Hindi to Dravidian Language Neural Machine Translation Systems

Proceedings of Recent Advances in Natural Language Processing, pages 1143–1150 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_121 1143 1 Abstract Neural machine translation (NMT) has achieved state-of-art performance in high- resource language pairs, but the performance of NMT drops in low- resource conditions. Morphologically rich languages are yet another challenge in NMT. The common strategy to handle this issue is to apply sub-word segmentation. In this work, we compare the morphologically inspired segmentation methods against the Byte Pair Encoding (BPE) in processing the input for building NMT systems for Hindi to Malayalam and Hindi to Tamil, where Hindi is an Indo- Aryan language and Malayalam and Tamil are south Dravidian languages. These two languages are low resource, morphologically rich and agglutinative. Malayalam is more agglutinative than Tamil. We show that for both the language pairs, the morphological segmentation algorithm out-performs BPE. We also present an elaborate analysis on translation outputs from both the NMT systems. 1 Introduction Machine translation has improved extensively using deep neural networks with the utilization of large dataset and high computational capacities. The successful works in Neural Machine Translation (NMT) started with the encoder- decoder based architecture presented by Kalchlorenner and Blunsom (2013), Sutskever et. al. (2014), and Cho et. al (2014). Sutskever et al (2014) built NMT system using Long short Term memory (LSTM) to overcome the fixed-length vector constraint in the previous architecture. Bahdanu et. al. (2015) introduced the attention mechanism, where bidirectional recurrent neural network (RNN) consisting of forward and backward RNN was used to focus around the word. This attention mechanism was simplified by considering the hidden states at the top layer of both encoder and decoder by Luong et. al. (2015). Transformer, an architecture where encoder and decoder completely relying on the attention machines was presented by Vaswani et. al. (2017). Though these NMT systems have achieved a state-of-art performance in high-resource, closely related languages, its performance drop significantly in low-resource and morphologically rich languages. Some of the techniques employed to mitigate challenges in handling the low- resource languages are as follows; increasing the data using back translation, utilisation of phrase tables generated in SMT, leveraging the pre- trained models, combining the similar language data and using transfer learning. The morphological rich languages are handled using different sub-word segmentation techniques, which helps in reducing the vocabulary size and increasing the number of examples of each tokens. In this work, we compare the morphologically inspired segmentation methods against the Byte Pair Encoding (BPE) in processing the input for building NMT systems for Hindi (Hi) to Malayalam (ML) and Hindi to Tamil (TA), where Hindi is an Indo-Aryan language and Malayalam and Tamil are south Dravidian languages. These two languages are low-resource, morphologically rich and agglutinative. Further the paper is organised as follows. In the following section, we present a summary on the different sub-word tokenisation works in NMT. This is followed by details on related works in Hindi to Dravidian Language Neural Machine Translation Systems Vijay Sundar Ram and Sobha Lalitha Devi AU-KBC Research …

Content truncated to 500 words