Slangscape

Slang doesn't die naturally—it's executed. TikTok kills language 3× faster than Reddit, and every 'cringe' tag is a weapon in generational warfare.

Comment

What Inspired Me

The inspiration hit during a family gathering when my younger cousin visibly cringed after I said "that's fire." A term that felt current just months ago had apparently crossed some invisible threshold into obsolescence. What changed? When? And who decided?

This wasn't just about slang—it was about power, identity, and cultural gatekeeping operating at internet speed. I realized platforms weren't neutral carriers of language evolution; they were active participants in linguistic death. TikTok's cringe/cool binary, Reddit's 15-year incubation cycles, Twitter's vocabulary vault—each platform shaped language mortality differently.

The Hex-a-thon's "Exploring the Evolution of Slang" prompt was the perfect excuse to transform this curiosity into data-driven discovery.

What I Learned

Technical Skills

  • Multi-platform data archaeology: Learned to extract, normalize, and reconcile slang usage patterns across Reddit (2009–2024), Twitter (2010/2020 snapshots), and TikTok (2023–2024 sentiment-tagged data)
  • Semantic ML at scale: Implemented SentenceTransformer embeddings + AgglomerativeClustering to discover latent categories in 6,737 unstructured slang descriptions without pre-defined taxonomies
  • Hex's hybrid SQL-Python workflow: Mastered chaining—SQL for warehouse queries → Python for ML/NLP → Explore cells for interactive visualization—creating atomic, debuggable analysis units
  • Sentiment trajectory modeling: Built predictive signals for cringe acceleration using TikTok's binary tags as death-switch indicators

Domain Insights

  • Platforms kill differently:
    • Reddit = utility decay (14-year gradual fade)
    • Twitter = 98% stability (linguistic vault)
    • TikTok = binary collapse (1–3 day cringe death)
  • The 44× hidden war: Only 0.7% of slang explicitly mentions generation, but 30.7% carries implicit age signals—generational warfare operates through coded language
  • Speed as weapon: 73% average compression (3.7× communication speed advantage) = conversation dominance in attention-scarce digital spaces
  • Language doesn't evolve, it's executed: Every cringe tag, compression protocol, and generational marker is a micro-battle in cultural control

How I Built It

Phase 1: Data Discovery & Wrangling (The Hard Part)

Finding usable data took 40% of project time. The provided VIRTUAL_HACKATHON.SLANG.SLANG_DATA (6,737 terms) was excellent baseline metadata, but lacked temporal usage data—the DNA of evolution tracking.

The challenge: Source platform-specific slang mentions across 15 years while maintaining:

  • Temporal granularity (year/month-level tracking)
  • Cross-platform comparability (equivalent sampling methods)
  • Sentiment signals (TikTok's cringe/cool tags)

The solution:

1. Reddit (2009–2024)

Uploaded combined_reddit_data_cleaned.csv

  • Challenge: 15 years × multiple subreddits = massive volume
  • Solution: Batch processing with streaming regex (cell [hexCell:019be011-b22a-7441-a97c-5136703c64c3:Analyze top 5 slang by year]) to avoid memory overflow
  • Output: Top 15 slangs tracked annually showing organic growth/decay

2. Twitter (2010 vs 2020)

Uploaded slang_2010_tweets.csv + slang_2020_tweets.csv

  • Challenge: Decadal comparison required consistent sampling methodology
  • Solution: Frequency analysis on curated snapshots (cells
    [hexCell:019be023-ff95-7ee4-96a8-6082b42d7da9:Top 10 slangs in 2010] –
    [hexCell:019be024-571c-7001-9e67-3233c86a407c:Compare slang trends 2010 vs 2020])
  • Output: 98% stability discovery (only 3 deaths out of 150 terms)

