Semi-supervised learning (SSL) has demonstrated high performance in image classification tasks by effectively utilizing both labeled and unlabeled data. However, existing SSL methods often suffer from poor calibration, with models yielding overconfident predictions that misrepresent actual prediction likelihoods. Recently, neural networks trained with mixup that linearly interpolates random examples from the training set have shown better calibration in supervised settings. However, calibration of neural models remains under-explored in SSL settings. Although effective in supervised model calibration, random mixup of pseudolabels in SSL presents challenges due to the overconfidence and unreliability of pseudolabels. In this work, we introduce CalibrateMix, a targeted mixup-based approach that aims to improve the calibration of SSL models while maintaining or even improving their classification accuracy. Our method leverages training dynamics of labeled and unlabeled samples to identify ''easy-to-learn'' and ''hard-to-learn'' samples, which in turn are utilized in a targeted mixup of easy and hard samples. Experimental results across several benchmark datasets show that our method achieves lower expected calibration error (ECE) and superior accuracy compared to existing SSL approaches.

Detecting hate speech on social media platforms is vital to mitigate technology-facilitated violence. Extensive research has targeted widely spoken languages like English, but there is a notable gap in studying hate speech detection in low-resource languages like Tamil. Additionally, with social media platforms now supporting various modalities, including text, speech, and video, effective techniques for hate speech detection in multimodal formats, especially videos, are crucial. However, detecting hate speech in Tamil presents unique challenges due to its morphology and code-mixing nature. This article presents a comprehensive approach for detecting hate speech in Tamil, with a focus on multimodal data. We introduce a new dataset, the MultimodAl Tamil Hate dataset, comprising videos along with their audio and textual transcripts, annotated with four categories of hate speech: offensive, sexist, racist, and casteist. To classify hate speech categories, we leverage transformer-based models. Through a series of experiments, we evaluated the performance of each modality individually and explored their fusion using a multimodal approach. BERT-based models were used for textual analysis to extract informative features, the TimeSformer model was employed for video modality, and Wav2Vec2 was used for audio modality. Specifically, we attained 81.82% accuracy and a 68.65% F1-score for the text modality, 63.63% accuracy and a 50.60% F1-score for the audio modality, and 45.45% accuracy and a 33.64% F1-score for the video modality. By integrating optimal combinations of models from each modality and employing machine learning classifiers, we achieved an accuracy of 81.82% and an F1-score of 66.67% in our hate speech classification task. Our research findings highlight the effectiveness of employing a multimodal approach for hate speech detection in Tamil, showcasing its efficacy in curbing the dissemination of hateful content on social media platforms.