Named Entity Recognition by Neural Prediction

Named entity recognition (NER) remains a very challenging problem essentially when the document, where we perform it, is handwritten and ancient. Traditional methods using regular expressions or those based on syntactic rules, work but are not generic because they require, for each dataset, additional work of adaptation. We propose here a recognition method by context exploitation and tag prediction. We use a pipeline model composed of two consecutive BLSTMs (Bidirectional Long-Short Term Memory). The first one is a BLSTM-CTC coupling to recognize the words in a text line using a sliding window and HOG features. The second BLSTM serves as a language model. It cleverly exploits the gates of the BLSTM memory cell by deploying some syntactic rules in order to store the content around the proper nouns. This operation allows it to predict the tag of the next word, depending on its context, which is followed gradually until the discovery of the named entity (NE). All the words of the context are used to help the prediction. We have tested this system on a private dataset of Philharmonie de Paris, for the extraction of proper nouns within sale music transactions as well as on the public IAM dataset. The results are satisfactory, compared to what exists in the literature.