Minority shareholders protection and the incremental textual information in analysts' reports: A quasi-natural experiment in China

Automatic real image editing offers unprecedented freedom to modify the appearance of the image or to edit a few objects through natural language. Recent scalable model families such as diffusion models have showcased remarkable proficiency in editing highly realistic images due to the introduction of vast amounts of training data and large pretrained language models. However, these large diffusion models require iterative evaluation that would significantly hinder the pace of image editing. Moreover, the pioneering work in this field necessitates the learning of a unique textual token that corresponds to each input image, or a group of images containing the same object, leading to the generation of redundant and fragmented models. Given the aforementioned problems, we suggest a novel Large pretrained models Assistant text-guided image Editing adversarial Network (LAE-Net) in this paper. More concretely, we introduce a deep semantic editing network to globally transfer text information among different isolated editing blocks, which would extract features from the source image to differentiate text-required areas from text-irrelevant ones. Furthermore, based on idea that the multi-modal CLIP model, leveraging vision-language alignment, captures comprehensive global semantic cues, whereas the vision-centric DINO model specializes in delivering intricate, fine-grained pixel-level details, the powerful discriminator of LAE-Net is designed by harnessing the visual embeddings derived from both the CLIP and DINO models separately to boost the visual discriminant capability and facilitate training a strong generator for conditioning image generation. Comprehensive experimental evaluations show that our LAE-Net not only delivers outstanding performance but also surpasses several cutting-edge models.

• We introduce PGSFNet for real-world scene text image super-resolution. • Adaptive Frequency Modulator is proposed to extract informative frequency components. • Text Information Enhancement module is designed to incorporate text priors. • We develop a Sobel loss to guide optimization towards sharper text details. • PGSFNet is shown to achieve superior performance on public text image datasets. Real-world scene text image super-resolution is challenging due to complex writing strokes, random text distribution, and diverse scene degradations. Existing text super-resolution methods focus on pure text images or fixed-size single-line text, which limits their practical utility. To address that, we propose a Prompt-Guided Selective Frequency super-resolution Network (PGSFNet). Our unique bicephalous neural model comprises a super-resolution branch and a prompt guidance branch. The latter specifically helps in leveraging text content-aware information priors. To that end, we propose a Text Information Enhancement module. To exploit selective frequency information present in the image, PGSFNet employs a proposed Adaptive Frequency Modulator fused with multi-attention structures. Considering the criticality of text edges in our task, we also propose a tailored text edge perception loss. Extensive experiments on the standard open real-world scene text image datasets demonstrate remarkable performance of our method, achieving up to 8.75% PNSR gain for × 2 and 2.28% SSIM gain for × 4 super-resolution on the Real-CE dataset. Our code will be made public at https://github.com/holastq/PGSFNet .

ABSTRACT Self‐regulated learning (SRL) components, such as motivation, cognitive skills and learning strategies, play a critical role in mathematics (math) learning. Using Bayesian Networks (BN), this study examines which SRL components predict membership in high (75th percentile) and low‐performing (25th percentile) groups in factual/procedural and conceptual math among middle school students ( N = 658). The results revealed that cognitive skills—specifically identifying key information in texts, scientific thinking and attention—were direct predictors of factual/procedural and conceptual knowledge. Self‐efficacy also emerged as a consistent motivational predictor across all models. Quick and frequent help‐seeking during tasks was found to predict belonging to a group of low conceptual knowledge, while text comprehension and working memory were specifically related to factual/procedural skills. Indirect influences among SRL components were also identified. The study highlights how different SRL variables contribute to various dimensions of mathematical competence and contrasts the most influential predictors of belonging to either high‐ or low‐achieving student groups.

Toward adaptive and high‑precision Integrated Pest Management in the big data era.

