GHaLIB is a multilingual framework for hope speech detection designed to address linguistic nuance, cultural variation, and data scarcity in low-resource languages. The framework combines language-specific encoders with a shared multilingual transformer backbone to improve both binary and fine-grained (multiclass) hope speech classification across multiple languages.
This work was developed and evaluated as part of the PolyHope-M 2025 shared task and accepted at ACIT 2025.
Hope speech is a distinct NLP task that goes beyond sentiment analysis. While sentiment focuses on present emotional polarity, hope speech captures future-oriented belief, often expressed through subtle, context-dependent language. This makes it especially difficult to detect in multilingual and low-resource settings.
Key challenges addressed by GHaLIB include:
- Over-reliance of existing research on English
- Limited datasets for hope speech in low-resource languages
- Lexical overlap between hope, sarcasm, and negative statements
- Failure of traditional machine learning methods to capture context and pragmatics
- Lack of scalable multilingual frameworks for nuanced hope detection
The framework follows the standard four-class taxonomy used in PolyHope:
-
Generalized Hope: Broad belief that things will improve
Example: “Things will get better over time.” -
Realistic Hope: Hope grounded in plausible conditions or effort
Example: “I can pass the exam if I study hard.” -
Unrealistic Hope: Hope for an implausible or impossible outcome
Example: “If I jump high enough, I can reach the moon.” -
Not Hope: Absence of hope or unrelated content
Example: “Nothing is going to change.”
GHaLIB integrates language-aware and multilingual representations in a single pipeline:
- Language Identification routes input text to an appropriate encoder.
- Language-Specific Encoders capture morphological, syntactic, and cultural features:
- Urdu: UrduBERT
- English: RoBERTa
- German & Spanish: EuroBERT
- Shared Multilingual Backbone: XLM-RoBERTa provides cross-lingual semantic alignment.
- Fusion Strategy: CLS embeddings from the language-specific encoder and XLM-RoBERTa are concatenated and passed to a context-aware classifier.
This design enables strong performance in both high-resource and low-resource settings while remaining computationally feasible.
Experiments are conducted on the PolyHope-M 2025 multilingual corpus, which includes annotated social media text in:
- Urdu
- English
- German
- Spanish
The dataset exhibits notable class imbalance, with Not Hope being the dominant category, and significant variation in text length and vocabulary across languages.
- Stratified 70/15/15 train–validation–test split
- Weighted cross-entropy loss to mitigate class imbalance
- Hyperparameter tuning using Optuna (30 trials)
- Maximum sequence length: 128 tokens
- Training performed on 2×16GB NVIDIA T4 GPUs
- Fixed random seeds for reproducibility
- Final evaluation conducted on the official hidden PolyHope-M benchmark
| Language | F1-score |
|---|---|
| Urdu | 95.0 |
| English | 86.3 |
| German | 87.4 |
| Spanish | 85.0 |
Binary classification benefits from shared multilingual representations and shows strong generalization across languages, with particularly high performance for Urdu.
| Language | Macro F1-score |
|---|---|
| Urdu | 65.2 |
| English | 71.0 |
| German | 70.1 |
| Spanish | 68.5 |
Multiclass classification remains challenging due to overlapping vocabulary and subtle pragmatic distinctions between hope categories. Nevertheless, GHaLIB achieves state-of-the-art results across all evaluated languages.
- Language-specific encoders significantly improve performance for morphologically rich and low-resource languages, particularly Urdu and German.
- Multilingual transformers generalize well in binary classification tasks.
- English benefits from richer pretraining resources, but hybrid modeling reduces the performance gap for low-resource languages.
- Most classification errors arise from code-mixed text and contextual ambiguity rather than model instability.
- Code-mixed inputs remain difficult to classify reliably.
- Multiclass hope detection is inherently ambiguous due to subjective and pragmatic factors.
- Language identification errors can occur, though their impact is mitigated by the multilingual backbone.
These limitations reflect open challenges in hope speech detection rather than architectural deficiencies.
- Extension to additional low-resource languages such as Punjabi, Sindhi, and Seraiki
- Improved handling of code-mixing and pragmatic ambiguity
- Exploration of parameter-efficient fine-tuning and domain-specific pretraining
- Investigation of more advanced representation fusion strategies
If you use this work, please cite:
@inproceedings{abdullah2025ghalib,
title={GHaLIB: A Multilingual Framework for Hope Speech Detection in Low-Resource Languages},
author={Ahmed Abdullah and Haroon Mahmood and Sana Fatima},
booktitle={International Arab Conference on Information Technology (ACIT)},
year={2025}
}