Neural IR-based Approaches for Bangla Text Retrieval

Abstract

Information retrieval (IR) for Bangla text has received relatively little attention de spite the widespread global usage of the language. The rich morphology and lack of capitalization for Bangla presents challenges for direct application of standard IR models developed predominantly for English text. This report explores the gradual development of IR methods for text retrieval including works on Bangla texts. The unsupervised nature of the methods used for Bangla text retrieval makes the methods unsuitable for specific domains. To mitigate the problems faced due to lack of domain specific training, different modern neural information retrieval techniques need to be explored that can handle different data availability scenarios. In this report, we exper iment with different neural information retrieval techniques on different percentage of available data and provide guidelines on building information retrieval pipelines for Bangla language. We also introduce a dataset containing rice-related scientific texts along with human annotated questions, which we used to train and evaluate the performance of domain-specific neural information retrieval architectures.