LSTM-Attention Model for Language Use

The goals of this project is to identify discriminative patterns in surgeon's language use to locate the fractures of their communication in surgical telementoring.

Model Construction:

•Communicative event = speaker * dialogue act * communication channel

•Embedding: input dimension = 42, output dimension = 100

•LSTM with 100 units

•Extract attention values:

•Get the hidden representation from LSTM

•Transform the hidden representation at each time step into 1 value with dense layer

•Normalize the values in the sequence into probabilities that sum to 1 with Softmax

•Multiply and sum up the attention weights with the LSTM output

•Transform the output into either 0 or 1 with Sigmoid function

•Compile: optimizer: Adam; loss function: Binary Cross-Entropy