Computational Linguistics, Speech and Image Processing for Arabic Language

The following sections are included:

• Preface
• Contents

The Arabic language has many features such as the phonology and the syntax that make it an easy language for developing automatic speech recognition systems. Many standard techniques for acoustic and language modeling such as context dependent acoustic models and n-gram language models can be easily applied to Arabic. Some aspects of the Arabic language such as the nearly one-to-one letter-to-phone correspondence make the construction of the pronunciation lexicon even easier than in other languages. The most difficult challenges in developing speech recognition systems for Arabic are the dominance of non-diacritized text material, the several dialects, and the morphological complexity. In this chapter, we review the efforts that have been done to handle the challenges of the Arabic language for developing automatic speech recognition systems. This includes methods for automatic generation for the diacritics of the Arabic text and word pronunciation disambiguation. We also review the used approaches for handling the limited speech and text resources of the different Arabic dialects. Finally, we review the approaches used to deal with the high degree of affixation, derivation that contributes to the explosion of different word forms in Arabic.

In this chapter an introduction is given on the Human Language Technologies (HLT). Over 20 technologies including natural language processing such as information retrieval, machine translation, text mining, speech processing that includes speech recognition, and text to speech and optical character recognition have been concisely described. The challenges facing this important area of research have been underlined. Arabic HLT has more challenges compared, for example, with those of the English language due to the features of the Arabic language. The last section of the chapter contains a good reference for most of the organizations, research centers and companies that are working on the Arabic human language technologies.

Natural Language Processing (NLP) has increased significance in machine interpretation and different type of applications like discourse combination and acknowledgment, limitation multilingual data frameworks, and so forth. Arabic Named Entity Recognition, Information Retrieval, Machine Translation and Sentiment Analysis are a percentage of the Arabic apparatuses, which have indicated impressive information in knowledge and security organizations. NLP assumes a key part in the preparing stage in Sentiment Analysis, Information Extraction and Retrieval, Automatic Summarization, Question Answering, to name a few. Arabic is a Semitic language, which contrasts from Indo-European lingos phonetically, morphologically, syntactically and semantically. This paper discusses different challenges of NLP in Arabic. In addition, it inspires scientists in this field and others to take measures to handle Arabic dialect challenges.

Arabic recognition is still a big challenge to the scientific community. Several approaches to address this challenge have been attempted in the last ten years but significant work remains before large scale commercially viable systems can be built. In this chapter, we first discuss the characteristics of Arabic script and give a brief overview of the features extraction techniques proposed in the past works to characterize and recognize Arabic script. These techniques attempt to extract the feature vector that will be used in the recognition engine. We then investigate the use of Machine learning techniques: some statistical methods, mainly generative and discriminative based models, for Arabic recognition. As generative methods, we propose one-dimensional, two-dimensional and planar Hidden Markov Models (HMMs). To increase the representational power of the HMM, Dynamic Bayesian Networks (DBNs) are explored. In an attempt to benefit from the advantages of the dimensionality and the temporality of the models, a novel approach is proposed which integrates causal Markov Random Field in two dimensional modeling and HMMs. We then show different applications of this model for analytic recognition and syntactic analysis. As discriminative methods, we used Transparent Neural Networks (TNNs) to recognize a large vocabulary of Arabic words, based on a cognitive model where the learning is replaced by an activation process considering the nodes neighborhood.

The effective retrieval of information from scanned handwritten documents is becoming essential with the increasing volumes of digitized documents, and therefore developing efficient means for analysis and recognition of documents is of significant interest. Among these methods is word spotting, which has recently become an active research area. These systems have been implemented for Latin-based and Chinese languages, while few of them have been implemented for Arabic handwriting. The fact that Arabic writing is cursive by nature and unconstrained, with no clear white spaces between words, makes the processing of Arabic handwritten documents a more challenging problem.

This chapter introduces and discusses Arabic word spotting approaches and challenges. This includes the definition of word spotting, performance measure and approaches. Then, the Arabic language characteristics are introduced, the most commonly used Arabic Word spotting databases are summarized, and finally some extracted features for Arabic word spotting are presented.

Analyzing Arabic language sentences grammatically is the key to understand their meaning. E‘raab is the process of syntactically analyzing an Arabic sentence, and for many it is the most daunting task when studying Arabic grammar in school. In this paper we develop and implement a system that automates the task of analyzing Arabic sentences syntactically. Our system, which we named A‘rib as it is the imperative verb of e‘raab, is composed of three subsystems that imitate the human e‘raab process. These are: the lexical analyzer, the syntactic analyzer, and the results builder. The lexical analyzer identifies words in the input sentence based on their type and properties, and outputs them as tokens. While the syntactic analyzer parses the tokens out of the previous subsystem and tries to identify the sentence structure using rules expressed in a context-free grammar format. Finally, we combine the results of both subsystems (tokens and suitable rules), and output the complete e‘raab of the sentence along with a fully vowelized sentence. The system is intended for school children up to junior high level.

