Input Text Research Articles

RB-TRNet: a regularity-guided and boundary-aware architecture for toponym recognition from Chinese text

ABSTRACT Extracting geographic information from texts contributes to both geographic information science research and various practical applications, but extracting fine-grained and complex location descriptions from Chinese text is still challenging, due to flexible word construction and lack of clear boundaries in Chinese place names. In this paper, we propose a regularity-guided and boundary-aware architecture for toponym recognition from Chinese text (RB-TRNet), achieving complex place name recognition by learning the internal compositional patterns of various place name constructions and automatically perceiving the boundaries and types of Chinese place name entities. First, RoBERTa is used to represent the input text containing Chinese place names. Then, two BiLSTM layers are fed with text representation sequences, with one processed sequence entering the toponym regularity-guided module to obtain the composition patterns of Chinese place name entities and the other sequence entering the toponym regularity-discriminant module to soften an excessive reliance on contextual information for recognizing patterns of Chinese place name entities. Additionally, an orthogonal space is established after the BiLSTM network to facilitate the learning of different rule features by the two modules. Finally, after joint optimization training of the three modules, the toponym regularity perception module is used to predict the Chinese place name entities. To validate the results, we established a new complex Chinese place name text (CCPNT) dataset for complex Chinese place name recognition. The CCPNT dataset, along with three other public datasets, were used for performance evaluation, and compared to eight baseline models, RB-TRNet exhibited state-of-the-art performance in recognizing complex Chinese place names.

An Adaptive Multimodal Fusion Network Based on Multilinear Gradients for Visual Question Answering

As an interdisciplinary field of natural language processing and computer vision, Visual Question Answering (VQA) has emerged as a prominent research focus in artificial intelligence. The core of the VQA task is to combine natural language understanding and image analysis to infer answers by extracting meaningful features from textual and visual inputs. However, most current models struggle to fully capture the deep semantic relationships between images and text owing to their limited capacity to comprehend feature interactions, which constrains their performance. To address these challenges, this paper proposes an innovative Trilinear Multigranularity and Multimodal Adaptive Fusion algorithm (TriMMF) that is designed to improve the efficiency of multimodal feature extraction and fusion in VQA tasks. Specifically, the TriMMF consists of three key modules: (1) an Answer Generation Module, which generates candidate answers by extracting fused features and leveraging question features to focus on critical regions within the image; (2) a Fine-grained and Coarse-grained Interaction Module, which achieves multimodal interaction between question and image features at different granularities and incorporates implicit answer information to capture complex multimodal correlations; and (3) an Adaptive Weight Fusion Module, which selectively integrates coarse-grained and fine-grained interaction features based on task requirements, thereby enhancing the model’s robustness and generalization capability. Experimental results demonstrate that the proposed TriMMF significantly outperforms existing methods on the VQA v1.0 and VQA v2.0 datasets, achieving state-of-the-art performance in question–answer accuracy. These findings indicate that the TriMMF effectively captures the deep semantic associations between images and text. The proposed approach provides new insights into multimodal interaction and fusion research, combining domain adaptation techniques to address a broader range of cross-domain visual question answering tasks.

HARNESSING CORPUS LINGUISTICS AND DATA-DRIVEN LEARNING APPROACH FOR ARABIC MULTI-CLASS DIALECT DETECTION AND CLASSIFICATION

Arabic dialect identification (ADI) is a specific task of natural language processing (NLP) that intends to forecast the Arabic language dialect of the input text automatically. ADI is the preliminary step toward establishing many NLP applications, including cross-language text generation, multilingual text-to-speech synthesis, and machine translation. The automatic classification of the Arabic dialect is the first step in various dialect-sensitive Arabic NLP tasks. ADI includes predicting the dialects related to the textual input and classifying them on their respective labels. As a result, increased interest has been gained in the last few decades to address the problems of ADI through deep learning (DL) and machine learning (ML) algorithms. The study develops an Arabic multi-class dialect recognition using fast random opposition-based fractals learning aquila optimizer with DL (FROBLAO-DL) technique. The FROBLAO-DL technique utilizes the optimal DL model to identify distinct types of Arabic dialects. In the FROBLAO-DL technique, data preprocessing is involved in cleaning the input Arabic dialect dataset. In addition, the ROBERTa word embedding process is used to generate word embedding. The FROBLAO-DL technique uses attention bidirectional long short-term memory (ABiLSTM) network to identify distinct Arabic dialects. Moreover, the ABiLSTM model’s hyperparameter tuning is implemented using the FROBLOA method. The performance evaluation of the FROBLAO-DL method is tested under the Arabic dialect dataset. The empirical analysis implies the supremacy of the FROBLAO-DL technique over recent approaches under various measures.

