Sumanth Doddapaneni, Mohammed Safi Ur Rahman Khan, Dilip Venkatesh, Raj Dabre, Anoop Kunchukuttan, Mitesh M Khapra. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Ashwin Sankar, Sparsh Jain, Nikhil Narasimhan, Devilal Choudhary, Dhairya Suman, Mohammed Safi Ur Rahman Khan, Anoop Kunchukuttan, Mitesh M Khapra, Raj Dabre. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

This work introduces PrahokBART, a compact pre-trained sequence-to-sequence model trained from scratch for Khmer using carefully curated Khmer and English corpora. We focus on improving the pre-training corpus quality and addressing the linguistic issues of Khmer, which are ignored in existing multilingual models, by incorporating linguistic components such as word segmentation and normalization. We evaluate PrahokBART on three generative tasks: machine translation, text summarization, and headline generation, where our results demonstrate that it outperforms mBART50, a strong multilingual pre-trained model. Additionally, our analysis provides insights into the impact of each linguistic module and evaluates how effectively our model handles space during text generation, which is crucial for the naturalness of texts in Khmer.

Recently, chrF and chrF++ have become the preferred metric over BLEU for automatic n-gram evaluation of machine translation. Since they focus on character-level n-grams, it appears to have better correlations with human judgments for translating into morphologically rich languages compared to word-level metrics. However, for non-Latin languages with sub-character-level structures, we can go one step further namely bytes. To this end, we propose bytF to capture sub-character-level information, where we consider byte-level n-grams. Furthermore, we augment it to bytF+ and bytF++ where we consider character and word n-gram backoffs. On machine translation metric meta-evaluation datasets from English into 5 Indian languages, Chinese and Japanese, we show that bytF and its variants are comparable (give minimum difference) or significantly better (give maximum difference) correlated than chrF and chrF++ with human judgments at the segment level. We often observe that backing off to characters and words for bytF and to words for chrF does not have the highest correlation with humans. Furthermore, we also observe that using default n-gram values often leads to scores having poorer correlations with humans, indicating the need for well studied and tuned n-gram metrics for efficacy.

Recently, there have been increasing efforts on Quality Estimation (QE) and Post-Editing (PE) using Large Language Models (LLMs) for Machine Translation (MT). However, the focus has mainly been on high resource languages and the approaches either rely on prompting or combining existing QE models with LLMs, instead of single end-to-end systems. In this paper, we investigate the efficacy of end-to-end QE and PE systems for low-resource languages taking 5 Indian languages as a use-case. We augment existing QE data containing multidimentional quality metric (MQM) error annotations with explanations of errors and PEs with the help of proprietary LLMs (GPT-4), following which we fine-tune Gemma-2-9B, an open-source multilingual LLM to perform QE and PE jointly. While our models attain QE capabilities competitive with or surpassing existing models in both referenceful and referenceless settings, we observe that they still struggle with PE. Further investigation reveals that this occurs because our models lack the ability to accurately identify fine-grained errors in the translation, despite being excellent indicators of overall quality. This opens up opportunities for research in end-to-end QE and PE for low-resource languages.

Genta Indra Winata, Frederikus Hudi, Patrick Amadeus Irawan, David Anugraha, Rifki Afina Putri, Wang Yutong, Adam Nohejl, Ubaidillah Ariq Prathama, Nedjma Ousidhoum, Afifa Amriani, Anar Rzayev, Anirban Das, Ashmari Pramodya, Aulia Adila, Bryan Wilie, Candy Olivia Mawalim, Cheng Ching Lam, Daud Abolade, Emmanuele Chersoni, Enrico Santus, Fariz Ikhwantri, Garry Kuwanto, Hanyang Zhao, Haryo Akbarianto Wibowo, Holy Lovenia, Jan Christian Blaise Cruz, Jan Wira Gotama Putra, Junho Myung, Lucky Susanto, Maria Angelica Riera Machin, Marina Zhukova, Michael Anugraha, Muhammad Farid Adilazuarda, Natasha Christabelle Santosa, Peerat Limkonchotiwat, Raj Dabre, Rio Alexander Audino, Samuel Cahyawijaya, Shi-Xiong Zhang, Stephanie Yulia Salim, Yi Zhou, Yinxuan Gui, David Ifeoluwa Adelani, En-Shiun Annie Lee, Shogo Okada, Ayu Purwarianti, Alham Fikri Aji, Taro Watanabe, Derry Tanti Wijaya, Alice Oh, Chong-Wah Ngo. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Jaavid J, Raj Dabre, Aswanth M, Jay Gala, Thanmay Jayakumar, Ratish Puduppully, Anoop Kunchukuttan. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Mohammed Safi Ur Rahman Khan, Priyam Mehta, Ananth Sankar, Umashankar Kumaravelan, Sumanth Doddapaneni, Suriyaprasaad B, Varun G, Sparsh Jain, Anoop Kunchukuttan, Pratyush Kumar, Raj Dabre, Mitesh M. Khapra. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

