GHaLIB: A Multilingual Framework for Hope Speech Detection in Low-Resource Languages
Ahmed Abdullah, Haroon Mahmood, Sana Fatima · 2025
Read GHaLIB: A Multilingual Framework for Hope Speech Detection in Low-Resource Languages by Ahmed Abdullah, Haroon Mahmood, Sana Fatima (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2512-22705.txt.
Introduction
Ahmed Abdullah, Sana Fatima, and Haroon Mahmood present GHaLIB, a multilingual system for detecting hope speech, the encouraging counterpart of hate speech that natural language processing has mostly ignored. Research concentrates on English, leaving low-resource languages like Urdu without tools for fostering positive online discussion. The authors focus on Urdu while testing whether proven transformer architectures transfer to this kinder classification task.
They fine-tuned pretrained multilingual models, XLM-RoBERTa, mBERT, EuroBERT, and UrduBERT, with simple preprocessing. On the PolyHope-M 2025 benchmark, the system reached an F1 score above ninety-five percent for Urdu binary classification and about sixty-five percent for the harder multi-class version. Spanish, German, and English results were similarly competitive, suggesting current multilingual models can serve low-resource communities for hope speech detection.
Catalog record
| CREATOR | Ahmed Abdullah, Haroon Mahmood, Sana Fatima |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 26 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2512-22705
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
GHaLIB: A Multilingual Framework for Hope Speech Detection in Low-Resource Languages Ahmed Abdullah, Haroon Mahmood, Sana Fatima Abstract Hope speech has been relatively underrepresented in Natural Language Processing (NLP). Current studies are largely focused on English, which has resulted in a lack of resources for low-resource languages such as Urdu. As a result, the creation of tools that facilitate positive online communication remains limited. Although transformer-based architectures have proven to be effective in detecting hate and offensive speech, little has been done to apply them to hope speech or, more generally, to test them across a variety of linguistic settings. This paper presents a multilingual framework for hope speech detection with a focus on Urdu. Using pretrained transformer models such as XLM-RoBERTa, mBERT, EuroBERT, and UrduBERT, we apply simple preprocessing and train classifiers for improved results. Evaluations on the PolyHope-M 2025 benchmark demonstrate strong performance, achieving F1-scores of 95.2% for Urdu binary classification and 65.2% for Urdu multi-class classification, with similarly competitive results in Spanish, German, and English. These results highlight the possibility of implementing existing multilingual models in low-resource environments, thus making it easier to identify hope speech and helping to build a more constructive digital discourse. I Introduction Understanding public sentiment across multiple domains, such as social media, news, and customer feedback, is an important activity in sentiment analysis within Natural Language Processing (NLP). However, analyzing sentiments that involve sarcasm and hope speech presents a different type of challenge due to the mixed nature of natural language. Sarcasm is particularly elusive, as it involves utterances that convey the contrary of what is literally meant [ 15 ] . Most sentiment analysis models crumble when tackling sarcasm and false hope; they frequently misclassify the underlying sarcasm as emotions that are neutral or positive, while the statement actually conveys a very negative connotation [ 12 ] . As in the case of the examples in Figure 1 , ”can win” and ”still win,” expressions may be conveyed in a contrary manner within any given context. Hence, the author’s intentions may often not be detected from surface features of the text. The same words may give encouragement in one instance and sarcasm or hopelessness in another, making it difficult for a classifier to determine whether a piece of text truly expresses hope or the exact opposite [ 14 ] [ 3 ] . Therefore, the classification usually leads to greater uncertainties and misclassifications, which emphasizes the need for deeper contextual and pragmatic understanding in automated systems. Fig. 1: Uncertain prediction caused by shared wording in hope sentiments. The following categories of Hope Speech are generally used in the literature: • Generalized Hope: A belief that things will get better in future. Example: ”Things will improve over time.” This type of hope reflects a belief that things will improve, without pointing to a specific action or result. It is often expressed in uncertain