Abstracts:Current large language models display limited emotional intelligence, often mimicking affective patterns without genuine understanding, which raises manipulation and safety risks. We argue that artificial intelligence (AI) should prioritize long-term human well-being over short-term engagement, and that advancing toward artificial general intelligence (AGI) requires embedding fundamental drives and artificial emotions in model architectures. Building on Lazarus’s cognitive-rational theory, we propose a framework with an emotional module and a rational module, where artificial drives guide affective appraisal and decision-making. This enables alignment of artificial emotions with core values—such as human well-being, fairness, and environmental preservation—anchoring AI safety at the architectural level rather than through post hoc fixes. We discuss technical and ethical challenges, including data needs and reward modeling, and call for open science and regulation to ensure human-centered AGI.

Abstracts:We concisely review the history of large language models in the business context, and then examine some of the problems that this technology brings when applied to generative AI, notably from an environmental perspective. We highlight the implications and impacts of these problems in different communities locally and globally, and suggest possible solutions, mostly advocating for more international regulations. We urge caution with the widespread application of the technology, given the potential for negative impact, especially affecting marginalized populations and the Global South.

Abstracts:Recent advances in large language models (LLMs) have revolutionized natural language processing, achieving impressive performance across a wide range of linguistic tasks. However, these successes often mask a critical limitation: Current evaluation paradigms provide little insight into how well LLMs handle linguistic abstractions, the very cognitive capability that underlies generalization, analogy-making, and systematic reasoning. Without a principled framework for evaluating abstraction, it remains unclear whether LLMs truly engage in abstractions and their nature, how consistently they do so, and to what extent these behaviors reflect genuine abstraction capabilities versus surface-level pattern matching. We propose a structured taxonomy of linguistic abstractions in natural language processing, spanning levels from morphology to knowledge graphs (KGs), organized along two key dimensions: linguistic granularity and contextual dependence. This taxonomy supports a more nuanced evaluation of LLMs’ abstraction capability and helps identify where current models fall short. In particular, we highlight the limitations of LLMs at higher levels of abstraction—such as semantic, topical, taxonomic, and KG levels—where relational composition, context sensitivity, and symbolic structures are critical. To remedy these weaknesses, we advocate for the integration of neurosymbolic artificial intelligence (AI) systems that combine neural representations with symbolic reasoning.

Abstracts:This position article examines trustworthy machine learning with foundation models by investigating the four essential aspects. In learning, we describe how pretraining, fine-tuning, and reinforcement learning enable models to acquire generalizable knowledge and emphasize the importance of high-quality, unbiased training data, as well as robust training methods. In reasoning, we summarize the fundamental methodology of training-free, post-training, and test-time scaling methods that enhance logical deduction, reasoning transparency, and systematic safety. In planning, we incorporate neurosymbolic methods that combine adaptable neural capabilities with formally verifiable symbolic reasoning, ensuring safe and accountable decision making. In multimodality, we investigate the need for multimodal integration, where aligning information from different sensory input sources is important to mitigate biases and errors. The article presents an interdisciplinary vision for incorporating capability, robustness, safety, and explainability to establish trustworthy foundation models, paving the way for their reliable deployment in real-world applications, such as financial and clinical decision making.

Abstracts:This study addresses the unique challenges of sentiment analysis in Chinese massive open online course (MOOC) reviews, where pedagogically embedded language, intra-sentence sentiment shifts, and class imbalance complicate classification tasks. To tackle these domain-specific issues, we integrated ChatGPT-based data augmentation with contrastive learning within a Bidirectional Encoder Representations from Transformers (BERT)–Chinese framework. We evaluated ChatGPT-based augmentation (GPTaug), similar word replacement, and random word deletion under a dual-loss setup that combines supervised cross-entropy and InfoNCE (information noise-constrastive estimation) contrastive learning, focusing on how they enhance model performance across sentiment categories. The results revealed that the integration of contrastive learning with data augmentation strategies substantially improved sentiment classification in Chinese MOOC reviews. Especially, GPTaug demonstrated robust and balanced performance across polarity categories, particularly enhancing the detection of underrepresented neutral sentiments. These findings suggest that generative augmentation, when aligned with contrastive objectives, mitigates data sparsity and semantic ambiguity in educational sentiment analysis.