The Nguni languages have over 20 million home language speakers in South Africa. There has been considerable growth in the datasets for Nguni languages, but so far no analysis of the performance of NLP models for these languages has been reported across languages and tasks. In this paper we study pretrained language models for the 4 Nguni languages - isiXhosa, isiZulu, isiNdebele, and Siswati. We compile publicly available datasets for natural language understanding and generation, spanning 6 tasks and 11 datasets. This benchmark, which we call NGLUEni, is the first centralised evaluation suite for the Nguni languages, allowing us to systematically evaluate the Nguni-language capabilities of pretrained language models (PLMs). Besides evaluating existing PLMs, we develop new PLMs for the Nguni languages through multilingual adaptive finetuning. Our models, Nguni-XLMR and Nguni-ByT5, outperform their base models and large-scale adapted models, showing that performance gains are obtainable through limited language group-based adaptation. We also perform experiments on cross-lingual transfer and machine translation. Our models achieve notable cross-lingual transfer improvements in the lower resourced Nguni languages (isiNdebele and Siswati). To facilitate future use of NGLUEni as a standardised evaluation suite for the Nguni languages, we create a web portal to access the collection of datasets and publicly release our models.

Subword regularized models leverage multiple subword tokenizations of one target sentence during training. However, selecting one tokenization during inference leads to the underutilization of knowledge learned about multiple tokenizations.We propose the SubMerge algorithm to rescue the ignored Subword tokenizations through merging equivalent ones during inference.SubMerge is a nested search algorithm where the outer beam search treats the word as the minimal unit, and the inner beam search provides a list of word candidates and their probabilities, merging equivalent subword tokenizations. SubMerge estimates the probability of the next word more precisely, providing better guidance during inference.Experimental results on six low-resource to high-resource machine translation datasets show that SubMerge utilizes a greater proportion of a model’s probability weight during decoding (lower word perplexities for hypotheses). It also improves BLEU and chrF++ scores for many translation directions, most reliably for low-resource scenarios. We investigate the effect of different beam sizes, training set sizes, dropout rates, and whether it is effective on non-regularized models.

In this paper, we explore the utility of Translationese as synthetic data created using machine translation for pre-training language models (LMs) for low-resource languages (LRLs).Our simple methodology consists of translating large amounts of web-crawled monolingual documents (clean) into the LRLs, followed by filtering the translated documents using tiny LMs trained on small but clean LRL data.Taking the case of Indian languages, we pre-train LMs from scratch with 28M and 85M parameters, and then fine-tune them for 5 downstream natural language understanding (NLU) and 4 generative (NLG) tasks.We observe that pretraining on filtered synthetic data leads to relative performance drops of only 0.87% for NLU and 2.35% for NLG, compared to pre-training on clean data, and this gap further diminishes upon the inclusion of a small amount of clean data.We also study the impact of synthetic data filtering and the choice of source language for synthetic data generation.Furthermore, evaluating continually pre-trained larger models like Gemma-2B and Llama-3-8B in few-shot settings, we observe that using synthetic data is competitive with using clean data.Our findings suggest that synthetic data shows promise for bridging the pre-training gap between English and LRLs.

