We introduce HATELEXICON, a lexicon of slurs and targets of hate speech for Brazil, Germany, India and Kenya, to aid model development and interpretability. First, we demonstrate how HATELEXICON can be used to interpret model predictions, showing that models developed to classify extreme speech rely heavily on target group names. Further, we propose a culturally-informed method to aid shot selection for training in low-resource settings. In few-shot learning, shot selection is of paramount importance to model performance and we need to ensure we make the most of available data. We work with HASOC German and Hindi data for training and the Multilingual HateCheck (MHC) benchmark for evaluation. We show that selecting shots based on our lexicon leads to models performing better than models trained on shots sampled randomly. Thus, when given only a few training examples, using HATELEXICON to select shots containing more sociocultural information leads to better few-shot performance. With these two use-cases we show how our HATELEXICON can be used for more effective hate speech detection.

Dehumanization is a mental process that enables the exclusion and ill treatment of a group of people. In this paper, we present two data sets of dehumanizing text, a large, automatically collected corpus and a smaller, manually annotated data set. Both data sets include a combination of political discourse and dialogue from movie subtitles. Our methods give us a broad and varied amount of dehumanization data to work with, enabling further exploratory analysis as well as automatic classification of dehumanization patterns. Both data sets will be publicly released.

This study investigates the robustness and generalization of transformer-based models for automatic media bias detection. We explore the behavior of current bias classifiers by analyzing feature attributions and stress-testing with adversarial datasets. The findings reveal a disproportionate focus on rare but strongly connotated words, suggesting a rather superficial understanding of linguistic bias and challenges in contextual interpretation. This problem is further highlighted by inconsistent bias assessment when stress-tested with different entities and minorities. Enhancing automatic media bias detection models is critical to improving inclusivity in media, ensuring balanced and fair representation of diverse perspectives.

In the field of spoken language understanding, systems like Whisper and Multilingual Massive Speech (MMS) have shown state-of-the-art performances. This study is dedicated to a comprehensive exploration of the Whisper and MMS systems, with a focus on assessing biases in automatic speech recognition (ASR) inherent to casual conversation speech specific to the Portuguese language. Our investigation encompasses various categories, including gender, age, skin tone color, and geo-location. Alongside traditional ASR evaluation metrics such as Word Error Rate (WER), we have incorporated p-value statistical significance for gender bias analysis. Furthermore, we extensively examine the impact of data distribution and empirically show that oversampling techniques alleviate such stereotypical biases. This research represents a pioneering effort in quantifying biases in the Portuguese language context through the application of MMS and Whisper, contributing to a better understanding of ASR systems’ performance in multilingual settings.

Natural Language Processing techniques have been developed to assist in simplifying online content while preserving meaning. However, for low-resource languages, like Maltese, there are still numerous challenges and limitations. Lexical Simplification (LS) is a core technique typically adopted to improve content accessibility, and has been widely studied for high-resource languages such as English and French. Motivated by the need to improve access to Maltese content and the limitations in this context, this work set out to develop and evaluate an LS system for Maltese text. An LS pipeline was developed consisting of (1) potential complex word identification, (2) substitute generation, (3) substitute selection, and (4) substitute ranking. An evaluation data set was developed to assess the performance of each step. Results are encouraging and will lead to numerous future work. Finally, a single-blind study was carried out with over 200 participants, where the system’s perceived quality in text simplification was evaluated. Results suggest that meaning is retained about 50% of the time, and when meaning is retained, about 70% of system-generated sentences are either perceived as simpler or of equal simplicity to the original. Challenges remain, and this study proposes a number of areas that may benefit from further research.

Large language models are prone to internalize social biases due to the characteristics of the data used for their self-supervised training scheme. Considering their recent emergence and wide availability to the general public, it is mandatory to identify and alleviate these biases to avoid perpetuating stereotypes towards underrepresented groups. We present a novel prompt-tuning method for reducing biases in encoder models such as BERT or RoBERTa. Unlike other methods, we only train a small set of additional reusable token embeddings that can be concatenated to any input sequence to reduce bias in the outputs. We particularize this method to gender bias by providing a set of templates used for training the prompts. Evaluations on two benchmarks show that our method is on par with the state of the art while having a limited impact on language modeling ability.

