Character Level Research Articles

Holistic Persian Handwritten Word Recognition Using Convolutional Neural Network

Due to the cursive-ness and high variability of Persian scripts, the segmentation of handwritten words into sub-words is still a challenging task. These issues could be addressed in a holistic approach by sidestepping segmentation at the character level. In this paper, an end-to-end holistic method based on deep convolutional neural network is proposed to recognize off-line Persian handwritten words. The proposed model uses only five convolutional layers and two fully connected layers for classifying word images effectively, which can lead to a substantial reduction in parameters. The effect of various pooling strategies is also investigated in this paper. The primary goal of this article is to ignore handcrafted feature extraction and attain a generalized and stable word recognition system. The presented model is assessed using two famous handwritten Persian word databases called Sadri and IRANSHAHR. The recognition accuracies were obtained at 98.6% and 94.6%, on Sadri and IRANSHAHR datasets respectively, and outperformed the state-of-the-art methods.

A study of the performance of embedding methods for Arabic short-text sentiment analysis using deep learning approaches

Sentiment analysis aims to classify a text according to sentimental polarities of people’s opinions, such as positive, negative, or neutral. While most of the studies focus on eliciting features from English text, the research on Arabic is limited due to the morphological and grammatical complexity of Arabic language. In this paper, we proposed an optimized sentiment classification for dialectal Arabic short text at the document level using deep learning (DL). The contributions of this paper are in three areas. First, we extracted semantic features for Arabic short text at the word level and character level. Second, we used three DL topologies for classification models: a long short-term memory recurrent neural network (LSTM); a convolutional neural network (CNN); and an ensemble model combining both models’ advantages to improve the prediction performance. Third, we used a hyperparameter tuning estimation method to optimize the neural network performance. We trained and tested our proposed models on a dataset that consists of Modern Standard Arabic and dialectal Arabic corpus collected from Twitter. The results showed significant improvement in Arabic text classification in term of classification accuracy that ranges between 88% and 69.7%. The ensemble model scored the highest accuracy of 96.7% on the test set.

ArCAR: A Novel Deep Learning Computer-Aided Recognition for Character-Level Arabic Text Representation and Recognition

Arabic text classification is a process to simultaneously categorize the different contextual Arabic contents into a proper category. In this paper, a novel deep learning Arabic text computer-aided recognition (ArCAR) is proposed to represent and recognize Arabic text at the character level. The input Arabic text is quantized in the form of 1D vectors for each Arabic character to represent a 2D array for the ArCAR system. The ArCAR system is validated over 5-fold cross-validation tests for two applications: Arabic text document classification and Arabic sentiment analysis. For document classification, the ArCAR system achieves the best performance using the Alarabiya-balance dataset in terms of overall accuracy, recall, precision, and F1-score by 97.76%, 94.08%, 94.16%, and 94.09%, respectively. Meanwhile, the ArCAR performs well for Arabic sentiment analysis, achieving the best performance using the hotel Arabic reviews dataset (HARD) balance dataset in terms of overall accuracy and F1-score by 93.58% and 93.23%, respectively. The proposed ArCAR seems to provide a practical solution for accurate Arabic text representation, understanding, and classification.

Cyberbullying detection from tweets using deep learning

PurposeWith the rise of social media platforms, an increasing number of cases of cyberbullying has reemerged. Every day, large number of people, especially teenagers, become the victim of cyber abuse. A cyberbullied person can have a long-lasting impact on his mind. Due to it, the victim may develop social anxiety, engage in self-harm, go into depression or in the extreme cases, it may lead to suicide. This paper aims to evaluate various techniques to automatically detect cyberbullying from tweets by using machine learning and deep learning approaches.Design/methodology/approachThe authors applied machine learning algorithms approach and after analyzing the experimental results, the authors postulated that deep learning algorithms perform better for the task. Word-embedding techniques were used for word representation for our model training. Pre-trained embedding GloVe was used to generate word embedding. Different versions of GloVe were used and their performance was compared. Bi-directional long short-term memory (BLSTM) was used for classification.FindingsThe dataset contains 35,787 labeled tweets. The GloVe840 word embedding technique along with BLSTM provided the best results on the dataset with an accuracy, precision and F1 measure of 92.60%, 96.60% and 94.20%, respectively.Research limitations/implicationsIf a word is not present in pre-trained embedding (GloVe), it may be given a random vector representation that may not correspond to the actual meaning of the word. It means that if a word is out of vocabulary (OOV) then it may not be represented suitably which can affect the detection of cyberbullying tweets. The problem may be rectified through the use of character level embedding of words.Practical implicationsThe findings of the work may inspire entrepreneurs to leverage the proposed approach to build deployable systems to detect cyberbullying in different contexts such as workplace, school, etc and may also draw the attention of lawmakers and policymakers to create systemic tools to tackle the ills of cyberbullying.Social implicationsCyberbullying, if effectively detected may save the victims from various psychological problems which, in turn, may lead society to a healthier and more productive life.Originality/valueThe proposed method produced results that outperform the state-of-the-art approaches in detecting cyberbullying from tweets. It uses a large dataset, created by intelligently merging two publicly available datasets. Further, a comprehensive evaluation of the proposed methodology has been presented.

