Reliability estimation of large language models is in many cases as crucial as their accuracy, as reliable models are more trustworthy, robust, and suitable for practical applications. Recent advancements in natural language processing (NLP), particularly those based on transformer architectures, have significantly accelerated progress across various NLP tasks. This study focuses on the reliability of transformer-based question answering (QA) models, specifically BERT models and its variants (RoBERTa, ALBERT, DistilBERT). These encoder-only pretrained transformers have demonstrated remarkable accuracy in QA tasks that can be treated as classification tasks. However, their reliability remains underexplored. This study evaluates the reliability of four BERT-based models by assessing response stability under two conditions: (1) internal model variations induced via Monte Carlo Dropout (MCD) and (2) input perturbations through paraphrasing. Using the SQuAD and QuAC datasets, we investigate how dropout rates affect prediction consistency and whether lexical changes impact answer stability. Our findings reveal that RoBERTa maintains higher reliability, whereas AlBERT and DistilBERT exhibit significant inconsistencies. Statistical analyses confirm that enabling MCD during prediction does not disrupt inference dynamics, validating its effectiveness as a reliability metric. These findings underscore the importance of evaluating both accuracy and stability in QA models to ensure stability in real-world applications.
Mining emotional information plays an important role in improving MOOC quality. However, the word embedding method, which has been greatly improved, has insufficient ability to express emotional semantics, and it cannot extract more semantic information from MOOC comments. To solve these problems, this study proposes a TCN-BiGRU network model with attention embedding (TGAE) based on the mechanism of multi-head attention to improve the accuracy and robustness of MOOC comment sentiment analysis. In the word-embedding stage, the TGAE combines BERT embedding with improved GloVe embedding. The TGAE uses a TCN to extract local semantic features and a BiGRU to acquire contextual semantic information. The TCN enhances the computational efficiency and long-distance dependency extraction of the dilated convolution. The BiGRU processes forward and backward contexts simultaneously to enhance semantic understanding. The mechanism of multi-head attention is used to select some heads and then assign different weights to the features of the TCN and BiGRU features, which can focus on the key information of sentiment from multiple views. The experiment shows that TGAE achieves better performance in the sentiment classification of MOOC comments, and the accuracy and F1-score are improved.
A major current approach in theoretical linguistics proposes that there is only a partial, defeasible correspondence between syntax and semantics: meaning and grammar are autonomous, parallel components in the architecture of language. This fact still lacks a satisfactory account. One type of explanation could be provided in terms of computational capability: architectures that separate out meaning and grammar into two processing streams might perform better in tasks that require the system to attend to syntactic vs semantic aspects of sentences. We compare single-stream vs dual-stream transformer architectures, which we train and test on data sets from two event-related brain potential (ERP) studies, with a variety of syntactically and semantically challenging sentences. The single-stream model showed greater sensitivity (the probability that a given sentence s from set S is labelled as belonging to S) with syntactic anomalies, whereas the dual-stream model showed greater sensitivity with semantic anomalies. However, overall, the two models classified sentences with similar accuracy: there were no general advantages for either architecture. Our results are a counterexample to the notion that the functional separation of meaning and grammar in the architecture of human language can be explained in strictly computational terms.