This study pioneers the use of synthetically generated data for training generative models in document-level text simplification of German texts. We demonstrate the effectiveness of our approach with real-world online texts. Addressing the challenge of data scarcity in language simplification, we crawled professionally simplified German texts and synthesized a corpus using GPT-4. We finetune Large Language Models with up to 13 billion parameters on this data and evaluate their performance. This paper employs various methodologies for evaluation and demonstrates the limitations of currently used rule-based metrics. Both automatic and manual evaluations reveal that our models can significantly simplify real-world online texts, indicating the potential of synthetic data in improving text simplification.

Large Language models (LLMs), while powerful, exhibit harmful social biases. Debiasing is often challenging due to computational costs, data constraints, and potential degradation of multi-task language capabilities. This work introduces a novel approach utilizing ChatGPT to generate synthetic training data, aiming to enhance the debiasing of LLMs. We propose two strategies: Targeted Prompting, which provides effective debiasing for known biases but necessitates prior specification of bias in question; and General Prompting, which, while slightly less effective, offers debiasing across various categories. We leverage resource-efficient LLM debiasing using adapter tuning and compare the effectiveness of our synthetic data to existing debiasing datasets. Our results reveal that: (1) ChatGPT can efficiently produce high-quality training data for debiasing other LLMs; (2) data produced via our approach surpasses existing datasets in debiasing performance while also preserving internal knowledge of a pre-trained LLM; and (3) synthetic data exhibits generalizability across categories, effectively mitigating various biases, including intersectional ones. These findings underscore the potential of synthetic data in advancing the fairness of LLMs with minimal retraining cost.

In this paper, we report on a new corpus of simplified German. It is recently requested from public agencies in Germany to provide information in easy language on their outlets (e.g. websites) so as to facilitate participation in society for people with low-literacy levels related to learning difficulties or low language proficiency (e.g. L2 speakers). While various rule sets and guidelines for Easy German (a specific variant of simplified German) have emerged over time, it is unclear (a) to what extent authors and other content creators, including generative AI tools consistently apply them, and (b) how adequate texts in authentic Easy German really are for the intended audiences. As a first step in gaining insights into these issues and to further LT development for simplified German, we compiled DE-Lite, a corpus of easy-to-read texts including Easy German and comparable Standard German texts, by integrating existing collections and gathering new data from the web. We built n-gram models for an Easy German subcorpus of DE-Lite and comparable Standard German texts in order to identify typical features of Easy German. To this end, we use relative entropy (Kullback-Leibler Divergence), a standard technique for evaluating language models, which we apply here for corpus comparison. Our analysis reveals that some rules of Easy German are fairly dominant (e.g. punctuation) and that text genre has a strong effect on the distinctivity of the two language variants.

As a large how-to website, wikiHow’s mission is to empower every person on the planet to learn how to do anything. An important part of including everyone also linguistically is the use of gender-neutral language. In this short paper, we study in how far articles from wikiHow fulfill this criterion based on manual annotation and automatic classification. In particular, we employ a classifier to analyze how the use of gender-neutral language has developed over time. Our results show that although about 75% of all articles on wikiHow were written in a gender-neutral way from the outset, revisions have a higher tendency to add gender-specific language than to change it to inclusive wording.

This paper provides a comprehensive summary of the “Homophobia and Transphobia Detection in Social Media Comments” shared task, which was held at the LT-EDI@EACL 2024. The objective of this task was to develop systems capable of identifying instances of homophobia and transphobia within social media comments. This challenge was extended across ten languages: English, Tamil, Malayalam, Telugu, Kannada, Gujarati, Hindi, Marathi, Spanish, and Tulu. Each comment in the dataset was annotated into three categories. The shared task attracted significant interest, with over 60 teams participating through the CodaLab platform. The submission of prediction from the participants was evaluated with the macro F1 score.

The overview of the shared task on speech recognition for vulnerable individuals in Tamil (LT-EDI-2024) is described in this paper. The work comes with a Tamil dataset that was gath- ered from elderly individuals who identify as male, female, or transgender. The audio sam- ples were taken in public places such as marketplaces, vegetable shops, hospitals, etc. The training phase and the testing phase are when the dataset is made available. The task required of the participants was to handle audio signals using various models and techniques, and then turn in their results as transcriptions of the pro- vided test samples. The participant’s results were assessed using WER (Word Error Rate). The transformer-based approach was employed by the participants to achieve automatic voice recognition. This overview paper discusses the findings and various pre-trained transformer- based models that the participants employed.