situations where the future is unpredictable and serves to comfort others [ 5 ] . • Realistic Hope: Hope for something possible to achieve. Example: ”I can pass the exam if I work hard for it.” Realistic hope is based on goals that are possible to reach. It connects optimism with effort or conditions that make success likely, making it practical and useful in real situations [ 5 ] . • Unrealistic Hope: Hope for something unlikely or impossible to achieve. Example: ”If I jump high enough, I can reach the moon.” This type of hope focuses on outcomes that cannot realistically happen. While it may still give emotional relief, it often creates false expectations [ 5 ] . • Not Hope: Texts that do not show hope. Example: ”Nothing is going to change.” This category includes statements that lack optimism and instead show negativity, doubt, or resignation. It is important for separating hopeful from non-hopeful content in analysis. It can also refer to as the text that not even related to hopelessness context [ 5 ] . The main contributions of this work are: • We propose a multilingual framework for hope speech detection with a focus on Urdu and English. • We evaluate multiple transformer models and machine learning classifiers on the PolyHope-M 2025 dataset. • We achieve strong results for Urdu and English, and show competitive performance in other languages. • We release our code and resources at GHaLIB . The rest of the paper is organized as follows. Section II reviews related work and literature. Section III presents exploratory data analysis, highlighting how different features can contribute to the task. Section IV describes our methodology. Section V reports experiments and results. Section VI discusses future work and provides the conclusion. II Literature Review II-A Early Feature-Based Approaches to Hope Speech Detection Lexicon-based approaches have been future-oriented in speech evaluation with a focus on the initial development of the lexicon and have relied heavily on pre-defined lists of positive terms. The limitations of hard-coded vocabulary included failing to perceive the language’s contextual sense, open-world benchmarks and the various nuances, which led the lexicon-based approaches lead to very high false positive incidents [ 33 ] . Palakodety et al. (2019) [ 22 ] suggested one of the first computational studies in detecting hope speech, focusing on comments on Hindi and English YouTube channels in response to the 2019 India-Pakistan crisis. The constructed hope speech corpus was trained with a logistic regression classifier using L2-regularization by the authors, having n-gram features along with another linguistic indicator. Palakodety ran experiments using an 80/10/10 split for training, validation, and testing, and averaged the results over 100 random splits to make sure the results were reliable. The configuration that worked ideally combined n-grams, indicator features, and FastText embeddings, leading to an F1-score of 78.5%, precision of 82.0%, recall of 75.3%, and an AUC of 95.5% on self-curated benchmark. These results overlapped an arbitrary standard for a general-purpose sentiment analysis system. These results significantly outperformed a general-purpose sentiment analysis baseline of Stanford CoreNLP, highlighting that hope speech detection is a distinct challenge requiring dedicated methods. Their work established early benchmarks and demonstrated that targeted feature engineering combined with classical machine learning can bring strong results on the hope classification tasks. II-B Traditional Machine Learning Models With the emergence of traditional Machine Learning (ML), researchers began employing algorithms such as Support Vector Machines (SVM), Naive Bayes, and logistic regression for hope speech detection. Balouchzahi et al. (2023) [ 5 ] implemented a logistic regression model for Urdu hope speech classification, attaining an F1-score of 75.9% in binary classification and a F1 (macro) score of 48.0% in multiclass settings. Similar to this, Sidorov et al. (2024) [ 28 ] demonstrated the use of various ML techniques, in addition to deep learning, for multiclass hope speech detection in Spanish and German. Their experiments reported F1-scores of 68.0% for Spanish and 69.8% for German through classical machine learning classifiers, proving their viability in the task. Yet, although these are still promising results, traditional ML models in general suffered in multilingual and low-resource contexts, often relying heavily on handcrafted features and language-specific preprocessing. These limitations prompted the shift of the field toward embedding-based and deep learning architectures for more robust and scalable hope speech detection [ 28