Contemporary deep learning models effectively handle languages with diverse morphology despite not being directly integrated into them.Morphology and word order are closely linked, with the latter incorporated into transformerbased models through positional encodings.This prompts a fundamental inquiry: Is there a correlation between the morphological complexity of a language and the utilization of positional encoding in pre-trained language models?In pursuit of an answer, we present the first study addressing this question, encompassing 22 languages and 5 downstream tasks.Our findings reveal that the importance of positional encoding diminishes with increasing morphological complexity in languages.Our study motivates the need for a deeper understanding of positional encoding, augmenting them to better reflect the different languages under consideration.

IEEE Automatic Mean Opinion Score (MOS) prediction is employed to evaluate the quality of synthetic speech. This study extends the application of predicted MOS to the task of Fake Audio Detection (FAD) as we expect that MOS can be used to assess how close synthesized speech is to the natural human voice. We propose MOS-FAD, where MOS can be leveraged at two key points in FAD: training data selection and model fusion. In training data selection, we demonstrate that MOS enables effective filtering of samples from unbalanced datasets. In the model fusion, our results demonstrate that incorporating MOS as a gating mechanism in FAD model fusion enhances overall performance.

Nathaniel Robinson, Raj Dabre, Ammon Shurtz, Rasul Dent, Onenamiyi Onesi, Claire Monroc, Loïc Grobol, Hasan Muhammad, Ashi Garg, Naome Etori, Vijay Murari Tiyyala, Olanrewaju Samuel, Matthew Stutzman, Bismarck Odoom, Sanjeev Khudanpur, Stephen Richardson, Kenton Murray. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Visual Question Answering~(VQA) is an important task in multimodal AI, which requires models to understand and reason on knowledge present in visual and textual data. However, most of the current VQA datasets and models are primarily focused on English and a few major world languages, with images that are Western-centric. While recent efforts have tried to increase the number of languages covered on VQA datasets, they still lack diversity in low-resource languages. More importantly, some datasets extend the text to other languages, either via translation or some other approaches, but usually keep the same images, resulting in narrow cultural representation. To address these limitations, we create CVQA, a new Culturally-diverse Multilingual Visual Question Answering benchmark dataset, designed to cover a rich set of languages and regions, where we engage native speakers and cultural experts in the data collection process. CVQA includes culturally-driven images and questions from across 28 countries in four continents, covering 26 languages with 11 scripts, providing a total of 9k questions. We benchmark several Multimodal Large Language Models (MLLMs) on CVQA, and we show that the dataset is challenging for the current state-of-the-art models. This benchmark will serve as a probing evaluation suite for assessing the cultural bias of multimodal models and hopefully encourage more research efforts towards increasing cultural awareness and linguistic diversity in this field.

Abstract Creoles represent an under-explored and marginalized group of languages, with few available resources for NLP research. While the genealogical ties between Creoles and a number of highly resourced languages imply a significant potential for transfer learning, this potential is hampered due to this lack of annotated data. In this work we present CreoleVal, a collection of benchmark datasets spanning 8 different NLP tasks, covering up to 28 Creole languages; it is an aggregate of novel development datasets for reading comprehension relation classification, and machine translation for Creoles, in addition to a practical gateway to a handful of preexisting benchmarks. For each benchmark, we conduct baseline experiments in a zero-shot setting in order to further ascertain the capabilities and limitations of transfer learning for Creoles. Ultimately, we see CreoleVal as an opportunity to empower research on Creoles in NLP and computational linguistics, and in general, a step towards more equitable language technology around the globe.

Ananya Sai B, Tanay Dixit, Vignesh Nagarajan, Anoop Kunchukuttan, Pratyush Kumar, Mitesh M. Khapra, Raj Dabre. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Knowledge distillation (KD) is a well-known method for compressing neural models. However, works focusing on distilling knowledge from large multilingual neural machine translation (MNMT) models into smaller ones are practically nonexistent, despite the popularity and superiority of MNMT. This paper bridges this gap by presenting an empirical investigation of knowledge distillation for compressing MNMT models. We take Indic to English translation as a case study and demonstrate that commonly used language-agnostic and language-aware KD approaches yield models that are 4-5x smaller but also suffer from performance drops of up to 3.5 BLEU. To mitigate this, we then experiment with design considerations such as shallower versus deeper models, heavy parameter sharing, multistage training, and adapters. We observe that deeper compact models tend to be as good as shallower non-compact ones and that fine-tuning a distilled model on a high-quality subset slightly boosts translation quality. Overall, we conclude that compressing MNMT models via KD is challenging, indicating immense scope for further research.

