RNN Architecture

[sfvnielsen]

We will (probably) use an RNN to train the claim-detection system.

How should the architecture look like?

• Many-to-one...? (cf. Andrej Kaparthy blog-post )
• Types of units (GRU, LSTM), etc?
• Output layer
• Number of hidden units?