GHaLIB: A Multilingual Framework for Hope Speech Detection in Low-Resource Languages

1. Problem Statement

The paper addresses a significant gap in Natural Language Processing (NLP): the underrepresentation of Hope Speech detection, particularly in low-resource languages like Urdu.

• Current Limitations: Most existing research focuses on English or negative sentiment analysis (hate speech, sarcasm). While transformer models excel at detecting hate speech, they are rarely applied to hope speech.
• The Challenge: Hope speech is linguistically complex. It often shares vocabulary with sarcasm or despair (e.g., "can win" can be hopeful or sarcastic depending on context). Furthermore, low-resource languages lack the large, annotated datasets required to train robust deep learning models from scratch.
• The Goal: To develop a scalable, multilingual framework capable of accurately detecting hope speech across diverse linguistic settings, with a specific focus on bridging the resource gap for languages like Urdu.

2. Methodology

The authors propose GHaLIB, a hybrid framework that combines language-specific encoders with a multilingual transformer backbone to handle cross-lingual generalization and class imbalance.

A. Dataset

• Source: The PolyHope-M 2025 benchmark corpus.
• Languages: Urdu, English, German, and Spanish.
• Classes: Four categories:
  1. Generalized Hope (belief things will improve).
  2. Realistic Hope (achievable goals).
  3. Unrealistic Hope (impossible outcomes).
  4. Not Hope (negativity, doubt, or unrelated text).
• Data Characteristics: The dataset exhibits class imbalance (dominated by "Not Hope"). Exploratory analysis revealed cultural and stylistic differences; for instance, English "Realistic Hope" tends to be longer and action-oriented, while Urdu "Hope" relies more on religious and motivational terminology.

B. Architecture & Pipeline

1. Preprocessing: Text is normalized and tokenized using the XLM-RoBERTa-base tokenizer. Sequences are truncated or padded to 128 tokens.
2. Hybrid Encoder Strategy:
   • The system identifies the input language and assigns it to a specialized encoder:
     • Urdu: UrduBERT (bilingual Roman Urdu BERT).
     • English: RoBERTa.
     • German/Spanish: EuroBERT.
   • These encoders generate dense embeddings which are then concatenated and fed into a multilingual backbone (XLM-RoBERTa-base) for final classification.
   • Note: Larger variants (XLM-RoBERTa-large) were tested but the Base version was selected for its superior cost-efficiency on low-end GPUs.
3. Training & Optimization:
   • Class Imbalance Handling: A custom WeightedTrainer applies a 1.5x weight to the positive (hope) class in the cross-entropy loss function to reduce false negatives.
   • Hyperparameter Tuning: Utilized Optuna over 30 trials to optimize learning rates (5e-6 to 5e-5), batch sizes, warm-up ratios, weight decay, and dropout rates.
   • Threshold Tuning: Classification thresholds (0.3–0.8) were tuned on the validation set to balance False Positives (FP) and False Negatives (FN).
4. Evaluation: Models were evaluated using Accuracy, Precision, Recall, and Weighted F1-score on a hidden benchmark via CodaLab.

3. Key Contributions

• Multilingual Framework: Introduced GHaLIB, a novel architecture that integrates language-specific encoders with a multilingual backbone, specifically targeting low-resource scenarios.
• Focus on Urdu: Provided the first robust deep learning approach for Urdu hope speech detection, addressing a critical gap in South Asian NLP.
• Comprehensive Benchmarking: Evaluated multiple architectures (XLM-RoBERTa, mBERT, EuroBERT, UrduBERT, RoBERTa, DistilBERT, GloVe) across both binary (Hope vs. Not Hope) and multi-class (4 categories) tasks.
• Open Source: Released code, preprocessing scripts, and resources to facilitate reproducibility and future research.

4. Experimental Results

The framework demonstrated state-of-the-art performance, particularly in binary classification.

Binary Classification (Hope vs. Not Hope):

• Urdu: Achieved 95.0% Accuracy and 95.0% F1-score using XLM-RoBERTa + UrduBERT.
• German: 87.4% F1-score.
• English: 86.3% F1-score.
• Spanish: 85.0% F1-score.

Multi-Class Classification (4 Categories):

• Urdu: Achieved 65.2% F1 (macro).
• English: 71.0% F1 (macro).
• German: 70.1% F1 (macro).
• Spanish: 68.5% F1 (macro).

Key Observations:

• Language-Specific Advantage: Language-specific models (e.g., UrduBERT for Urdu, EuroBERT for German/Spanish) significantly outperformed generic multilingual models in morphologically rich and low-resource languages.
• Resource Disparity: Performance correlated with dataset size; English and European languages (with larger annotated corpora) generally outperformed Urdu in multi-class settings.
• Robustness: The model maintained high performance despite challenges like overlapping vocabulary and code-mixing, though multi-class tasks remained harder due to ambiguity.

5. Significance and Future Work

• Constructive Digital Discourse: By enabling the detection of hope speech, this framework supports the creation of tools that foster positive online communication and counteract the dominance of hate speech detection tools.
• Low-Resource Viability: The study proves that existing multilingual models, when fine-tuned with language-specific encoders and weighted loss functions, can effectively operate in low-resource environments without requiring massive new datasets.
• Future Directions:
  • Expansion to other Pakistani regional languages (Punjabi, Seraiki, Sindhi).
  • Implementation of domain-specific pretraining and parameter-efficient fine-tuning (PEFT) to further enhance scalability and adaptability.

In conclusion, GHaLIB establishes a new benchmark for hope speech detection, demonstrating that a hybrid approach combining specialized encoders with multilingual transformers is the most effective strategy for handling the linguistic and cultural nuances of hope in diverse, low-resource languages.