This paper offers a detailed overview of the first shared task on “Multitask Meme Classification - Unraveling Misogynistic and Trolls in Online Memes,” organized as part of the LT-EDI@EACL 2024 conference. The task was set to classify misogynistic content and troll memes within online platforms, focusing specifically on memes in Tamil and Malayalam languages. A total of 52 teams registered for the competition, with four submitting systems for the Tamil meme classification task and three for the Malayalam task. The outcomes of this shared task are significant, providing insights into the current state of misogynistic content in digital memes and highlighting the effectiveness of various computational approaches in identifying such detrimental content. The top-performing model got a macro F1 score of 0.73 in Tamil and 0.87 in Malayalam.

We present an overview of the first shared task on “Caste and Migration Hate Speech Detection.” The shared task is organized as part of LTEDI@EACL 2024. The system must delineate between binary outcomes, ascertaining whether the text is categorized as a caste/migration hate speech or not. The dataset presented in this shared task is in Tamil, which is one of the under-resource languages. There are a total of 51 teams participated in this task. Among them, 15 teams submitted their research results for the task. To the best of our knowledge, this is the first time the shared task has been conducted on textual hate speech detection concerning caste and migration. In this study, we have conducted a systematic analysis and detailed presentation of all the contributions of the participants as well as the statistics of the dataset, which is the social media comments in Tamil language to detect hate speech. It also further goes into the details of a comprehensive analysis of the participants’ methodology and their findings.

This paper introduces an approach to stress identification in Tamil and Telugu, leveraging traditional machine learning models—Fasttext for Tamil and Naive Bayes for Telugu—yielding commendable results. The study highlights the scarcity of annotated data and recognizes limitations in phonetic features relevant to these languages, impacting precise information extraction. Our models achieved a macro F1 score of 0.77 for Tamil and 0.72 for Telugu with Fasttext and Naive Bayes, respectively. While the Telugu model secured the second rank in shared tasks, ongoing research is crucial to unlocking the full potential of stress identification in these languages, necessitating the exploration of additional features and advanced techniques specified in the discussions and limitations section.

This research focuses on Homophobia and Transphobia Detection in Dravidian languages, specifically Telugu, Kannada, Tamil, and Malayalam. Leveraging the Homophobia/ Transphobia Detection dataset, we propose an innovative approach employing a custom-designed tokenizer with a Bidirectional Long Short-Term Memory (BiLSTM) architecture. Our distinctive contribution lies in a tokenizer that reduces model sizes to below 7MB, improving efficiency and addressing real-time deployment challenges. The BiLSTM implementation demonstrates significant enhancements in hate speech detection accuracy, effectively capturing linguistic nuances. Low-size models efficiently alleviate inference challenges, ensuring swift real-time detection and practical deployment. This work pioneers a framework for hate speech detection, providing insights into model size, inference speed, and real-time deployment challenges in combatting online hate speech within Dravidian languages.

In this paper, we describe our approaches and results for Task 2 of the LT-EDI 2024 Workshop, aimed at detecting homophobia and/or transphobia across ten languages. Our methodologies include monolingual transformers and ensemble methods, capitalizing on the strengths of each to enhance the performance of the models. The ensemble models worked well, placing our team, MasonTigers, in the top five for eight of the ten languages, as measured by the macro F1 score. Our work emphasizes the efficacy of ensemble methods in multilingual scenarios, addressing the complexities of language-specific tasks.

Stress detection from social media texts has proved to play an important role in mental health assessments. People tend to express their stress on social media more easily. Analysing and classifying these texts allows for improvements in development of recommender systems and automated mental health assessments. In this paper, a GPT model is used for classification of social media texts into two classes - stressed and not-stressed. The texts used for classification are in two Dravidian languages - Tamil and Telugu. The results, although not very good shows a promising direction of research to use GPT models for classification.