Think Twice: A Post-Processing Approach for the Chinese Spelling Error Correction

Chinese Spelling Error Correction is a hot subject in the field of natural language processing. Researchers have already produced many great solutions, from the initial rule-based solution to the current deep learning method. At present, SpellGCN, proposed by Alibaba’s team, achieves the best results of which character level precision over SIGHAN2013 is 98.4%. However, when we apply this algorithm to practical error correction tasks, it produces many false error correction results. We believe that this is because the corpus used for model training contains significantly more errors than the text used for model correcting. In response to this problem, we propose performing a post-processing operation on the error correction tasks. We employ the initial model’s output as a candidate character, obtain various features of the character itself and its context, and then use a classification model to filter the initial model’s false error correction results. The post-processing idea introduced in this paper can apply to most Chinese Spelling Error Correction models to improve their performance over practical error correction tasks.

A two-phase gradient based feature embedding approach

Over the years, researchers attempted to maximize the capacity, imperceptibility, and robustness of Steganography algorithms. Nowadays, the challenge is not only to maximize these parameters rather identify new ideas toward the security aspect. Keeping this requirement in mind, in the present scope of the work, we introduce a novel Steganographic approach mingled with optical character recognition (OCR) techniques to maximize the security against intruders who aim to extract the hidden data. We introduce two novel feature descriptors, the modified shadow and percentage of light source (PLS) feature specifically keeping in mind the embedding requirements. We train our model on a standard publicly available dataset, also reasoning the idea behind doing so. In our embedding method, we vary the quantity of the embedded message for each pixel, based on the gradient value of that pixel, to ensure the least distortion which human beings cannot differentiate. Here, the message image (text image) is not embedded in its entirety, rather character level features, extracted from the text message, are then embedded in the cover image. This very phase ensures that even if the intruder has the access to the stego image, s/he cannot extract the data as it is not having the raw data rather some features. Even the feature values alone would not give the notion of a secret image as the intruder requires the trained OCR model, which the receiver has, to recognize the text. Furthermore, for every character image extracted from the text, we choose between two candidate feature sets and embed the one which causes the least distortion. The feature extraction step generates high recognition accuracy, which in turn completes the extraction process. The security model goes on to show good results in terms of the Steganography metrics. We also perform a fusion of standard robustness and statistical tests, whereby our embedded image successfully performs well.

A Baybayin word recognition system.

Baybayin is a pre-Hispanic Philippine writing system used in Luzon island. With the effort in reintroducing the script, in 2018, the Committee on Basic Education and Culture of the Philippine Congress approved House Bill 1022 or the ”National Writing System Act,” which declares the Baybayin script as the Philippines’ national writing system. Since then, Baybayin OCR has become a field of research interest. Numerous works have proposed different techniques in recognizing Baybayin scripts. However, all those studies anchored on the classification and recognition at the character level. In this work, we propose an algorithm that provides the Latin transliteration of a Baybayin word in an image. The proposed system relies on a Baybayin character classifier generated using the Support Vector Machine (SVM). The method involves isolation of each Baybayin character, then classifying each character according to its equivalent syllable in Latin script, and finally concatenate each result to form the transliterated word. The system was tested using a novel dataset of Baybayin word images and achieved a competitive 97.9% recognition accuracy. Based on our review of the literature, this is the first work that recognizes Baybayin scripts at the word level. The proposed system can be used in automated transliterations of Baybayin texts transcribed in old books, tattoos, signage, graphic designs, and documents, among others.