This study investigates machine translation between related languages i.e., languages within the same family that share linguistic characteristics such as word order and lexical similarity.Machine translation through few-shot prompting leverages a small set of translation pair examples to generate translations for test sentences.This procedure requires the model to learn how to generate translations while simultaneously ensuring that token ordering is maintained to produce a fluent and accurate translation.We propose that for related languages, the task of machine translation can be simplified by leveraging the monotonic alignment characteristic of such languages.We introduce DecoMT, a novel approach of fewshot prompting that decomposes the translation process into a sequence of word chunk translations.Through automatic and human evaluation conducted on multiple related language pairs across various language families, we demonstrate that our proposed approach of decomposed prompting surpasses multiple established few-shot baseline approaches.For example, DecoMT outperforms the strong fewshot prompting BLOOM model with an average improvement of 8 chrF++ scores across the examined languages.

In this paper we evaluate the utility of large language models (LLMs) for translation of text with markup in which the most important and challenging aspect is to correctly transfer markup tags while ensuring that the content, both, inside and outside tags is correctly translated. While LLMs have been shown to be effective for plain text translation, their effectiveness for structured document translation is not well understood. To this end, we experiment with BLOOM and BLOOMZ, which are open-source multilingual LLMs, using zero, one and few-shot prompting, and compare with a domain-specific in-house NMT system using a detag-and-project approach for markup tags. We observe that LLMs with in-context learning exhibit poorer translation quality compared to the domain-specific NMT system, however, they are effective in transferring markup tags, especially the large BLOOM model (176 billion parameters). This is further confirmed by our human evaluation which also reveals the types of errors of the different tag transfer techniques. While LLM-based approaches come with the risk of losing, hallucinating and corrupting tags, they excel at placing them correctly in the translation.

This article investigates approaches to effectively harness source-side linguistic features for low-resource multilingual neural machine translation (MNMT). Previous works focus on using various features of a word such as lemma, part-of-speech tag, dependency label, and so on, to improve translation quality in a low-resource scenario. However, these studies deal with bilingual translation and do not focus on using features in multilingual training setups. Our work focuses on this particular point and experiments with low-resource multilingual models incorporating source-side linguistic features. Although techniques for integrating features into an NMT model such as concatenation and feature relevance perform quite well in bilingual settings, they do not work well in multilingual settings. To remedy this, we propose the use of dummy features and language indicator features in MNMT models. Experiments are conducted on English to Asian language translation on a multilingual, multi-parallel corpus spanning English and eight Asian languages where for each language pair, the training data size does not exceed 20,000 parallel sentences. After establishing strong bilingual baselines using feature relevance mechanisms and multilingual baselines without any features, we show that our proposed dummy features and language indicator features, in combination with feature relevance mechanisms, yield significant improvements in BLEU points for all language pairs. We then analyze our models from the perspectives of model sizes, the impact of individual linguistic features, validation perplexity computed during training, visualization of the activations of the relevance mechanisms, and exhaustive tuning of hyperparameters. We also report preliminary results for multilingual multi-way models using linguistic features.