This paper describes our homophobia/transphobia in social media comments detection system developed as part of the shared task at LT-EDI-2024. We took a transformer-based approach to develop our multiclass classification model for ten language conditions (English, Spanish, Gujarati, Hindi, Kannada, Malayalam, Marathi, Tamil, Tulu, and Telugu). We introduced synthetic and organic instances of script-switched language data during domain adaptation to mirror the linguistic realities of social media language as seen in the labelled training data. Our system ranked second for Gujarati and Telugu with varying levels of performance for other language conditions. The results suggest incorporating elements of paralinguistic behaviour such as script-switching may improve the performance of language detection systems especially in the cases of under-resourced languages conditions.

The exponential rise in social media users has revolutionized information accessibility and exchange. While these platforms serve various purposes, they also harbor negative elements, including hate speech and offensive behavior. Detecting hate speech in diverse languages has garnered significant attention in Natural Language Processing (NLP). This paper delves into hate speech detection in Tamil, particularly related to migration and refuge, contributing to the Caste/migration hate speech detection shared task. Employing a Convolutional Neural Network (CNN), our model achieved an F1 score of 0.76 in identifying hate speech and significant potential in the domain despite encountering complexities. We provide an overview of related research, methodology, and insights into the competition’s diverse performances, showcasing the landscape of hate speech detection nuances in the Tamil language.

Speech recognition is known to be a specialized application of speech processing. Automatic speech recognition (ASR) systems are designed to perform the speech-to-text task. Although ASR systems have been the subject of extensive research, they still encounter certain challenges when speech variations arise. The speaker’s age, gender, vulnerability, and other factors are the main causes of the variations in speech. In this work, we propose a fine-tuned speech recognition model for recognising the spoken words of vulnerable individuals in Tamil. This research utilizes a dataset sourced from the LT-EDI@EACL2024 shared task. We trained and tested pre-trained ASR models, including XLS-R and Whisper. The findings highlight that the fine-tuned Whisper ASR model surpasses the XLSR, achieving a word error rate (WER) of 24.452, signifying its superior performance in recognizing speech from diverse individuals.

We describe the second-place submission for the shared task organized at the Fourth Workshop on Language Technology for Equality, Diversity, and Inclusion (LT-EDI-2024). The task focuses on detecting caste/migration hate speech in Tamil. The included texts involve the Tamil language in both Tamil script and transliterated into Latin script, with some texts also in English. Considering different scripts, we examined the performance of 12 transformer language models on the dev set. Our analysis revealed that for the whole dataset, the model google/muril-large-cased performs the best. We used an ensemble of several models for the final challenge submission, achieving 0.81 for the test dataset.

Our work addresses the growing concern of abusive comments in online platforms, particularly focusing on the identification of Homophobia and Transphobia in social media comments. The goal is to categorize comments into three classes: Homophobia, Transphobia, and non-anti LGBT+ comments. Utilizing machine learning techniques and a deep learning model, our work involves training on a English dataset with a designated training set and testing on a validation set. This approach aims to contribute to the understanding and detection of Homophobia and Transphobia within the realm of social media interactions. Our team participated in the shared task organized by LTEDI@EACL 2024 and secured seventh rank in the task of Homophobia/Transphobia Detection in social media comments in Tamil with a macro- f1 score of 0.315. Also, our run was submitted for the English language and secured eighth rank with a macro-F1 score of 0.369. The run submitted for Malayalam language securing fourth rank with a macro- F1 score of 0.883 using the Random Forest model.

Commonly used language defines “hate speech” as objectionable statements that may jeopardize societal harmony by singling out a group or a person based on fundamental traits (including gender, caste, or religion). Using machine learning techniques, our research focuses on identifying hate speech in social media comments. Using a variety of machine learning methods, we created machine learning models to detect hate speech. An approximate Macro F1 of 0.60 was attained by the created models.

Hate speech refers to the offensive remarks against a community or individual based on inherent characteristics. Hate speech against a community based on their caste and native are unfortunately prevalent in the society. Especially with social media platforms being a very popular tool for communication and sharing ideas, people post hate speech against caste or migrants on social medias. The Shared Task LT–EDI 2024: Caste and Migration Hate Speech Detection was created with the objective to create an automatic classification system that detects and classifies hate speech posted on social media targeting a community belonging to a particular caste and migrants. Datasets in Tamil language were provided along with the shared task. We experimented with several traditional models such as Naive Bayes, Support Vector Machine (SVM), Logistic Regression, Random Forest Classifier and Decision Tree Classifier out of which Support Vector Machine yielded the best results placing us 8th in the rank list released by the organizers.

