Transformer-based models are highly vulnerable to adversarial attacks, where even small perturbations can cause significant misclassifications. This paper introduces *I-Guard*, a defense framework to increase the robustness of transformer-based models against adversarial perturbations. *I-Guard* leverages model interpretability to identify influential parameters responsible for adversarial misclassifications. By selectively fine-tuning a small fraction of model parameters, our approach effectively balances performance on both original and adversarial test sets. We conduct extensive experiments on English and code-mixed Hinglish datasets and demonstrate that *I-Guard* significantly improves model robustness. Furthermore, we demonstrate the transferability of *I-Guard* in handling other character-based perturbations.

Whilst large language models (LLMs) have made significant advances in every natural language processing task, studies have shown that these models are vulnerable to small perturbations in the inputs, raising concerns about their robustness in the real-world. Given the rise of misinformation online and its significant impact on society, fact verification is one area in which assessing the robustness of models developed for this task is crucial. However, the robustness of LLMs in fact verification remains largely unexplored. In this paper, we introduce FactEval, a novel large-scale benchmark for extensive evaluation of LLMs in the fact verification domain covering 17 realistic word-level and character-level perturbations and 4 types of subpopulations. We investigate the robustness of several LLMs in zero-shot, few-shot, and chain-of-thought prompting. Our analysis using FEVER, one of the largest and most widely-used datasets for fact verification, reveals that LLMs are brittle to small input changes and also exhibit performance variations across different subpopulations.

Toxic content detection plays a vital role in addressing the misuse of social media platforms to harm people or groups due to their race, gender or ethnicity. However, due to the nature of the datasets, systems develop an unintended bias due to the over-generalization of the model to the training data. This compromises the fairness of the systems, which can impact certain groups due to their race, gender, etc.Existing methods mitigate bias using data augmentation, adversarial learning, etc., which require re-training and adding extra parameters to the model.In this work, we present a robust and generalizable technique BiasWipe to mitigate unintended bias in language models. BiasWipe utilizes model interpretability using Shapley values, which achieve fairness by pruning the neuron weights responsible for unintended bias. It first identifies the neuron weights responsible for unintended bias and then achieves fairness by pruning them without loss of original performance. It does not require re-training or adding extra parameters to the model. To show the effectiveness of our proposed technique for bias unlearning, we perform extensive experiments for Toxic content detection for BERT, RoBERTa, and GPT models. .

With the growing popularity of code-mixed data, there is an increasing need for better handling of this type of data, which poses a number of challenges, such as dealing with spelling variations, multiple languages, different scripts, and a lack of resources. Current language models face difficulty in effectively handling code-mixed data as they primarily focus on the semantic representation of words and ignore the auditory phonetic features. This leads to difficulties in handling spelling variations in code-mixed text. In this paper, we propose an effective approach for creating language models for handling code-mixed textual data using auditory information of words from SOUNDEX. Our approach includes a pre-training step based on masked-language-modelling, which includes SOUNDEX representations (SAMLM) and a new method of providing input data to the pre-trained model. Through experimentation on various code-mixed datasets (of different languages) for sentiment, offensive and aggression classification tasks, we establish that our novel language modeling approach (SAMLM) results in improved robustness towards adversarial attacks on code-mixed classification tasks. Additionally, our SAMLM based approach also results in better classification results over the popular baselines for code-mixed tasks. We use the explainability technique, SHAP (SHapley Additive exPlanations) to explain how the auditory features incorporated through SAMLM assist the model to handle the code-mixed text effectively and increase robustness against adversarial attacks.

Deep learning models have been proven vulnerable towards small imperceptible perturbed input, known as adversarial samples, which are indiscernible by humans. Initial attacks in Natural Language Processing perturb characters or words in sentences using heuristics and synonyms-based strategies, resulting in grammatical incorrect or out-of-context sentences. Recent works attempt to generate contextual adversarial samples using a masked language model, capturing word relevance using leave-one-out (LOO). However, they lack the design to maintain the semantic coherency for aspect based sentiment analysis (ABSA) tasks. Moreover, they focused on resource-rich languages like English. We present an attack algorithm for the ABSA task by exploiting model explainability techniques to address these limitations. It does not require access to the training data, raw access to the model, or calibrating a new model. Our proposed method generates adversarial samples for a given aspect, maintaining more semantic coherency. In addition, it can be generalized to low-resource languages, which are at high risk due to resource scarcity. We show the effectiveness of the proposed attack using automatic and human evaluation. Our method outperforms the state-of-art methods in perturbation ratio, success rate, and semantic coherence.

Social media platforms such as Twitter have evolved into a vast information sharing platform, allowing people from a variety of backgrounds and expertise to share their opinions on numerous events such as terrorism, narcotics and many other social issues. People sometimes misuse the power of social media for their agendas, such as illegal trades and negatively influencing others. Because of this, sentiment analysis has won the interest of a lot of researchers to widely analyze public opinion for social media monitoring. Several benchmark datasets for sentiment analysis across a range of domains have been made available, especially for high-resource languages. A few datasets are available for low-resource Indian languages like Hindi, such as movie reviews and product reviews, which do not address the current need for social media monitoring. In this paper, we address the challenges of sentiment analysis in Hindi and socially relevant domains by introducing a balanced corpus annotated with the sentiment classes, viz. positive, negative and neutral. To show the effective usage of the dataset, we build several deep learning based models and establish them as the baselines for further research in this direction.

Nowadays, the spread of Internet memes on online social media platforms such as Instagram, Facebook, Reddit, and Twitter is very fast. Analyzing the sentiment of memes can provide various useful insights. Meme sentiment classification is a new area of research that is not explored yet. Recently SemEval provides a dataset for meme sentiment classification. As this dataset is highly imbalanced, we extend this dataset by annotating new instances and use a sampling strategy to build a meme sentiment classifier. We propose a multi-modal framework for meme sentiment classification by utilizing textual and visual features of the meme. We found that for meme sentiment classification, only textual or only visual features are not sufficient. Our proposed framework utilizes textual as well as visual features together. We propose to use the attention mechanism to improve meme classification performance. Our proposed framework achieves macro F1 and accuracy of 34.23 and 50.02, respectively. It increases the accuracy by 6.77 and 7.86 compared to only textual and visual features, respectively.

Due to the phenomenal growth of online content in recent time, sentiment analysis has attracted attention of the researchers and developers. A number of benchmark annotated corpora are available for domains like movie reviews, product reviews, hotel reviews, etc. The pervasiveness of social media has also lead to a huge amount of content posted by users who are misusing the power of social media to spread false beliefs and to negatively influence others. This type of content is coming from the domains like terrorism, cybersecurity, technology, social issues, etc. Mining of opinions from these domains is important to create a socially intelligent system to provide security to the public and to maintain the law and order situations. To the best of our knowledge, there is no publicly available tweet corpora for such pervasive domains. Hence, we firstly create a multi-domain tweet sentiment corpora and then establish a deep neural network based baseline framework to address the above mentioned issues. Annotated corpus has Cohen’s Kappa measurement for annotation quality of 0.770, which shows that the data is of acceptable quality. We are able to achieve 84.65% accuracy for sentiment analysis by using an ensemble of Convolutional Neural Network (CNN), Long Short Term Memory (LSTM), and Gated Recurrent Unit(GRU).