An approach for output decoding of neural networks
Tobias StraußRoger, Labahn Gundram, Leifert Welf Wustlich · 2013
We review two decoding methods for recurrent neural networks ‐ based on the well known weighted Levenshtein distance and the connectionist temporal classification (CTC) ‐ and we introduce a new method ‐ the dynamic Levenstein distance (DynWL). We compare these three methods analytically in terms of time complexity and error performance. Although the approaches are dierent, there are deep connections between these ways of decoding. Finally, we test on the Arabic and French ICDAR data sets. Our experiments show that CTC yields the smallest error rates. Nevertheless, there are scenarios where DynWL is a good choice between performance and time complexity.