Named entity recognition: a semi-supervised learning approach

Abstract

Named entity recognition (NER) application development for under-resourced (i.e. NLP resource) language is usually obstructed by lack of named entity tagged dataset and this led to performance deterioration. Similarly, in Amharic language getting annotated training dataset for named entity recognition problem is extortionate, though an enormous amount of untagged data is easily accessible. Fortunately, the performance of NER possibly be boosted via encompassing a few labeled data with an oversized collection of unlabeled data. Based on this premise, this paper tend to investigate graph-based label propagation algorithm for the Amharic NER problem, a simple semi-supervised, iterative algorithm, to propagate labels through the dataset. In addition, it is aimed at making a rigorous comparison with expectation–maximization with semi-supervised learning approaches. The experiment reveals, label propagation based NER achieves superior performance compared to expected maximization using a few labeled training data. Since expectation maximization algorithm demands a moderate labeled example to be learned, meant very few labeled examples are not enough to generate adequate parameters for recognition of named entities, consequently it couldn’t perform great as the label propagation algorithm.