Sentiment Attribution & Methodology

PublicMoodTracker generates political sentiment scores using a reproducible, version-controlled AI pipeline. This document provides complete transparency on how scores are calculated, what they represent, and their known limitations — in compliance with emerging AI transparency obligations and good-faith commitment to responsible AI deployment in Kenya's democratic context.

1. What a Sentiment Score Represents

A PublicMoodTracker sentiment score for a politician represents the weighted average sentiment of public media discourse about that individual over a defined time window (24 hours, 7 days, or 30 days). It is derived from news articles, social media posts, and broadcast content — not from direct polling, surveys, or voter intention data.

A score of +0.72 means: on average, public media coverage of this politician is moderately positive. It does not mean 72% of Kenyans support them.

2. Data Ingestion Pipeline

2.1 Sources (9 active scrapers)

2.2 Content Preprocessing

1. De-duplication: Near-duplicate articles (cosine similarity > 0.92) are collapsed to prevent single stories dominating scores.
2. Spam filtering: Content flagged by our spam classifier (precision: 94%) is excluded.
3. Sentence segmentation: Articles are split into sentences; sentiment is scored per sentence, not per article.
4. Language detection: langdetect library classifies language; Swahili and Sheng trigger the multilingual model path.

3. Named Entity Recognition (NER)

Our custom NER model identifies Kenyan political entities within text — politicians, counties, parties, and institutions. It is fine-tuned from bert-base-multilingual-casedon a labelled Kenyan political corpus of 28,000 sentences.

• Mention linking: Aliases are resolved (e.g., "DP Ruto", "His Excellency Ruto", "Hustler" → William Ruto).
• Ambiguity handling: Ambiguous names (e.g., common surnames) require county or party context; unresolvable ambiguities are discarded.
• Minimum threshold: A politician requires ≥3 mentions within a scoring window to appear in rankings.

4. Sentiment Classification Model

4.1 Base Model

XLM-RoBERTa-large (Phase 13.1) — a transformer model supporting 100 languages, fine-tuned on 12,000 labelled Kenyan political sentences (English, Swahili, Sheng). Labels: POSITIVE, NEGATIVE, NEUTRAL.

4.2 Custom Lexicon Enhancement

A supplemental lexicon of 40 Swahili political terms and150 Sheng political terms (e.g., "hustler", "dynasty", "kamwana") modifies token embeddings before classification to improve accuracy for Kenyan-specific discourse.

4.3 Output

Each classified mention produces:

• sentiment_label: POSITIVE / NEGATIVE / NEUTRAL
• confidence: 0.0–1.0 (model softmax probability)
• raw_score: +1.0 (positive), −1.0 (negative), 0.0 (neutral)

5. Score Aggregation Formula

5.1 Weighted Average Score

The displayed score for a politician over a time window is a confidence-weighted mean:

score = Σ(raw_score_i × confidence_i) / Σ(confidence_i)

Mentions with confidence < 0.55 are excluded to reduce noise from ambiguous text.

5.2 Polarity Confidence

Measures how consistently directional public discourse is (used in the Rivals comparison table):

polarity_confidence = |positive_count − negative_count| / total_mentions

A value near 1.0 means near-unanimous sentiment; near 0.0 means deeply polarised.

5.3 Mention Velocity (Spike Detection)

velocity = mentions_in_last_hour / avg_mentions_per_hour_last_7d spike_flag = True if velocity ≥ 3.0 and |score_delta_1h| ≥ 0.4

6. Score Display Rules

7. Known Limitations

• Media ≠ public opinion: Scores reflect what journalists and social users write, not what the broader public thinks.
• Sheng coverage gaps: Rapidly evolving Sheng slang may lag our lexicon updates by 2–4 weeks.
• Satire misclassification: The model may misread satirical content as genuine sentiment; estimated error rate ~3%.
• Source bias: Not all Kenyan media outlets are included; sources skew toward urban, online audiences.
• Coordinated inauthentic behaviour: Social media manipulation (bot networks) can temporarily distort scores; we apply velocity filters but cannot guarantee full detection.

8. Model Versioning and Change Log

9. Requesting Source Evidence

Any registered user may click the transparency icon (🔍) on a politician's profile to view the source articles contributing to the current score. Users with a Daily Access subscription can export these citations as a reference list.

Academic researchers may request full methodology documentation by contacting research@siasaiq.com.