sushildalavi/WB-Project

Social listening pilot — gender norm discourse on social media around popular TV in Nigeria, India, and Kenya. WB / Penn. # WB MIP Social Listening — Gender Norms Across Three Reality TV Audiences How do online audiences talk about gender norms when they discuss popular TV? This repo holds the data pipeline, code, and coded outputs for a World Bank pilot study answering that question across three countries, three shows, three platforms. The full write-up: `WB_MIP Social Listening Report Draft - WB_Penn Feedback.docx`. --- ## What is studied ```mermaid flowchart LR A([Nigeria BBNaija S9 ‘Her Money Her Power’]):::nig B([India Made in Heaven S2]):::ind C([Kenya Real Housewives of Nairobi]):::ken A -.thematic.-> A1[Nairaland forum] A -.reach.-> A2[TikTok] B -.thematic + reach.-> B1[YouTube] C -.thematic + reach.-> C1[Twitter / X] classDef nig fill:#fde68a,stroke:#92400e,stroke-width:1px,color:#000; classDef ind fill:#bfdbfe,stroke:#1e3a8a,stroke-width:1px,color:#000; classDef ken fill:#bbf7d0,stroke:#065f46,stroke-width:1px,color:#000; ``` Each country picks one show whose narrative explicitly engages with women's agency, marriage, sexuality, or status — and one platform where audience discussion is sustained enough to study at scale. --- ## The pipeline (5 stages) ```mermaid flowchart TD R[Raw scrapes Apify + Selenium] --> S[Standardize common schema] S --> C[Clean URL strip, dedup, retain emojis & code-switch] C --> K[Keyword-family filter Boolean, high-recall] C --> E[Semantic rerank text-embedding-3-large cosine similarity] K --> A[Analysis-ready set Top-N, calibrated threshold] E --> A A --> L[LLM thematic coding gpt-5.1, temp=0, strict JSON] A --> H[Focused human coding 3 trained annotators ~100 per country] L --> EV[Evaluation: F1, Cohen's κ, Krippendorff's α] H --> EV style R fill:#f3f4f6,stroke:#6b7280 style A fill:#fef3c7,stroke:#b45309,stroke-width:2px style L fill:#dbeafe,stroke:#1d4ed8 style H fill:#dcfce7,stroke:#15803d style EV fill:#fee2e2,stroke:#b91c1c,stroke-width:2px ``` The keyword and semantic stages run in parallel; their union is sampled to produce the …