BiLSTM-CRF Manipuri NER with Character-Level Word Representation

Named Entity Recognition and Classification (NER) serves as a foundation for many natural language processing tasks such as question answering, text summarization, news/document clustering and machine translation. Manipuri's early NER systems are based on machine learning approaches and employ handcrafted morphological features and domain-specific rules. The domain-specific rules for Manipuri NER are hard to extract as the language is highly agglutinative, inflectional and falls in the category of low resource language. In recent years, deep learning, empowered by continuous vector representation and semantic composition through non-linear processing, has been employed in the various NER task yielding state-of-the accuracy. In this paper, we propose a Manipuri NER model using Bidirectional Long Short Term Memory (BiLSTM) deep neural network in unison with an embedding technique. The embedding technique is a BiLSTM character-level word representation in conjunction with word embedding, which acts as a feature for the Bi-LSTM NER model. The proposed model also employs a Conditional Random Field (CRF) classifier to capture the dependency among output NER tags. Various Gradient Descent (GD) optimizers for the neural model were experimented with to establish an efficient GD optimizer for accurate NER. The NER model with RMSprop GD optimizer achieved an F-Score measure of approximately 98.19% at learning rate eta = 0.001 and with decay constant of rho = 0.9. Further, while performing an intrinsic evaluation on the word embedding, it is found that the proposed embedding technique as a feature can capture the semantic and syntactic rule of the language with 88.14% average clustering accuracy for all NE classes.