Hope speech detection in social media plays a critical role in fostering positivity, resilience, and inclusivity across online platforms. Traditional lexical-based approaches, while eficient, often struggle with the nuanced and context-dependent nature of hope expressions, particularly when masked by sarcasm or cultural references. In this paper, we present a fine-grained hope speech detection system leveraging a transformer-based architecture, specifically BERT (Bidirectional Encoder Representations from Transformers). Our approach is evaluated within the framework of the PolyHope Shared Task at IberLEF 2025, targeting both binary and multiclass classification subtasks in English texts. By fine-tuning pretrained BERT models and applying robust preprocessing techniques, our system efectively captures subtle linguistic cues and contextual semantics that are critical for distinguishing between genuine hope, unrealistic expectations, and sarcastic expressions. Experimental results demonstrate significant improvements over traditional baselines, highlighting the potential of deep contextual models for nuanced hope speech detection. We further provide a detailed error analysis and outline directions for future enhancements, emphasizing scalability, multilingual adaptability, and real-world deployment considerations.
In recent years, the detection of hope speech has emerged as a critical research area within the broader ifeld of afective computing and social media analysis. Hope speech, characterized by expressions of optimism, encouragement, and resilience, contributes positively to online discourse, fostering inclusivity, mental well-being, and community support. As social media platforms increasingly influence societal narratives, the ability to automatically identify and promote hopeful content has gained significant sociocultural and technological relevance. Beyond academic interest, detecting hope speech has tangible applications in mental health monitoring, crisis intervention, and fostering supportive online communities—areas where scalable and interpretable AI systems can drive meaningful societal impact.
Traditional machine learning methods, often relying on lexical features such as TF-IDF representations combined with linear classifiers, have demonstrated promising results in early hope speech detection tasks. However, these approaches typically struggle to capture the nuanced semantics, implicit sentiment structures, and contextual dependencies inherent in human expressions of hope—particularly when hope is conveyed through sarcasm, metaphor, or culturally specific references. Furthermore, traditional pipelines often falter under severe class imbalance, a persistent challenge in social media datasets where non-hopeful content dominates.
The advent of transformer-based architectures, such as BERT (Bidirectional Encoder Representations from Transformers), has revolutionized natural language processing (NLP) by enabling models to learn deep contextual representations of text. Leveraging self-attention mechanisms, BERT can efectively model complex linguistic phenomena, making it particularly well-suited for fine-grained classification tasks where subtle diferences in meaning are crucial. However, deploying large transformer models in resource-constrained environments remains challenging due to their computational overhead. This work bridges this gap by demonstrating that fine-tuned BERT, despite its complexity, can be optimized for real-world deployment through strategic hyperparameter tuning and class-balancing techniques.
In this work, we propose a transformer-based system for fine-grained hope speech detection, utilizing BERT to address the limitations of traditional lexical methods. Our approach was developed in the context of the PolyHope Shared Task at IberLEF 2025, which challenges participants to detect hope speech across both binary (Hope vs. Not Hope) and multiclass (Generalized Hope, Realistic Hope, Unrealistic Hope, Not Hope, and Sarcasm) classification settings, in English and Spanish social media texts. Our key contributions include: 1. A systematic evaluation of BERT’s capabilities for distinguishing nuanced hope categories, including sarcastic and metaphorical expressions. 2. A robust preprocessing and class-balancing pipeline tailored to social media data, addressing severe label imbalance through weighted loss functions and oversampling. 3. An actionable error analysis identifying persistent challenges (e.g., code-switching, overlapping class definitions) and proposing hybrid solutions for future work.
Through careful preprocessing, fine-tuning, and evaluation, we demonstrate that BERT-based models significantly outperform traditional pipelines, particularly in recognizing nuanced hope categories and handling sarcastic expressions. Our results highlight the potential of deep contextual models for advancing the state of hope speech detection and ofer insights into future directions for building more robust, scalable, and inclusive NLP systems.The remainder of this paper is structured as follows: Section 2 reviews related work, Section 3 details the dataset and preprocessing steps, Section 4 describes our methodology, Section 5 presents experimental results, Section 6 discusses implications and limitations, and Section 7 concludes with future directions.
Hope speech detection bridges sentiment analysis, emotion recognition, and low-resource NLP. Early
traditional lexical approaches relied heavily on TF-IDF features paired with linear classifiers like SVMs.
While these achieved moderate success (e.g., macro-F1: 0.817 in the PolyHope Shared Task [
The emergence of transformer-based models revolutionized the field. Fine-tuned BERT architectures
dramatically outperformed traditional methods by capturing contextual dependencies essential for
resolving ambiguities, such as distinguishing sincere hope from sarcastic remarks [
Recent work prioritizes multilingual and cultural inclusivity. Spanish social media studies revealed
unique challenges in LGBTQ+ communities, where hope intertwines with activism [
Hybrid and lightweight solutions balance accuracy and eficiency. Combining TF-IDF with distilled
transformers (e.g., TinyBERT) reduced inference latency by 60% while retaining 90% of BERT’s
performance [
The foundation of our study is the PolyHope Shared Task dataset (IberLEF 2025), comprising 12,988 English and Spanish social media posts curated from Twitter, Reddit, and Facebook. This corpus supports two annotation schemes: a binary classification distinguishing hopeful expressions from non-hopeful content, and a finer-grained multiclass categorization identifying five distinct hope-related phenomena—Generalized Hope, Realistic Hope, Unrealistic Hope, Not Hope, and Sarcasm.
