News articles play a crucial role in helping humans know about many important events, developments, and inventions worldwide. The busy chores of our day-to-day life have made it quite challenging to consume important information from lengthy news articles. Therefore, short summaries of news articles are not only crucial but essential as well. Deep learning has revolutionized the field of natural language processing research. A lot of research has been done using pre-trained transformer-based models, and it has significantly improved the text sum-marization performance. In this paper, efforts were made to analyze transformer-based models such as BERT, GPT-2, XL Net, BART, and T5 for extractive and abstractive summarizations. This research investigates various methods through observation and experimentation. It also proposes methods that produce better summaries than comparable methods.

Fatigued driving has been reported to be a major cause of road accidents claiming millions of lives worldwide. Studies have shown that most road accidents occur either at night or early morning when the driver is already fatigued and there is insufficient light to notice obstacles. Some of the automated fatigue detection systems use physiological signals like EEG, ECG, and blood pressure movements. But, in most cases, the invasive nature of obtaining these signals makes them non-ideal. The recently developed computer vision based fatigue detection systems are too bulky or have limited accuracy due to prediction using single facial features or low-light conditions. Hence, the proposed method first enhances low-light images by improving the overall saturation and creating a uniform image using Gamma Correction. The enhanced images are then fed to a modified Multi-Task Cascaded Convolutional Neural Network for face detection and facial landmark extraction. Finally, the extracted eye state and mouth state features are fed to the LSTM network for fatigue classification. The output of this model decides whether the driver is fatigued or alert. The Mirror subset of the publicly available YawDD data set has been used for effective training and evaluation of the proposed model. The model achieved an exceptionally high F1 score of 0.98 and a Recall of 0.99 on the validation set.

Mental health of a person plays equivalent significant role in ensuring their wellbeing as their physical health. A great deal of work and e ffort has gone into increasing awareness of this issue. One su ch effort is made by the discipline of computer science, whic h makes use of social media data to give more information in identifying these mental illnesses. People are increasingly usi ng internet platforms to voice our suicide ideas as technology advances quickly. The purpose of the study is to identify a person's indicators of depression based on their social media postings, where users express their feelings and emotions. The goal of this study is to develop three models-Naive Bayes, Pre-Trained Model BERT, and XLNET-and compare their performance in identifying depression from messages on Twitter. These models are pre-processed using the Tweet preprocessor and BERT embeddings, and then the pretrained models are fine-tuned. With an accuracy of 0.9942, it was found that Bert performed better than the other two models.

COVID-19 crisis has led to an outburst of information that needs to be organized, validated, and made available to the seekers. Despite the rapid growth and success of BERT models in the last 3 years, COVID QA is a difficult task due to the lack of applicable datasets and a relevant language representation. Therefore, this study proposes a transformer-based Question Answering (QA) model for COVID-19 questions from the biomedical domain. Further, explored several datasets, and models required for question type prediction, no-answer prediction, and answer extraction and transfer learning strategies. It has been demonstrated that the exact match score can be significantly improved with limited amounts of training data from the biomedical domain. Finally, the findings of the study have been summarized as Factoid QA Finetuning Framework (FQFF), which can provide initial direction for domain-specific QA tasks with a limited amount of data.

Breast Cancer is the second most leading cause of death among women. The early detection of the disease increases the chances of survival of the patient. Therefore, there is always a need for techniques that can accurately predict the presence of cancer. Data Mining is one such powerful technique that can assist clinicians to effectively use the data for timely prediction of the disease. In the medical domain, data is usually imbalanced with unequal distribution of the positive and negative classes. Imbalanced datasets introduce a bias in the model and can thus reduce the accuracy of the minority class predictions. In the case of cancer detection, the mammographic data is highly imbalanced, and predicting the positive (minority) class is of the utmost importance. To achieve this, different models using various class balancing techniques are built and evaluated. The experiments show that the performance of the weighted approach and the undersampling technique is better than oversampling and hybrid techniques. The best performing classifiers are the weighted XGBoost model and Stacking ensemble with the average AUC of 0.78 and 0.76 respectively.

This paper presents an approach to the problem of paraphrase identification in English and Indian languages using Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN). Traditional machine learning approaches used features that involved using resources such as POS taggers, dependency parsers, etc. for English. The lack of similar resources for Indian languages has been a deterrent to the advancement of paraphrase detection task in Indian languages. Deep learning helps in overcoming the shortcomings of traditional machine Learning techniques. In this paper, three approaches have been proposed, a simple CNN that uses word embeddings as input, a CNN that uses WordNet scores as input and RNN based approach with both LSTM and bi-directional LSTM.