Detection and Identification of Obfuscated Obscene Language with Character Level Transformers

We present O-Norm (Offense Normalizer), a system for normalizing offensive words and phrases by reversing user-created obfuscations. O-Norm is intended to be a preprocessing tool for pipelines that run on user-created text, particularly those tasks where obscenities may be of significant value such as sentiment, toxicity, and abuse detection. O-Norm is constructed with a purely generated, context-free dataset derived from a curse dictionary. The generative dataset allows O-Norm to be flexible to the addition of new words as language adapts, and also allows it to be readily retrained in other Latin alphabet-based languages as it does not require manual annotations to train. O-Norm is based on a character-level, transformer network which attempts de-obfuscation only on out-of-vocabulary (OOV) tokens. In an 80/20 train-test split, O-Norm achieves an F1 score of 89.6% over 141 curses in a generated dataset with 2.16 million unique training points. An inspection of O-Norm’s output on a sample of social media posts from Kaggle’s Jigsaw corpus reveals accuracy of 95.7% on de-obfuscating transformations in toxic user-created text.

Predicting Polymers' Glass Transition Temperature by a Chemical Language Processing Model.

We propose a chemical language processing model to predict polymers’ glass transition temperature () through a polymer language (SMILES, Simplified Molecular Input Line Entry System) embedding and recurrent neural network. This model only receives the SMILES strings of a polymer’s repeat units as inputs and considers the SMILES strings as sequential data at the character level. Using this method, there is no need to calculate any additional molecular descriptors or fingerprints of polymers, and thereby, being very computationally efficient. More importantly, it avoids the difficulties to generate molecular descriptors for repeat units containing polymerization point ‘*’. Results show that the trained model demonstrates reasonable prediction performance on unseen polymer’s . Besides, this model is further applied for high-throughput screening on an unlabeled polymer database to identify high-temperature polymers that are desired for applications in extreme environments. Our work demonstrates that the SMILES strings of polymer repeat units can be used as an effective feature representation to develop a chemical language processing model for predictions of polymer . The framework of this model is general and can be used to construct structure–property relationships for other polymer properties.

Recognition of Devanagari Scene Text Using Autoencoder CNN

Scene text recognition is a well-rooted research domain covering a diverse application area. Recognition of scene text is challenging due to the complex nature of scene images. Various structural characteristics of the script also influence the recognition process. Text and background segmentation is a mandatory step in the scene text recognition process. A text recognition system produces the most accurate results if the structural and contextual information is preserved by the segmentation technique. Therefore, an attempt is made here to develop a robust foreground/background segmentation(separation) technique that produces the highest recognition results. A ground-truth dataset containing Devanagari scene text images is prepared for the experimentation. An encoder-decoder convolutional neural network model is used for text/background segmentation. The model is trained with Devanagari scene text images for pixel-wise classification of text and background. The segmented text is then recognized using an existing OCR engine (Tesseract). The word and character level recognition rates are computed and compared with other existing segmentation techniques to establish the effectiveness of the proposed technique.

In Codice Ratio: A crowd-enabled solution for low resource machine transcription of the Vatican Registers

In Codice Ratio is a research project to study techniques for analyzing the contents of historical documents conserved in the Vatican Apostolic Archives. In this paper, we present our efforts to develop a system to support the automatic transcription of medieval manuscripts, while maintaining the training data collection effort minimal. We focus on crowdsourcing as a means for scalable, expertless training data collection: using crowdsourced character symbols, we train a custom convolutional neural network able to jointly learn correct character shape identification and character recognition. Our approach generates candidate transcriptions by submitting over-segmented character strokes and their combinations to this classifier, while ranking and choosing the most promising outputs by combining the recognition confidence with character and word level statistical language models.We conducted experiments on an unreleased corpus, the Vatican Registers: training our system on 20 pages annotated by the crowd, we were able to obtain good results (19% CER); comparisons to an off-the-shelf system trained with 20 pages annotated with the same process, and to a professional system trained with more than 300 pages transcribed by skilled paleographers demonstrate the opportunities of the proposed approach.

