Search Results (356)

Aspect-based sentiment analysis (ABSA) aims to identify the sentiment polarity expressed toward a specific aspect in a sentence. Existing sequential and Transformer-based methods can effectively capture contextual semantics, but they often lack explicit modeling of aspect–opinion relations. Graph-based approaches partially address this limitation by incorporating syntactic dependency structures; however, most rely on deterministic parser-derived graphs or fixed relation weights, which may introduce noisy edges and unstable message propagation for ambiguous, informal, or domain-shifted text. To address these issues, this paper proposes a Bayesian Biaffine Variational Graph Convolutional Network (BBV-GCN) for ABSA. Specifically, a contextual encoder first generates token-level representations for each sentence–aspect pair. A biaffine relation scorer then estimates aspect-aware pairwise token interactions and constructs a soft latent relation graph. Rather than treating relation weights as deterministic values, BBV-GCN introduces latent relation variables and learns their posterior distributions through variational inference, thereby enabling uncertainty-aware graph construction. Based on the learned graph, a variational graph convolutional network performs multi-hop message passing to aggregate opinion cues, modifiers, negation patterns, and contrastive signals toward the target aspect representation. Experiments on five benchmark datasets, including Twitter, Laptop14, Restaurant14, Restaurant15, and Restaurant16, demonstrate that BBV-GCN achieves competitive and well-balanced performance relative to representative attention-based, Transformer-based, and graph-based baselines. Ablation studies further confirm the contributions of Bayesian relation modeling, KL regularization, and variational graph propagation. Visualization results illustrate how uncertainty-aware weighting can attenuate spurious relations and produce more interpretable aspect-specific latent graphs. Overall, BBV-GCN provides a robust and uncertainty-aware graph reasoning framework for fine-grained sentiment analysis. Full article

In the digital education system, the increasing demand for measuring lexical and semantic text similarity has led to the need for developing scalable and intelligent automated grading systems. Traditional assessment methods used for short-answer grading are labor-intensive, time-consuming and prone to human-evaluator bias. Furthermore, computer science curricula have different evaluation challenges as programming-centric courses require rigorous syntactical and structural validation, whereas theoretical courses require deep conceptual and semantic comprehension. To address these dual challenges, we introduce the Virtual University Automatic Short-Answer Grading (VUASAG) framework, a model addressing both statistical-based lexical analysis and Transformer-based semantic architectures. Utilizing an expanded Mohler dataset spanning multiple computer science domains (including Data Structures, Introduction to Programming, Object-Oriented Programming, and Software Engineering), we evaluated the efficiency of traditional string-matching metrics alongside state-of-the-art Transformer-based models (T5, BERT, XLNet, and SBERT). The empirical results demonstrate that token-based statistical models perform robustly on keyword-dependent syntax validation while Transformer-based models excel by capturing deep conceptual context. Specifically, Sentence-BERT (SBERT) achieved superior predictive accuracy across courses, yielding a minimum Root Mean Square Error (RMSE) of 0.9511 for the Introduction to Programming course, paired with moderate positive Pearson and Spearman correlation coefficients. Confidence intervals are estimated using Fisher’s z-transformation with a 95 percentile confidence level. Question-level five-fold and two-fold cross-validation is conducted for all four courses, and SBERT is found to perform better. Finally, to demonstrate the practical utility, we deployed a web-based portal for students’ evaluation leveraging a LLaMA-based pre-trained model to deliver real-time, personalized pedagogical feedback measuring code functionality, quality, documentation, and error handling. This reduces the educator workload while preserving grading integrity. Full article

Traditional methods of citizen deliberation, such as mini-publics, have been criticized for their limited scope. However, the implementation of massive deliberative processes creates new socio-technical challenges, particularly at the systematization stage. In general, the task of processing citizen participation is typically met with the general tools of textual analysis and Natural Language Processing (NLP). However, processing participation is not just a technical problem, but fundamentally a problem of normative design whose solution must be coherent with the philosophical, ethical, and institutional decisions that the participatory process embodies. In this article, we show how, using basic syntactic rules, we can produce simple and highly explainable reconstructions of public opinions. This methodology was tested on datasets from Chilean participatory processes, specifically dialogs from We have to talk about Chile and the 2023 Constituent Process. We show how our strategy enables further information extraction by identifying different syntactic components, providing deeper insights into the answers, while aligning with the goals of citizen participation. Overall, this study presents a compelling alternative for interpreting large volumes of citizen input, particularly in ways that preserve context and advance a more coherent analytical framework aligned with the normative ideals of deliberative democracy. Full article

