This would be quite a cool problem to attack using RNN. Perhaps using an attention-following mechanism like the one used for house numbers [1]. It's fairly well-defined. You would need a ton of example data though.

[1] http://arxiv.org/abs/1412.7755