Abstracts:Large language models have transformed sentiment analysis, yet balancing accuracy, efficiency, and explainability remains a critical challenge. This study presents the first comprehensive evaluation of DeepSeek-R1—an open source reasoning model—against OpenAI’s GPT-4o and GPT-4o-mini. We test the full 671B model and its distilled variants, systematically documenting few-shot learning curves. Our experiments show DeepSeek-R1 achieves a 91.39% F1 score on five-class sentiment and 99.31% accuracy on binary tasks with just five shots, an eightfold improvement in few-shot efficiency over GPT-4o. Architecture-specific distillation effects emerge, where a 32B Qwen2.5-based model outperforms the 70B Llama-based variant by 6.69 percentage points. While its reasoning process reduces throughput, DeepSeek-R1 offers superior explainability via transparent, step-by-step traces, establishing it as a powerful, interpretable open source alternative.

Abstracts:Detecting aggressive driving is challenging but crucial for public safety. Existing methods rely on time-series data of drivers’ physiology, behavior, and vehicle movement but overlook driver’s emotion and environmental influences. We propose ARISE, a multisource aggregation model integrating physiological, behavioral, and emotional data, vehicle sensor inputs, and environmental conditions. ARISE employs multisource feature extraction, multimodal fusion, and a classifier to detect aggressive driving. Unlike graph-based methods that fails to detect gradual aggression shifts or transformer-based methods prone to delays, ARISE explicitly models vehicle state continuity and the aggressive driving environment. Motion similarity descriptor tracks state transitions, while aggression descriptor quantifies environmental aggression. Additionally, a driving performance descriptor assesses driving workload and stability. Experiments show that ARISE significantly outperforms state-of-the-art methods in aggressive driving detection.

Abstracts:Countering online hate speech is essential for creating a safer digital space where positive interactions can thrive. As central hubs of global communication, platforms like social media platforms require effective moderation through explainable and affective computing approaches. This study introduces a novel artificial intelligence-driven system for detecting misogynstic discourse. We collected 11,245 YouTube video uniform resource locators using specific keywords, then extracted audio to create Urdu transcripts and transliterated them into Roman Urdu, resulting in two distinct datasets. Various feature sets were explored using classic machine learning and deep learning algorithms. The results showed that classical models achieved 0.90 accuracy on the Urdu dataset, while deep learning models reached 0.96 accuracy on Roman Urdu. The corpus is publicly available to promote transparency and further research. Comparative evaluations against existing English hate speech dataset demonstrate the effectiveness of the proposed approach. This work lays the foundation for more ethical and transparent content moderation systems.

Abstracts:Aspect-based sentiment analysis (ABSA) aims to identify aspect terms in texts and determine their sentiment polarities. The in-context learning paradigm, powered by large language models, has proven effective in low-resource scenarios, where the retrieval of effective demonstration examples is crucial. Existing retrieval methods prioritize semantic and syntactic similarities, overlooking the fact that sentiment is often driven by its underlying causes. Recognizing that similar causes tend to yield similar sentiments, we propose the semantic-causal contextual demonstration retrieval (SCCDR), a demonstration retriever that integrates semantic and syntactic information while explicitly modeling sentiment causes. SCCDR was trained using contrastive learning based on rich contextual signals, including semantics, aspect-sentiment relationships, syntactic structures, and sentiment causes. Experiments on four datasets show that SCCDR outperforms other retrieval methods, thereby effectively improving ABSA performance under the ICL paradigm.

Abstracts:Depression is a prevalent mental health issue, and early detection is crucial for effective intervention. In this article, we propose the depression large language model (DLLM), a novel two-stage fine-tuning framework designed to enhance the accuracy and robustness of depression detection using multimodal data from social media. In the first stage, we design specific prompts to incorporate various types of multimodal data and collect diverse instruction data. The DLLM is then fine-tuned for depression detection based on these data. In the second stage, we enhance the model’s robustness and generalization by performing behavioral alignment. This involves a deep understanding of user actions to improve behavior perception, enabling the policy model to distinguish between positive and negative behaviors for individual users. Experiments on the WU3D dataset show that DLLM outperforms state-of-the-art baselines (e.g., +6.2% accuracy over ALBERT, +2.4% F1 over EKG-MDDM) and demonstrates strong generalization in ablation studies.