Handwritten word recognition using Web resources and recurrent neural networks

Handwriting recognition systems usually rely on static dictionaries and language models. Full coverage of these dictionaries is generally not achieved when dealing with unrestricted document corpora due to the presence of Out-Of-Vocabulary (OOV) words. We propose an approach which uses the World Wide Web as a corpus to improve dictionary coverage. We exploit the very large and freely available Wikipedia corpus in order to obtain dynamic dictionaries on the fly. We rely on recurrent neural network (RNN) recognizers, with and without linguistic resources, to detect words that are non-reliably recognized within a word sequence. Such words are labeled as non-anchor words (NAWs) and include OOVs and In-Vocabulary words recognized with low confidence. To recognize a non-anchor word, a dynamic dictionary is built by selecting words from the Web resource based on their string similarity with the NAW image, and their linguistic relevance in the NAW context. Similarity is evaluated by computing the edit distance between the sequence of characters generated by the RNN recognizer exploited as a filler model, and the Wikipedia words. Linguistic relevance is based on an N-gram language model estimated from the Wikipedia corpus. Experiments conducted on a word-segmented version of the publicly available RIMES database show that the proposed approach can improve recognition accuracy compared to systems based on static dictionaries only. The proposed approach shows even better behavior as the proportion of OOVs increases, in terms of both accuracy and dictionary coverage.