Current technology could already treat the handwriting-script recognition problem very well in certain domain. On other domains, such as recognizing damaged, faded, or stained scripts, the recognition rate are still not perfect. This paper intends to compare two language models (n-gram and LSTM network) that could potentially be used in enhancing the recognition rate for such domain by making prediction to the missing word. We found that with limited context given, the traditional n-gram model with Katz smoothing performs around 10 times better than the LSTM network. As a conclusion when the context given is limited, one should avoid using LSTM network on word prediction, and only use it when enough context is given. In the case of enhancing handwriting script recognition, this means to avoid using LSTM network when too many words are damaged or missing from the script.

Recommended Citation: Xiao Y., Guinn C., Kline D., Ricanek K., (2017). Enhance Handwriting Script Recognition with Language Modeling. UNCW MS CSIS Proceedings. V. 11 , N. 1 .