The analysis of emotional speech has gained significant attention in the fields of s peech r ecognition a nd natural language processing. From emotion recognition to emotional text-to-speech synthesis, emotional speech plays a crucial role, particularly in areas such as human-computer interaction and intelligent robotics. However, this area remains underexplored. Recent research trends emphasize the use of multimodal data, such as emotional audio and video recordings. Although effective, these approaches require additional resources, which can be time-consuming and costly, especially for low-resource languages such as Serbian. On the other hand, a significant g ap exists in understanding cognitive processes involved in human emotional speech production. To address this, emotional speech from an information-theoretic perspective was explored. Specifically, surprisal values, estimated using five s tate-of-the-art language models were analyzed for their correlation with spoken word duration. The results indicated variations in Pearson's coefficient between these parameters in different emotional states, with general multilingual models outperforming Serbian-specific models in surprising estimation. These results can offer valuable insights into emotional speech production in other South Slavic languages as well, such as in Croatian, Bosnian, and Montenegrin.

Social networks have become an integral part of modern society, allowing users to express their thoughts, opinions, and feelings, and engage in discussions on various topics. The vast amount of user-generated content on these platforms provides a valuable source of data for sentiment analysis (SA), which is the computational analysis of opinions and sentiments expressed in text. However, most existing deep learning models for SA rely on minimizing the cross-entropy loss, which does not incorporate any knowledge of the sentiment of labels themselves. To address this limitation, a novel approach that utilizes an optimal transport-based loss function to improve sentiment analysis performance was proposed. Optimal transport (OT) metrics are fundamental theoretical properties for histogram comparison, and the proposed loss function uses the cost of the OT plan between ground truth and outputs of the classifier. The experimental results demonstrate that this approach can significantly reduce miss detections between positive and negative classes and suggest that using an OT-based loss function can effectively overcome the deficiency of existing SA models and improve their performance in real-world applications.

Intimacy is one of the fundamental aspects of our social life. It relates to intimate interactions with others, often including verbal self-disclosure. In this paper, we researched machine learning algorithms for quantification of the intimacy in the tweets. A new multilingual textual intimacy dataset named MINT was used. It contains tweets in 10 languages, including English, Spanish, French, Portuguese, Italian, and Chinese in both training and test datasets, and Dutch, Korean, Hindi, and Arabic in test data only. In the first experiment, linear regression models combine with the features and word embedding, and XLM-T deep learning model were compared. In the second experiment, cross-lingual learning between languanges was tested. In the third experiments, data was clustered using K-means. The results indicate that XLM-T pre-trained embedding might be a good choice for an unsupervised learning algorithm for intimacy detection.