Abstract:
The use of Short Message Service (SMS) is increasing due to the rapid increase in mobile
phone usage and the simplicity in sending SMS messages. With the increasing complexity
of human lives, people are seeking more efficient activities to save time. This research
proposes an end-to-end method that automatically generates short responses known as
Smart Replies by identifying the content of an SMS using natural language processing.
There are a few pieces of research done on the topic of Smart Reply. Most of them are
carried out for the emails. And the efficiency and the size of those existing models cannot
be used in an offline mobile device. The application will use Natural Language
Processing to process an incoming message and then uses a neural network to predict the
most likely responses which will allow us to send it directly or edit it before sending it to
the recipient. The Ubuntu Corpus dataset was used for training and testing the model by
analysing its properties. It is identified that there are three main approaches: TF-IDF,
Recurrent Neural networks (RNN), and Long Short-Term Memory (LSTM) that can be
used in the model. After a performance test, identified the most suitable approach is
LSTM. Accordingly built a Sequential Neural Network with a Dense with sigmoid
activation using LSTM. Finally, extract the highest three responses from the trained
model to show in the SMS application. This proposed model achieved around 92%
percent of accurate results and it can be used offline and also it is a lightweight file that
can be easily handled in a mobile device.
Keywords: Smart reply, SMS, Natural language processing, Long short-term memory,
Sequential neural network