An LLM‐Based Method for Quality Information Extraction From Web Text for Crowed‐Sensing Spatiotemporal Data

ABSTRACTKnowledge‐driven GIS increasingly requires multi‐source, multi‐type, and multi‐model crowd‐sensing spatiotemporal data, whose data quality is difficult to guarantee and determine. Hence, extracting quality indicator information, widely present in various unstructured web texts, is crucial to providing supplementary quality information for crowd‐sensing spatiotemporal data. Recent advances in large language models show potential in extracting quality indicator information. However, it is still hard to get accurate results from large language models that use different quality indicators for crowd‐sensing spatiotemporal data. Therefore, we have designed a large language model that is fine‐tuned for the extraction of spatiotemporal quality information from quality description text (LLMFT‐STQIE). Firstly, we establish a quality indicator vocabulary to determine whether the text includes quality indicator information from the spatiotemporal data. Then, we create a two‐stage prompt model with QILE and QIVE prompts that include input text, task type, instructions, the quality indicator vocabulary, output format, and a reference case. This model is based on the fine‐tuning technology of large language models. The results show that our LLMFT‐STQIE achieves an accuracy of 91% and a recall rate of 80%, respectively, representing improvements of 23% and 38% compared to untuned large language models. These results further show that the suggested method easily and accurately extracts quality indicator information from web texts for crowd‐sensing spatiotemporal data. The study helps investigate strategies for optimizing huge language models for specific scenarios or task specifications.

AI-Based Hate Speech Detection in Albanian Social Media: New Dataset and Mobile Web Application Integration

This paper aims to advance AI-based hate speech (HS) detection in the Albanian language, which is resource-limited in natural language processing (NLP). Addressing the challenge of limited data, we developed a human-annotated dataset of over 11,000 comments, carefully curated from various Albanian social media platforms, containing a substantial number of HS instances. The dataset was annotated using a detailed two-layer taxonomy to capture the complex dimensions of HS. To ensure high-quality annotations, three expert annotators applied a majority voting system, achieving a substantial Fleiss’s kappa coefficient of 0.62, underscoring the reliability and consistency of the annotations. We conducted a comparative analysis of several machine learning (ML) algorithms, including support vector machine (SVM), Naïve Bayes (NB), XGBoost, and random forest (RF), paired with various text vectorisation techniques and pre-processing methods. In binary classification, the NB model with term frequencyinverse document frequency (TF-IDF) vectorization achieved the highest performance, with an F1 score of 0.80. For multiclass classification, XGBoost outperformed other models, achieving an F1 score of 0.77. Interestingly, our experiments revealed that pre-processing steps generally reduced model performance, suggesting that raw text inputs work better for the Albanian language. Through error analysis using local interpretable model-agnostic explanations (LIME), we identified key challenges, such as polysemy and irony, which contributed to misclassifications. To demonstrate the practical applicability of our work, we developed a user-friendly mobile web application based on the best-performing model, providing realtime HS detection with the potential for integration into social media platforms.