We study how sustainability preferences and information about climate impacts shape investment choices. In a sequential discrete-choice experiment, participants repeatedly chose between a sustainable and an unsustainable asset, with the sustainable option offering equal, lower, or higher returns. In a representative US sample (N=1,003), respondents were assigned to a control group or to four treatments providing climate-consequence information in unspecific text, specific text, graphics, or an external webpage. Text and webpage information lowered the probability of choosing the unsustainable asset by 0.109–0.431 percentage points in an unincentivized round. Effects attenuate when choices affect payoffs. We discuss implications for ESG disclosure.

ABSTRACT Challenges faced by students and professionals with visual impairment in reading and understanding digital content have always been a great area of concern, particularly in subjects like mathematics and science due to their high dependency on special symbols. On one hand, useful tools are limited by complex formats; on the other hand, most available tools are directional dependent as standard textual information does not support an effective tactile form of knowledge. This research therefore, developed Optical Braille Recognition Methodology (OBRM) for transforming printed documents into Braille language files. The proposed methodology integrates image segmentation with multistage AI processing to enhance recognition of textual content and interpretation of mathematical and special symbols. A novel key-element quantification strategy reduces overall complexity and minimizes memory usage. Multi-format document acquisition followed pre-processing steps to enhance quality and remove noises before starting OCR processing that again used additional Neural Network backend for structured parsing which finally ended the tactile symbol generation stage, where new 3-bit encoding implemented readable form creation in a BRF extension file output. Even the semantic structure organized text making similar word or symbol placed together assists easier understanding during the later tactile reading learning process. A user-experience survey involving target users was organized and carried out on the system as a beta test.

Multimodal emotion detection systems commonly integrate visual, audio, and textual information to improve recognition accuracy in human–computer interaction and effective computing applications. While such approaches perform well under con-trolled conditions, they typically assume the constant availability of all modalities during inference. In real-world deployments, however, one or more modalities may be unavailable due to technical, environmental, or privacy-related constraints. This study investigates the robustness of multimodal emotion detection under missing modality conditions. A modular framework is proposed in which visual, audio, and text modali-ties are processed independently using transformer-based emotion detection models, and their outputs are combined through a late fusion strategy. The system is evaluated on the IEMOCAP dataset by simulating various missing modality scenarios during in-ference while keeping the training process unchanged. Experimental results show that the complete multimodal configuration achieves the highest performance, while partial modality setups exhibit a gradual and predictable performance degradation rather than abrupt failure. Relative performance retention analysis indicates that a large propor-tion of the full-system performance is preserved even when operating with reduced modality availability. These findings demonstrate that late fusion-based multimodal systems can provide reliable emotion detection in practical settings where complete multimodal input cannot always be guaranteed. Keywords: multimodal emotion detection, missing modalities, late fusion, ro-bustness.

BackgroundArtificial intelligence–based skin cancer screening apps (AISCSAs) offer diagnostic potential but face limited adoption. App store cues, such as ratings, may influence acceptance; yet, little is known about how users cognitively process app store information in high-stakes health contexts. To address this gap, eye-tracking was used to measure visual attention while participants evaluated a mock AISCSA app store listing.ObjectiveThis study aimed to test whether a single negative rating captures visual attention and whether an extended technology acceptance model (TAM) can predict behavioral intention to use (BI) AISCSAs.MethodsParticipants (N=76) evaluated a mock app store listing for an AISCSA under positive (n=42) or negative (n=34) rating conditions while their eye movements were recorded. Analyses combined fixation durations in defined areas of interest (AOIs) with self-reported measures of perceived usefulness (PU), perceived ease of use (PEOU), trust, BI, willingness to pay, and the self-rated importance of app attributes.ResultsNormalized fixation durations (seconds per square pixel) revealed the highest attention to the description (0.166 s/px2), followed by the reviews (0.11 s/px2) and the ratings (0.04 s/px2), while the price and the data protection received the least attention. Of the 5 self-rated app attributes, only reviews correlated positively with fixation durations on the reviews-AOI (r=0.28; P=.01). Rating valence had no significant effect on gaze patterns, PU, PEOU, trust, BI, or willingness to pay (all Ps>.05). However, PEOU (P=.001), PU (P<.001), and trust (P<.001) were significantly correlated with BI.ConclusionsAlthough the expected attentional capture effect of the negative rating was not observed, the weak or nonexistent associations between fixation durations on the AOIs and the self-rated importance of app attributes suggest that eye-tracking captures aspects of information processing that are not directly reflected in self-reported evaluations. These findings indicate that eye-tracking provides a more direct approximation of actual user behavior by revealing implicit attentional processes beyond what is captured by questionnaires. While the technology acceptance model constructs and trust predicted BI, rating valence alone did not affect acceptance or gaze behavior. In high-stakes health contexts, textual information may outweigh rating valence in driving adoption. Future research should explore conditions under which rating valence matters, including more extreme rating contrasts, variations in accompanying review texts, and the influence of individual differences such as preexisting attitudes toward artificial intelligence and levels of artificial intelligence literacy.