Neural OCR Post-Hoc Correction of Historical Corpora

Abstract Optical character recognition (OCR) is crucial for a deeper access to historical collections. OCR needs to account for orthographic variations, typefaces, or language evolution (i.e., new letters, word spellings), as the main source of character, word, or word segmentation transcription errors. For digital corpora of historical prints, the errors are further exacerbated due to low scan quality and lack of language standardization. For the task of OCR post-hoc correction, we propose a neural approach based on a combination of recurrent (RNN) and deep convolutional network (ConvNet) to correct OCR transcription errors. At character level we flexibly capture errors, and decode the corrected output based on a novel attention mechanism. Accounting for the input and output similarity, we propose a new loss function that rewards the model’s correcting behavior. Evaluation on a historical book corpus in German language shows that our models are robust in capturing diverse OCR transcription errors and reduce the word error rate of 32.3% by more than 89%.

Automatic Integrated Scoring Model for English Composition Oriented to Part-Of-Speech Tagging

Part-of-speech tagging for English composition is the basis for automatic correction of English composition. The performance of the part-of-speech tagging system directly affects the performance of the marking and analysis of the correction system. Therefore, this paper proposes an automatic scoring model for English composition based on article part-of-speech tagging. First, use the convolutional neural network to extract the word information from the character level and use this part of the information in the coarse-grained learning layer. Secondly, the word-level vector is introduced, and the residual network is used to establish an information path to integrate the coarse-grained annotation and word vector information. Then, the model relies on the recurrent neural network to extract the overall information of the sequence data to obtain accurate annotation results. Then, the features of the text content are extracted, and the automatic scoring model of English composition is constructed by means of model fusion. Finally, this paper uses the English composition scoring competition data set on the international data mining competition platform Kaggle to verify the effect of the model.

STDNet: A CNN-based approach to single-/mixed-script detection

Script identification serves as a guide to the detection of the text of the scene through optical character recognition (OCR). But this is not a principal concern for the OCR engine. Until script identification, it is important to identify the script-type because today the text of the scene in natural images does not consist only of a single script, rather mixed-script words at character level are very often encountered. These words are also used in various ways, such as signboards, t-shirt graffiti, hoardings, and banners and often written in artistic way. In this work, a CNN-based deep learning framework, named as STDNet: Script-Type detection Network, was developed to detect single-/mixed-script images. To determine the feasibility of the system presented, tests were also undertaken with an outlier which is composed of a wide range of single scripts. Experiments were performed with over 20K images and 99.53% highest accuracy was reached. This approach was compared to a state-of-the-art deep learning techniques and handcrafted feature-based methodologies where the proposed approach obtained a better performance.

Quantifying the efficiency of written language

Abstract Information theory can be used to assess how efficiently a message is transmitted on the basis of different symbolic systems. In this paper, I estimate the information-theoretic efficiency of written language for parallel text data in more than 1000 different languages, both on the level of characters and on the level of words as information encoding units. The main results show that (i) the median efficiency is ∼29% on the character level and ∼45% on the word level, (ii) efficiency on both levels is strongly correlated with each other and (iii) efficiency tends to be higher for languages with more speakers.

Position-Aware Deep Character-Level CTR Prediction for Sponsored Search

Predicting the click-through rate of an advertisement is a critical component of online advertising platforms. In sponsored search, the click-through rate estimates the probability that a displayed advertisement is clicked by a user after she submits a query to the search engine. Commercial search engines typically rely on machine learning models trained with a large number of features to make such predictions. This inevitably requires a lot of engineering efforts to define, compute, and select the appropriate features. In this paper, we propose two novel approaches (one working at character level and the other working at word level) that use deep convolutional neural networks to predict the click-through rate of a query-advertisement pair. Specifically, the proposed architectures consider as input only the textual content appearing in a query-advertisement pair and the page position at which the advertisement appears on the search result page of the query, and produce as output a click-through rate prediction. By comparing the character-level model with the word-level model, we show that language representation can be learnt from scratch at character level when trained on enough data. Through extensive experiments using billions of query-advertisement pairs of a popular commercial search engine, we demonstrate that both approaches significantly outperform a baseline model built on well-selected text features and a state-of-the-art word2vec-based approach. We also show the importance of the position feature in the proposed approaches in improving the prediction accuracy. When combining the predictions of the deep models introduced in this study with the prediction of the model in production of the same commercial search engine, we significantly improve the accuracy and the calibration of the click-through rate prediction of the production system. We also show the potential of leveraging the CTR prediction of the proposed deep learning models for query-ad relevance modeling and query-ad matching tasks in sponsored search.

