Joint Extraction of Entities and Relations Using Reinforcement Learning and Deep Learning

Algorithm 1

Training procedure for -Learning.

Initialize BILSTM, the Attention Layer, and Tree-LSTM with random parameters

Pre-train BILSTM, the Attention Layer, and Tree-LSTM respectively

for epoch = do

for each input sentence do

Use the deep learning models above to automatically extract features of , and generate and .

for = 1, 2 do

= the reward and state after taking the action

Perform gradient descent step:

The update rule is

Where is update step, and is the reward function (Section 3.1), and is the state-action pair of next time.

,

end for

end for

end for