Using ChatGPT and Other AI Engines to Vocalize Medieval Hebrew

Hebrew is usually written without vowel points, making it challenging for some readers to decipher. This is especially true of medieval Hebrew, which can have nonstandard grammar and orthography. This paper tested four artificial intelligence (AI) tools by asking them to add vowel points to an unpublished medieval Hebrew translation of the Lord’s Prayer. The vocalization tools tested were OpenAI’s ChatGPT-3.5 and ChatGPT-4, Pellaworks’ DoItInHebrew, and Dicta’s Nakdan. ChatGPT-3.5 freely changed the text, even rewriting some phrases and adding an entire sentence. ChatGPT-3.5 also provided erroneous vowels in its rewritten Hebrew text. ChatGPT-4 did a moderately good job with only a few errors, but also modified the orthography. One of ChatGPT-4’s errors was not trivial, resulting in the invention of a word. When challenged, ChatGPT-4 corrected this confabulation by inventing another word, which it claimed was a “rare form” for which it provided a fictitious derivation. When challenged on this second made-up word, ChatGPT-4 replaced the word from the input text with a word based on an entirely different root. DoItInHebrew inserted vowels that produced a gibberish text. In contrast, Dicta’s Nakdan provided near perfect vocalization, with only one genuine error, but like ChatGPT-4 it modified the orthography. ChatGPT-3.5, ChatGPT-4, and DoItInHebrew exhibited serious “hallucinations,” of both the “factual” and the “untruthful” varieties, typical of other AIs, making them counterproductive for vocalizing historic Hebrew texts. Nakdan can be a powerful tool but still requires someone with expertise in Hebrew grammar to verify and correct the vocalization. Nakdan’s interface simplified correcting the vocalization, although it required its user to have advanced knowledge of Hebrew.

Efficient DNA Cryptography Using One-Time Pad and Run-Length Encoding for Optimized Ciphertext Storage

Cryptography ensures data privacy by transforming data into unreadable formats that only authorized individuals can decrypt. With the increase in electronically stored and transmitted data, enhanced methods for data protection are required. DNA cryptography, leveraging the genetic structure of DNA, provides a promising approach for secure communication and data storage. This paper introduces a novel DNA-based cryptographic method employing a DNA one-time pad (OTP) combined with modified run-length encoding to reduce ciphertext size. Unlike traditional cryptography, which often results in a larger ciphertext than plaintext, our proposed method demonstrates a significant reduction in ciphertext size. Experimental results reveal that for input text files of 1MB, 2MB, 3MB, 5MB, and 10MB, the ciphertext sizes were reduced by up to 20KB, 40KB, 60KB, 100KB, and 200KB, respectively. This reduction not only enhances storage efficiency but also minimizes transmission costs, marking a substantial advancement over existing DNA and classical cryptography methods. Future work will explore the application of this technique for encrypting biological data and incorporating DNA barcoding for improved data authentication and reliability.

Comparing human-labeled and AI-labeled speech datasets for TTS

As the output quality of neural networks in the fields of automatic speech recognition (ASR) and text-to-speech (TTS) continues to improve, new opportunities are becoming available to train models in a weakly supervised fashion, thus minimizing the manual effort required to annotate new audio data for supervised training. While weak supervision has recently shown very promising results in the domain of ASR, speech synthesis has not yet been thoroughly investigated regarding this technique despite requiring the equivalent training dataset structure of aligned audio-transcript pairs. In this work, we compare the performance of TTS models trained using a well-curated and manually labeled training dataset to others trained on the same audio data with text labels generated using both grapheme- and phoneme-based ASR models. Phoneme-based approaches seem especially promising, since even for wrongly predicted phonemes, the resulting word is more likely to sound similar to the originally spoken word than for grapheme-based predictions. For evaluation and ranking, we generate synthesized audio outputs from all previously trained models using input texts sourced from a selection of speech recognition datasets covering a wide range of application domains. These synthesized outputs are subsequently fed into multiple state-of-the-art ASR models with their output text predictions being compared to the initial TTS model input texts. This comparison enables an objective assessment of the intelligibility of the audio outputs from all TTS models, by utilizing metrics like word error rate and character error rate. Our results not only show that models trained on data generated with weak supervision achieve comparable quality to models trained on manually labeled datasets, but can outperform the latter, even for small, well-curated speech datasets. These findings suggest that the future creation of labeled datasets for supervised training of TTS models may not require any manual annotation but can be fully automated.

Multimodal Contrastive Learning for Remote Sensing Image Feature Extraction Based on Relaxed Positive Samples.