ABSTRACT Combining hedge funds’ quarterly position information with their access records on the SEC's EDGAR server, we explore whether hedge funds proactively gather and analyze textual information in 10‐K filings related to their stock holdings, and how these behaviors affect their positions. We find that hedge funds adjust their positions according to the textual information in the annual reports they download. Meanwhile, analyzing these reports helps them to generate excess returns. Overall, our evidence suggests that the textual content of annual reports contains crucial company insights, prompting a subset of hedge funds that engage in bulk downloads from the SEC's website to trade based on diligent analysis of 10‐K filings.

General-purpose vision-language models can analyze medical images without task-specific training, but their value for pediatric abdominal ultrasound is unknown. To investigate the zero-shot performance of a general-purpose vision-language model to diagnose pediatric appendicitis using multimodal inputs, including images, report text, and clinical information. In this retrospective study, diagnostic capabilities of Llama 4 Maverick were evaluated on the Regensburg Pediatric Appendicitis Dataset. Five experiments were conducted using the following inputs as well as input combinations: images only, ultrasound report text only, images and ultrasound text, images and clinical data, entire multimodal input. Reference standard of appendicitis was defined based on histopathology in patients who underwent surgical resection and on clinical follow-up in patients managed conservatively. Performance was assessed at the patient level using accuracy, sensitivity, specificity, positive predictive value, negative predictive value, and F1-score. Area under the receiver operating characteristic curve (AUROC) was evaluated secondarily as a measure of discrimination. Of 782 patients, 294 met inclusion criteria; based on data availability, four experiments are conducted using 293 patients and in the fifth experiment, 284. Images-only experiment showed very low specificity (13.5%) and had the poorest discrimination (AUROC, 0.567). Ultrasound report text improved classification performance, with specificity increasing to 63.5% while maintaining high sensitivity; discrimination was also strong (AUROC, 0.883). Adding images to ultrasound text or combining all available inputs did not yield further meaningful improvement. Differences between images-only experiment and text-containing experiments were statistically significant (P≤0.002). Zero-shot visual interpretation of pediatric abdominal ultrasound by a general-purpose vision-language model is inadequate for safe appendicitis diagnosis. Performance is driven primarily by a structured sonographic imaging report, supporting a role for such models as text-based decision support tools rather than autonomous ultrasound image interpreters.

ABSTRACT This paper proposes a dictionary tailored for volatility analysis in finance research. We investigate the comovement between corporate textual information and option-implied volatility, via robust multinomial inverse regression. The volatility dictionary contains vastly different words from the sentiment dictionary and the colour dictionary. The signals distilled from the volatility dictionary explain the cross-sectional variation in implied volatility dynamics, as well as the levels of implied and realized volatility. Further, the long-short strategies of Straddle options based on the signals deliver sizeable and statistically significant expected returns. Compared to Large Language Models, we provide statistical inference on the properties of vocabularies in the dictionary.

