Le lancement de ChatGPT et de Gemini marque le début d’une nouvelle ère de l’interaction homme / machine. Les espoirs sont immenses à l’égard de ces robots conversationnels qui donnent des conseils, conçoivent des contenus dans tous les domaines de connaissance, résolvent des problèmes et aussi créent des algorithmes. Actuellement, ChatGPT et Gemini sont intégrés dans divers domaines : l’enseignement/apprentissage, la littérature et les arts, le tourisme, le journalisme, les sciences, l’informatique, etc. Cependant, certains spécialistes en informatique ainsi que des utilisateurs commencent à parler des erreurs ou des hallucinations de ChatGPT et de Gemini. Des erreurs illogiques parfois futiles et parfois graves mettent en question la performance de leurs systèmes informatiques. Au cours du présent article nous présenterons deux types d’erreurs commises par les deux agents conversationnels les plus utilisés à l’échelle mondiale en traduction et en génération de texte. Nous essayerons de comprendre pourquoi ces systèmes linguistiques et informatiques rigoureux se trompent-ils. De même, nous suggérerons des moyens permettant d’éviter ce type d’erreurs.

Paraphrasing is one of the major yet the most challenging tasks of the deep semantic analysis of natural languages. In this paper we present a novel algorithm that operates on a big parallel text corpus and automatically generates the paraphrases of the two natural languages of the corpus. Like several previously crafted algorithms in this regard, our algorithm exploits the bidirectional translation provided by the big parallel text corpora to infer couples of synonymous phrases, however, our algorithm is simpler and more efficient. Moreover, our algorithm is the only one that constructs the whole paraphrase through its run without any need for further post processing. We implemented and ran our algorithm on the English-Arabic text corpora from the 2018 version of the OpenSubtitles (OPUS) parallel text corpora, and through the statistical evaluation of random samples we found that the semantic quality among the phrases of the automatically generated paraphrases to be interestingly superb.

La mathématisation des langues naturelles était née du besoin de faire comprendre à la machine la langue humaine pour qu’elle puisse accomplir les tâches exécutées en principe par l’homme dans un laps de temps relativement court et avec des prix réduits. Les recherches qui traitent de la formalisation de la langue arabe ont commencé dès les années quatre-vingt du siècle précédent. Les linguistes arabes modernes cherchent à créer un système formel rigoureux qui est inspiré des systèmes de formalisation des langues européennes et qui sera apte à gérer les différents phénomènes lexicologiques arabes. Le présent article abordera la formalisation de la langue arabe d’un point de vue linguistico-informatique : nous jetterons la lumière sur l’importance de l’analyse linguistique dans le processus de mathématisation des langues naturelles en général. Le système de formalisation que nous adopterons dans la recherche est celui des fonctions lexicales qui a été élaboré dans le cadre de la Théorie Sens-Texte. Nous soulignerons les atouts qui le distinguent des autres systèmes de formalisation en l’appliquant sur quatre phénomènes lexicologiques arabes : le cliché, l’hyperbole, le diminutif et le complément absolu.

This work presents a novel emotion recognition via automatic speech recognition (ASR) using a deep feed-forward neural network (DFFNN) for Arabic speech. We present results for the recognition of the three emotions happy, angry, and surprised. The Arabic natural audio dataset (ANAD) is used. Twenty-five low-level descriptors (LLDs) are extracted from the audio signals. Different combination of extracted features is examined. Also, the effect of using the principal component analysis (PCA) technique for dimensionality reduction is examined. For the classification stage, DFFNN is used. Also, the problem of imbalances samples in the dataset is managed by using the borderline-synthetic minority over-sampling technique (B-SMOTE). It is shown from the results thatthe best accuracy is obtained when applying PCA on the extracted features is 98.56 %.Also, the accuracy is 98.33 % when using the combination of all the extracted features.This result is not too much different from the accuracy of using PCA.It is followed by the accuracy of using MFCC and LSF which is 97.79 %.It is noticed that the accuracy is 95.63 % when using LSF features which shows that they are dominant features.The obtained results showed an improvement compared to previous studies.

This work is focused on single word Arabic automatic speech recognition (AASR). Two techniques are used during the feature extraction phase; Log frequency spectral coefficients (MFSC) and Gammatone-frequency cepstral coefficients (GFCC) with their first and second-order derivatives. The convolutional neural network (CNN) is mainly used to execute feature learning and classification process. CNN achieved performance enhancement in automatic speech recognition (ASR). Local connectivity, weight sharing, and pooling are the crucial properties of CNNs that have the potential to improve ASR. We tested the CNN model using an Arabic speech corpus of isolated words. The used corpus is synthetically augmented by applying different transformations such as changing the pitch, the speed, the dynamic range, adding noise, and forward and backward shift in time. It was found that the maximum accuracy obtained when using GFCC with CNN is 99.77 %. The outcome results of this work are compared to previous reports and indicate that CNN achieved better performance in AASR.

