This research aims to automatically classify social support threads shared by users on public social network platforms. Recognizing the growing role of social platforms in providing emotional support, this study focuses on identifying different support types using sentiment analysis. By incorporating data augmentation, three datasets were created for this analysis. We used existing models to compare classification performance. The results indicate that while data augmentation can sometimes improve model performance, the quality and context of the original data are crucial for accurately identifying emotional support. The study also found that current models struggle to distinguish subtle differences between support categories, especially when the data is imbalanced. These findings highlight the limitations of current AI models and the need for more balanced datasets and better classification techniques. This research contributes to the development of automated social support classification and provides a foundation for future studies to improve model accuracy using data augmentation and advanced methods.

Online social networking can be done through traditional media such as online forums or new media such as Twitter. Multi-party conversation is the most common form of discussion thread. This makes the social process very crowded and complicated, which is also a key challenge in applying natural language processing. To analyze the patterns of online social networking more precisely, this study builds a social intent dataset of the Chinese corpus of multi-party conversation. After performing the annotation task, the Fleiss kappa score in each social intent is acceptable. In this study, we design a social intent detection model with Transformers architecture, which is based on the interaction of responses. The decoder decodes the post and the response sequentially to enable the model learning interactions between the responses. Finally, we analyze whether the decoder improves the effectiveness of the model in detecting the social intents of the responses. As confirmed by the results in this study, the Transformers model achieved the best macro F1 scores for detecting most social intents.

Traditional comment sentiment analysis focuses solely on text and often results in inconsistencies between textual meanings and actual sentiments. A multimodal approach that considers both text and video content is employed to address this issue, recognizing that the video’s content influences comments. This study constructed a dataset comprising video content and corresponding comments from a video platform to detect valence and arousal sentiment. The video content was aligned and segmented into visual, audio, and dialogue segments to extract hidden features of multimodal models from a pretrained multimodal model. The cosine similarity approach has been used to identify the most relevant features of multimodal, which served as features for sentiment analysis. The proposed model enhanced with multimodal features demonstrated a 1.7% increase in macro-F1 for valence and a 3.2% increase in arousal compared to models without these multimodal features. This suggests that employing a multimodal approach significantly benefits video comment processing and enhances sentiment analysis accuracy.

Communicating in online social media has become the main way of socializing today. Different from face-to-face conversations in the real world, the meaning of online conversations can only be interpreted through text, and it is difficult to observe the hidden meaning. In addition, users often use pronouns or omit subject words in online social media. In this way, the topic of the discussion thread may be led in a direction unrelated to the article content due to wrong interpretation. Most previous studies have explored the topics contained in long texts, such as the entire article content, but have not conducted research on topic clustering of discussion thread data. Therefore, this paper uses the data from social media platforms to identify five entity categories, such as people, events, times, locations, and things, and cluster the discussion thread data into the topic. Finally establish a Chinese discussion thread topic dataset. In addition, this paper uses two methods to obtain response representation and six existing topic clustering models to conduct preliminary topic hybrid clustering experiments on the Chinese discussion thread topic dataset.

Cyberbullying is an important issue in recent years because the opinion of everyone can be easily circulated and discussed in the era of social media and Internet celebrity. However, this communication of convenience on the Internet exacerbates the diffusion of malicious comments and makes cyberbullying easier to happen. Cyberbullying leads to psychological trauma or suicide to victims. Cyberbullying belongs to a complex problem with many different factors which is quite difficult to detect. Most studies have focused on whether a comment contains cyberbullying behaviors or key words in an article or a sentence. Their research has not considered cyberbullying factors for each comment, and cannot explain what kind of reason to make the cyberbullying in the real world. Therefore, this paper proposes annotation rules to define six factors of cyberbullying, and these factors can help with follow-up analysis of cyberbullying patterns. The Chinese factors of cyberbullying corpus is collected from the largest forum, and manually annotated by the three experts. The Chinese factors of cyberbullying corpus will be used to train deep learning models to detect six factors of cyberbullying on each comment. In addition, the paper analyzes the cyberbullying factors on other larger data using the proposed model.

This research treats the stock turning point prediction as the imbalanced data classification problems and proposes the evolving weighted support vector machines (EW-SVM) system that leads to superior predictions upon the direction-of-change of the market. However, many parameters of the w-SVM model have to be decided by the user beforehand. Therefore, the EW-SVM system combining both w-SVM with GA is applied to forecast stock turning points. In the experimental results, the EW-SVM system is used to predict stock turning points and is compared to other prediction models including the SVM, DT, NB and k-NN models. These experimental results show that our EW-SVM system has the better performance among all the different approaches.

Information technology such as data mining or text mining, can extract useful data structure from data. But extracted data structure was hardly used in explaining the meaning of data itself. In reality, a laborer can hardly catch any meaningful job information from members of his work group (strong tie) when it comes to changing jobs. However, some of his friends (weak tie), although not so close to him, associated to other work groups, can bring him information about new jobs. This information is very important in job hunting. This kind of personal central base social networks can be used to realize a person's motivation and enrich his social capital, extends his resources to the resources of his linked friends. This process is called social chance discovery. Base on this concept, in this study, we build model of qualitative chance discovery and apply it on DVD news. Experiment results show this model can be used as an interactive tool to help company understands the way of breaking entry barrier to compete with existing companies in the harsh environment of DVD industry.

To analyze the competitive advantage, the analyst has to collect and analyze useful news. Therefore, the text mining technology is necessary for extracting information from the collected news. Both data mining and text mining are, as we know, frequency based methods; the low frequency data are usually deleted to discover the associative rules. As a result, some rare but actually important events for future are also deleted. Chance discovery, a discipline that focuses on revealing rare but important events, can solve the problem mentioned above to some extent. However, the canonical scenarios, which are used to assist in identifying rare events within the chance discovery process, are manually determined by participators. The work would be very complicated while data are numerous. A framework that fuses Porter's five forces and the strategy of competitive advantage has thus been proposed in this paper. Both concepts fused are utilized to automatically define scenarios. Furthermore, the scenarios determined by the proposed framework could be zoomed in / out under the direction of participators. In other words, the amount of data viewed by participators at a time is under control. Cases concerning the optical disk industry in Taiwan have been studied by using our framework. The experimental results indicated that CMC was as good as the Ritek and Prodisk in both CD and DVD techniques. Subsequently, both CMC and Ritek were better than Prodisk in the double layer products. Finally, CMC became the best in developing Blue-Ray products. These analyses verified the company whose competitive advantages are superior to others will have a great chance to success in the market.