Notable challenges emerged during initial analysis. The multiclass distribution exhibits significant imbalance, with Not Hope representing 65% of samples while critical minority categories like Sarcasm (5%) and Unrealistic Hope (8%) remain substantially underrepresented. Linguistic complexity further complicates analysis, as posts contain code-switching patterns, metaphorical language, sarcastic expressions, and informal platform-specific vernacular. To maintain benchmarking integrity, we preserved the dataset’s predefined 70% 15% 15% train/validation/test partitions throughout our experiments.
Social media text necessitates specialized preprocessing to handle noise while preserving linguistic signals. Our domain-adapted workflow executes six sequential operations:
First, anonymization removes user identifiers and URLs to eliminate non-linguistic artifacts. Next, encoding normalization converts HTML entities to standard characters and enforces UTF-8 consistency. Text sanitization follows, retaining alphanumeric characters, basic punctuation, and apostrophes (to preserve contractions) while discarding ambiguous symbols and emojis. Notably, hashtags are preserved as potential hope indicators.
Subsequent standardization applies lowercase conversion and whitespace reduction. The pipeline concludes with linguistic refinement: stopword filtration using NLTK’s language-specific lexicons, followed by WordNet lemmatization. This final step reduces inflectional variance while prioritizing semantic integrity—a critical consideration for distinguishing nuanced hope categories.
To address distribution skew, we implemented complementary countermeasures during training preparation. Class-weighted loss functions assign 20× higher penalty to misclassified Sarcasm instances versus Not Hope samples. Concurrently, SMOTE-based oversampling synthesizes new minority-class examples through feature-space interpolation, increasing Unrealistic Hope and Sarcasm representation by 30% in the training partition.
Input processing leverages the BERT-base-uncased tokenizer, selected for compatibility with informal social media text. Each post undergoes subword decomposition into WordPiece tokens, with sequences truncated or padded to 128 tokens—a threshold capturing 95% of posts while balancing computational load and context preservation. Attention masks isolate padding tokens during model training to prevent noise propagation.
The preprocessing architecture prioritizes accessibility and reproducibility. Entirely executable on CPU systems with 4GB RAM, our implementation eliminates GPU dependencies. A modular design decouples processing stages, enabling straightforward adaptation to new languages or platforms. By building exclusively on open-source tools (spaCy, NLTK, Hugging Face), we ensure researchers with limited computational resources can replicate and extend our workflow.
Our system leverages BERT-base-uncased, a 12-layer transformer model with 110M parameters, pretrained on English text. We fine-tune BERT for both binary and multiclass hope speech detection tasks using the following architecture: 1. BERT Encoder: Processes input tokens to generate contextualized embeddings. 2. Classification Head: A fully connected layer maps the pooled [CLS] token output to class probabilities.
• Binary Task: Sigmoid activation for Hope/Not Hope prediction.
• Multiclass Task: Softmax activation for 5-class categorization.
Rationale for BERT: Unlike static embeddings (e.g., Word2Vec), BERT’s bidirectional attention mechanism captures context-dependent semantics critical for distinguishing sarcasm (e.g., "Sure, I’m hopeful") from genuine hope.
To optimize performance on imbalanced social media data, we implement: 1. Dynamic Learning Rate Scheduling: 2. Class-Balanced Loss Functions: • Warm-up Phase: Linear increase from 0 to 2e-5 over 500 steps to stabilize early training. • Decay Phase: Linear reduction to 0 over subsequent steps to avoid overfitting. • Weighted Cross-Entropy: Penalize misclassifications of minority classes (e.g., Sarcasm) 20x more heavily than dominant classes (Not Hope). • Focal Loss: Experimentally applied to down-weight well-classified majority samples, though final results favored standard weighted loss. • Fixed batch size of 16 to balance GPU memory constraints and gradient stability. • Stratified sampling to ensure proportional representation of minority classes.
Key hyperparameters were selected via grid search on the validation set: Training Infrastructure:
• Hardware: NVIDIA Tesla T4 GPU (16GB VRAM). • Software: Hugging Face Transformers, PyTorch.
• Training Time: 45 minutes for binary task, 70 minutes for multiclass.
To mitigate bias toward the dominant Not Hope class: 1. Oversampling with SMOTE: Generated synthetic Sarcasm and Unrealistic Hope samples by interpolating nearest neighbors in TF-IDF space. 2. Threshold Adjustment: Optimized decision thresholds for minority classes during inference (e.g., Sarcasm threshold lowered to 0.3).
Impact: Oversampling improved multiclass F1 by 12% (from 0.108 to 0.120), while threshold adjustment reduced false negatives by 18%.
