Temporal reasoning in financial texts is essential for understanding event timing and claim validity, especially in earnings conference calls and social media discussions. While transformer-based models have advanced natural language processing, the comparative performance of fine-tuned encoder models and prompt-based decoder models in multilingual temporal classification remains underexplored. This study systematically compares model types, model sizes, and prompting strategies across two tasks: detecting temporal references in English texts and assessing claim validity in Chinese posts. Encoder models such as RoBERTa and BERT and decoder models such as GPT-4o, Mistral, and Gemma are evaluated using fine-tuning and few-shot prompting approaches. Results show that fine-tuned encoder models achieve consistently strong performance across both English and Chinese datasets. Mid-sized prompt-based decoder models also perform competitively under well-designed prompts, offering a practical alternative when fine-tuning is not feasible. In addition, decoder models are more robust to class imbalance, as reflected by smaller gaps between Micro-F1 and Macro-F1 scores. However, decoder models perform less effectively on Chinese tasks, indicating the need for language-specific adaptation. These findings provide practical guidance for selecting models and designing prompts for financial natural language processing under resource constraints.
Large Language Models (LLMs) exhibit impressive capabilities in natural language understanding and generation; however, their ability to interpret and integrate nonverbal emotional cues, such as facial expressions, while maintaining security and interpretability remains underexplored. This study investigates how multimodal LLMs, specifically Qwen2.5-VL and Deepseek-VL, respond to conversational prompts paired with facial expression images through the lens of neural-symbolic integration. We constructed a dataset of 10,000 conversational lines combined with real and synthetic facial expressions depicting various emotional states. Using both automated sentiment analysis and human evaluations based on a 5-point Likert scale, we assessed model responses for tone appropriateness, helpfulness, and emotional alignment. Our results indicate that neural-symbolic integration significantly enhances interpretability and robustness against adversarial inputs, enabling models to achieve a higher average similarity with human interpretations (cosine similarity peaking around 0.7–0.9) compared to Qwen2.5-VL (0.45–0.55). However, both models struggled to accurately interpret subtle or mismatched emotional cues. These findings highlight the potential of neural-symbolic integration to improve the security and emotional reasoning of AI systems.
This study examines whether targeted feedback can lead to improvements in nonverbal communication, specifically in speech delivery, by analyzing 3-min videos of three management consultants. The analysis focused on facial expressions, utilizing the Facial Action Coding System (FACS) and vocal characteristics, as represented by Mel-Frequency Cepstral Coefficients (MFCCs). For five months, the authors provided individualized feedback focusing on facial expression and speech delivery, which involves key elements such as speaking pace, intonation, and pitch variation. While the sample size was limited to three individuals, the study leveraged a year’s worth of accumulated video data. The findings indicate that feedback contributed to observable improvements in facial expressions, whereas vocal features derived from MFCCs remained largely unaffected. Furthermore, this study suggests that combining FACS with real-time speech content extraction offers a deeper understanding of how message content and emotional expression interact, and using the T5 model to generate feedback on speech content automatically improves the speech.
Social media platforms like X (formerly Twitter) play a central role in public discourse but are also exploited for influence operations (IO) through coordinated inauthentic behavior (CIB). This study proposes a method to detect IO-related coordinated communities during the August 2023 release of Advanced Liquid Processing System (ALPS)-treated water from the Fukushima Daiichi Nuclear Power Plant. Using reposting data, we construct a graph of user communities based on network science techniques, incorporating both intra- and inter-community features. A Graph Neural Networks (GNN) is trained on these structures to classify communities as abnormal or normal. The model achieves F1 = 0.97, outperforming baseline methods. By automating early detection of coordinated communities, our method supports timely countermeasures in IO. While effective, further work is needed to capture communities with varied intents and improve attribution. The results demonstrate the potential of graph-based learning in real-time monitoring of influence activities on social platforms.
Managing the privacy of social media posts remains a complex task, especially as audience diversity and content sensitivity grow. We propose a comprehensive privacy management framework that combines post content features with behavioral signals from social interactions to deliver personalized audience recommendations. Leveraging Facebook and Reddit datasets, we implement five core modules: post privacy classification, persona contradiction detection, interaction and privacy alignment scoring, expectation mismatch analysis, and privacy-aware friend grouping. Our post classifier achieves F1-scores of 0.76 (Facebook) and 0.72 (Reddit); contradiction detection yields an F1-score of 0.80 by combining behavioral and BERT-based features. Friend clustering based on interaction and alignment scores results in silhouette scores of 0.65 (Facebook) and 0.60 (Reddit), while expectation mismatch analysis reveals stronger emotional alignment in highly private posts. Overall, our approach enables explainable and behavior-sensitive audience control for social platforms, improving upon prior content- or interaction-only methods.
