Myanmar News Headline Generation with Sequence-to-Sequence model

Abstract

News Headline generation is one of the most valuable research recently in NLP area. Generation of News headline means by learning to map articles to headlines using Sequence-to-Sequence model. Headline Generator that used an encoder and a decoder designed using Long Short-Term Memory (LSTM) was applied in this work. In this paper, an automatic headline generation for Myanmar News article using Seq2Seq model is implemented. There are various ways to generate a headline for news. In this paper, headline was generated using Seq2Seq with one-hot encoding and described about the comparative analysis results. While constructing the model, there are some challenges such as vocabulary counting and find out unknown terms in word embedding. In order to get more meaningful results, used the error analysis to typical neural headline generation system and evaluated based on machine generated headlines and actual headlines using ROUGE evaluation metric. The experiments have been conducted on Myanmar News dataset of 7000 pairs of news articles and their corresponding headlines. According to the evaluation, Seq2Seq with one-hot encoding outperforms than other Seq2Seq with word embedding (GloVe) and Recursive Recurrent Neural Network (Recursive RNN).