Text Detection in Trademark Images under Semi-Supervision

The texts in trademark image usually contain characters, text lines and other different parts. Therefore, compared with the detectors that can detect text lines directly, the detector that can detect characters is more suitable for the detection of texts in trademark images However, training character detectors needs a huge number of characters with position annotation, the con-struction process of this type of dataset will take a lot of time. In order to overcome the limitations of the detectors trained by datasets with word-level annotations and the shortage of datasets with character-level annotations, we propose a text detect method that can detect the single character in trademark images. Our proposed model is pretrained by the synthetic image dataset with character level annotation, then the pretrained model is used to detect the unannotated trademark images dataset to find more annotations for retrain the pretrained model. We also construct a trademark text dataset, which includes 4500 images. Experimental results on our dataset demonstrate our method can achieve the state-of-the-art performance compared to other methods. The results show our detector can detect complicated texts in trademark images.

Mining the web to discover acronym‐definitions based on sequence labeling and iterative query expansion model

SummaryFinding definitions associated with an acronym is a constructive task to many applications like web search, information retrieval, natural language processing, ontology mapping, and question answering. A large‐scale manually built acronym‐definition repositories are available online but updating is a difficult task. To address this problem, previous research works used either heuristics or a machine learning approach to automate the detection of acronym‐definition pairs. This article presents a heuristics approach based on a rule‐based sequence‐labeling model, for finding the list of definitions associated with an acronym from the Web. The organic search result of the web page includes website title, URL, Meta description, and site links. In the proposed work, web pages of the search engine are used as corpus and sequence‐labeling task is done through character and word level mapping schemes. In these schemes, the desirable properties of the definition are expressed by using rules. Each identified definition is validated through a collocation measure. Besides, the obtained results are assessed against a manually built acronym/definitions repository: Acronym Finder. The proposed model certainly improves the coverage of manual repositories maintaining high precision and recall.

Development of Chemistry Practicum Guidelines with the Support of STEM (Science, Technology, Engineering, and Mathematics) Integrating Character Education

The implementation of the STEM-based learning approach is able to shape student character. One of them is using a chemistry lab guide based on the STEM approach. The purpose of this study was to determine the feasibility level of a chemistry lab guide based on the integrated STEM approach to character education based on the criteria of the National Education Standards Agency (BSNP) and the character level of students after using a validated practicum guide. This type of research uses the Research and Development (R & D) method with the ADDIE model development stages (Analysis, Design, Development, Implementation, and Evaluation). The data collection instrument consisted of a practicum guide validation sheet based on BSNP criteria and student character observation sheets. The data source in this study was 40 students of class XII SMA N 1 Tanjungbalai academic year 2020/2021 using purposive sampling technique of collecting samples. The results showed that the validation test for chemistry lab guides by expert validators stated that the chemistry practicum guidance based on the integrated STEM approach to character education was valid with an average value of 3.53 from a maximum scale of 4 which means it is feasible and does not need to be revised. Furthermore, the results of observations of the character level of students using a practical guide based on the STEM integrated approach to character education are very good with an average of 85.8%. Based on the results of character validation and observation, it shows that chemistry practicum speakers based on the STEM integrated approach to character education are feasible to be implemented in schools.

Tri-level handwritten text segmentation techniques for Gujarati language

Objectives: To improve the efficiency of tri-level segmentation tasks for handwritten Gujarati text. Methods: Using hybrid methods for tri-level segmentation, we have used line, word and character segmentation from the image. This study presents a segmentation paradigm that works with touching characters, slop of the line written on the page, character overlapping, etc. It evaluated on the dataset of 500+ images created by us on different writing sentences by different people. We have used the Horizontal projection technique for line segmentation, Scale-space technique for word segmentation and the Vertical projection technique for character segmentation. Findings: The experimental results show that the proposed method is more efficient for handwritten Gujarati text with diacritics. We have obtained the accuracy for character level segmentation is 82%, word-level is 90% and for the line-level segmentation is 87%. Novelty: We have designed a methodology to segment Gujarati handwritten text with diacritics at all three levels including characters, words and lines. Applications: We have proposed tri-level segmentation which is pre-processing task that can be used in any character recognition systems i.e. OCR. Keywords: Deep learning; trilevel segmentation; handwritten Gujarati text