The Register Gap: A Meaning Intelligence Framework for Nigerian Public Discourse
Proposes the nine-dimensional Meaning Intelligence Framework (MIF) to distinguish surface sentiment from true intent in Nigerian discourse; zero-shot accuracy 33.3%, schema-guided 73.3%.
Key Findings
Methodology
This paper introduces the Meaning Intelligence Framework (MIF), a comprehensive nine-dimension annotation and evaluation schema designed to analyze Nigerian public discourse. The framework includes dimensions such as register, surface sentiment, true intent, irony, coded subtext, risk tier, annotator confidence, speaker emotion, and recommended communication action. The authors constructed a 30-item calibration dataset covering Standard English, Nigerian English, Nigerian Pidgin, and code-mixed registers. Using the state-of-the-art language model Gemini 2.5 Flash, they conducted evaluations under two conditions: zero-shot prompting and schema-informed prompting, where the latter incorporates full MIF guidance. Results demonstrate a significant performance gap: zero-shot register classification accuracy was only 33.3%, which increased to 73.3% (+40 points) with schema guidance. The overall Meaning Intelligence Score improved by 5.4 points (73.2 to 78.6), with notable gains in register identification, coded subtext detection (+10 points), and strategic action recommendation (+10.3 points). The study also identified a critical blind spot: the model's failure to distinguish mobilization signals disguised as humor, which has implications for media monitoring and crisis response.
Key Results
- The model's register classification accuracy improved from 33.3% in zero-shot to 73.3% with schema guidance, illustrating the framework's effectiveness in capturing pragmatic nuances related to register shifts in Nigerian discourse.
- The composite Meaning Intelligence Score increased from 73.2 to 78.6, driven by substantial improvements in register identification, coded subtext detection (+10 points), and strategic communication action (+10.3 points), demonstrating the framework's capacity to enhance multi-dimensional understanding.
- The model consistently misclassified a mobilization signal disguised as humor as a routine warning in both conditions, revealing a critical failure mode with potential real-world consequences for media surveillance and crisis management.
Significance
This research addresses a longstanding gap in NLP for African languages by moving beyond polarity-based sentiment analysis to pragmatic, context-aware understanding. The MIF framework captures the complex interplay of register, intent, irony, and subtext, which are crucial for accurate interpretation of Nigerian discourse. Its application can significantly improve media intelligence, political analysis, and crisis response in low-resource settings. The framework's design emphasizes operational relevance, integrating risk assessment and action recommendations, thus bridging the gap between linguistic analysis and real-world decision-making. This approach sets a new standard for culturally nuanced NLP evaluation, with potential to influence multilingual NLP research globally.
Technical Contribution
Technically, the paper introduces a novel nine-dimensional annotation schema that operationalizes pragmatic analysis in low-resource languages. The framework employs a structured scoring system, including flags for divergent and deceptive positive utterances, and computes a composite Meaning Intelligence Score (MIS) based on weighted per-dimension accuracies. The evaluation leverages prompt engineering with the Gemini 2.5 Flash model, demonstrating substantial improvements in registration and subtext detection through schema-guided prompting. The dataset construction involved stratified sampling across difficulty levels and sectors, ensuring comprehensive coverage of discourse phenomena. The methodology exemplifies how multi-dimensional annotation can enhance large language model interpretability in culturally complex settings.
Novelty
This work is pioneering in formalizing a multi-dimensional pragmatic analysis framework tailored for Nigerian and African discourse, explicitly differentiating surface sentiment from true intent—a departure from conventional polarity-based sentiment models. Unlike prior datasets focused solely on polarity or discourse relations, MIF integrates dimensions like irony, coded subtext, and strategic action, operationalized through scoring rules and flags. Its emphasis on context-dependent pragmatics and operational actionability distinguishes it from existing sentiment and intent detection frameworks, marking a significant advancement in culturally aware NLP research.
Limitations
- The calibration dataset comprises only 30 items, authored by a single annotator, which may limit the diversity and representativeness of discourse phenomena. The dataset's controlled nature may not fully capture the variability of real-world social media content.
- Evaluation was conducted solely on the Gemini 2.5 Flash model under prompt-based settings, without fine-tuning or multi-model comparison, leaving open questions about generalizability and robustness across different architectures and real social media data.
- The framework relies heavily on rich contextual information, which may be scarce or noisy in practical applications, potentially reducing performance in real-time social media monitoring scenarios.