This paper presents our submission for Shared task on Stress Identification in Dravidian Languages: StressIdent LT-EDI@EACL2024. The objective of this task is to identify stress levels in individuals based on their social media content. The system is tasked with analysing posts written in a code-mixed language of Tamil and Telugu and categorising them into two labels: “stressed” or “not stressed.” Our approach aimed to leverage feature extraction and juxtapose the performance of widely used traditional, deep learning and transformer models. Our research highlighted that building a pipeline with traditional classifiers proved to significantly improve their performance (0.98 and 0.93 F1-scores in Telugu and Tamil respectively), surpassing the baseline as well as deep learning and transformer models.

Meme is a very popular term prevailing among almost all social media platforms in recent days. A meme can be a combination of text and image whose sole purpose is meant to be funny and entertain people. Memes can sometimes promote misogynistic content expressing hatred, contempt, or prejudice against women. The Shared Task LT–EDI 2024: Multitask Meme Classification: Unraveling Misogynistic and Trolls in Online Memes Task 1 was created with the purpose to classify social media memes as “misogynistic” and “Non - Misogynistic”. The task encompassed Tamil and Malayalam datasets. We separately classified the textual data using Multinomial Naive Bayes and pictorial data using ResNet50 model. The results of from both data were combined to yield an overall result. We were ranked 2nd for both languages in this task.

Homophobia and transphobia are terms which are used to describe the fear or hatred towards people who are attracted to the same sex or people whose psychological gender differs from his biological sex. People use social media to exert this behaviour. The increased amount of abusive content negatively affects people in a lot of ways. It makes the environment toxic and unpleasant to LGBTQ+ people. The paper talks about the classification model for classifying the contents into 3 categories which are homophobic, transphobic and nonhomophobic/ transphobic. We used many traditional models like Support Vector Machine, Random Classifier, Logistic Regression and KNearest Neighbour to achieve this. The macro average F1 scores for Malayalam, Telugu, English, Marathi, Kannada, Tamil, Gujarati, Hindi are 0.88, 0.94, 0.96, 0.78, 0.93, 0.77, 0.94, 0.47 and the rank for these languages are 5, 6, 9, 6, 8, 6, 6, 4.

This paper presents our submission for the shared task on Caste and Migration Hate Speech Detection: LT-EDI@EACL 20241 . This text classification task aims to foster the creation of models capable of identifying hate speech related to caste and migration. The dataset comprises social media comments, and the goal is to categorize them into negative and positive sentiments. Our approach explores back-translation for data augmentation to address sparse datasets in low-resource Dravidian languages. While Part-of-Speech (POS) tagging is valuable in natural language processing, our work highlights its ineffectiveness in Dravidian languages, with model performance drastically reducing from 0.73 to 0.67 on application. In analyzing boosting and ensemble methods, the voting classifier with traditional models outperforms others and the boosting techniques, underscoring the efficacy of simper models on low-resource data despite augmentation.

The widespread use of online communication has caused a significant increase in the spread of hate speech on social media. However, there are also hate crimes based on caste and migration status. Despite several nations efforts to bring equality among their citizens, numerous crimes occur just based on caste. Migration-based hostility happens both in India and in developed countries. A shared task was arranged to address this issue in a low-resourced language such as Tamil. This paper aims to improve the detection of hate speech and hostility based on caste and migration status on social media. To achieve this, this work investigated several Machine Learning (ML), Deep Learning (DL), and transformer-based models, including M-BERT, XLM-R, and Tamil BERT. Experimental results revealed the highest macro f1-score of 0.80 using the M-BERT model, which enabled us to rank 3rd on the shared task.

In this paper, the main goal of the study is to create an automatic speech recognition (ASR) system that is tailored to the Tamil language. The dataset that was employed includes audio recordings that were obtained from vulnerable populations in the Tamil region, such as elderly men and women and transgender individuals. The pre-trained model Rajaram1996/wav2vec2- large-xlsr-53-tamil is used in the engineering of the ASR system. This existing model is finetuned using a variety of datasets that include typical Tamil voices. The system is then tested with a specific test dataset, and the transcriptions that are produced are sent in for assessment. The Word Error Rate is used to evaluate the system’s performance. Our system has a WER of 37.733.

