def __init__(self,
args,
embedding_matrix=None,
aspect_embedding_matrix=None):
super(TD_LSTM, self).__init__()
self.args = args
self.encoder = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.encoder_aspect = nn.Embedding.from_pretrained(
torch.tensor(aspect_embedding_matrix, dtype=torch.float))
self.lstm_l = DynamicRNN(args.embed_dim,
args.hidden_dim,
num_layers=1,
batch_first=True,
dropout=args.dropout,
rnn_type="LSTM")
self.lstm_r = DynamicRNN(args.embed_dim,
args.hidden_dim,
num_layers=1,
batch_first=True,
dropout=args.dropout,
rnn_type="LSTM")
self.dense = nn.Linear(args.hidden_dim * 2, args.polarities_dim)
self.dropout = nn.Dropout(args.dropout)
self.softmax = nn.Softmax()
def __init__(self,
args,
embedding_matrix=None,
aspect_embedding_matrix=None):
super(RAM, self).__init__()
self.args = args
self.encoder = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.encoder_aspect = nn.Embedding.from_pretrained(
torch.tensor(aspect_embedding_matrix, dtype=torch.float))
self.bi_lstm_context = DynamicRNN(args.embed_dim,
args.hidden_dim,
num_layers=1,
batch_first=True,
bidirectional=True,
dropout=args.dropout,
rnn_type="LSTM")
self.bi_lstm_aspect = DynamicRNN(args.embed_dim,
args.hidden_dim,
num_layers=1,
batch_first=True,
bidirectional=True,
dropout=args.dropout,
rnn_type="LSTM")
self.attention = Attention(args.hidden_dim * 2, score_function='mlp')
self.gru_cell = nn.GRUCell(args.hidden_dim * 2, args.hidden_dim * 2)
self.dense = nn.Linear(args.hidden_dim * 2, args.polarities_dim)
self.softmax = nn.Softmax()
self.dropout = nn.Dropout(args.dropout)
def __init__(self, args, embedding_matrix=None, aspect_embedding_matrix=None):
super(HAN, self).__init__()
self.position_dim = 100
self.args = args
self.encoder = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float))
self.encoder_aspect = nn.Embedding.from_pretrained(torch.tensor(aspect_embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding(batch_first=True)
self.word_rnn = DynamicRNN(args.embed_dim + self.position_dim, args.hidden_dim, num_layers=1, batch_first=True,
dropout=args.dropout, bidirectional=True, rnn_type="LSTM")
self.sentence_rnn = DynamicRNN(args.hidden_dim * 2 + self.position_dim, args.hidden_dim, num_layers=1,
batch_first=True,
dropout=args.dropout, bidirectional=True, rnn_type="LSTM")
self.word_W = nn.Parameter(
torch.Tensor(args.hidden_dim * 2 + self.position_dim, args.hidden_dim * 2 + self.position_dim))
self.word_bias = nn.Parameter(torch.Tensor(args.hidden_dim * 2 + self.position_dim, 1))
self.word_weight_proj = nn.Parameter(torch.Tensor(args.hidden_dim * 2 + self.position_dim, 1))
self.word_attention = BasicAttention(hidden_dim=args.hidden_dim * 2, score_function="basic")
self.sentence_W = nn.Parameter(
torch.Tensor(args.hidden_dim * 2 + self.position_dim, args.hidden_dim * 2 + self.position_dim))
self.sentence_bias = nn.Parameter(torch.Tensor(args.hidden_dim * 2 + self.position_dim, 1))
self.sentence_weight_proj = nn.Parameter(torch.Tensor(args.hidden_dim * 2 + self.position_dim, 1))
self.sentence_attention = BasicAttention(hidden_dim=args.hidden_dim * 2, score_function="basic")
self.dense = nn.Linear(args.hidden_dim * 4, args.polarities_dim)
self.word_position_embed = nn.Embedding(1005, self.position_dim)
self.segment_position_embed = nn.Embedding(25, self.position_dim)
self.softmax = nn.Softmax()
self.dropout = nn.Dropout(args.dropout)
def __init__(self,
args,
embedding_matrix=None,
aspect_embedding_matrix=None):
super(ATAE_BiGRU, self).__init__()
self.args = args
self.encoder = nn.Embedding.from_pretrained(
torch.tensor(embedding_matrix, dtype=torch.float))