3. TikTok (2023–2024)

Uploaded tiktok_filtered_with_slangs.csv

  • Challenge: Raw sentiment tags ("cool"/"cringe") needed temporal correlation
  • Solution: Monthly aggregation + binary sentiment tracking (cells
    [hexCell:019be05f-6287-7aa4-afee-5a1b81e14e7e:Identify rising and declining slangs] –
    [hexCell:019be05e-e192-7992-b5d4-f6fe9c431bcd:Analyze slang by sentiment (Cringe vs Cool)])
  • Output: Cringe trajectory prediction model (±3.4% weekly acceleration signals)

Processing complexity: Each platform required custom extraction logic:

  • Reddit: Regex-based slang extraction from comments with STOPWORDS filtering
  • Twitter: Count-based frequency analysis with 60% decline threshold for "death" detection
  • TikTok: Sentiment-aware aggregation with generational markers (Gen Z vs Alpha)

Phase 2: Cross-Platform Death Analysis

Built a unified lifecycle tracking system (cells Prepare cross-platform lifecycle data → Analyze platform death patterns) that:

  • Normalized temporal scales (Reddit years → TikTok months → Twitter decades)
  • Calculated death mechanisms:
    • Utility Decay (Reddit): 88% decline from peak over 14 years
    • Cringe-Tagged (TikTok): 82% cringe ratio = permanent death
    • Meme Exhaustion (Twitter): 73% decline (only 2% death rate)
  • Quantified kill speeds: TikTok 3× faster than Reddit

Key technical innovation:
Pre-calculated sentiment counts (cell Pre-calculate TikTok sentiment counts (optimized)) using Counter objects instead of repeated DataFrame queries—reduced 52-second query times by 98%.

Phase 3: Deep Semantic Analysis (The ML Heavy-Lifting)

For each insight (Cultural Origins, Age War, Semantic Evolution, Emotion, Speed, Social Control), I built a 3-stage ML pipeline:

Stage 1: Signal Extraction (regex patterns)

# Example: Generational markers
gen_signals = extract_signals(df, patterns=[
    'generation', 'Gen Z', 'Boomer', 'Millennial', 'youth', 'old'
])