Traditional multimodal contrastive learning brings text and its corresponding image closer together as a positive pair, where the text typically consists of fixed sentence structures or specific descriptive statements, and the image features are generally global features (with some fine-grained work using local features). Similar to unimodal self-supervised contrastive learning, this approach can be seen as enforcing a strict identity constraint in a multimodal context. However, due to the inherent complexity of remote sensing images, which cannot be easily described in a single sentence, and the fact that remote sensing images contain rich ancillary information beyond just object features, this strict identity constraint may be insufficient. To fully leverage the characteristics of remote sensing images, we propose a multimodal contrastive learning method for remote sensing image feature extraction, based on positive sample tripartite relaxation, where the model is relaxed in three aspects. The first aspect of relaxation involves both the text and image inputs. By introducing learnable parameters in the language and image branches, instead of relying on fixed sentence structures and fixed image features, the network can achieve a more flexible description of remote sensing images in text and extract ancillary information from the image features, thereby relaxing the input constraints. Second relaxation is achieved through multimodal alignment of various features. By aligning semantic information with the corresponding semantic regions in the images, the method allows for the relaxation of local image features under semantic constraints. This approach addresses the issue of selecting image patches in unimodal settings, where there is no semantic constraint. The proposed method for remote sensing image feature extraction has been validated on four datasets. On the PatternNet dataset, it achieved a 91.1% accuracy with just one-shot.

Imaginique Expressions: Tailoring Personalized Short-Text-to-Image Generation Through Aesthetic Assessment and Human Insights

The text-to-image task, a critical branch of computer vision and image processing, has witnessed remarkable advancements fueled by the abundance of realistic data and rapid AI innovation. However, existing research often overlooks scenarios involving sparse textual input and fails to incorporate human personalized preferences into the generative process. To address these gaps, we propose a novel AI methodology: personalized short-text-to-image generation through aesthetic assessment and human insights. Our approach introduces a symmetry between personalized aesthetic preferences and the generated images by leveraging a data-driven personality encoder (PE) to extract personal information and embed it into a Big Five personality trait-based image aesthetic assessment (BFIAA) model. This model harmonizes aesthetic preferences with the generative process by adapting the stable diffusion framework to align with personalized assessments. Experimental results demonstrate the effectiveness of our method: the PE module achieves an accuracy of 98.1%, while the BFIAA model surpasses the baseline by 13% on the PLCC metric, accurately reflecting human aesthetic preferences. Furthermore, our adapted generation model improves convergence loss by over 10% compared to the base model, consistently producing personalized images that are more aligned with human preferences.

Designing a novel framework of email spam detection using an improved heuristic algorithm and dual-scale feature fusion-based adaptive convolution neural network

ABSTRACT Nowadays, e-mail is becoming more prevalent among individuals. In recent days, it has been declared to be the least expensive and quickest mode of communicating. In recent years, e-mail spam has become a big problem, so the number of e-mail spam has also increased, and they are used for unethical and illegal conduct, scams, and theft. As a result, the main objective of the study is to detect e-mail spam using an adaptive deep-learning model with feature fusion. Initially, the input image and text are collected from benchmark datasets. Further, the raw text undergoes the pre-processing stage, and it is then followed by extracting the features using Bidirectional Encoder Representations from Transformers (BERT). Similarly, the Vision Transformer (ViT) is employed for extracting the features from raw images. Finally, these features are given in Dual Scale Feature Fusion-based Adaptive Convolutional Neural Network (DFF-ACNNet), and it is then fused for performing the classification in Convolutional Neural Network (CNN). Here, the hyper-parameters are tuned by using Modified Squid Game Optimizer (MSGO). Finally, the evaluation is done with various metrics to estimate the performance of the model and the accuracy of the developed system is 98.46, which is higher than the other conventional approaches.

Comparing Large Language Model and Human Reader Accuracy with New England Journal of Medicine Image Challenge Case Image Inputs.