Code authorship attribution is an important research area for uncovering the source, affiliation, and purpose of software and malware. It encompasses a wide range of techniques for identifying code authorship based on the analysis of authors’ coding styles. These approaches play a critical role not only in combating copyright infringement and software piracy but also in the field of malware analysis and cybersecurity defense. Traditional code authorship attribution techniques mainly focus on semantic and lexical features while overlooking structural features. In contrast, graph neural networks (GNNs) can effectively learn from graph-structured data, enabling the incorporation of structural information into the learning process. This study investigates the authorship attribution of Android applications using GNN-based algorithms with graph representations that allow code to be expressed in terms of lexical, syntactic, structural, and semantic features. Experimental evaluations were conducted using four GNN models across the Genome and malware datasets, alongside a broader set of conventional machine learning baselines, including Random Forest, Support Vector Machine (SVM), Logistic Regression, and XGBoost. With 10-fold stratified cross-validation, Logistic Regression using graph-level features only achieved the highest mean accuracy of 97.47% on the Genome dataset and Random Forest using graph-level features only achieved the highest mean accuracy of 93.44% on the malware dataset. Among the GNN models, GraphSAGE achieved the most consistent performance across both datasets, with accuracies of 97.26% and 90.67% on the Genome and malware datasets, respectively. These findings demonstrate that GNN-based approaches can provide competitive performance for Android authorship attribution. Full article

Syntactic foams combine low density and thermal-insulation capability, but their dynamic flexural properties at subzero temperatures remain insufficiently characterized. This study applied an Euler–Bernoulli beam-based inverse framework to natural frequencies and damping ratios of four epoxy syntactic-foam specimens representing four formulations at 0, −20, and −40 °C. Apparent storage, loss, and complex flexural moduli and equivalent modal loss factors were identified for the first two bending modes, with intermodal consistency and analytical sensitivity evaluated. The storage flexural modulus increased with decreasing temperature for every specimen and mode, with first-mode increases of approximately 5.8–11.0% from 0 to −40 °C. At equal nominal weight fractions, the S38HS-based specimens exhibited higher storage flexural moduli than the corresponding S28HS-based specimens. First- and second-mode estimates agreed within 5% for most conditions, while S4 showed a 10.93% discrepancy at −40 °C. Equivalent modal loss factors were non-monotonic and more variable in the second mode; the maximum values, ${\eta }_2$ = 0.1040 and $E_2^{\prime \prime }$ = 0.298 GPa, occurred for S3 at −40 °C. Effective free length showed the greatest normalized influence on the inverse estimate. The identified properties provide temperature- and mode-specific, coupon-level effective inputs for preliminary vibration analyses of moderately low-temperature marine insulation components under comparable conditions. Full article

Java is one of the most widely used object-oriented programming languages, making accurate and efficient defect localization essential for improving software quality and reliability. Conventional static analysis techniques primarily rely on predefined rules and localized syntactic matching, which may limit their ability to capture complex structural and semantic relationships among program entities. To address these limitations, this paper proposes JDQuery, a query-driven defect localization framework for Java source code based on a code knowledge graph. The framework parses Java source code into abstract syntax trees (ASTs), extracts software entities and their semantic relationships according to a formalized domain ontology, and constructs a unified code knowledge graph that integrates syntactic and semantic information. Based on the structural characteristics of Java defects, defect patterns are translated into Cypher queries, enabling flexible defect localization through graph pattern matching. Experiments on multiple open-source Java projects, including both injected defects and native real-world defects, demonstrate that JDQuery achieves precision values of 97.20% and 92.87% on two projects of different code sizes. A comparative evaluation with PMD further shows that JDQuery achieves substantially higher recall while maintaining comparable precision for the evaluated defects. Efficiency experiments demonstrate that JDQuery maintains millisecond-level query latency even when processing large-scale Java projects. Full article

