In retrospect, I believe that this early draft is a much better document than the published papers discussing the powerful Graph Transformer Networks.

Abstract: We present here a general architecture for building Automatic Document Analysis Systems. This architecture is composed of a succession of modules transforming graphs describing lower-level hypotheses on the documents into graphs describing higher level hypotheses. This architecture generalizes techniques used in Neural Networks, Optical Character Recognition, Natural Language Processing and Speech Recognition.

@misc{bottou-1996,
  author = {Bottou, L\'{e}on and Bengio, Yoshua and {LeCun}, Yann},
  title = {Document Analysis with Transducers},
  year = {1996},
  month = {July},
  optnote = {Tech report available on http://leon.bottou.com/publications},
  url = {http://leon.bottou.org/papers/bottou-1996},
}