Paradigm Shift in Language Modeling: Revisiting CNN for Modeling Sanskrit Originated Bengali and Hindi Language

TL;DR

This paper introduces CoCNN, a memory-efficient CNN architecture tailored for Bengali and Hindi, demonstrating superior performance over Transformer and LSTM models in low-resource language modeling.

Contribution

The study presents a novel end-to-end trainable CNN model that outperforms state-of-the-art Transformer and LSTM models for Bengali and Hindi language modeling.

Findings

CoCNN outperforms pretrained BERT with 16X fewer parameters.

CoCNN achieves better results than SOTA LSTM models.

First comprehensive comparison of CNN, RNN, and Transformer architectures for Bengali and Hindi.

Abstract

Though there has been a large body of recent works in language modeling (LM) for high resource languages such as English and Chinese, the area is still unexplored for low resource languages like Bengali and Hindi. We propose an end to end trainable memory efficient CNN architecture named CoCNN to handle specific characteristics such as high inflection, morphological richness, flexible word order and phonetical spelling errors of Bengali and Hindi. In particular, we introduce two learnable convolutional sub-models at word and at sentence level that are end to end trainable. We show that state-of-the-art (SOTA) Transformer models including pretrained BERT do not necessarily yield the best performance for Bengali and Hindi. CoCNN outperforms pretrained BERT with 16X less parameters, and it achieves much better performance than SOTA LSTM models on multiple real-world datasets. This is the first study on the effectiveness of different architectures drawn from three deep learning paradigms - Convolution, Recurrent, and Transformer neural nets for modeling two widely used languages, Bengali and Hindi.