Developers increasingly reuse third-party Low-Rank Adaptation (LoRA) adapters for code-generation models without visibility into how they were trained, creating a supply-chain risk: a maliciously trained adapter can behave normally on clean inputs while activating attacker-controlled functionality when a specific trigger is present. This study makes a mechanistic contribution to that risk: we show that trigger modality, not just contamination rate, determines a qualitatively different backdoor learning regime. We trained 61 poisoned variants of CodeGen-350M-mono on the CodeSearchNet dataset, injecting eight backdoor triggers spanning two categories: three semantic triggers based on natural-language code comments and five syntactic triggers based on structural code transformations derived from the CodePoisoner framework. Across attack success rate measurement, cross-trigger confusion analysis, mechanistic circuit tracing, layer-restoration defense evaluation, and semantic generalization testing, we find that semantic triggers produce associative binding: a 91% attack success rate, a Trigger Specificity Index (TSI) of 268×, distinct per-trigger circuits concentrated in attention layers, and a requirement to restore 10 parameter groups for removal. Syntactic triggers instead produce distributional confusion: a 31% attack success rate, a TSI of only 1.25× (1.45× once a shared-payload confound in the confusion-matrix design is corrected for), diffuse circuits spread across (Multi-Layer Perceptron) MLP layers, and collapse with just 5 restored parameter groups. Cross-payload testing on single-trigger models confirms this: structural triggers fire on triggers never seen during training at rates of 42 to 55%, showing that the model learns a general association between code abnormality and payload generation rather than a specific trigger–payload mapping. Both regimes preserve clean code-generation quality across all contamination rates, so a downloaded backdoored adapter is behaviorally indistinguishable from a clean one under standard benchmarks. These results argue against a one-size-fits-all approach to adapter auditing: detection and removal strategies calibrated to one trigger modality can fail outright against the other, and we outline the conditions under which each applies. Full article

Large Language Models (LLMs) have been rapidly evolving lately, resulting in the need for strong, explainable models to detect the difference between human-generated and machine-generated articles. Existing approaches which are mostly based on fine-tuned transformers suffer from several drawbacks such as rapid obsolescence, paraphrasing attacks, and lack of interpretability. To improve their ability to detect, this paper proposes a novel paradigm called Human vs. LLM Identification (HLI) which introduces a Retrieval-Augmented Generation (RAG)-inspired evidence-based detection strategy alongside a fine-tuned transformer classifier. Our core model, DeBERTa-Sentinel, is built on top of a fine-tuned Microsoft DeBERTa-v3-small model, which uses a disentangled attention mechanism to better capture subtle syntactic and stylistic deviations characteristic of AI-generated text. We evaluate our framework on a balanced dataset of 43,456 text samples, curated from the OpenGPTText corpus and covering AI-generated and human-authored content across diverse domains including news, education, and creative text. The experimental results show improved performance over the selected baselines, with our framework achieving an accuracy of 97.53%, precision of 95.89%, recall of 99.34%, and ROC-AUC of 99.53%. In addition, explainability is integrated into our framework through Local Interpretable Model-agnostic Explanations (LIME) analysis, providing token-level insight into classification decisions. This study establishes a benchmark for scalable, explainable AI text detection, with implications for academic integrity, content moderation, and combating misinformation. Full article

Individual differences in statistical learning (SL) constitute an important cognitive mechanism that supports second language (L2) syntactic processing. However, existing research has primarily focused on adult populations, while potential differences across input modalities remain understudied. This study investigated the associations between SL and implicit L2 morphosyntactic processing among Chinese adolescent English learners, comparing results across visual and auditory modalities. Participants completed visual and auditory SL tasks, self-paced reading (SPR), self-paced listening (SPL), as well as working memory (WM) and attention tasks. L2 morphosyntactic processing ability was measured via grammatical sensitivity indicators extracted from SPR and SPL data. Mixed-effects analyses did not detect statistically reliable group-level grammaticality effects in either task, though SPR exhibited a descriptive pattern where ungrammatical sentences elicited longer reading times at the third word. Participants achieved higher accuracy on the visual SL task than the auditory one. In exploratory regression analyses, visual SL accuracy was found to positively predict grammatical sensitivity measured at Word 3 in SPR. Still, this association was not confirmed in a follow-up sensitivity analysis. No corresponding significant association was observed between auditory SL and SPL sensitivity, and neither working memory nor attention showed consistent independent contributions to the outcome. These findings extend individual-differences research on this topic to adolescent classroom L2 learners and underscore the potentially important role of visual regularity learning in online L2 reading processing. Full article