Video summarisation, a critical task in multimedia analysis, aims to condense lengthy videos into concise representations while preserving essential content. As the volume of video data grows exponentially, efficient and informative summarisation becomes increasingly crucial. Striking the right balance between preserving essential content and adhering to length constraints remains an emerging field of study. In this context, we explore the fusion of 3D VGG16 (Visual Geometry Group 16-layer network) and BERT (Bidirectional Encoder Representations from Transformers) for video summarisation (3D-BERSUM). This method offers comprehensive coverage by leveraging 3D VGG16 to extract visual features and BERT to understand textual context, ensuring that the generated summaries capture both the visual elements of the footage and the key information conveyed through narration. BERT's contextual understanding enhances the summarisation process by grasping nuances, relationships, and sentiment in the text, enriching the summary with a deeper understanding of the narrative. By integrating visual and textual information, we generate summaries that convey more than just visual scenes. To assess the performance of 3D-BERSUM, we compared it with existing methods using the content coverage measures (recall, precision, and F1-score) and textual quality metrics (ROUGE and BLEU). The findings indicate that the 3D-BERSUM yields enhancements in recall, precision, and F1-score by 5.6%, 3.9%, and 5.2% for the SumMe dataset, and 10.1%, 7.2%, and 6.1% for the TVSum dataset, respectively. Finally, comparison is performed with other state-of-the-art methods, and the result shows that the superiority of 3D-BERSUM. This underscores the efficacy and potential of our method in achieving enhanced video summarisation performance across diverse datasets and scenarios.
Naive Bayes is renowned for its simplicity and efficiency, occupying a significant position within the domains of data mining and machine learning. However, its performance optimisation is constrained by the adequacy of instance training and the effectiveness of feature selection. To address these limitations, this paper first proposes a novel algorithm called Misclassified Instances Re-training Naive Bayes (MIReNB). This methodology employs Leave-One-Out Cross Validation (LOOCV) to meticulously discern and iteratively reuse misclassified instances, dynamically adjusting the frequency table of the conditional probability distribution for each attribute-value and class, followed by retraining procedures to augment predictive performance. Building upon MIReNB, we subsequently propose two extensions. The first extension, designated as MIReNB$<^>{{\rm{Select}}}$ Select, incorporates an additional feature selection stage before identifying misclassified instances to determine the optimal feature subset for retraining, thereby improving prediction accuracy and robustness. The second extension, termed MIReNB$<^>{{\rm{Twice}}}$ Twice, involves a dual iterative process that deeply integrates misclassified instances, achieving adaptive refinement of the conditional probability distribution while enhancing classification accuracy and generalisation ability. Numerous empirical investigations have shown that MIReNB, along with its two extensions, can enhance classification accuracy notably while preserving the inherent simplicity and efficiency of the Naive Bayes classifier.
The integration of Artificial Intelligence (AI) into smart farming, particularly Crop Recommendation System (CRS), has propelled significant advancements but is often hindered by the 'black box' nature of models, which limits transparency and trust. This study aims to enhance smart farming by embedding explainable Artificial Intelligence (XAI) techniques - specifically Contrastive Explanation Method (CEM) and Accumulated Local Effects (ALE) - within CRS, empowering farmers to understand AI-generated crop suggestions. Implemented an XAI-driven CRS, utilising Local Interpretable Model-agnostic Explanations (LIME), SHapley Additive exPlanations (SHAP), CEM, and ALE for comprehensive explainability. Notably, CEM provides farmers with actionable contrastive explanations, while ALE details the average influence of environmental factors. To address data scarcity, Generative Adversarial Networks (GANs) were used to augment the dataset with synthetic data, and an interactive, explainable interface was developed using Streamlit. Results show a substantial improvement in system interpretability and user trust, evidenced by clearer, actionable explanations for farmers. Quantitatively, incorporating GAN-augmented data improved the Random Forest model's Area Under the Receiver Operating Characteristic curve (AUROC) from 0.94 to 0.985 and F1-score from 0.93 to 0.98. This research is the first to integrate CEM and ALE in CRS, establishing a new benchmark for transparent and effective AI-powered agricultural decision-making.
