We want to design a neural network that segments an English word into prefixes, root, and suffixes using a BIO labelling scheme.
For example, the word “unprepossessing” has the labelling: (“u”, B-pre), (“n”, I-pre), (“p”, B-pre), (“r”, I-pre), (“e”, I-pre), (“p”, B-root), (“o”, Iroot), (“s”, I-root), (“s”, I-root), (“e”, I-root), (“s”, I-root), (“s”, I-root), (“i”, B-suf), (“n”, I-suf), (“g”, I-suf). Note that due to the nature of the application, O will not be used.
Fully specify a neural network to solve this problem. Describe:
• how the inputs and outputs are encoded
• the structure of the network
• the loss function used
Describe the network in enough detail that one could implement it using PyTorch. You may describe it in terms of common abstractions (e.g. “use a standard LSTM cell of such-and-such size”) if that’s useful.