This paper extends an earlier work on designing a speech recognition system based on Hidden Markov Model (HMM) classification technique of using visual modality in addition to audio modality[1]. Improved off traditional HMM-based Automatic Speech Recognition (ASR) accuracy is achieved by implementing a technique using either RNN-based or CNN-based approach. This research is intending to deliver two contributions: The first contribution is the methodology of choosing the visual features by comparing different visual features extraction methods like Discrete Cosine Transform (DCT), blocked DCT, and Histograms of Oriented Gradients with Local Binary Patterns (HOG+LBP), and applying different dimension reduction techniques like Principal Component Analysis (PCA), auto-encoder, Linear Discriminant Analysis (LDA), t-distributed Stochastic Neighbor Embedding (t-SNE) to find the most effective features vector size. Then the obtained visual features are early integrated with the audio features obtained by using Mel Frequency Cepstral Coefficients (MFCCs) and feed the combined audio-visual feature vector to the classification process. The second contribution of this research is the methodology of developing the classification process using deep learning by comparing different Deep Neural Network (DNN) architectures like Bidirectional Long-Short Term Memory (BiLSTM) and Convolution Neural Network (CNN) with the traditional HMM. The proposed model is evaluated on two multi-speakers AV-ASR datasets named AVletters and GRID with different SNR. The model performs speaker-independent experiments in AVlettter dataset and speaker-dependent in GRID dataset.

Abstract: This paper presents a novel method of classifying speech phonemes. Four hybrid techniques based on the acoustic-phonetic approach and pattern recognition approach are used to emphasize the principle idea of this research. The first hybrid model is constructed of fixed state, structured Hidden Markov Model, Gaussian Mixture, Mel scaled Best Tree Image, Convolution Neural network, Vector Quantization (FS-HMM-GM-MBTI-CNN-VQ). The second hybrid model is constructed of variable state, dynamically structured Hidden Markov Model, Gaussian Mixture, Mel scaled Best Tree Image, Convolution Neural network, Vector Quantization (VS-HMM-GM-MBTI-CNN-VQ). The third hybrid model is constructed of fixed state, structured Hidden Markov Model, Gaussian Mixture, Mel scaled Best Tree Image, Convolution Neural network (FS-HMM-GM-MBTI-CNN). The fourth hybrid model is constructed of variable state, dynamically structured Hidden Markov Model, Gaussian Mixture, Mel scaled Best Tree Image, Convolution Neural network (VS-HMM-GM-MBTI-CNN). TIMIT database is used in this paper. All phones are classified into five classes and segregated into Vowels, Plosives, Fricatives, Nasals, and Silences. The results show that using (VS-HMM-GM-MBTI-CNN-VQ) is an available method for classification of phonemes, with the potential for use in applications such as automatic speech recognition and automatic language identification. Competitive results are achieved especially in nasals, plosives, and silence high successive rates than others.

Language and psychology have much in common in the sense that each of these two disciplines can be used to study thefeatures of the other one. Words that people use every day carry a lot of explicit linguistic features which signal some implicit psychological traits, personal characteristics, social relations and cognitive processes. Since more attention has been recently paidto such an area of research, this study investigates linguistic and psychological features used as indicators of veracity or deception. The study examines selected speeches of the last five Egyptian and American presidents using a computerized content analysis tool called LIWC (Linguistic Inquiry and Word Count). Drawing on previous researches that examine the psychometrics of the language, this study follows an eclectic approach adopting Newman et al.’s model combined with other cues to deception concluded by other studies. The analysis detects the frequency of using pronouns, negation, exclusive words, details, conjunctions and big words in the speeches in order to reveal instances of deception.

Drawing on the classified prototypes of semantic categories, this paper uses the principles of the prototype theoryto cross-culturally explore the hierarchical prototypes of crimes in web-booted Arabic and English corpora. The studycompares the conceptualization and categorization of crimes in the Arabic and English worlds. For doing so, four domains‘com, org, info and edu’ are explored over the past seven years for differentiating the individual mentality from theinstitutional mentalities. The study uses the web as corpus to investigate the inductive pattern ‘crime* such as’ in three web domains: ‘.com’, ‘.org’ and ‘.edu’ in Arabic and English. The collected data is analyzed using AntConc software program. Indicated statistical tests are calculated to measure the universality of conceptualizing emotions across Arabic and English speaking worlds. The study explores the validity of using Rosch’s prototype theory (1975) in determining the dynamically changing categorization of criminal acts. It also tests the authenticity of using Web-as-Corpus, which is a verystraightforward tool, in attaining so. Results reveal that there is an Arabic-English agreement on defining the basic levelsof the concept crime. However, few sporadic differences exist and are subject of cultural differences between the orientaland occidental moralities. The principles of prototype theory are also valid as regards the metamorphosis of conceptualizing a concept at the folkloric and expert levels. This holds true as regards the Arabic and English data of this study.

منطلقاً مهما في هذا البحث؛ إذ خلَصَت إلى عدد من القواعد التي تحکم تآلف الأصوات العربية وتنافرها، وهو ما يشير إلى أن وراء السلوک اللغوي التلقائي للعربية نسقا ضمنيا يحدد النماذج المقبولة وغير المقبولة، وهذا ما أتاح للعربي القديم أن يقبل منها ما هو جدير بالقبول، ويعرض عمّا سواه.