Training Artificial Neural Networks (ANNs) with Stochastic Gradient Descent (SGD) frequently encounters difficulties, including substantial computing expense and the risk of converging to local optima, attributable to its dependence on partial weight gradients. Therefore, this work investigates Particle Swarm Optimization (PSO) and Genetic Algorithms (GAs) - two population-based Metaheuristic Optimizers (MHOs) - as alternatives to SGD to mitigate these constraints. A hybrid PSO-SGD strategy is developed to improve local search efficiency. The findings indicate that the hybrid PSO-SGD technique decreases the median training MSE by 90 to 95 percent relative to conventional GA and PSO across various network sizes (e.g., from around 0.02 to approximately 0.001 in the Sphere function). RMHC attains substantial enhancements, reducing MSE by roughly 85 to 90 percent compared to GA. Simultaneously, RS consistently exhibits errors exceeding 0.3, signifying subpar performance. These findings underscore that hybrid and evolutionary procedures significantly improve training efficiency and accuracy compared to conventional optimization methods and imply that the Building Block Hypothesis (BBH) may still be valid, indicating that advantageous weight structures are retained during evolutionary search.
The rapid integration of Artificial Intelligence (AI) in Higher Education (HE) is transforming personalized learning, administrative automation, and decision-making. However, this progress presents a duality, as AI adoption also introduces ethical and institutional challenges, including algorithmic bias, data privacy risks, and governance inconsistencies. To address these concerns, this study introduces the Human-Driven AI in Higher Education (HD-AIHED) Framework, ensuring compliance with UNESCO and OECD ethical standards. This conceptual research employs a qualitative meta-synthesis approach, integrating qualitative and quantitative studies to identify patterns, contradictions, and gaps in AI adoption within HE. It reinterprets existing datasets through theoretical and ethical lenses to develop governance frameworks. The study applies a participatory integrated co-system, Phased Human Intelligence, SWOC analysis, and AI ethical review boards to assess AI readiness and governance strategies for universities and HE institutions. The HD-AIHED model bridges AI research gaps, addresses global real-time challenges, and provides tailored, scalable, and ethical strategies for diverse educational contexts. By emphasizing interdisciplinary collaboration among stakeholders, this study envisions AIHED as a transparent and equitable force for innovation. The HD-AIHED framework ensures AI acts as a collaborative and ethical enabler rather than a disruptive replacement for human intelligence while advocating for responsible AI implementation in HE.
The challenge of reliable alignment of artificial intelligence systems has drawn increasing attention from both researchers and policymakers. While current alignment efforts often presume the feasibility of specific solutions for today's systems, this study addresses a further question: is robust alignment compatible with the essential cognitive capacities required for general intelligence? Based on the analysis of necessary characteristics of general intelligence, such as autonomous exploration of sensory environment and others, our analysis shows that under natural assumptions, systems advancing towards general intelligence can develop goals and perspectives that diverge from human norms and expectations. We argue that such divergence is not a failure of design but a logical outcome of the very properties that enable general intelligence. These findings imply both technological limits to alignment programs based on rigid control, and ethical and societal challenges including the risks of cognitive divergence, miscommunication, and limits of enforceable compliance. They also carry policy implications, highlighting the need for monitoring for early signs of misalignment, responsive governance, and frameworks for long-term human - AI coexistence. The study emphasises a structural tension between open-ended intelligence and robust alignment, underscoring the necessity of rethinking not only technical solutions but also societal approaches to managing advanced AI systems.
Proximal Policy Optimisation (PPO) is a well-known deep reinforcement learning (RL) algorithm that relies on a multilayer perceptron (MLP) to approximate PPO's policy and value functions of the PPO. However, the representational capacity of MLPs may limit the performance of the PPOs. Recent research has shown that the Kolmogorov-Arnold Network (KAN) is a powerful substitute for MLPs, outperforming them across various domains. Although conventional KAN has demonstrated exceptional improvements over MLP-based PPO, it still has limitations, particularly when capturing complex correlations in input observations. Additionally, KAN heavily relies on the grid hyperparameter, and the performance is susceptible to the grid selection. To overcome these limitations, we propose a convolution-based Kolmogorov-Arnold network (Conv-KAN) as an alternative to MLP and conventional KAN in PPO. The novelty of the proposed approach lies in the integration of convolutional layers with the Kolmogorov-Arnold framework, which significantly enhances the model's ability to capture feature correlations. Convolutional layers are inherently effective in modelling spatial dependencies and structured local features in the input observations. By leveraging this capability, the Conv-KAN aligns more naturally with the Kolmogorov-Arnold decomposition and improves the representation of complex patterns in the input space. As a result, this integration leads to more effective policy learning and decision making in RL tasks. Furthermore, the proposed approach is insensitive to the selection of grid size. Experimental results show that our Conv-KAN-based PPO outperforms MLP and recent KAN-based PPO across several RL benchmarks by approximately 12.9% and 108.7% in HalfCheetah and Hopper environments, respectively.
