Ahmad A. Al Sallab

Deep learning framework with confused sub-set resolution architecture for automatic Arabic diacritization

The Arabic language belongs to a group of languages that require diacritization over their characters. Modern Standard Arabic (MSA) transcripts omit the diacritics, which are essential for many machine learning tasks like Text-To-Speech (TTS) systems. In this work Arabic diacritics restoration is tackled under a deep learning framework that includes the Confused Sub-set Resolution (CSR) method to improve the classification accuracy, in addition to an Arabic Part-of-Speech (PoS) tagging framework using deep neural nets. Special focus is given to syntactic diacritization, which still suffers low accuracy as indicated in prior works. Evaluation is done versus state-of-the-art systems reported in literature, with quite challenging datasets collected from different domains. Standard datasets like the LDC Arabic Tree Bank are used in addition to custom ones we have made available online to allow other researchers to replicate these results. Results show significant improvement of the proposed techniques over other approaches, reducing the syntactic classification error to 9.9% and morphological classification error to 3% compared to 12.7% and 3.8% of the best reported results in literature, improving the error by 22% over the best reported systems.

Automatic Arabic diacritics restoration based on deep nets

In this paper, Arabic diacritics restoration problem is tackled under the deep learning framework presenting Confused Subset Resolution (CSR) method to improve the classification accuracy, in addition to Arabic Part-of-Speech (PoS) tagging framework using deep neural nets. Special focus is given to syntactic diacritization, which still suffer low accuracy as indicated by related works. Evaluation is done versus state-of-the-art systems reported in literature, with quite challenging datasets, collected from different domains. Standard datasets like LDC Arabic Tree Bank is used in addition to custom ones available online for results replication. Results show significant improvement of the proposed techniques over other approaches, reducing the syntactic classification error to 9.9% and morphological classification error to 3% compared to 12.7% and 3.8% of the best reported results in literature, improving the error by 22% over the best reported systems

Self learning machines using Deep Networks

Self learning machines as defined in this paper are those learning by observation under limited supervision, and continuously adapt by observing the surrounding environment. The aim is to mimic the behavior of human brain learning from surroundings with limited supervision, and adapting its learning according to input sensory observations. Recently, Deep Belief Nets (DBNs) [1] have made good use of unsupervised learning as pre-training stage, which is equivalent to the observation stage in humans. However, they still need supervised training set to adjust the network parameters, as well as being nonadaptive to real world examples. In this paper, Self Learning Machine (SLM) is proposed based on deep belief networks and deep auto encoders

RDI_Team at SemEval-2016 Task 3: RDI Unsupervised Framework for Text Ranking.

Ranking is an important task in the field of information retrieval. Ranking may be used in different modules in natural language processing such as search engines. In this paper, we introduce a competitive ranking system which combines three different modules. The system participated in SemEval 2016 question ranking task for the Arabic language. The task is a ranking task that targets reordering results retrieved from search engine. Results reordering is done based on relevancy between search result and the original query issued. The data provided in the competition is in the form of question (query) and 30 question answer pairs retrieved from search engine. For each question retrieved from the search engine the system generates a relevancy score that is to be used for ranking. The proposed system came in the third position in the Competition. Since the majority of modules are unsupervised the unsupervised naming was used.

Deep learning with ensemble classification method for sensor sampling decisions

Modern mobile pervasive applications focus on context awareness that monitors a diverse range of personal domains. In order to infer contextual information, most of these applications require the collection of raw data from sensors which are either embedded in personal smartphones or worn by the user. Critical context-aware applications rely on continuous accurate monitoring of the users current context. Continuous sensing mechanisms in sensors cost high energy consumption to support accurate contextual detection. Hence, there is a trade-off between the classification accuracy and the energy consumption. In this paper, we exploit the advantages of Deep Neural Network (DNN) with ensemble classification of other complementary machine learning approaches to determine the best sensor sampling frequency for the recognition of a given context. DNN relies on raw data for classification while the other complementary methods (such as Decision Tree and Naïve Bayes) use feature recognition to classify data. Therefore, our approach provides a range of granularity from raw data. We prove the robustness of our approach in experiments which show high accuracy in context recognition. In addition, real experiments demonstrate the energy gains of the proposed algorithm which reach 87% reduction in energy consumption when compared to continuous sensing.

Hardware implementation of distributed speech recognition system front end

Modern speech recognition applications are heading towards embedded systems and hand-held devices. Distributed Speech Recognition (DSR) system architecture emerged to address this kind of applications. Most of the existing implementations of this system are presented in software fashion, with little consideration to the end product platform in which the system will be deployed. In this paper, an optimized hardware implementation of the front end part of the DSR specified in the basic ETSI Aurora standard ETSI ES 201 108 is presented in FPGA platform prototype, with consideration of migration to structured ASIC in case of mass-production. Main design issues and tips are highlighted. Results are presented in terms of hardware resources utilization, comparison of some basic system components to third party reference designs and compliance to the Aurora standard.