### Stage 2: Semantic Clustering (unsupervised)

```python
# Turn unstructured signals into categories
embedding_model = SentenceTransformer('all-MiniLM-L6-v2')
signal_embeddings = embedding_model.encode(signals)
clustering = AgglomerativeClustering(n_clusters=8)

Stage 3: Assignment & Analysis

  • Assigned 6,737 terms to discovered categories
  • Calculated distributions, co-occurrence patterns, transformation narratives

Result: Discovered categories NO human taxonomy predicted:

  • "Traditional-Coded" vs "Youth-Coded" (hidden age war)
  • 8 evolution pattern types (Historical Origin, Viral Spread, Contemporary Shift...)
  • 14 social label types (gender-based labels = 5.1%, exceeding behavior/appearance combined)

Phase 4: Interactive Visualization

Created 51+ Explore/Chart cells using Hex's native visualization tools:

  • Line charts: Reddit 15-year trends, TikTok monthly volatility
  • Bar charts: Platform death rates, cringe trajectories, compression champions
  • Pie charts: Emotional valence distributions, ambiguity breakdowns
  • Sankey diagram: Platform migration flows (Reddit → TikTok, bypassing Twitter)

Design principle: Every chart = one key insight, fully styled with titles, axis labels, formatting, and data callouts.

Challenges I Faced

1. Data Volume vs Memory Constraints

Problem: Processing 500K+ TikTok posts + 15 years Reddit data exceeded notebook memory limits

Solution:

  • Batch processing with BATCH_SIZE=10000 chunks (cell Analyze top 5 slang by year)
  • Pre-computed sentiment counts instead of joins (cell Pre-calculate TikTok sentiment counts (optimized))
  • Strategic memory cleanup checkpoints (cells Memory cleanup checkpoint, ⚠️ FINAL CLEANUP: Delete large datasets)
  • Deleted intermediate DataFrames after downstream cells consumed them

Impact: Reduced memory footprint by 60%, kept notebook under 8GB RAM

2. Cross-Platform Temporal Normalization

Problem: Reddit (yearly), Twitter (decadal), TikTok (monthly)—how to compare "death speed"?

Solution:

  • Normalized to "decline rate per unit time"

    • Reddit: 88% decline ÷ 14 years = 6.3% per year
    • TikTok: 82% cringe ratio ÷ 18 months = 4.6% per month = 55% per year

  • 3× multiplier validated through unit-normalized comparison

3. Unstructured Text → Structured Insights

Problem: 6,737 slang descriptions are free-text—no pre-defined categories for "age coding" or "emotional valence"

Solution:

  • Built regex signal extraction to find linguistic markers
  • Used sentence embeddings + clustering to discover latent categories
  • Validated with example analysis (cell Stage 3.5: Example Validation)—achieved 95%+ category assignment confidence

Example: Discovered "Cheugy" operates as meta-weapon accelerating death cycles—insight impossible without semantic clustering revealing the "generational mockery" cluster

4. TikTok Sentiment Collapse Timeline

Problem: Initial model showed gradual cringe growth—contradicted platform behavior observation

Breakthrough: Realized binary collapse happens in 1–3 days (cell Calculate Cool→Cringe transition timelines analysis):

  • "lol" flipped cool → cringe in 1 day (June 5 → 6, 2023)
  • "yeet" flipped in 3 days (June 16 → 19, 2023)

Implication: Changed entire narrative from "sentiment decay" to "collective binary death switch"—viral cringe-tagging creates feedback loops triggering platform-wide execution

5. Twitter Death Paradox

Problem: Initial analysis showed 0 Twitter deaths—seemed impossible

Root cause: Testing Gen Z-era terms ("yeet," "sus") that didn't exist in 2010 baseline

Fix: Expanded to all 150 common slangs (cell Analyze ALL Twitter slangs for deaths)—discovered true 2% death rate (3 terms: "hypermiler," "skeevy," "celebutante")

Lesson: Sample bias = narrative destruction. Full-corpus validation crucial.

Technical Stack

Data Layer

  • Snowflake warehouse: VIRTUAL_HACKATHON.SLANG.SLANG_DATA
  • Uploaded files: Reddit, Twitter, TikTok CSVs (custom-sourced)

Processing

  • SQL: Warehouse queries, dataframe chaining (40+ cells)
  • Python: Pandas, NumPy, regex, batch processing
  • ML: SentenceTransformer (embeddings), AgglomerativeClustering (unsupervised discovery)

Visualization

  • Hex Explore cells (bar, line, pie, horizontal bar)
  • Plotly Sankey (platform migration flows)

Infrastructure

  • Hex's reactive execution model for atomic cell chaining
  • Memory-optimized batch processing
  • Strategic cleanup for 6,737-term corpus

What Makes This "Only Possible in Hex"

  • SQL → Python → Viz chaining: Query warehouse → ML processing → interactive charts in seamless atomic cells
  • Reactive updates: Change Reddit filter in cell Filter top 15 slangs → entire downstream analysis (15 cells) auto-updates
  • Explore cells: Created 51+ publication-ready charts WITHOUT writing matplotlib/plotly code—focused 100% on insights, not boilerplate
  • Collaboration-ready: Project is fully executable—anyone can run, modify filters, test hypotheses without setup friction
  • Scale: Processed 500K+ records, ran ML clustering on 6,737 terms, generated 51+ visualizations—all in one unified notebook

Impact

This project reveals language death is engineered, not natural. Platforms optimize for engagement through linguistic obsolescence, generations weaponize vocabulary for tribal boundaries, and compression efficiency determines conversation dominance.

The implications:

  • Brands: Twitter vocabulary stable (predictable), TikTok volatile (14-day cringe cycles)
  • Linguists: 44× hidden generational coding vs explicit age mentions
  • Technologists: Platform architecture shapes cultural evolution at internet speed

Language isn't evolving—it's being executed. And now we have the data to prove it.

Share this project:

Updates