_tags=None,
n_chars=None,
n_lemmas=None,
feat='tag,char,lemma',
n_embed=100,
n_embed_proj=125,
n_feat_embed=100,
n_char_embed=50,
char_pad_index=0,
bert=None,
n_bert_layers=4,
mix_dropout=.0,
bert_pad_index=0,
embed_dropout=.2,
n_lstm_hidden=600,
n_lstm_layers=3,
lstm_dropout=.33,
n_mlp_edge=600,
n_mlp_label=600,
edge_mlp_dropout=.25,
label_mlp_dropout=.33,
interpolation=0.1,
pad_index=0,
unk_index=1,
**kwargs):
super().__init__()
self.args = Config().update(locals())
# the embedding layer
self.word_embed = nn.Embedding(num_embeddings=n_words,
embedding_dim=n_embed)
self.embed_proj = nn.Linear(n_embed, n_embed_proj)
self.n_input = n_embed + n_embed_proj
if 'tag' in feat:
self.tag_embed = nn.Embedding(num_embeddings=n_tags,
embedding_dim=n_feat_embed)
self.n_input += n_feat_embed
if 'char' in feat:
self.char_embed = CharLSTM(n_chars=n_chars,
n_embed=n_char_embed,
n_out=n_feat_embed,
pad_index=char_pad_index)
self.n_input += n_feat_embed
if 'lemma' in feat:
self.lemma_embed = nn.Embedding(num_embeddings=n_lemmas,
embedding_dim=n_feat_embed)
self.n_input += n_feat_embed
if 'bert' in feat:
self.bert_embed = BertEmbedding(model=bert,
n_layers=n_bert_layers,
pad_index=bert_pad_index,
dropout=mix_dropout)
self.n_input += self.bert_embed.n_out
self.embed_dropout = IndependentDropout(p=embed_dropout)
# the lstm layer
self.lstm = LSTM(input_size=self.n_input,
hidden_size=n_lstm_hidden,
num_layers=n_lstm_layers,
bidirectional=True,
dropout=lstm_dropout)
self.lstm_dropout = SharedDropout(p=lstm_dropout)
# the MLP layers
self.mlp_edge_d = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_edge, dropout=edge_mlp_dropout, activation=False)
self.mlp_edge_h = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_edge, dropout=edge_mlp_dropout, activation=False)
self.mlp_label_d = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_label, dropout=label_mlp_dropout, activation=False)
self.mlp_label_h = MLP(n_in=n_lstm_hidden*2, n_out=n_mlp_label, dropout=label_mlp_dropout, activation=False)
# the Biaffine layers
self.edge_attn = Biaffine(n_in=n_mlp_edge, n_out=2, bias_x=True, bias_y=True)
self.label_attn = Biaffine(n_in=n_mlp_label, n_out=n_labels, bias_x=True, bias_y=True)
self.criterion = nn.CrossEntropyLoss()
self.interpolation = interpolation
self.pad_index = pad_index
self.unk_index = unk_index