Future Work
Future research will expand the calibration dataset to 500 real-world Nigerian discourse examples with triple annotation and inter-annotator agreement to improve robustness. Multi-model evaluations, including fine-tuning approaches, are planned to assess the framework's adaptability across architectures. Additionally, extending the framework to other West African Pidgins and code-mixed varieties will enhance its cross-linguistic applicability. The authors also aim to develop automated tools for scalable annotation and real-time deployment, facilitating broader adoption in media and crisis monitoring contexts.
AI Executive Summary
In recent years, artificial intelligence has revolutionized natural language processing, enabling machines to interpret and generate human language with unprecedented accuracy. However, most advances have focused on high-resource languages like English, leaving low-resource and culturally nuanced languages underexplored. Nigerian discourse exemplifies this challenge: its rich tapestry of registers—including Standard English, Nigerian English, Nigerian Pidgin, and code-mixed speech—presents unique pragmatic phenomena such as sarcasm, emotional escalation, and coded subtext. Traditional sentiment analysis models, which classify text into positive, negative, or neutral categories, are inadequate for capturing these subtleties, often leading to misinterpretations with serious real-world consequences.
Addressing this gap, the authors introduce the Meaning Intelligence Framework (MIF), a novel nine-dimensional annotation and evaluation schema designed specifically for Nigerian public discourse. Unlike conventional sentiment models, MIF explicitly separates surface sentiment from true communicative intent, incorporating dimensions such as register, irony, coded subtext, risk tier, and recommended action. This multi-layered approach recognizes that the same utterance can carry vastly different pragmatic meanings depending on context, speaker, and audience. For example, the phrase “You don try well well” can be genuine praise or sarcastic condemnation, depending on situational cues.
To validate the framework, the researchers constructed a 30-item calibration dataset spanning various registers and sectors, annotated according to detailed guidelines. They evaluated a state-of-the-art language model, Gemini 2.5 Flash, under two conditions: zero-shot prompting and schema-informed prompting with full MIF guidance. Results showed a dramatic improvement in register classification accuracy—from 33.3% to 73.3%—demonstrating the importance of contextual and pragmatic cues. The overall Meaning Intelligence Score increased by 5.4 points, with notable gains in detecting coded subtext and recommending appropriate communication actions.
The study also uncovered critical failure modes, such as the model’s inability to distinguish mobilization signals disguised as humor, which could have severe implications for crisis management and media monitoring. These findings highlight the importance of multi-dimensional pragmatic analysis for operational AI systems in complex socio-cultural environments. The authors provide comprehensive documentation, including framework specifications, annotation guidelines, and a public calibration set, to facilitate reproducibility and further research.
Overall, this work marks a significant step toward culturally aware NLP, emphasizing the necessity of context-sensitive, pragmatic understanding for real-world applications. Its implications extend beyond Nigeria, offering a blueprint for developing AI systems capable of nuanced interpretation in diverse linguistic and cultural settings. Future efforts will focus on scaling the dataset, evaluating multiple models, and extending the framework to other African languages, ultimately fostering more accurate and responsible AI deployment in multilingual, low-resource environments.
Deep Dive
Abstract
We introduce the Meaning Intelligence Framework (MIF), a nine-dimension annotation and evaluation schema for Nigerian public discourse that separates surface sentiment from true communicative intent. Existing benchmarks for Nigerian languages, including NaijaSenti and AfriSenti, treat sentiment classification as a three-way polarity task (positive, negative, neutral). We argue that the dominant failure mode of AI systems on Nigerian discourse is not translation failure but context failure: the same utterance carries opposite pragmatic force depending on speaker, audience, and situation. The MIF operationalises this insight across nine scored dimensions: register, surface sentiment, true intent, irony, coded subtext, risk tier, annotator confidence, speaker emotion, and recommended communications action. We construct a 30-item calibration dataset spanning Standard English, Nigerian English, Nigerian Pidgin, and code-mixed registers, and evaluate a frontier language model (Gemini 2.5 Flash) under zero-shot and schema-informed prompting conditions. The headline finding is the Register Gap: zero-shot register classification accuracy is 33.3%, rising to 73.3% (+40 points) when the model receives the MIF schema in-context. The composite Meaning Intelligence Score increases by 5.4 points (73.2 to 78.6) under schema-informed prompting, with the largest practical gains in register identification, coded-subtext detection (+10 points), and strategic action recommendation (+10.3 points). We release the framework specification, annotation guidelines, and the 30-item public calibration set to support reproducibility, while retaining a private holdout corpus for contamination-protected evaluation.