Aspect-based Sentiment Analysis (ABSA) is a fine-grained sentiment classification task that aims to predict the sentiment polarity associated with aspect terms in sentences. Traditional methods based on syntactic and semantic dependency trees are insufficient for capturing contextual sentence features. To address this, we propose a Knowledge-Enhanced and Multidimensional Feature Fusion (KEMFF) model for ABSA, which captures sentiment feature representations across multiple dimensions, including syntax, semantics, and knowledge. First, the pre-trained model RoBERTa is used to obtain embeddings of sentences and aspect terms. Then, a syntactic dependency parser and a graph convolutional network are utilized to learn syntactic features. Meanwhile, an Abstract Meaning Representation (AMR)-based parser is employed to construct semantic relations, and axial attention is used to aggregate incoming and outgoing semantic dependencies. Furthermore, external knowledge is embedded, and an attention mechanism is employed to obtain aspect-specific knowledge representations, thereby complementing syntactic and semantic representations with external lexical knowledge. Finally, multidimensional features are fused and passed to a softmax classifier for predicting sentiment polarities. Unlike previous models that mainly focus on either syntax–semantic fusion or knowledge-enhanced graph propagation, KEMFF explicitly models syntax, semantics, and lexical knowledge in three parallel branches and aligns them into a unified aspect-level representation. Experiments on Laptop14, Restaurant14, and Twitter datasets show that KEMFF achieves the best performance among the compared baselines on Laptop14 and Restaurant14, and it obtains competitive results on Twitter. Full article

Current advances in neural network models have improved state-of-the-art performance in natural language processing tasks such as named-entity recognition, sentiment analysis, and machine translation. In particular, neural language models are applied to encode information in word embeddings. These approaches are generally trained on large corpora using semi-supervised learning. Word embeddings encode the syntactic and semantic properties of words as dense vectors. In agglutinative languages such as Turkish, Finnish, and Hungarian, word-embedding construction is challenging because extensive suffixation and polysemy can cause information loss. To overcome these limitations, character n-grams are often preferred for embedding representations. Nevertheless, character n-grams do not guarantee the capture of information in long word sequences. In this study, a method that partitions word sequences according to frequent patterns within a given context is proposed for training a neural language model. In this respect, likelihood- and ranking-based inference are combined with n-gram and syllable partitioning for word-embedding generation from a text corpus. The proposed approach provides a language-agnostic, context-sensitive segmentation mechanism that can complement language processing methods such as lemmatization, morphological analysis, and stemming. For embedding generation, the SkipGram and FastText models are used, and the effects of word partitioning are evaluated using analogy, named-entity recognition, POS tagging, sentiment analysis, and morphological disambiguation datasets for Turkish. The results indicate task-dependent and generally limited improvements over traditional token-based word-embedding extraction. In particular, skip n-gram partitioning produces a substantial improvement over partitioning based on frequent-ngrams, sentencepiece-bpe, sentence-unigram and morfessor. No consistent relationship was observed across tasks between performance and either graph density or the average number of distinct n-grams per sentence. Full article

Traditional space syntax research mainly focuses on the overall topology of street networks, with insufficient attention to street segments and spatial skeletons. To fill this gap, this study innovatively integrates segment and skeleton analysis into space syntax and proposes a multilevel segment–skeleton–network analytical framework. Taking Beijing’s Capital Core Area (CCA) as a case, we analyze its street network evolution, spanning more than 600 years from the Yuan Dadu to the present, based on multi-temporal historical maps and syntactic indicators. The results indicate that main streets and alleys present polarized development; traditional commercial centres gradually decline while maintaining high accessibility; and a distinctive ring-shaped public skeleton centred on the Forbidden City has taken shape. This integrated framework remedies the defects of a single network analysis. The outcomes offer practical implications for the protection and sustainable development of historic street systems. Full article

