This paper presents our system developed for the SemEval-2026 Task 7: Everyday KnowledgeAcross Diverse Languages and Cultures. on Subtask 1: Short Answer Questions (SAQ). on Subtask 2: Multiple-Choice Questions (MCQ). To this end, we focus on models’ cultural competence across 26 languages and 30 countries using four different versions large language models (LLMs): deepseek-v3.2-exp, qwen-max, qwen-plus, and qwen3-next-80ba3b-instruct. We experiment with 1) the trialand test dataset is analyzed visually, 2) use the large language generative model to perform generate or select the answer that it deems correct on the trial and test dataset through prompts, and 3) many prompt engineering approaches of generative models are evaluated on the trial dataset. We further study the influence of different hyperparameters on the generative model and select the best single model for the prediction of the test dataset. Our submission achieved the good ranking place in the test dataset leaderboard. For Subtask 1 (SAQ), the evaluation criteria for this task mainly consistof the aggregate results of the 23 languages: ar-EG, ar-MA, ar-SA, bg-BG, el-GR, en-AU, and so on, and they are measured using the accuracy score. For Subtask 2 (MCQ), this task is essentially a multiple-choice task for questions text. Performance will be evaluated using accuracy score. In other words, this subtask evaluated using accuracy score based on the correctness of the selected answer across different languages and cultural contexts. For Subtask 1 (SAQ) and Subtask 2 (MCQ), our best approach is to obtain the results in test dataset are accuracy score 51.4689 and accuracy score 80.26 separately. For the final ranking, organizers will use the aggregate results of accuracy score. Even so,our approach has yielded good results.

This paper presents our system developed for the SemEval-2026 Task 9: Detecting Multilingual, Multicultural and Multievent Online Polarization. on Subtask 1: Multilingual Text Classification Challenge - Polarization Detection. on Subtask 2: Multilingual Text Classification Challenge - Polarization Type Classification. on Subtask 3: Multilingual Text Classification Challenge - Manifestation Identification. For Subtask 1, we explored classical text representation approaches including Bag-of-Words, Word2Vec Average Vectors, and Bag-of-Centroids. Among these methods, the Bag-of-Centroids model achieved the best performance on both development and test datasets. For Subtask 2 and Subtask 3, we fine-tuned four different pre-trained language models: google-bert, FacebookAI-roberta, dccuchile-bert, and distilbert-multi. We experiment with 1) the training set data is analyzed visually, 2) multiple numbers of single models are trained on the training set data, and 3) multiple number of single models for voting weight ensemble learning. We further study the influence of different hyperparameters on the integrated model and select the best integration model for the prediction of the test set. On the official test set, our system achieved Macro-F1 scores of 0.6882 (EN) and 0.6711 (SP) for Subtask 1, 0.3752 (EN) and 0.6386 (SP) for Subtask 2, and 0.3561 (EN) and 0.4366 (SP) for Subtask 3. For the final ranking, organizers will use the Macro F1 score. These approachs has yielded good results.

This paper presents our system developed for the SemEval-2026 Task 1: MWAHAHA-Competition on Humor Generation. on Subtask A: Text-based Humor Generation. Given a set of text-based constraints, generate a joke. This subtask A will be conducted in English, Spanish, and Chinese. on Subtask B: Image-Based Caption Generation. This subtask explores humor in a multimodal context, combining visual inputs with text generation. This subtask B is in English only. To this end, we mainly focus on Subtask A: Text-based Humor Generation in English and Chinese, Subtask B: Image-BasedCaption Generation in English language to use two important languages models: BLIP and Qwen series LLM. For Task B1: Image-only Humor Generation and Task B2: Image and Prompt Humor Generation. Our submission achieved the good ranking place in the test set. All subtasks evaluated using Rating (95% CI) score across different languages and modality contexts. For Subtask A in English and Chinese, Rating score 950 and 1054, 95% CI [ 922, 982] and [1024, 1104], ranked 16th and 1st respectively. For Subtask B in B1 and B2, Rating score 976 and 987, 95% CI [ 941, 1007] and[948, 1016], ranked 5th and 3rd respectively. For the final ranking, organizers will use the Rating (95% CI) score. Even so, our approach still has yielded good results.

We describe our system for SemEval-2026 Task 10 on psycholinguistic conspiracy marker extraction and conspiracy detection from English texts. The shared task consists of two subtasks: (1) extracting conspiracy-related markers—actor, action, effect, victim, and evidence—evaluated using an overlap-based macro F1-score, and (2) detecting conspiracy content as a binary text classification problem evaluated using macro-averaged F1-score. Our approach relies on fine-tuning pre-trained transformer encoders, including multilingual DistilBERT variants and DeBERTa-v3, without using external corpora or data augmentation techniques. Experimental results show that our best models achieve a macro-F1 score of 0.1476 for Subtask~1 and a Weighted-F1 score of 0.7267 for Subtask~2. These results show that simple fine-tuning of pre-trained models provides a strong baseline for both marker extraction and conspiracy detection.

Our team was interested in content classification and labeling from multimodal detection of Hate speech, Humor, and Stance in marginalized socio-political movement discourse. We joined the task: Subtask A-Detection of Hate Speech and Subtask B-Classifying the Targets of Hate Speech. In this two task, our goal is to assign a content classification label to multimodal Hate Speech. Detection of Hate Speech: The aim is to detect the presence of hate speech in the images. The dataset for this task will have binary labels: No Hate and Hate. Classifying the Targets of Hate Speech: Given that an image is hateful, the goal here is to identify the targets of hate speech. The dataset here will have four labels: Undirected, Individual, Community, and Organization. Our group used a supervised learning method and a text prediction model. The best result on the test set for Subtask-A and Subtask-B were F1 score of 0.6209 and 0.3453, ranking twentieth and thirteenth among all teams.