Background Application of multimodal large language models (LLMs) with both textual and visual capabilities has been steadily increasing, but their ability to interpret radiologic images is still doubted. Purpose To evaluate the accuracy of LLMs and compare it with that of human readers with varying levels of experience and to assess the factors affecting LLM accuracy in answering New England Journal of Medicine Image Challenge cases. Materials and Methods Radiologic images of cases from October 13, 2005, to April 18, 2024, were retrospectively reviewed. Using text and image inputs, LLMs (Open AI's GPT-4 Turbo with Vision [GPT-4V] and GPT-4 Omni [GPT-4o], Google's DeepMind Gemini 1.5 Pro, and Anthropic's Claude 3) provided answers. Human readers (seven junior faculty radiologists, two clinicians, one in-training radiologist, and one medical student), blinded to the published answers, also answered. LLM accuracy with and without image inputs and short (cases from 2005 to 2015) versus long text inputs (from 2016 to 2024) was evaluated in subgroup analysis to determine the effect of these factors. Factor analysis was assessed using multivariable logistic regression. Accuracy was compared with generalized estimating equations, with multiple comparisons adjusted by using Bonferroni correction. Results A total of 272 cases were included. GPT-4o achieved the highest overall accuracy among LLMs (59.6%; 162 of 272), outperforming a medical student (47.1%; 128 of 272; P < .001) but not junior faculty (80.9%; 220 of 272; P < .001) or the in-training radiologist (70.2%; 191 of 272; P = .003). GPT-4o exhibited similar accuracy regardless of image inputs (without images vs with images, 54.0% [147 of 272] vs 59.6% [162 of 272], respectively; P = .59). Human reader accuracy was unaffected by text length, whereas LLMs demonstrated higher accuracy with long text inputs (all P < .001). Text input length affected LLM accuracy (odds ratio range, 3.2 [95% CI: 1.9, 5.5] to 6.6 [95% CI: 3.7, 12.0]). Conclusion LLMs demonstrated substantial accuracy with text and image inputs, outperforming a medical student. However, their accuracy decreased with shorter text lengths, regardless of image input. © RSNA, 2024 Supplemental material is available for this article.

Large Language Models with Vision on Diagnostic Radiology Board Exam Style Questions

Rationale and ObjectivesThe expansion of large language models to process images offers new avenues for application in radiology. This study aims to assess the multimodal capabilities of contemporary large language models, which allow analysis of image inputs in addition to textual data, on radiology board style examination questions with images. Materials and Methods280 questions were retrospectively selected from the AuntMinnie public test bank. The test questions were converted into three formats of prompts; 1) Multimodal, 2) Image only, and 3) Text only input. Three models, GPT-4V, Gemini 1.5 Pro, and Claude 3.5 Sonnet, were evaluated using these prompts. The Cochran Q test and pairwise McNemar test were used to compare performances between prompt formats and models. ResultsNo difference was found for the performance in terms of % correct answers between the text, image, and multimodal prompt formats for GPT-4V (54%, 52%, and 57%, respectively; p=.31) and Gemini 1.5 Pro (53%, 54%, and 57%, respectively; p=.53). For Claude 3.5 Sonnet, the image input (48%) significantly underperformed compared to the text input (63%, p<.001) and the multimodal input (66%, p<.001), but no difference was found between the text and multimodal inputs (p=.29). Claude significantly outperformed GPT and Gemini in the text and multimodal formats (p<.01). ConclusionVision capable large language models cannot effectively use images to increase performance on radiology board-style examination questions. When using textual data alone, Claude 3.5 Sonnet outperforms GPT-4V and Gemini 1.5 Pro, highlighting the advancements in the field and its potential for the use in further research.

Using word cloud as a formative assessment classroom technique in physiology.

Formative assessment is a valuable tool in medical education as it allows students to reflect on their own learning and make necessary corrections. The use of innovative techniques in formative assessment can make it more engaging and effective for students, particularly utilizing electronic tools. A word cloud is a visual representation of data that is created using text inputs from a group of individuals. The resulting image shows the most frequently used words in a larger font, allowing quick understanding of the common themes and concepts in the group's responses. Mentimeter is an online platform for interactive presentations where students can provide their responses to a question, which can then be visualized as a word cloud. The study aims to explore the benefits and drawbacks of using word cloud formation as a formative assessment technique in the classroom, utilizing the online platform Mentimeter.NEW & NOTEWORTHY The current study was conducted to present a novel formative assessment classroom technique using Mentimeter. The benefit of this novel technique is that it is easy to implement, quick, anonymous, and strengthens student's conceptual learning. It also gives the teacher an opportunity to give feedback and stress commonly committed mistakes.

Towards Visual-Prompt Temporal Answer Grounding in Instructional Video.