Cross-model diffusion: Mitigating hallucination in large language models for rumor detection.

MotivationMultiomics data analysis is essential for scientific discovery in precision medicine. However, translating analysis results of omics data analysis into novel scientific hypotheses remains a significant challenge. Human experts must manually review analysis results and generate new hypotheses based on extensive and interconnected biomedical prior knowledge, which is subjective and not scalable. While large language models can accelerate the discovery, their reasoning improves when grounded in structured, auditable, and comprehensive biomedical prior knowledge. However, biomedical knowledge is scattered across heterogeneous databases that use diverse and inconsistent nomenclature systems, making it difficult to integrate resources into a unified format for scalable analysis. This fragmentation limits the ability of artificial intelligence systems to fully leverage biomedical data for scientific discovery.ResultsWe developed BioMedGraphica, a novel all-in-one platform that harmonizes fragmented biomedical resources by integrating 11 entity types and 30 relation types from 43 databases into a unified textual prior knowledge graph containing 2 306 921 entities and 27 232 091 relations. In addition, we present a novel textual-numeric graph (TNG) data structure concept, where textual information captures prior biological knowledge (e.g. transcription start sites, functions, mechanisms), numeric values represent quantitative biomedical features, and the integrated relations can help uncover mechanisms. By bridging prior knowledge with user-specific data, TNG is a novel and ideal data structure for developing novel graph analysis models.Availability and implementationThe code is available at: https://github.com/FuhaiLiAiLab/BioMedGraphica and BioMedGraphica knowledge graph database can be downloaded from huggingface dataset: https://huggingface.co/datasets/FuhaiLiAiLab/BioMedGraphica

ABSTRACT Tourism plays a vital role in economic development, but fluctuations in demand, particularly during crises such as the COVID-19 pandemic, highlight the limitations of traditional data sources for timely and detailed monitoring. This study investigates the potential of social media data (Twitter now X) as a complementary data source for estimating domestic and international tourist arrivals. A three-step pipeline was developed to identify tourists from other non-tourist social media users: (1) users were classified using a Large Language Model (LLM) to exclude non-individual accounts; (2) home location was inferred through geocoding of self-reported textual information; and (3) tourists were classified based on the United Nations World Tourism Organization (UNWTO) criteria of spatial origin, trip duration, and travel purpose. Social media derived estimates of tourist arrivals in Bali between 2019 and 2021 were compared with official statistics and showed strong correlations, with coefficients of 0.999 for international tourist arrivals and 0.896 for domestic arrivals. As well as monthly counts, the approach provided fine-grained insights into tourist behaviour, including inferred home origins, destination hotspots, and daily and hourly arrival patterns – dimensions that are typically absent from conventional datasets. While the consistency of social media data declined during the COVID-19 pandemic, the findings demonstrate that social media can be used to generate timely, flexible, and spatially detailed observations that complement official statistics. A number of discussion points and areas of further work are identified to refine classification methods, to enhance scalability and the applicability of the approach in diverse contexts.

LLM-QueryBC: An LLM-Based Regulation Query System for Textual and Tabular Information in Building Codes

A deep reinforcement learning approach for emotion recognition from unaligned multimodal inputs

Optimizing 2D bridge engineering drawing digitization: A comparative study of text recognition tools and development of lightweight post-recognition structured information extraction methods

ABSTRACT Inflation forecasting is a central issue in applied macroeconomics, particularly in data-rich environments where information arrives from multiple sources. This paper examines whether combining unstructured media news data with structured macroeconomic indicators can improve inflation forecasts for China. We construct a data fusion framework that allows information from different sources to be combined through early, hybrid, and late fusion strategies. Our findings demonstrate that the hybrid fusion approach substantially outperforms traditional time series models and machine learning methods across multiple forecasting horizons. We further find that the way textual information is represented matters for forecasting performance: topic-based measures extracted using Latent Dirichlet Allocation (LDA) capture relevant inflation-related signals more effectively than alternative text representations. Overall, the findings provide new empirical evidence on the predictive value of news text in inflation forecasting and highlight the usefulness of data fusion frameworks in applied macroeconomic analysis.

