SIBI (Sign System for Indonesian Language) is an official sign language system used in school for hearing impairment students in Indonesia. This work uses the skeleton and hand shape features to classify SIBI gestures. In order to improve the performance of the gesture classification system, we tried to fuse the features in several different ways. The accuracy results achieved by the feature fusion methods are, in descending order of accuracy: 88.016%, when using sequence-feature-vector concatenation, 85.448% when using Conneau feature vector concatenation, 83.723% when using feature-vector concatenation, and 49.618% when using simple feature concatenation. The sequence-feature-vector concatenation techniques yield noticeably better results than those achieved using single features (82.849% with skeleton feature only, 55.530% for the hand shape feature only). The experiment results show that the combined features of the whole gesture sequence can better distinguish one gesture from another in SIBI than the combined features of each gesture frame. In addition to finding the best feature combination technique, this study also found the most suitable Recurrent Neural Network (RNN) model for recognizing SIBI. The models tested are 1-layer, 2-layer LSTM, and GRU. The experimental results show that the 2-layer bidirectional LSTM has the best performance.

-

-

-

The number of legal documents is already huge and growth is occurring at a swift pace; yet, the need to procure information from these documents for various purposes persists. One of the technologies available for this purpose is information extraction (IE), which has been applied widely in different domains to date. However, few research reviews have specifically discussed the development of legal IE. For this reason, it is considered important to conduct a literature review to identify and analyse the research, topics, datasets and methods used in legal IE research. This study used a standard systematic review method as stipulated by the PRISMA guidelines and collected papers from five main databases published between January 1990 and December 2020. A total of 107 articles were selected for inclusion in this review. Legal IE has been employed for various applications to date, but there is a wide range of different conditions, problems, and solutions reported by various studies in applying legal IE in various countries. This study offers a more concrete understanding, a broader perspective and provides an overview of various open problems in this regard so that it can serve as a guide in the development of future legal IE research.

Sinkhorn divergence is a symmetric normalization of entropic regularized optimal transport. It is a smooth and continuous metrized weak-convergence with excellent geometric properties. We use it as an alternative for the minimax objective function in formulating generative adversarial networks. The optimization is defined with Sinkhorn divergence as the objective, under the non-convex and non-concave condition. This work focuses on the optimization’s convergence and stability. We propose a first order sequential stochastic gradient descent ascent (SeqSGDA) algorithm. Under some mild approximations, the learning converges to local minimax points. Using the structural similarity index measure (SSIM), we supply a non-asymptotic analysis of the algorithm’s convergence rate. Empirical evidences show a convergence rate, which is inversely proportional to the number of iterations, when tested on tiny colour datasets Cats and CelebA on the deep convolutional generative adversarial networks and ResNet neural architectures. The entropy regularization parameter ε is approximated to the SSIM tolerance ϵ . We determine that the iteration complexity to return to an ϵ -stationary point to be O(κlog(ϵ−1)) , where κ is a value that depends on the Sinkhorn divergence’s convexity and the minimax step ratio in the SeqSGDA algorithm.

Named Entity Recognition (NER) is often used to acquire important information from text documents as a part of the Information Extraction (IE) process. However, the text documents quality affects the accuracy of the data obtained, especially for text documents acquired involving the Optical Character Recognition (OCR) process, which never reached 100% accuracy. This research tried to examine which OCR engine with the highest performance for IE using NER by comparing three OCR engines (Foxit, PDF2GO, Tesseract) over 8,562 government human resources documents within six document categories, two document structures, and four measurements. Several essential entities such as name, employee ID, document number, document publishing date, employee rank, and family member's name were trying to be extracted automatically from the documents. NER processes were done using Python programming language, and the preprocessing tasks were done separately for Foxit, PDF2GO, and Tesseract. In summary, each OCR engine has its drawbacks and benefit, such as Tesseract has better NER extraction and conversion time with better accuracy but lack in the number of entities acquired.