Sub-word segmentation is an essential pre-processing step for Neural Machine Translation (NMT). Existing work has shown that neural sub-word segmenters are better than Byte-Pair Encoding (BPE), however, they are inefficient, as they require parallel corpora, days to train, and hours to decode. This article introduces SelfSeg, a self-supervised neural sub-word segmentation method that is much faster to train/decode and requires only monolingual dictionaries instead of parallel corpora. SelfSeg takes as input a word in the form of a partially masked character sequence, optimizes the word generation probability, and generates the segmentation with the maximum posterior probability, which is calculated using a dynamic programming algorithm. The training time of SelfSeg depends on word frequencies, and we explore several word frequency normalization strategies to accelerate the training phase. Additionally, we propose a regularization mechanism that allows the segmenter to generate various segmentations for one word. To show the effectiveness of our approach, we conduct MT experiments in low-, middle-, and high-resource scenarios, where we compare the performance of using different segmentation methods. The experimental results demonstrate that, on the low-resource ALT dataset, our method achieves more than 1.2 BLEU score improvement compared with BPE and SentencePiece, and a 1.1 score improvement over Dynamic Programming Encoding (DPE) and Vocabulary Learning via Optimal Transport (VOLT), on average. The regularization method achieves approximately a 4.3 BLEU score improvement over BPE and a 1.2 BLEU score improvement over BPE-dropout, the regularized version of BPE. We also observed significant improvements on IWSLT15 Vi→En, WMT16 Ro→En, and WMT15 Fi→En datasets and competitive results on the WMT14 De→En and WMT14 Fr→En datasets. Furthermore, our method is 17.8× faster during training and up to 36.8× faster during decoding in a high-resource scenario compared to DPE. We provide extensive analysis, including why monolingual word-level data is enough to train SelfSeg.

In this paper we present FeatureBART, a linguistically motivated sequence-to-sequence monolingual pre-training strategy in which syntactic features such as lemma, part-of-speech and dependency labels are incorporated into the span prediction based pre-training framework (BART). These automatically extracted features are incorporated via approaches such as concatenation and relevance mechanisms, among which the latter is known to be better than the former. When used for low-resource NMT as a downstream task, we show that these feature based models give large improvements in bilingual settings and modest ones in multilingual settings over their counterparts that do not use features.

Aman Kumar, Himani Shrotriya, Prachi Sahu, Amogh Mishra, Raj Dabre, Ratish Puduppully, Anoop Kunchukuttan, Mitesh M. Khapra, Pratyush Kumar. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing. 2022.

We participated in the mean opinion score (MOS) prediction challenge, 2022.This challenge aims to predict MOS scores of synthetic speech on two tracks, the main track and a more challenging sub-track: out-of-domain (OOD).To improve the accuracy of the predicted scores, we have explored several model fusion-related strategies and proposed a fused framework in which seven pretrained self-supervised learned (SSL) models have been engaged.These pretrained SSL models are derived from three ASR frameworks, including Wav2Vec, Hubert, and WavLM.For the OOD track, we followed the 7 SSL models selected on the main track and adopted a semi-supervised learning method to exploit the unlabeled data.According to the official analysis results, our system has achieved 1 st rank in 6 out of 16 metrics and is one of the top 3 systems for 13 out of 16 metrics.Specifically, we have achieved the highest LCC, SRCC, and KTAU scores at the system level on main track, as well as the best performance on the LCC, SRCC, and KTAU evaluation metrics at the utterance level on OOD track.Compared with the basic SSL models, the prediction accuracy of the fused system has been largely improved, especially on OOD sub-track.

Neural machine translation (NMT) models are typically trained using a softmax cross-entropy loss where the softmax distribution is compared against the gold labels. In low-resource scenarios and NMT models tend to perform poorly because the model training quickly converges to a point where the softmax distribution computed using logits approaches the gold label distribution. Although label smoothing is a well-known solution to address this issue and we further propose to divide the logits by a temperature coefficient greater than one and forcing the softmax distribution to be smoother during training. This makes it harder for the model to quickly over-fit. In our experiments on 11 language pairs in the low-resource Asian Language Treebank dataset and we observed significant improvements in translation quality. Our analysis focuses on finding the right balance of label smoothing and softmax tempering which indicates that they are orthogonal methods. Finally and a study of softmax entropies and gradients reveal the impact of our method on the internal behavior of our NMT models.

In a real-time simultaneous translation setting and neural machine translation (NMT) models start generating target language tokens from incomplete source language sentences and making them harder to translate and leading to poor translation quality. Previous research has shown that document-level NMT and comprising of sentence and context encoders and a decoder and leverages context from neighboring sentences and helps improve translation quality. In simultaneous translation settings and the context from previous sentences should be even more critical. To this end and in this paper and we propose wait-k simultaneous document-level NMT where we keep the context encoder as it is and replace the source sentence encoder and target language decoder with their wait-k equivalents. We experiment with low and high resource settings using the ALT and OpenSubtitles2018 corpora and where we observe minor improvements in translation quality. We then perform an analysis of the translations obtained using our models by focusing on sentences that should benefit from the context where we found out that the model does and in fact and benefit from context but is unable to effectively leverage it and especially in a low-resource setting. This shows that there is a need for further innovation in the way useful context is identified and leveraged.