To enhance scalability, we: 1. Quantized the Model: Reduced BERT’s memory footprint by 4x (1.2GB → 300MB) via 8-bit quantization with minimal accuracy loss (<1%). 2. Cached Embeddings: Precomputed BERT embeddings for frequent terms (e.g., hope, wish) to accelerate inference by 40%.
Our BERT-based model achieves state-of-the-art performance on the PolyHope Shared Task, surpassing traditional methods in both binary and multiclass settings. Key results are summarized below:
To contextualize our model’s practicality, we compare its computational demands with TF-IDF + SVM: Table 4: Resource Consumption Comparison Implications: • BERT’s higher accuracy comes at a cost: 5.8x slower inference and 8x higher memory usage than traditional methods. • However, our optimizations (e.g., 8-bit quantization) reduce inference latency to 150 ms and
RAM usage to 4 GB, making real-time deployment feasible on mid-tier hardware.
Manual inspection of 200 misclassified samples reveals persistent challenges: 1. Sarcasm and Irony: 2. Code-Switching: 3. Short Texts: • Example: "Sure, I’m totally hopeful about this disaster" → Misclassified as Generalized Hope. • Root Cause: BERT struggles with implicit sentiment inversion without explicit negations. • Example: "InshaAllah, we’ll survive this storm " → Misclassified as Not Hope. • Root Cause: Limited multilingual pretraining data for Arabic-English code-switching. • Example: "Hope!" → Ambiguous between Generalized and Unrealistic Hope. • Root Cause: Insuficient context for nuanced classification.
Our results underscore two critical insights for hope speech detection: 1. Contextual Embeddings Are Essential: BERT’s ability to model phrases like "light at the end of the tunnel" (metaphorical hope) explains its 12% F1 gain over TF-IDF in binary classification. 2. Class Imbalance Requires Hybrid Solutions: While oversampling improved Sarcasm recall by 18%, integrating synthetic data generation (e.g., GPT-3 paraphrasing) could further bridge the gap.
• Computational Cost: BERT’s resource demands limit deployment in edge devices. • Cultural Bias: Performance drops on non-Western hope expressions (e.g., "Inshallah") highlight the need for multilingual adaptation.
Our experimental results establish that fine-tuned BERT models significantly advance hope speech detection capabilities in social media contexts, outperforming traditional TF-IDF-based approaches across both binary and multiclass classification tasks. The contextual depth of transformer embeddings proves particularly advantageous for interpreting semantic nuances—enabling more accurate identification of metaphorical expressions, culturally grounded hope, and sarcastic undertones that frequently challenge lexical methods.
Three critical findings emerge from the PolyHope evaluation: 1. Binary Classification Robustness: The model achieves a macro-F1 score of 0.825 in distinguishing hopeful from non-hopeful content, demonstrating reliable performance where simpler approaches falter. 2. Multiclass Recognition Gains: Despite the inherent complexity of fine-grained categorization, our BERT-based system triples the macro-F1 performance of conventional baselines, confirming its superiority in diferentiating subtle hope categories. 3. Persistent Linguistic Challenges: Error analysis reveals ongoing dificulties with implicitly expressed hope, sarcasm resolution, and ambiguous class boundaries. These limitations suggest that integrating specialized modules for sarcasm detection or domain-adaptive pretraining could yield further improvements.
These insights align with broader trends in afective computing, where transformer architectures consistently outperform handcrafted feature engineering in emotion recognition tasks. Nevertheless, the computational demands of large language models remain problematic for resource-constrained environments, underscoring the need for eficiency-focused adaptations in future work.
This study demonstrates the eficacy of BERT-based architectures for fine-grained hope speech detection across English and Spanish social media texts. Our approach consistently surpasses traditional lexical pipelines, achieving substantial improvements in both binary and multiclass classification tasks within the PolyHope Shared Task framework. By leveraging contextual language understanding, the system successfully identifies subtle distinctions between authentic hope expressions, unrealistic aspirations, and ironic statements. These findings afirm the transformative potential of pretrained language models in advancing nuanced emotion analysis.
To improve hope speech detection systems, we can focus on several key areas. First, we need to enhance how we model sarcasm, possibly by using new training methods that specifically target the way sentiment gets subtly inverted. We should also consider domain-specialized pretraining by creating large datasets focused on hopeful language. This will help our models adapt better to the unique characteristics of social media. Furthermore, expanding our systems to handle multiple languages is crucial, especially improving their ability to recognize code-switching and non-English expressions.
Another important direction is eficiency optimization . We can explore techniques like knowledge distillation, where a smaller model learns from a larger one (similar to DistilBERT), and quantization to make our systems more resource-eficient for widespread deployment. Finally, data enrichment through generating synthetic examples for categories that are currently underrepresented will help improve the robustness of our models. Addressing these areas will lead to more accurate, eficient, and culturally inclusive hope speech detection systems that can perform efectively in various real-world scenarios.
Generative AI tools were used for minor linguistic assistance. All intellectual, analytical, and interpretative contributions are those of the authors. No data, figures, or claims were generated by AI.
• GitHub,