With the rapid development of the Internet and social media, the multidimensional subjective sentiments embedded in massive volumes of user-generated content have become a critical data foundation for business intelligence, public opinion monitoring, and public decision-making, thereby driving the rapid advancement of sentiment analysis technologies. Traditional models typically treat sentiment analysis as an end-to-end pattern matching or classification task, lacking explicit modeling of the intrinsic logic underlying sentiment formation. As a result, their robustness and interpretability remain limited when handling challenging scenarios such as implicit sentiments and complex sentence structures. In particular, when confronted with cases that require deep semantic understanding and multi-step reasoning to accurately determine sentiment polarity, existing methods often prove inadequate. To address these issues, this paper proposes a sentiment analysis method based on a sentiment chain-of-thought framework combined with LoRA fine-tuning. The proposed approach leverages large language models to generate high-quality reasoning data that conform to a five-step protocol, transforming implicit sentiment judgments into explicit, structured chains of thought encompassing entity localization, contextual analysis, and transition detection. In addition, quantized low-rank adaptation is employed to conduct full-sequence joint modeling training, which significantly reduces computational and memory overhead while minimizing performance degradation, thereby enabling cost-efficient training of large models. Experimental results demonstrate that the proposed method outperforms traditional deep learning baselines and generic chain-of-thought fine-tuning approaches on the Laptop and Restaurant datasets of SemEval-2014 Task 4, validating its effectiveness.

Aspect-based sentiment analysis has attracted extensive attention from both academia and industry. Existing approaches often suffer from a large number of noise nodes unrelated to aspect-term sentiment judgment when processing complex syntactic structures, while long-distance dependencies and intricate syntactic-semantic relations are difficult to capture effectively. To address these issues, this paper proposes a dual-graph collaborative model for aspect-based sentiment analysis based on adaptive pruning and key-path enhancement. First, for the syntactic dependency tree, the model introduces adaptive distance pruning and a key-path enhancement strategy, which removes global noise while significantly amplifying the structural signals between aspect terms and opinion words. Second, a semantic similarity dual graph is constructed to capture implicit long-range dependencies. Finally, a multi-source output fusion module is introduced to deeply integrate the refined syntactic features with dense semantic features, thereby enabling efficient interaction and collaborative modeling of multi-source information. Experimental results on several public datasets demonstrate that the proposed model outperforms existing mainstream baseline methods in both accuracy and Macro-F1 score, validating the effectiveness of the pruning-enhancement and dual cross-fusion strategies.

In the post-Moore's Law era, traditional bulk semiconductor devices are approaching fundamental physical limits, making it difficult to meet the demands for high integration, miniaturization and multifunctionality of next-generation electronics. Semiconductor nanostructures, with unique quantum size, surface, and confinement effects, have become the core means of breaking through these bottlenecks. This review systematically summarizes the latest research progress in the rational design, controllable fabrication and multi-field applications of semiconductor nanostructures, and focuses on the technical characteristics and pros and cons of precise preparation of single-unit nanostructures, directional construction of composite heterostructures, preparation of doped/modified nanostructures and epitaxial growth of low-dimensional single-crystal nanostructures. It elaborates on the performance optimization mechanisms and practical application values of nanostructures in optoelectronic, chemical/gas sensing and energy conversion devices, and deeply analyzes the core challenges in wafer-scale fabrication integration, device performance reliability and interface compatibility. Corresponding targeted interface regulation strategies are further proposed and discussed, including δ-doping engineering, van der Waals epitaxy and low-temperature etching technology. This review clarifies the key technical paths for the performance regulation of nanostructured semiconductor devices and provides a comprehensive theoretical and technical reference for the industrial development of high-performance nanostructured semiconductor devices in the post-Moore era.

