A Comparative study of Long Short-Term Memory and Gated Recurrent Unit

Abstract

In natural language processing (NLP), the assumption that a neural network has an independent state among data samples does not apply to sequential data.Hence recurrent neural networks (RNN) have played a key role in sequential dependency in natural language processing with the key features of providing context to the processing and tackling vanishing gradient.Long Short-Term Memory Units (LSTM) are RNN blocks that can retain essential information even if it is far from the current point of analysis (extended memory).Still, they also have a fading effect that favours closer information (short memory).Despite this, they still need to remember vital details far from their current position, which goes against the intent of the extended memory effect.Gated Recurrent Units (GRU) have shown excellent results in sequential data and were introduced to overcome the limitations of LSTM by using two vectors (update gate and reset gate) to decide what information passes to the output.They can also train to keep data for a long time without it washing it through time or removing information irrelevant to the prediction.Some scholars suggest that gated recurrent units could be a suitable replacement for long short term memory.This comparative study presented the performance difference between LSTM and GRU and their bi-directional-based neural networks when they face the task of classifying text data.The evaluation of Gated Recurrent Unit (GRU) versus Long Short Term Memory (LSTM) and their bi-directional versions were carried out on a task of a website based on its content.Our analysis showed that a gated recurrent unit (GRU) is a good substitute for long short-term memory for text data classification.The bi-directional GRU outperformed the bi-direction LSTM.We recommend a gated recurrent unit as a better alternative to Long Short Term Memory on text data classification.