In recent years, there has been a persistent focus on developing systems that can automatically identify the hate speech content circulating on diverse social media platforms. This paper describes the team “Transformers” submission to the Caste and Migration Hate Speech Detection in Tamil shared task by LT-EDI 2024 workshop at EACL 2024. We used an ensemble approach in the shared task, combining various transformer-based pre-trained models using majority voting. The best macro average F1-score achieved was 0.82. We secured the 1st rank in the Caste and Migration Hate Speech in Tamil shared task.

Caste and Migration speech refers to the use of language that distinguishes the offense, violence, and distress on their social, caste, and migration status. Here, caste hate speech targets the imbalance of an individual’s social status and focuses mainly on the degradation of their caste group. While the migration hate speech imposes the differences in nationality, culture, and individual status. These speeches are meant to affront the social status of these people. To detect this hate in the speech, our task on Caste and Migration Hate Speech Detection has been created which classifies human speech into genuine or stimulate categories. For this task, we used multiple classification models such as the train test split model to split the dataset into train and test data, Logistic regression, Support Vector Machine, MLP (multi-layer Perceptron) classifier, Random Forest classifier, KNN classifier, and Decision tree classification. Among these models, The SVM gave the highest macro average F1 score of 0.77 and the average accuracy for these models is around 0.75.

Detection of Homophobia and Transphobia in social media comments serves as an important step in the overall development of Equality, Diversity and Inclusion (EDI). In this research, we describe the system we formulated while participating in the shared task of Homophobia/ Transphobia detection as a part of the Fourth Workshop On Language Technology For Equality, Diversity, Inclusion (LT-EDI- 2024) at EACL 2024. We used an ensemble of three state-of-the-art multilingual transformer models, namely Multilingual BERT (mBERT), Multilingual Representations for Indic Languages (MuRIL) and XLM-RoBERTa to detect the presence of Homophobia or Transphobia in YouTube comments. The task comprised of datasets in ten languages - Hindi, English, Telugu, Tamil, Malayalam, Kannada, Gujarati, Marathi, Spanish and Tulu. Our system achieved rank 1 for the Spanish and Tulu tasks, 2 for Telugu, 3 for Marathi and Gujarati, 4 for Tamil, 5 for Hindi and Kannada, 6 for English and 8 for Malayalam. These results speak for the efficacy of our ensemble model as well as the data augmentation strategy we adopted for the detection of anti-LGBT+ language in social media data.

The pervasive impact of stress on individuals necessitates proactive identification and intervention measures, especially in social media interaction. This research paper addresses the imperative need for proactive identification and intervention concerning the widespread influence of stress on individuals. This study focuses on the shared task, “Stress Identification in Dravidian Languages,” specifically emphasizing Tamil and Telugu code-mixed languages. The primary objective of the task is to classify social media messages into two categories: stressed and non stressed. We employed various methodologies, from traditional machine-learning techniques to state-of-the-art transformer-based models. Notably, the Tamil-BERT and Telugu-BERT models exhibited exceptional performance, achieving a noteworthy macro F1-score of 0.71 and 0.72, respectively, and securing the 15^th position in Tamil code-mixed language and the 9^th position in the Telugu code-mixed language. These findings underscore the effectiveness of these models in recognizing stress signals within social media content composed in Tamil and Telugu.

Machine learning and deep learning models have shown great potential in detecting hate speech from social media posts. This study focuses on the homophobia and transphobia detection task of LT-EDI-2024 in English. Several machine learning models, a Deep Neural Network (DNN), and the Bidirectional Encoder Representations from Transformers (BERT) model have been trained on the provided dataset using different feature vectorization techniques. We secured top rank with the best macro-F1 score of 0.4963, which was achieved by fine-tuning the BERT model on the English test set.

