Generating Myanmar News Headlines using Recursive Neural Network

Abstract

Text summarization in the form of Headline prediction for written articles becomes a popular research recently. This paper presents a headline prediction model using Recursive Recurrent Neural Network (Recursive RNN) for Myanmar articles and evaluates the performance by comparing with sequence-to-sequence models. Recursive RNN model use encoder-decoder LSTM model that generates a single word forecast and call it recursively to text summarization in the form of news headlines for Myanmar news articles. A Recursive RNN, which takes input as the sequence of variable length and has to generate the variable length output by taking into account the previous input. 5000 articles of Myanmar news were collected to train headline prediction model. The performance of Recursive RNN, Seq2Seq with one-hot encoding and Seq2Seq with word embedding (GloVe) were evaluated in terms of ROUGE score values. The experimental results show that Recursive RNN model significantly better than other two models.