An extended and revised form of Tim Buckwalter’s Arabic lexical and morphological resource AraMorph, named eXtended Revised AraMorph (XRAM), is presented. A number of weaknesses and inconsistencies of the original model are addressed by allowing a wider coverage of real-world classical and contemporary (both formal and informal) Arabic texts. Building upon previous research, XRAM enhancements include (i) flag-selectable usage markers, (ii) probabilistic mildly context-sensitive POS tagging, filtering, disambiguation and ranking of alternative morphological analyses, and (iii) semi-automatic increments of lexical coverage through the extraction of lexical and morphological information from existing lexical resources. Testing XRAM through a front-end Python module showed a remarkable success level.

Arabic sentiment analysis, has been consistently gaining attention over the past couple of years. While research in the area of English Sentiment Analysis has been often aided by the presence of sentiment lexicons, such lexicons are very scarce for the Arabic language. Furthermore, lexicons that do exist are not scored or weighted. This paper describes the process by which entries in an existing Arabic Sentiment Lexicon built by the author, were assigned scores. Through a number of experiments on different datasets, it also shows that the use of the scored lexicon almost always improves the accuracy of sentiment analysis.

Multi-label classification (MLC) is concerned with learning from examples where each example is associated with a set of labels in opposite to traditional single-label classification where an example typically is assigned a single label. MLC problems appear in many areas, including text categorization, protein function classification, and semantic annotation of multimedia. The religious domain has become an interesting and challenging area for machine learning and natural language processing. A “fatwa” in the Islamic religion represents the legal opinion or interpretation that a qualified scholar (mufti) can give on issues related to the Islamic law. It is similar to the issue of legal opinions from courts in common-law systems. In this paper, a hierarchical classification system is introduced to automatically route incoming fatwa requests to the most relevant mufti. Each fatwa is associated to multiple categories by mufti where the categories can be organized in a hierarchy. The results on fatwa requests routing have confirmed the effective and efficient predictive performance of hierarchical ensembles of multi-label classifiers trained using the HOMER method and its variations compared to binary relevance which simply trains a classifier for each label independently.

Most existing studies on typeface design have primarily focused on typeface legibility and readability. This paper concentrates on the perception of typefaces and consists of two experimental studies. The first study investigates the perception of Arabic typefaces and the second contains an analysis of English typeface design. The Arabic typefaces study began by designing a font survey. Eighty-two people were asked to rate their perceptions of 20 Arabic typefaces based on 5 personality traits (Legible, Attractive, Comfortable, Artistic, and Formal). The correlation between typefaces and their personality traits were discovered by performing a series of statistical analyses. Based on the results, the studied typefaces were categorized into 3 groups according to their personality traits and typographical features. The second study focused on English typefaces and their personality traits. A font survey was designed. Seventy-one people rated their perceptions and preferences of 24 English typefaces based on six personality traits (Legible, Artistic, Formal, Sloppy, Readable, and Attractive). The correlation between typefaces and their personality traits were discovered by performing a series of statistical analyses. Based on the results, the number of studied typefaces was reduced from 24 to 16, and these 16 typefaces were categorized into four groups, according to their personality traits and typographical features.

This chapter presents the first end-to-end recipe for an Arabic speech-to-text transcription system using the lexicon free Recurrent Neural Networks (RNNs). The developed approach does not depend on Hidden Markov Models (HMMs), Gaussian Mixture Models (GMMs), or decision trees. In addition, a character based decoder is used for searching to avoid using a word lexicon. The Connectionist Temporal Classification (CTC) objective function is used to maximize the output character sequences given the acoustic features as an input. The recipe was evaluated using 1200 hours corpus of Aljazeera multi-Genre broadcast programs. On the development set, we report Word Error Rate (WER) of 12.03% for non-overlapped speech.

Arabic handwritten letter recognition is a challenging field in pattern recognition due to the different writing styles, variation in letters, and poor quality of manuscripts. There has not been such a system with sufficient accuracy until now. In this chapter, a letter recognition system based on swarm optimization algorithms with neural networks (NNs) was presented. The main purpose of the proposed approach is to improve the classification accuracy of NNs for Arabic handwritten letters recognition. Evolutionary Strategy (ES), Probability Based Incremental Learning (PBIL), Particle Swarm Optimization (PSO) and Moth-flame optimization (MFO) algorithms were applied to train NNs by generating optimum weights and biases and to be fed into neural networks. The proposed approach consists of these main phases: pre-processing, feature extraction and classification by optimized NNs. The proposed approach shows a good ability to train NNs, as it achieved high accuracy and f-measure values (not less than 93% for accuracy and 96% for f-measure) in all experiments. Up to our knowledge, these results are the highest achieved on a published Arabic letter recognition system yet. In addition, when comparing the proposed approach to the classic NNs, advantages were shown in both accuracy and f-measure.

The following section is included:

• Index