Certain abilities of Transformer-based language models consistently emerge in their later layers. Previous research has leveraged this phenomenon to improve factual accuracy through self-contrast, penalizing early-exit predictions based on the premise that later-layer updates are more factually reliable than earlier-layer associations. We observe a similar pattern for fine-grained emotion classification in text, demonstrating that self-contrast can enhance encoder-based text classifiers. Additionally, we reinterpret self-contrast as a form of linear extrapolation, which motivates a refined approach that dynamically adjusts the contrastive strength based on the selected intermediate layer. Experiments across multiple models and emotion classification datasets show that our method outperforms standard classification techniques in fine-grained emotion classification tasks.

This paper describes the multimodal deep learning system proposed for SemEval 2022 Task 5: MAMI - Multimedia Automatic Misogyny Identification. We participated in both Subtasks, i.e. Subtask A: Misogynous meme identification, and Subtask B: Identifying type of misogyny among potential overlapping categories (stereotype, shaming, objectification, violence). The proposed architecture uses pre-trained models as feature extractors for text and images. We use these features to learn multimodal representation using methods like concatenation and scaled dot product attention. Classification layers are used on fused features as per the subtask definition. We also performed experiments using unimodal models for setting up comparative baselines. Our best performing system achieved an F1 score of 0.757 and was ranked 3^rd in Subtask A. On Subtask B, our system performed well with an F1 score of 0.690 and was ranked 10^th on the leaderboard. We further show extensive experiments using combinations of different pre-trained models which will be helpful as baselines for future work.

This paper describes our system used for SemEval 2022 Task 6: iSarcasmEval: Intended Sarcasm Detection in English and Arabic. We participated in all subtasks based on only English datasets. Pre-trained Language Models (PLMs) have become a de-facto approach for most natural language processing tasks. In our work, we evaluate the performance of these models for identifying sarcasm. For Subtask A and Subtask B, we used simple finetuning on PLMs. For Subtask C, we propose a Siamese network architecture trained using a combination of cross-entropy and distance-maximisation loss. Our model was ranked 7^th in Subtask B, 8^th in Subtask C (English), and performed well in Subtask A (English). In our work, we also present the comparative performance of different PLMs for each Subtask.

In this paper, we describe our system used for SemEval 2021 Task 5: Toxic Spans Detection. Our proposed system approaches the problem as a token classification task. We trained our model to find toxic words and concatenate their spans to predict the toxic spans within a sentence. We fine-tuned Pre-trained Language Models (PLMs) for identifying the toxic words. For fine-tuning, we stacked the classification layer on top of the PLM features of each word to classify if it is toxic or not. PLMs are pre-trained using different objectives and their performance may differ on downstream tasks. We, therefore, compare the performance of BERT, ELECTRA, RoBERTa, XLM-RoBERTa, T5, XLNet, and MPNet for identifying toxic spans within a sentence. Our best performing system used RoBERTa. It performed well, achieving an F1 score of 0.6841 and secured a rank of 16 on the official leaderboard.

In this paper, we describe our system used for SemEval 2021 Task 7: HaHackathon: Detecting and Rating Humor and Offense. We used a simple fine-tuning approach using different Pre-trained Language Models (PLMs) to evaluate their performance for humor and offense detection. For regression tasks, we averaged the scores of different models leading to better performance than the original models. We participated in all SubTasks. Our best performing system was ranked 4 in SubTask 1-b, 8 in SubTask 1-c, 12 in SubTask 2, and performed well in SubTask 1-a. We further show comprehensive results using different pre-trained language models which will help as baselines for future work.

In this paper, we describe our deep learning system used for SemEval 2020 Task 8: Memotion analysis. We participated in all the subtasks i.e Subtask A: Sentiment classification, Subtask B: Humor classification, and Subtask C: Scales of semantic classes. Similar multimodal architecture was used for each subtask. The proposed architecture makes use of transfer learning for images and text feature extraction. The extracted features are then fused together using stacked bidirectional Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU) model with attention mechanism for final predictions. We also propose a single model for predicting semantic classes (Subtask B) as well as their scales (Subtask C) by branching the final output of the post LSTM dense layers. Our model was ranked 5 in Subtask B and ranked 8 in Subtask C and performed nicely in Subtask A on the leader board. Our system makes use of transfer learning for feature extraction and fusion of image and text features for predictions.