The integration of artificial intelligence (AI) into the new product development (NPD) process has been shown to significantly enhance innovation success. Drawing on situated AI theory, this study investigates two core senior management behaviours-emphasis on AI adoption and reward systems for AI adoption - and their roles in catalysing AI-driven innovation and facilitating AI integration across the seven NPD stages. Based on empirical analysis of 558 AI-driven NPD projects, the findings reveal that both senior management's emphasis on AI adoption and reward systems positively impact innovation performance. Notably, senior management's emphasis on AI adoption, rather than reward systems, significantly improves innovation quality, innovation speed, and product innovativeness. The results further indicate that senior management's emphasis on AI adoption facilitates AI integration across six of the seven NPD stages, with the exception of the product testing stage. In contrast, reward systems for AI adoption positively influence AI usage in four stages, primarily within the early (idea development, business analysis, product design) and late (operations management) phases of the NPD process. This research advances situated AI theory to elucidate the mechanisms through which senior management behaviours drive AI-driven innovation success and adoption, and guides manager in effectively leveraging AI's potential within the NPD process.
2020 is the year of decisive victory in the comprehensive well-off society and the battle against poverty. While fully completing the task of poverty alleviation, consolidating the achievements of poverty alleviation, and researching and establishing a long-term mechanism to solve relative poverty, it is also necessary to do a good job in summarising poverty alleviation. The sorting and summary of China's poverty alleviation theory is not only an important content of poverty alleviation summary, but also a need to study and consolidate the achievements of poverty alleviation. The risk assessment of poverty alleviation households' return to poverty in Chongqing ethnic regions is the MAGDM. The MULTIMOORA approach is an efficient technique to solve the MAGDM, and CRITIC approach is an efficient technique to obtain the objective weight information. PDHLTS could excavate the uncertainty more effectively and deeply for risk assessment of poverty alleviation households' return to poverty in Chongqing ethnic regions. In this study, CRITIC technique is put forward with weight numbers along with PDHLTSs and PDHL-MULTIMORA approach is administrated for MAGDM with PDHLTSs. At last, numerical example for risk assessment of poverty alleviation households' return to poverty in Chongqing ethnic regions and some comparisons is put forward to illustrate the PDHL-MULTIMORA.
More courses are now readily available to university students when they begin their higher education studies in recent years. Universities, especially those seeking programme flexibility and quick responses to evolving market demands, may tremendously benefit from understanding more about student preferences. The goal of this study is to identify the factors influencing students' course choices and to evaluate learners' performance. The sample was collected from the e-KHOOL website. The methods of Regression analysis and exploratory factor analysis are used to evaluate the data collection of student's course selection. The study's findings and previous research served to contextualise this issue by showing not only how graduates choose their main study fields but also the variables that affect how many students decide not to enrol. Moreover, the student's course selection can be significantly predicted by influencing factors like career advancement, teaching standard, course quality standard, learning capability, and learner performance. The performed analysis revealed that learner's capability, course quality standard, and career advancement highly impact the course selection process of students. Our findings validate the significance of motivation as a source of educational efficiency by demonstrating that it is the primary variable influencing the Student's performance of learners.
