Sequence labelling in structured domains with hierarchical recurrent neural networks

Modelling data in structured domains requires establishing the relations among patterns at multiple scales. When these patterns arise from sequential data, the multiscale structure also contains a dynamic component that must be modelled, particularly, as is often the case, if the data is unsegmented. Probabilistic graphical models are the predominant framework for labelling unsegmented sequential data in structured domains. Their use requires a certain degree of a priori knowledge about the relations among patterns and about the patterns themselves. This paper presents a hierarchical system, based on the connectionist temporal classification algorithm, for labelling unsegmented sequential data at multiple scales with recurrent neural networks only. Experiments on the recognition of sequences of spoken digits show that the system outperforms hidden Markov models, while making fewer assumptions about the domain.