Traditional villages commonly face functional decline and vitality loss amid modernization, which severely threatens the long-term cultural and ecological sustainability of rural heritage spaces. Commercial settlements, evolved along ancient trade routes, follow distinct spatial logics, yet existing studies rely heavily on qualitative descriptions, lacking quantitative frameworks that integrate spatial structure and functional vitality to support targeted sustainable heritage renewal. Taking the main street of Lanche Village, a typical commercial settlement along the Taihang ancient trail in southeastern Shanxi, as a case study, this paper establishes an analytical framework combining Space Syntax, Platform-derived geospatial POI data, and agent-based simulation. Core syntactic indicators (integration, choice, intelligibility) are calculated, and POI kernel density analysis and pedestrian flow simulation are conducted. The results confirm the traditional commercial logic of “post road as the axis, shop-front and residence-back” with high integration and accessibility of the main street. However, three key problems are identified: low spatial intelligibility (R² = 0.43), a significant vitality gap between southern and central sections, and a mismatch between high spatial potential and low functional vitality. Agent-based simulation further reveals pedestrian hotspots and “invisible boundaries” caused by closed street interfaces. This study identifies a context-dependent partial decoupling between spatial potential and functional vitality in declining commercial settlements, expanding the applicable boundary of classical Space Syntax vitality theory in recession-stage settlements. While overall spatial accessibility and functional density show a significant positive correlation consistent with classical expectations, local mismatch zones emerge under the mediation of closed street interfaces and functional inertia. Based on quantitative diagnosis, four micro-regeneration strategies are proposed: structural infill, functional insertion, morphological restoration, and placemaking, forming a complete workflow from diagnosis to intervention. This study constructs a multi-scale spatial vitality mismatch diagnostic framework, and proposes targeted renewal strategies corresponding to the diagnostic conclusions. Centered on revealing the decoupling mechanism between spatial potential and functional vitality, this study builds a matched multi-scale diagnostic workflow as supporting methodology, and puts forward targeted renewal suggestions corresponding to the identified vitality barriers. It also provides an actionable approach for the sustainable renewal of traditional commercial settlements. Full article

This paper presents a quantitative evaluation of an automatic translation system between industrial robot programming languages based on a sequence-to-sequence (Seq2Seq) neural architecture using Long Short-Term Memory (LSTM) networks. The study addresses the interoperability problem between proprietary robot programming languages by proposing a data-driven approach capable of learning correspondences between structured code instructions. A parallel dataset of 28,000 aligned instruction pairs was constructed and preprocessed through tokenization and normalization to enable structured sequence learning. The model was trained under four configurations (50,100, 150 and 200 epochs) to analyze the impact of training duration on performance and generalization capability. The system was evaluated using multiple quantitative metrics, including accuracy, loss, BLEU, and Exact Match (EM), allowing assessment of both structural similarity and exact sequence correctness. Experimental results demonstrate that the 200-epoch configuration improves the performance across all metrics, achieving an accuracy of 0.9943, a BLEU score of 0.682, and an Exact Match of 0.970 on the test set. These results indicate that the model is capable of generating both structurally consistent and syntactically correct translations. The analysis shows that while BLEU captures structural similarity, EM provides a stricter evaluation of exact sequence correctness, which is critical in structured code translation tasks where minor variations may affect execution. The proposed approach demonstrates the feasibility of applying neural machine translation techniques to industrial robot programming, contributing to improved interoperability and reduced manual effort in multi-platform robotic environments. Full article

The growing use of large language models (LLMs) makes it essential to identify which system produced a given text. Existing stylometric detectors perform well on raw outputs; however, users routinely translate, paraphrase, or edit content before release. These transformations can erase the lexical and syntactic cues that support attribution. This study measures the resulting performance drop on a purpose-built benchmark of synthetic responses from two instruction-tuned models, Qwen 2.5 (32B) and Gemma 3 (27B). Each passage is subjected to common editing workflows ranging from machine translation to paraphrasing and grammar-focused rewriting. The edited variants are evaluated using a custom machine-learning classifier. Overall, attribution remains robust, but obfuscation leads to degradation that often renders the text unusable for professional purposes. Although the analysis is limited to data from two LLMs, the controlled setting shows how specific post-editing actions alter stylometric signals. Recognising this sensitivity is essential for academic integrity checks, platform-moderation tools, and the ongoing debate between watermarking and stylometry. The results provide practical guidance on developing attribution methods that remain reliable after routine text updates. Full article