9 February 2023 in ACM Computing Surveys
Neural Machine Translation (NMT) has seen tremendous growth in the last ten years since the early 2000s and has already entered a mature phase. While considered the most widely used solution for Machine Translation, its performance on low-resource language pairs remains sub-optimal compared to the high-resource counterparts due to the unavailability of large parallel corpora. Therefore, the implementation of NMT techniques for low-resource language pairs has been receiving the spotlight recently, thus leading to substantial research on this topic. This article presents a detailed survey of research advancements in low-resource language NMT (LRL-NMT) and quantitative analysis to identify the most popular techniques. We provide guidelines to select the possible NMT technique for a given LRL data setting based on our findings. We also present a holistic view of the LRL-NMT research landscape and provide recommendations to enhance the research efforts further.
18 January 2022 in ACM Transactions on Asian and Low-Resource Language Information Processing
Availability of corpora is a basic requirement for conducting research in a particular language. Unfortunately, for a morphologically rich language like Urdu, despite being used by over a 100 million people around the globe, the dearth of corpora is a major reason for the lack of attention and advancement in research. To this end, we present the first-ever large-scale publicly available Roman-Urdu parallel corpus, Roman-Urdu-Parl, with 6.37 million sentence-pairs. It is a huge corpus collected from diverse sources, annotated using crowd-sourcing techniques, and also assured for quality. It has a total of 92.76 million Roman-Urdu words, 92.85 million Urdu words, Roman-Urdu vocabulary of 42.9 K words, and Urdu vocabulary of 43.8 K words. Roman-Urdu-Parl has been built to ensure that it not only captures the morphological and linguistic features of the language but also the heterogeneity and variations arising due to demographic conditions. We validate the authenticity and quality of our corpus by using it to address two natural language processing research problems, that is, on learning word embeddings and building a machine transliteration system. Our contribution of the corpus leads to exceptional results in both settings, for example, our machine transliteration system sets a new state-of-the-art with a Bilingual Evaluation Understudy (BLEU) score of 84.67. We believe that Roman-Urdu-Parl can serve as fuel for igniting and advancing works in many research areas related to the Urdu language.
10 December 2020 in International Journal of Pattern Recognition and Artificial Intelligence
19 March 2025 in IEEE Transactions on Audio, Speech and Language Processing
We present a multi-way parallel corpus of Math Word Problems (MWPs) in nine languages, including six low-resource languages. To date, this is the largest multilingual MWP dataset available. We utilize this dataset and show the viability of using pre-trained multilingual sequence-sequence language models (prMSLMs) for autoregressive MWP generation in both monolingual and multilingual setups, particularly for low-resource languages. We also integrate a math constraint satisfaction module with autoregressive text generation. Our extensive evaluations identify several factors that affect autoregressive text generation on prMSLMs. These include language representation in the model, model size, existence of similar languages in the model, and language script. Overall, our results reveal that autoregressive MWP generation on top of prMSLMs is very promising, even for low-resource languages.