In this study, linguistic knowledge at different levels are incorporated into the neural machine translation (NMT) framework to improve translation quality for language pairs with extremely limited data.Integrating manually designed or automatically extracted features into the NMT framework is known to be beneficial.However, this study emphasizes that the relevance of the features is crucial to the performance.Specifically, we propose two methods, 1) self relevance and 2) word-based relevance, to improve the representation of features for NMT.Experiments are conducted on translation tasks from English to eight Asian languages, with no more than twenty thousand sentences for training.The proposed methods improve translation quality for all tasks by up to 3.09 BLEU points.Discussions with visualization provide the explainability of the proposed methods where we show that the relevance methods provide weights to features thereby enhancing their impact on low-resource machine translation.

Cognates are variants of the same lexical form across different languages; for example "fonema" in Spanish and "phoneme" in English are cognates, both of which mean "a unit of sound".The task of automatic detection of cognates among any two languages can help downstream NLP tasks such as Cross-lingual Information Retrieval, Computational Phylogenetics, and Machine Translation.In this paper, we demonstrate the use of cross-lingual word embeddings for detecting cognates among fourteen Indian Languages.Our approach introduces the use of context from a knowledge graph to generate improved feature representations for cognate detection.We then evaluate the impact of our cognate detection mechanism on neural machine translation (NMT), as a downstream task.We evaluate our methods to detect cognates on a challenging dataset of twelve Indian languages, namely, Sanskrit, Hindi, Assamese, Oriya, Kannada, Gujarati, Tamil, Telugu, Punjabi, Bengali, Marathi, and Malayalam.Additionally, we create evaluation datasets for two more Indian languages, Konkani and Nepali 1 .We observe an improvement of up to 18% points, in terms of F-score, for cognate detection.Furthermore, we observe that cognates extracted using our method help improve NMT quality by up to 2.76 BLEU.We also release 2 our code, newly constructed datasets and cross-lingual models publicly.

Neural machine translation (NMT) needs large parallel corpora for state-of-the-art translation quality. Low-resource NMT is typically addressed by transfer learning which leverages large monolingual or parallel corpora for pre-training. Monolingual pre-training approaches such as MASS (MAsked Sequence to Sequence) are extremely effective in boosting NMT quality for languages with small parallel corpora. However, they do not account for linguistic information obtained using syntactic analyzers which is known to be invaluable for several Natural Language Processing (NLP) tasks. To this end, we propose JASS, Japanese-specific Sequence to Sequence, as a novel pre-training alternative to MASS for NMT involving Japanese as the source or target language. JASS is joint BMASS (Bunsetsu MASS) and BRSS (Bunsetsu Reordering Sequence to Sequence) pre-training which focuses on Japanese linguistic units called bunsetsus. In our experiments on ASPEC Japanese–English and News Commentary Japanese–Russian translation we show that JASS can give results that are competitive with if not better than those given by MASS. Furthermore, we show for the first time that joint MASS and JASS pre-training gives results that significantly surpass the individual methods indicating their complementary nature. We will release our code, pre-trained models and bunsetsu annotated data as resources for researchers to use in their own NLP tasks.

Lectures translation is a case of spoken language translation and there is a lack of publicly available parallel corpora for this purpose. To address this, we examine a framework for parallel corpus mining which is a quick and effective way to mine a parallel corpus from publicly available lectures at Coursera. Our approach determines sentence alignments, relying on machine translation and cosine similarity over continuous-space sentence representations. We also show how to use the resulting corpora in a multistage fine-tuning based domain adaptation for high-quality lectures translation. For Japanese–English lectures translation, we extracted parallel data of approximately 40,000 lines and created development and test sets through manual filtering for benchmarking translation performance. We demonstrate that the mined corpus greatly enhances the quality of translation when used in conjunction with out-of-domain parallel corpora via multistage training. This paper also suggests some guidelines to gather and clean corpora, mine parallel sentences, address noise in the mined data, and create high-quality evaluation splits. For the sake of reproducibility, we have released our code for parallel data creation.