Grey-Wolf Optimizer (GWO) and Salp-Swarm Algorithm (SSA) are among the recent metaheuristic algorithms that have shown considerable potential in solving practical problems. Nevertheless, these optimisation methods usually suffer from a poor performance, especially when dealing with multimodal and high-dimensional problems. To overcome drawbacks of these algorithms, a new population-based global optimisation algorithm inspired by GWO and SSA is proposed in this paper. The basic idea behind the proposed approach is to divide the population into two groups that adopt new position updating strategies. The movement method of the first group is inspired by the hunting mechanism used in GWO algorithm. On the other hand, the population of the second group moves into the search space using a new strategy based on the position updating method employed by the follower salps in SSA. The performances of the proposed algorithm, named GW-SSA, have been studied through 23 benchmark functions of different types and dimensions. The obtained results were statistically analysed using Wilcoxon signed-rank. Furthermore, GW-SSA algorithm has been applied to solve four real engineering design problems. The experimental results show the superior performance of GW-SSA compared to GWO, SSA and other recently proposed optimisation algorithms.
The recent growth of live chatbots, email services, and direct text messages has been more helpful for students' education. However, providing irrelevant and inappropriate replies to a query remains a drawback in the existing chatbot system. Thus, to solve this issue, a deep learning-based chatbot system for educational purposes is designed. At first, the pre-processed data are given to the word-to-vector conversion stage, which is done by Bidirectional Encoder Representations from Transformers (BERT), Transformer net, and Text Convolutional Neural Network (CNN). The optimisation of the weights and the weighted feature selection are done by an advanced Averaging-based Driving Training-Barnacles Mating Optimizer (ADT-BMO). The weighted fused features are now given to the Adaptive and Attentive Deep Temporal Convolutional Network with Symmetric Convolution (AA-DTCN-SC) for intent classification. The parameters in the recommended AA-DTCN-SC model are optimised using the suggested ADT-BMO algorithm to provide a better classification performance. From the classified intents, a suitable reply to the query asked by the student is obtained. Various simulations are executed by comparing the implemented model with the traditional systems using several performance measures. Throughout the result analysis, the accuracy and precision rate of the recommended approach are 85.4% and 92.6%.
In this essay, some sufficient criteria guaranteeing the positive exponential stability of the positive periodic solution of recurrent neural networks involving mixed delays are obtained. The results are established under the hypothesis that the activation functions are monotone nondecreasing only on a local interval, which improves and supplements some existing ones in recent references. Furthermore, a numerical example is provided to demonstrate the reliability of the theoretical outcomes.
The rapid growth of Internet-based applications, exemplified by the widespread adoption of social media platforms and weblogs, has started an era where comments and evaluations of everyday activities have become pervasive. In this context, sentiment analysis uses the capabilities of natural language processing, text analysis, and computational techniques to automate the extraction and classification of emotions and sentiments expressed in textual reviews. The insights derived from public opinions hold substantial value for businesses, governmental bodies, and individuals, serving as a critical source of information to support informed decision-making. Consequently, sentiment analysis has become a focal point within the field of decision support systems. Moreover, sentiment analysis based on machine learning (ML) has emerged as a cross-disciplinary approach to exploring opinion mining, particularly within the domains of media and communication research. These ML-based models have demonstrated great promise in achieving the goals of social media analysis due to their ability to capture both the structural subtleties and semantic complexities of text, without the need for the extensive feature engineering that characterised earlier approaches. While marketing researchers have employed various methods to analyse textual reviews, there remains a lack of a comprehensive performance evaluation framework, which complicates the selection of appropriate methods for future applications. This situation raises a fundamental question: What aspects are most frequently examined by researchers using ML-based sentiment analysis methods? To address this question, we propose a systematic literature review focused on sentiment analysis using ML techniques. This comprehensive review examines recent research efforts, highlighting the contributions of various scholars and focusing on ML techniques categorised into four primary clusters. The findings demonstrate the effectiveness of ML methodologies in conducting sentiment analysis with greater efficiency and accuracy. Notably, these ML models outperform their simpler counterparts, leading to a strong endorsement of their superiority.