self.encoder_aspect = nn.Embedding.from_pretrained(
torch.tensor(aspect_embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding(batch_first=True)
self.gru = DynamicRNN(args.embed_dim,
args.hidden_dim,
num_layers=1,
batch_first=True,
dropout=args.dropout,
bidirectional=True,
rnn_type="GRU")
# self.attention = BasicAttention(hidden_dim=args.hidden_dim*2, score_function="aspect")
self.attention = AttentionAspect(embed_dim_k=args.hidden_dim * 2,
embed_dim_q=args.embed_dim,
score_function="mlp")
self.dense = nn.Linear(args.hidden_dim * 2, args.polarities_dim)
self.softmax = nn.Softmax()
self.dropout = nn.Dropout(args.dropout)
def __init__(self, args, embedding_matrix=None, aspect_embedding_matrix=None):
super(AT_GRU, self).__init__()
self.args = args
self.encoder = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float))
self.encoder_aspect = nn.Embedding.from_pretrained(torch.tensor(aspect_embedding_matrix, dtype=torch.float))
self.gru = DynamicRNN(args.embed_dim, args.hidden_dim, num_layers=1, dropout=args.dropout, rnn_type="GRU")
self.attention = BasicAttention(hidden_dim=args.hidden_dim)
self.dense = nn.Linear(args.hidden_dim, args.polarities_dim)
self.softmax = nn.Softmax()
self.dropout = nn.Dropout(args.dropout)
def __init__(self, args, embedding_matrix=None, aspect_embedding_matrix=None):
super(CABASC, self).__init__()
self.args = args
self.encoder = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float))
self.encoder_aspect = nn.Embedding.from_pretrained(torch.tensor(aspect_embedding_matrix, dtype=torch.float))
self.squeeze_embedding = SqueezeEmbedding(batch_first=True)
self.attention = Attention(args.embed_dim, score_function='mlp') # content attention
self.m_linear = nn.Linear(args.embed_dim, args.embed_dim, bias=False)
self.mlp = nn.Linear(args.embed_dim, args.embed_dim) # W4
self.dense = nn.Linear(args.embed_dim, args.polarities_dim) # W5
# context attention layer
self.rnn_l = DynamicRNN(args.embed_dim, args.hidden_dim, num_layers=1, batch_first=True, dropout=args.dropout,
rnn_type='GRU')
self.rnn_r = DynamicRNN(args.embed_dim, args.hidden_dim, num_layers=1, batch_first=True, dropout=args.dropout,
rnn_type='GRU')
self.mlp_l = nn.Linear(args.hidden_dim, 1)
self.mlp_r = nn.Linear(args.hidden_dim, 1)
self.dropout = nn.Dropout(args.dropout)
self.softmax = nn.Softmax()
def __init__(self, args, embedding_matrix, aspect_embedding_matrix=None, memory_weighter='no'):
super(LCRS, self).__init__()
self.args = args
self.memory_weighter = memory_weighter
self.encoder = nn.Embedding.from_pretrained(torch.tensor(embedding_matrix, dtype=torch.float))
self.encoder_aspect = nn.Embedding.from_pretrained(torch.tensor(aspect_embedding_matrix, dtype=torch.float))
self.blstm_l = DynamicRNN(args.embed_dim, args.hidden_dim, num_layers=1, batch_first=True, dropout=args.dropout,
rnn_type='LSTM')
self.blstm_c = DynamicRNN(args.embed_dim, args.hidden_dim, num_layers=1, batch_first=True, dropout=args.dropout,
rnn_type='LSTM')
self.blstm_r = DynamicRNN(args.embed_dim, args.hidden_dim, num_layers=1, batch_first=True, dropout=args.dropout,
rnn_type='LSTM')
self.dense = nn.Linear(args.hidden_dim * 4, args.polarities_dim)
# target to context attention
self.t2c_l_attention = Attention(args.hidden_dim, score_function='bi_linear')
self.t2c_r_attention = Attention(args.hidden_dim, score_function='bi_linear')
# context to target attention
self.c2t_l_attention = Attention(args.hidden_dim, score_function='bi_linear')
self.c2t_r_attention = Attention(args.hidden_dim, score_function='bi_linear')
self.softmax = nn.Softmax()
self.dropout = nn.Dropout(args.dropout)