Temporal answer grounding in instructional video (TAGV) is a new task naturally derived from temporal sentence grounding in general video (TSGV). Given an untrimmed instructional video and a text question, this task aims at locating the frame span from the video that can semantically answer the question, i.e., visual answer. Existing methods tend to solve the TAGV problem with a visual span-based predictor, taking visual information to predict the start and end frames in the video. However, due to the weak correlations between the semantic features of the textual question and visual answer, current methods using the visual span-based predictor do not work well in the TAGV task. In this paper, we propose a visual-prompt text span localization (VPTSL) method, which introduces the timestamped subtitles for a text span-based predictor. Specifically, the visual prompt is a learnable feature embedding, which brings visual knowledge to the pre-trained language model. Meanwhile, the text span-based predictor learns joint semantic representations from the input text question, video subtitles, and visual prompt feature with the pre-trained language model. Thus, the TAGV is reformulated as the task of the visual-prompt subtitle span localization for the visual answer. Extensive experiments on five instructional video datasets, namely MedVidQA, TutorialVQA, VehicleVQA, CrossTalk and Coin, show that the proposed method outperforms several state-of-the-art (SOTA) methods by a large margin in terms of mIoU score, which demonstrates the effectiveness of the proposed visual prompt and text span-based predictor.

Enhancing Video-Language Representations With Structural Spatio-Temporal Alignment.

While pre-training large-scale video-language models (VLMs) has shown remarkable potential for various downstream video-language tasks, existing VLMs can still suffer from certain commonly seen limitations, e.g., coarse-grained cross-modal aligning, under-modeling of temporal dynamics, detached video-language view. In this work, we target enhancing VLMs with a fine-grained structural spatio-temporal alignment learning method (namely Finsta). First of all, we represent the input texts and videos with fine-grained scene graph (SG) structures, both of which are further unified into a holistic SG (HSG) for bridging two modalities. Then, an SG-based framework is built, where the textual SG (TSG) is encoded with a graph Transformer, while the video dynamic SG (DSG) and the HSG are modeled with a novel recurrent graph Transformer for spatial and temporal feature propagation. A spatial-temporal Gaussian differential graph Transformer is further devised to strengthen the sense of the changes in objects across spatial and temporal dimensions. Next, based on the fine-grained structural features of TSG and DSG, we perform object-centered spatial alignment and predicate-centered temporal alignment respectively, enhancing the video-language grounding in both the spatiality and temporality. We design our method as a plug&play system, which can be integrated into existing well-trained VLMs for further representation augmentation, without training from scratch or relying on SG annotations in downstream applications. On 6 representative VL modeling tasks over 12 datasets in both standard and long-form video scenarios, Finsta consistently improves the existing 13 strong-performing VLMs persistently, and refreshes the current state-of-the-art end task performance significantly in both the fine-tuning and zero-shot settings.

Let’s Discover More API Relations: A Large Language Model-Based AI Chain for Unsupervised API Relation Inference

APIs have intricate relations that can be described in text and represented as knowledge graphs to aid software engineering tasks. Existing relation extraction methods have limitations, such as limited API text corpus, and are affected by the characteristics of the input text. To address these limitations, we propose utilizing large language models (LLMs) (e.g., GPT-3.5) as a neural knowledge base for API relation inference. This approach leverages the entire Web used to pre-train LLMs as a knowledge base and is insensitive to the context and complexity of input texts. To ensure accurate inference, we design an AI chain consisting of three AI modules: API Fully Qualified Name (FQN) Parser, API Knowledge Extractor, and API Relation Decider. The accuracy of the API FQN Parser and API Relation Decider is 0.81 and 0.83, respectively. Using the generative capacity of the LLM and our approach’s inference capability, we achieve an average F1 value of 0.76 under the three datasets, significantly higher than the state-of-the-art method’s average F1 value of 0.40. Compared to the original CoT and modularized CoT methods, our AI chain design has improved the performance of API relation inference by 71% and 49%, respectively. Meanwhile, the prompt ensembling strategy enhances the performance of our approach by 32%. The API relations inferred by our method can be further organized into structured forms to provide support for other software engineering tasks.

