Encoder Decoder based Nepali News Headline Generation

In this paper, a method for Nepali News Headline Generation is presented. The proposed method uses GRUs, in an encoder-decoder fashion, taking the news content as input and generating a headline as the output. The news is converted into word tokens, which are vectorized using FastText, trained on a corpus of Nepali news articles and headlines collected from several web portals. The headline generation model is also trained on the same corpus. A sequence to sequence model, with an encoder and a decoder GRU is used as the generation model. The model was able to attain a BLEU score of 22.1 on the test set.