Abstract This paper presents a set of effective approaches to handle extremely low-resource language pairs for self-attention based neural machine translation (NMT) focusing on English and four Asian languages. Starting from an initial set of parallel sentences used to train bilingual baseline models, we introduce additional monolingual corpora and data processing techniques to improve translation quality. We describe a series of best practices and empirically validate the methods through an evaluation conducted on eight translation directions, based on state-of-the-art NMT approaches such as hyper-parameter search, data augmentation with forward and backward translation in combination with tags and noise, as well as joint multilingual training. Experiments show that the commonly used default architecture of self-attention NMT models does not reach the best results, validating previous work on the importance of hyper-parameter tuning. Additionally, empirical results indicate the amount of synthetic data required to efficiently increase the parameters of the models leading to the best translation quality measured by automatic metrics. We show that the best NMT models trained on large amount of tagged back-translations outperform three other synthetic data generation approaches. Finally, comparison with statistical machine translation (SMT) indicates that extremely low-resource NMT requires a large amount of synthetic parallel data obtained with back-translation in order to close the performance gap with the preceding SMT approach.

In encoder-decoder based sequence-to-sequence modeling, the most common practice is to stack a number of recurrent, convolutional, or feed-forward layers in the encoder and decoder. While the addition of each new layer improves the sequence generation quality, this also leads to a significant increase in the number of parameters. In this paper, we propose to share parameters across all layers thereby leading to a recurrently stacked sequence-to-sequence model. We report on an extensive case study on neural machine translation (NMT) using our proposed method, experimenting with a variety of datasets. We empirically show that the translation quality of a model that recurrently stacks a single-layer 6 times, despite its significantly fewer parameters, approaches that of a model that stacks 6 different layers. We also show how our method can benefit from a prevalent way for improving NMT, i.e., extending training data with pseudo-parallel corpora generated by back-translation. We then analyze the effects of recurrently stacked layers by visualizing the attentions of models that use recurrently stacked layers and models that do not. Finally, we explore the limits of parameter sharing where we share even the parameters between the encoder and decoder in addition to recurrent stacking of layers.

Raj Dabre, Atsushi Fujita, Chenhui Chu. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

India is a country with 22 officially recognized languages and 17 of these have WordNets, a crucial resource.Web browser based interfaces are available for these WordNets, but are not suited for mobile devices which deters people from effectively using this resource.We present our initial work on developing mobile applications and browser extensions to access WordNets for Indian Languages.Our contribution is two fold: (1) We develop mobile applications for the Android, iOS and Windows Phone OS platforms for Hindi, Marathi and Sanskrit WordNets which allow users to search for words and obtain more information along with their translations in English and other Indian languages.(2) We also develop browser extensions for English, Hindi, Marathi, and Sanskrit WordNets, for both Mozilla Firefox, and Google Chrome.We believe that such applications can be quite helpful in a classroom scenario, where students would be able to access the WordNets as dictionaries as well as lexical knowledge bases.This can help in overcoming the language barrier along with furthering language understanding.

Raj Dabre, Fabien Cromieres, Sadao Kurohashi, Pushpak Bhattacharyya. Proceedings of the 2015 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2015.

The task of Word Sense Disambiguation (WSD) incorporates in its definition the role of 'context'.We present our work on the development of a tool which allows for automatic acquisition and ranking of 'context clues' for WSD.These clue words are extracted from the contexts of words appearing in a large monolingual corpus.These mined collection of contextual clues form a discrimination net in the sense that for targeted WSD, navigation of the net leads to the correct sense of a word given its context.Utilizing this resource we intend to develop efficient and light weight WSD based on look up and navigation of memoryresident knowledge base, thereby avoiding heavy computation which often prevents incorporation of any serious WSD in MT and search.The need for large quantities of sense marked data too can be reduced.