Switch from BiLSTM to the modern attention architecture

[vmarkovtsev]

Our current NN splitter is based on BiLSTM, which has problems with performance. We should leverage the recent advancements in deep learning and implement the new attention-based (seq2seq-like?) architecture of the model.

Stage 1 - research

Follow the paper, take the same dataset, and design the model. Calculate the metrics.

Stage 2 - production

Package the model, publish it on Modelforge.

[vmarkovtsev]

Assigning to you @zurk because you worked for solutions and missed interesting tasks.

[Guillemdb]

@vmarkovtsev I think it's time to close this issue 😉, for some reason I cannot do it myself.

[vmarkovtsev]

I'd rather leave these to indicate what was lacking in the project when we stopped. Thanks for pinging anyway!