A Partial String Matching Approach for Named Entity Recognition in Unstructured Bengali Data

In today's data driven, automated and digitized world, a significant stage of information extraction is to look for special keywords, more formally known as 'Named Entity'. This has been an active research topic for more than two decades and significant progresses have been made. Today we have models powered by deep learning that, although not perfect, have near human level accuracy on certain occasions. Unfortunately these algorithms require a lot of annotated training data, which we hardly have for Bengali language. This paper proposes a partial string matching approach to identify a named entity from an unstructured text corpus in Bengali. The algorithm is a partial string matching technique, based on Breadth First Search (BFS) search on a Trie data structure, augmented with dynamic programming. This technique is capable of not only identifying namedentities present on a text, but also estimating the actual named-entities from erroneous data. To evaluate the proposed technique, we conducted experiments in a closed domain where we employed this approach on a text corpus with some predefined named entities. The texts experimented on was both structured and unstructured, and our algorithm managed to succeed in both the cases.