The lack of parallel corpora remains challenging for multilingual neural machine translation (MNMT), particularly for low-resource languages. This article presents an unsupervised framework to utilize pre-trained cross-lingual encoders (XLM-R) in an unsupervised way and generates high-quality translations using monolingual corpora and bilingual dictionaries. The proposed method constructs pseudo-parallel corpora by combining word-by-word translation using bilingual dictionaries with contextual refinement via masked language modeling (MLM). To improve alignment quality, we propose a two-tier representation strategy: (1) word-level alignment that combines VecMap with FastText embeddings to address out-of-vocabulary (OOV) terms and capture morphological variations. (2) Sentence-level alignment using Adversarial Contrastive Learning (ACL) enhanced with Hard Negative Mining (HNM) to build semantically robust and discriminative sentence embeddings. Experimental results on the FLORES-101 dataset demonstrate that the proposed model outperforms the existing state-of-the-art models, with an average of +1.2 BLEU in bilingual settings and +0.9 BLEU in multilingual settings. Furthermore, the proposed model is evaluated on 4 low-resource Indian languages (e.g., Hindi, Urdu, Telugu, and Bengali), and it outperforms the state-of-the-art models with an average of +0.7 in bilingual and multilingual settings. Finally, evaluation in zero-shot and few-shot settings confirms the proposed approach’s robustness and generalization, demonstrating an effective solution for multilingual translation without using parallel corpora.

Neural Machine Translation (NMT) is one step ahead of traditional statistical phrase-based translation systems because of its better translation ability. But it requires a large amount of parallel training data, which can be challenging for languages with limited resources like many Indian languages. In the past, researchers have tried to address the issue using data augmentation. In this paper, we present a data augmentation technique for the Hindi language based on five phrases: noun phrases, verb phrases, prepositional phrases, adjective phrases, and adverb phrases. We augment the training corpus using parser-generated phrasal segments and evaluate the efficiency of the proposed work on the Hindi language. Further, the paper presents training in the NMT model at the character level instead of the word level. This approach can help overcome challenges associated with word-level translations, such as handling rare and out-of-vocabulary words and phrases, dealing with morphological complexity, and addressing languages with ambiguous word boundaries. The proposed work was evaluated on a low-resource language pair, Hindi-English, using the Google Transformer model as the baseline state-of-the-art. The experiments used two distinct datasets: WMT14 Hin-Eng and Samanantar Hin-Eng parallel corpus with character-level encoding for the translation task. The proposed model is able to surpass the cutting-edge baseline and saw an increase in BLEU scores for the WMT14 translation challenge with +2.52 on base paper using three phrase sentences with character-level encoding and +2.68 BLEU Score on base paper using five phrase sentences with character-level encoding. Further, character-level encoding is evaluated on non-augmented Samanantar dataset; it performs better in the baseline approach for translation purposes. It clearly shows that the proposed model outperforms in Hindi language translation.

Drivers undergo a lot of stress that might cause distraction and might lead to an unfortunate incident. Emotional recognition via facial expressions is one of the most important field in the human–machine interface. The goal of this paper is to analyze the drivers’ facial expressions in order to monitor their stress levels. In this paper, we propose FERNET — a hybrid deep convolutional neural network model for driver stress recognition through facial emotion recognition. FERNET is an integration of two DCNNs, pre-trained ResNet101V2 CNN and a custom CNN, ConvNet4. The experiments were carried out on the widely used public datasets CK[Formula: see text], FER2013 and AffectNet, achieving the accuracies of 99.70%, 74.86% and 70.46%, respectively, for facial emotion recognition. These results outperform the recent state-of-the-art methods. Furthermore, since a few specific isolated emotions lead to higher stress levels, we analyze the results for stress- and nonstress-related emotions for each individual dataset. FERNET achieves stress prediction accuracies of 98.17%, 90.16% and 84.49% for CK[Formula: see text], FER2013 and AffectNet datasets, respectively.