Computer Science

HYBRID END-TO-END APPROACH INTEGRATING ONLINE LEARNING WITH FACE-IDENTIFICATION SYSTEM

Son Van Nguyen — 2023-03-10

To date, facial recognition has been one of the most intriguing, interesting research topics over years. It requires some specific face-based algorithms such as facial detection, facial alignment, facial representation, and facial recognition as well; however, all of these algorithms derive from heavy deep learning architectures that cause limitations for development, scalability, flawed accuracy, and deployment into publicity with mere CPU servers. It also calls for large datasets containing hundreds of thousands of records for training purposes. In this paper, we propose a full pipeline for an effective face recognition application which only uses a small Vietnamese celebrity dataset and CPU for training that can solve the leakage of data and the need for GPU devices. It is based on a face vector-to-string tokens algorithm then saves face’s properties into Elasticsearch for future retrieval, so the problem of online learning in Facial Recognition is also tackled. Comparison with another popular algorithm on the dataset, our proposed pipeline not only outweighs the accuracy counterpart, but it also achieves a very speedy time inference for a real-time face recognition application.

ADAPTATION OF DOMAIN-SPECIFIC TRANSFORMER MODELS WITH TEXT OVERSAMPLING FOR SENTIMENT ANALYSIS OF SOCIAL MEDIA POSTS ON COVID-19 VACCINE

Anmol Bansal — 2023-03-10

Covid-19 has spread across the world and many different vaccines have been developed to counter its surge. To identify the correct sentiments associated with the vaccines from social media posts, this paper aims to fine-tune pre-trained transformer models on tweets associated with different Covid vaccines, specifically RoBERTa, XLNet and BERT which are recently introduced state-of-the-art bi-directional transformer models, and domain-specific transformer models BERTweet and CT-BERT that are pre-trained on Covid-19 tweets. We further explore the option of data augmentation by text oversampling using LMOTE to improve the accuracies of these models, specifically, for small sample datasets where there is an imbalanced class distribution among the positive, negative and neutral sentiment classes. Our results summarize our findings on the suitability of text oversampling for imbalanced, small sample datasets that are used to fine-tune state-of-the-art pre-trained transformer models, and the utility of having domain-specific transformer models for the classification task.

ARNLI: ARABIC NATURAL LANGUAGE INFERENCE ENTAILMENT AND CONTRADICTION DETECTION

Khloud Al Jallad — 2023-03-10

Natural Language Inference (NLI) is a hot topic research in natural language processing, contradiction detection between sentences is a special case of NLI. This is considered a difficult NLP task which has a big influence when added as a component in many NLP applications, such as Question Answering Systems, text Summarization. Arabic Language is one of the most challenging low-resources languages in detecting contradictions due to its rich lexical, semantics ambiguity. We have created a dataset of more than 12k sentences and named ArNLI, that will be publicly available. Moreover, we have applied a new model inspired by Stanford contradiction detection proposed solutions on English language. We proposed an approach to detect contradictions between pairs of sentences in Arabic language using contradiction vector combined with language model vector as an input to machine learning model. We analyzed results of different traditional machine learning classifiers and compared their results on our created dataset (ArNLI) and on an automatic translation of both PHEME, SICK English datasets. Best results achieved using Random Forest classifier with an accuracy of 99%, 60%, 75% on PHEME, SICK and ArNLI respectively.

TRANSFORMATION AND CLASSIFICATION OF ORDINAL SURVEY DATA

Roopam Sadh — 2023-03-10

Currently, Machine Learning is being significantly used in almost all of the research domains. However, its applicability in survey research is still in its infancy. We in this paper, attempt to highlight the applicability of Machine Learning in survey research while working on two different aspects in parallel. First, we introduce a pattern-based transformation method for ordinal survey data. Our purpose behind developing such a transformation method is twofold. Our transformation facilitates easy interpretation of ordinal survey data and provides convenience while applying standard Machine Learning approaches. Second, we demonstrate the application of various classification techniques over real and transformed ordinal survey data and interpret their results in terms of their suitability in survey research. Our experimental results suggest that Machine Learning coupled with the Pattern Recognition paradigm has a tremendous scope in survey research.

DEEP CONVOLUTIONAL NEURAL NETWORK USING A NEW DATASET FOR BERBER LANGUAGE

Mokrane Kemiche — 2023-03-10

Currently, Handwritten Character Recognition (HCR) technology has become an interesting and immensely useful technology. It has been explored with high
performance in many languages. However, a few HCR systems are proposed for the Amazigh (Berber) language. Furthermore, the validation of any Amazigh
handwritten recognition system remains a major challenge due to no availability of a robust Amazigh database. To address this problem, we first created two new datasets for Tifinagh and Amazigh Latin characters, by extending the well-known EMNIST database with the Amazigh alphabet. And then, we have proposed a handwritten character recognition system, which is based on a deep convolutional neural network to validate the created datasets. The proposed CNN has been trained and tested on our created datasets, and the experimental tests show that it achieves satisfactory results in terms of accuracy and recognition efficiency.

LEARNING-FREE DEEP FEATURES FOR MULTISPECTRAL PALM-PRINT CLASSIFICATION

hakim Bendjenna — 2023-03-10

The feature extraction step is a major and crucial step in analyzing and understanding raw data as it has a considerable impact on the system accuracy. Unfortunately, despite the very acceptable results obtained by many handcrafted methods, they can have difficulty representing the features in the case of large databases or with strongly correlated samples. In this context, we proposed a new, simple and lightweight method for deep feature extraction. Our method can be configured to produce four different deep features, each controlled to tune the system accuracy. We have evaluated the performance of our method using a multispectral palmprint based biometric system and the experimental results, using the CASIA database, have shown that our method has high accuracy compared to many current handcrafted feature extraction methods and many well known deep learning based methods.