The fast expansion of social media platforms has generated an unprecedented quantity of user-generated textual statistics containing reviews, emotions, reactions, and discussions related to politics, enjoyment, healthcare, commercial enterprise, schooling, and international events. Sentiment evaluation has therefore emerged as one of the maximum great studies areas in natural Language Processing (NLP) and synthetic Intelligence (AI). but, conventional sentiment evaluation systems frequently fail to correctly interpret sarcastic expressions due to the fact sarcasm regularly conveys meanings opposite to the literal sentiment expressed in a sentence. This creates semantic ambiguity, emotional contradiction, and contextual complexity that reduce sentiment type accuracy. This studies paper provides a comprehensive examine on transformer-primarily based sarcasm detection for boosting sentiment evaluation of social media text. The proposed framework investigates advanced transformer architectures inclusive of BERT, RoBERTa, XLNet, DistilBERT, and contextual attention mechanisms to discover sarcastic expressions and improve sentiment prediction performance. in contrast to traditional gadget learning strategies that depend upon hand made linguistic capabilities, transformer models utilize contextual embeddings and self-interest mechanisms to understand lengthy-variety semantic relationships and hidden emotional cues inside online conversations. The growing dependence on social media verbal exchange has substantially transformed the manner human beings express opinions, emotions, and reactions toward actual-international occasions and virtual interactions. tens of millions of customers constantly percentage remarks, opinions, and discussions across systems such as Twitter, fb, Instagram, Reddit, and YouTube, producing huge volumes of unstructured textual records every 2nd. This developing availability of user-generated content material has made sentiment evaluation an essential research area inside natural Language Processing and synthetic Intelligence. Sentiment analysis ambitions to identify emotional polarity, such as fine, poor, and neutral sentiments, from textual information to help selection-making in areas such as enterprise intelligence, healthcare monitoring, political forecasting, product advice systems, and purchaser comments evaluation. but, in spite of essential enhancements in deep getting to know and language modelling technologies, as it should be understanding sarcastic expressions within social media textual content remains one of the maximum tough demanding situations in sentiment evaluation. Sarcasm intentionally conveys meanings contrary to literal word interpretation, thereby growing contextual ambiguity and emotional contradiction that frequently mislead conventional sentiment type systems. for example, statements that seem linguistically effective may virtually explicit dissatisfaction, grievance, or frustration when interpreted contextually. Such complexities appreciably reduce type accuracy and restriction the reliability of traditional sentiment evaluation frameworks. To cope with these challenges, present day studies has an increasing number of focused on transformer-based totally architectures able to studying contextual and semantic representations of language with wonderful performance. Transformer fashions which includes BERT, RoBERTa, XLNet, and DistilBERT have revolutionized herbal Language Processing through introducing self-attention mechanisms and bidirectional contextual learning that allow models to seize long-variety dependencies and hidden emotional relationships inside textual facts. not like traditional device getting to know algorithms that depend closely on handcrafted linguistic functions, sentiment lexicons, or shallow statistical representations, transformer-based fashions mechanically analyse meaningful contextual embeddings from huge-scale corpora. those architectures are mainly effective in sarcasm detection because they could analyse contextual contradictions, tone variations, sentence dependencies, and implicit semantic cues found in social media conversations. by using incorporating sarcasm-aware contextual gaining knowledge of into sentiment evaluation structures, transformer models significantly enhance prediction overall performance, allowing extra correct interpretation of emotionally complicated and context-based online expressions. Experimental studies across benchmark datasets have verified that transformer-primarily based sarcasm detection fashions constantly outperform conventional classifiers and recurrent neural networks in phrases of accuracy, precision, recollect, and F1-rating, making them notably suitable for actual-world sentiment analytics packages.