Objective: This study aimed to evaluate the anti-obesity efficacy of Qiangli Dingxuan Tablets using a glucose-induced obese model inCaenorhabditis elegans (C. elegans)and to clarify its molecular mechanisms associated with lipid metabolism regulation. Methods: Obesity was induced in C. elegans using 2% glucose, followed by treatment with Qiangli Dingxuan Tablets at concentrations of 0.125, 0.25, and 0.5 mg/mL. To assess lipid accumulation and triglyceride content, we performed Oil Red O (ORO) staining, Nile Red staining, and triglyceride assays. Locomotion and body size were evaluated. The mRNA expression levels of lipid metabolism-, IIS pathway-, and stress-related genes were determined by RT-qPCR. Results: Qiangli Dingxuan Tablets significantly reduced lipid accumulation and triglyceride content in obese nematodes and improved obesity-associated locomotor impairment and body enlargement. RT-qPCR analysis showed that treatment downregulateddaf-2, sbp-1, fat-5, and fat-6, while upregulateddaf-16, sod-2, and lipl-4. Conclusion: Qiangli Dingxuan Tablets ameliorated glucose-induced obesity inC. elegans by regulating lipid metabolism, inhibiting lipogenesis, and promoting lipolysis. These findings provide experimental evidence for its potential application in obesity-related metabolic disorders.

Federated Learning has become an important distributed machine learning paradigm, enabling several clients to work together to jointly train a global model without exchanging raw data without sharing raw data. Although it improves data privacy, it also creates security vulnerabilities, particularly poisoning attacks conducted by malicious clients. Understanding these risks has become crucial for guaranteeing system robustness and dependability as federated learning is used more frequently in privacy-sensitive situations. This study investigates the impact of label flipping attacks in federated learning systems through experimental analysis. A distributed learning environment is constructed using multiple clients, and malicious participants are introduced to manipulate training data. This study evaluates the performance of the global model under different attack ratios, and further uses gradient analysis to observe abnormal update patterns caused by poisoned clients. Results show that poisoning attacks greatly affect model stability and accuracy. As the number of malicious clients increases, training becomes less stable and the convergence process is more easily disrupted. The findings highlight the vulnerability of federated learning systems and emphasize the necessity for robust defense mechanisms.

With growing global concern for women's safety, portable personal protection devices have gained increasing attention. However, traditional tools such as pepper spray and stun guns suffer from poor portability, high visibility, and complex operation, resulting in a real-world carrying rate below 15%. This study proposes a smart personal-safety hair tie integrating GPS, Bluetooth communication, and a high-decibel alarm into a fashionable hair accessory, enabling discreet and unobtrusive protection. We systematically analyze how five fabric materials—silk, linen, cotton, polyester, and wool—influence three core performances: Bluetooth Received Signal Strength Indicator (RSSI) signal strength, capacitive touch sensitivity, and wearing comfort based on pressure distribution. Using electromagnetic theory, capacitor physics, and solid mechanics, we establish quantitative correlations between fabric properties and device performance. We also verify a proximity detection method using Bluetooth RSSI fluctuations, achieving an overall accuracy of approximately 84.5%. The results indicate that silk is superior to all other fabrics across all three metrics, whereas wool is the least suitable. These findings provide a scientific basis for material selection in next-generation wearable safety devices.

Coherent Ising machines (CIMs) have emerged as a compelling hardware paradigm for accelerating the solution of complex combinatorial optimization problems. In this work, we demonstrate an opto-electronic CIM architecture implemented on a heterogeneous photonic integrated circuit (PIC) platform. Our design monolithically integrates silicon nitride (SiN) waveguides, for ultra-low optical loss, with lithium niobate (LN) for high-speed Mach-Zehnder modulation, alongside integrated germanium-silicon (Ge-Si) photodetectors (PDs). By leveraging the strong Pockels effect of the SiN-LN hybrid platform, we exploit the intrinsic nonlinear transfer function of the modulator to emulate Ising spin dynamics within an opto-electronic feedback loop, thereby bypassing the need for auxiliary nonlinear components. To validate this approach, we formulate a five-currency arbitrage problem as a quadratic unconstrained binary optimization (QUBO) task and map it onto the Ising Hamiltonian. Numerical simulations, driven by experimentally calibrated modulation characteristics, demonstrate robust convergence to the ground state, successfully identifying the optimal arbitrage path with a profit rate of 1.0793. These results highlight the potential of heterogeneous PICs for realizing compact, stable, and scalable CIMs, paving a promising way toward high-performance photonic accelerators for financial modeling and broader combinatorial optimization.