尝试提高bi-lstm+crf准确率的几个尝试

一.将lstm改为gru。

cell_forward = tf.contrib.rnn.BasicLSTMCell(unit_num)
cell_backward = tf.contrib.rnn.BasicLSTMCell(unit_num)

改为

cell_forward = tf.nn.rnn_cell.GRUCell(unit_num)
cell_backward = tf.nn.rnn_cell.GRUCell(unit_num)

二.多层RNN结构

_fw_cells = []
_bw_cells = []
for _ in range(2):
    cell_forward = tf.contrib.rnn.BasicLSTMCell(unit_num)
    cell_backward = tf.contrib.rnn.BasicLSTMCell(unit_num)
    _fw_cells.append(tf.nn.rnn_cell.DropoutWrapper(cell_forward, output_keep_prob=0.6))
    _bw_cells.append(tf.nn.rnn_cell.DropoutWrapper(cell_backward, output_keep_prob=0.6))
cell_forward = tf.nn.rnn_cell.MultiRNNCell(_fw_cells)
cell_backward = tf.nn.rnn_cell.MultiRNNCell(_bw_cells)

三.损失函数添加L2正则项

self.loss = tf.reduce_mean(-self.log_likelihood)
#L2正则
self.l2_loss = self._l2_rate * (tf.nn.l2_loss(W) + tf.nn.l2_loss(b))
self.loss = self.loss + self.l