МЕТОД ІНТЕРПРЕТАЦІЇ РЕЗУЛЬТАТІВ ВИЯВЛЕННЯ КІБЕРЗАЛЯКУВАНЬ У ТЕКСТОВОМУ КОНТЕНТІ ЗАСОБАМИ ШТУЧНОГО ІНТЕЛЕКТУ

The article proposes the method for interpreting the results of cyberbullying detection in textual content by means of artificial intelligence, which is intended to explain the decisions of the neural network model regarding the types of cyberbullying identified in the textual content. The method is original in that it interprets the results for each detected type of cyberbullying separately, which is achieved by using a multi-label classifier of a transformer neural network architecture and an interpretation model of a machine learning model. By using the trained BERT neural network model for multi-label classification of cyberbullying types in the input text sample, different types of cyberbullying are detected with the percentage of each of them. According to the developed method, an approach based on the use of a machine learning model for the local interpretability of LIME models is used for the visual interpretation of the results of cyberbullying detection, which allows you to visualize the impact of the use of individual words on the model's decision regarding whether the text belongs to different types of cyberbullying. The developed method provides three views of the interpretation of the results of cyberbullying detection: interpretation of the results according to the color palette, interpretation of the results according to the diagrams of the local importance of words, interpretation of the results according to the diagrams of the general importance of the words. The interpretation of the results according to the color palette consists in using the absolute value of the weights to determine the brightness of the color, where the brightest color indicates the greatest influence of the word on the decision made by the model, and the least bright color indicates the smallest influence, regardless of whether this influence was positive or negative. The interpretation of the results based on the diagrams of the local importance of words is provided by constructing diagrams of the influence of individual words of the text on the probability of assigning this text to a specific type of cyberbullying, which allows you to see how the model evaluates the weight of each word in the text depending on its contribution to the model's decision. The interpretation of the results from the charts of the overall importance of words is provided by forming a set of 10 words that the model considers important regardless of the specific type of cyberbullying. The results of the experiments showed that the created method provides interpretation of decisions regarding the results of neural network detection of cyberbullying at a level sufficient for a person to understand the features of the text, which resulted in decision-making by artificial intelligence regarding the detection of types of cyberbullying.

Transformer-based automatic Arabic text diacritization

In Arabic natural language processing (NLP), automatic text diacritization is a major obstacle, and progress has been slow when compared to other language processing tasks. Automatic diacritical marking of Arabic text is proposed in this work using the first transformer-based paradigm designed solely for this task. By taking advantage of the attention mechanism, our system is able to capture more of the innate patterns in Arabic, surpassing the performance of both rule-based alternatives and neural network techniques. The model trained with the Clean-50 dataset had a diacritic error rate (DER) of 2.03%, even though the model trained with the Clean-400 dataset had a DER of 1.37%. As compared to state-of-the-art results, the improvement for the Clean-50 dataset is minimal. However, for the larger Clean-400 dataset, it is a notable improvement, indicating that this approach can deliver more accurate solutions for applications requiring precise diacritical marks with larger datasets. Additionally, this method achieves a DER of 1.21% for the Clean-400 dataset, and it performs even better when given extended input text with overlapping windows.

Annotating publicly-available samples and studies using interpretable modeling of unstructured metadata.

Reusing massive collections of publicly available biomedical data can significantly impact knowledge discovery. However, these public samples and studies are typically described using unstructured plain text, hindering the findability and further reuse of the data. To combat this problem, we propose txt2onto 2.0, a general-purpose method based on natural language processing and machine learning for annotating biomedical unstructured metadata to controlled vocabularies of diseases and tissues. Compared to the previous version (txt2onto 1.0), which uses numerical embeddings as features, this new version uses words as features, resulting in improved interpretability and performance, especially when few positive training instances are available. Txt2onto 2.0 uses embeddings from a large language model during prediction to deal with unseen-yet-relevant words related to each disease and tissue term being predicted from the input text, thereby explaining the basis of every annotation. We demonstrate the generalizability of txt2onto 2.0 by accurately predicting disease annotations for studies from independent datasets, using proteomics and clinical trials as examples. Overall, our approach can annotate biomedical text regardless of experimental types or sources. Code, data, and trained models are available at https://github.com/krishnanlab/txt2onto2.0.