Identifying an individual where he/she is stressed or not stressed is our shared task topic. we have used several machine learning models for identifying the stress. This paper presents our system submission for the task 1 and 2 for both Tamil and Telugu dataset, focusing on us- ing supervised approaches. For Tamil dataset, we got highest accuracy for the Support Vector Machine model with f1-score of 0.98 and for Telugu dataset, we got highest accuracy for Random Forest algorithm with f1-score of 0.99. By using this model, Stress Identification System will be helpful for an individual to improve their mental health in optimistic manner.

Misogynistic memes are a category of memes which contain disrespectful language targeting women on social media platforms. Hence, detecting such memes is necessary in order to maintain a healthy social media environment. To address the challenges of detecting misogynistic memes, “Multitask Meme classification - Unraveling Misogynistic and Trolls in Online Memes: LT-EDI@EACL 2024” shared task organized at European Chapter of the Association for Computational Linguistics (EACL) 2024, invites researchers to develop models to detect misogynistic memes in Tamil and Malayalam. The shared task has two subtasks, and in this paper, we - team MUCS, describe the learning models submitted to Task 1 - Identification of Misogynistic Memes in Tamil and Malayalam. As memes represent multi-modal data of image and text, three models: i) Bidirectional Encoder Representations from Transformers (BERT)+Residual Network (ResNet)-50, ii) Multilingual Representations for Indian Languages (MuRIL)+ResNet-50, and iii) multilingual BERT (mBERT)+ResNet50, are proposed based on joint representation of text and image, for detecting misogynistic memes in Tamil and Malayalam. Among the proposed models, mBERT+ResNet-50 and MuRIL+ ResNet-50 models obtained macro F1 scores of 0.73 and 0.87 for Tamil and Malayalam datasets respectively securing 1st rank for both the datasets in the shared task.

Homophobic/Transphobic (H/T) content includes hatred and discriminatory comments directed at Lesbian, Gay, Bisexual, Transgender, Queer (LGBTQ) individuals on social media platforms. As this unfavourable perception towards LGBTQ individuals may affect them physically and mentally, it is necessary to detect H/T content on social media. This demands automated tools to identify and address H/T content. In view of this, in this paper, we - team MUCS describe the learning models submitted to “Homophobia/Transphobia Detection in social media comments:LT-EDI@EACL 2024” shared task at European Chapter of the Association for Computational Linguistics (EACL) 2024. The learning models: i) Homo_Ensemble - an ensemble of Machine Learning (ML) algorithms trained with Term Frequency-Inverse Document Frequency (TFIDF) of syllable n-grams in the range (1, 3), ii) Homo_TL - a model based on Transfer Learning (TL) approach with Bidirectional Encoder Representations from Transformers (BERT) models, iii) Homo_probfuse - an ensemble of ML classifiers with soft voting trained using sentence embeddings (except for Hindi), and iv) Homo_FSL - Few-Shot Learning (FSL) models using Sentence Transformer (ST) (only for Tulu), are proposed to detect H/T content in the given languages. Among the models submitted to the shared task, the models that performed better for each language include: i) Homo_Ensemble model obtained macro F1 score of 0.95 securing 4th rank for Telugu language, ii) Homo_TL model obtained macro F1 scores of 0.49, 0.53, 0.45, 0.94, and 0.95 securing 2nd, 2nd, 1st, 1st, and 4th ranks for English, Marathi, Hindi, Kannada, and Gujarathi languages, respectively, iii) Homo_probfuse model obtained macro F1 scores of 0.86, 0.87, and 0.53 securing 2nd, 6th, and 2nd ranks for Tamil, Malayalam, and Spanish languages respectively, and iv) Homo_FSL model obtained a macro F1 score of 0.62 securing 2nd rank for Tulu dataset.

The results of the Shared Task on Speech Recognition for Vulnerable Individuals in Tamil (LT-EDI-2024) are discussed in this paper. The goal is to create an automated system for Tamil voice recognition. The older population that speaks Tamil is the source of the dataset used in this task. The proposed ASR system is designed with pre-trained model akashsivanandan/wav2vec2-large-xls-r300m-tamil-colab-final. The Tamil common speech dataset is utilized to fine-tune the pretrained model that powers our system. The suggested system receives the test data that was released from the task; transcriptions are then created for the test samples and delivered to the task. Word Error Rate (WER) is the evaluation statistic used to assess the provided result based on the task. Our Proposed system attained a WER of 29.297%.