Your bio passed the parser. Did the claim?
Natural Language Processing in Profile Bios
SLOP1276 · Week 04 · Alex’s story
Fictional Alex has two bio drafts. A separate synthetic dashboard declares A the winner, although B leads in both zones. The letters invite a bad shortcut.
Investigate: What does this comparison actually establish?
What you will be able to do
Implement a declared text budget and phrase matcher.
Support a bio revision with evidence from Alex’s case.
Reproduce a pooled ranking reversal and explain its denominator.
Write a synthetic experiment protocol with a defensible conclusion.
How this pack fits the week
60 minutes · Monday’s selected lecture slides.
90 minutes · Thursday’s four practical blocks.
About 210 minutes · independent preparation, remaining examples and consolidation.
The complete 180-minute route is a resource guide; its break is for a full sitting.
Alex is a fixed input
Fictional adult CS student; board games and terrible puns.
Friday 5–7 pm · $20 total · bus travel.
Wants to meet someone, with the possibility of a relationship.
Adding a dog, car or mountain-climbing hobby would change the case.
Reuse the previous two artefacts
Week 2: your platform audit separates counts, denominators and unknown allocation rules.
Week 3: your photo manifest freezes A-portrait-v1.
Keep the same image bytes, crop, alt text and display treatment. The planned change is the bio text.
Two texts, one image input
A-portrait-v1
A bio becomes inspectable data
This small NLP baseline exposes its rules. It does not understand a person.
Define the budget’s unit
Lab limit: 150 Unicode code points in the original text, including spaces and punctuation.
JavaScript text.length counts UTF-16 code units; [...text].length counts code points.
We do not trim or normalise before counting. Visible characters can contain several code points.
The budget is a constraint
Filling the remaining space earns no automatic quality points.
Test the boundary · 6 min
Count 150 ASCII letters, then append one space.
In JavaScript, compare the length of "\u{1F600}" with its code-point count.
Compare "\u00E9" with "e\u0301" under our no-normalisation rule.
Does an empty string pass the budget? Does that make it a useful bio?
Budget answers
150 letters pass; an added space gives 151, so reject the budget check.
U+1F600 occupies two UTF-16 units, one code point.
The two accented forms count as one and two code points here.
Empty text passes the length ceiling but provides no clarity evidence.
Declare the tokenizer
Lowercase; take maximal runs of ASCII a–z and 0–9.
Spaces, punctuation, apostrophes and hyphens separate tokens.
Keep all words and numbers. No stemming or stop-word removal.
This deliberately narrow English exercise needs a different tokenizer for broader language coverage.
Trace a phrase match
- 0lookingunmatched
- 1forunmatched
- 2aunmatched
- 3partnermatched
- 4inmatched
- 5crimematched
Matched-token share measures this dictionary’s coverage, not originality or compatibility.
Trace the matcher · 6 min
Dictionary: “must love dogs” and “partner in crime.”
Tokenise “Must love dogs; partner-in-crime.” Count the matched share.
Then match “partner in crime” against both “partner in crime” and “in crime.”
Count each covered token once, even when phrase spans overlap.
Matching answers
First input: six tokens, six covered → 100%.
Overlapping phrases: three distinct covered tokens / three → 100%, not 5/3.
Full bio A: 3/14 ≈ 21.4%. Full bio B: 0/20 = 0%.
No tokens means an undefined share. Report it explicitly; do not call it 0% repetition.
Break the metric · 4 min
“No partner in crime; I prefer board games.”
“Looking for partners in crime.”
“I own three dogs and a sports car.”
Which strings expose missed meaning, a missed phrase or invented biography?
A dictionary is not a fact checker
The first still matches partner in crime, despite the negation.
The second misses: partners ≠ partner under exact matching.
The third matches neither phrase, yet invents Alex’s possessions.
A zero cliché score cannot certify truth, intent or a good invitation.
Review the evidence · 5 min
For each candidate, quote evidence for clarity, specificity, feasibility and exit.
Apply the toolkit’s 0–4 anchors consistently.
Does bio B specify Friday’s window? Does it account for the budget and bus?
A higher rubric score and a higher response rate are separate claims.
An evidence-backed reading
Both: relationship possibility gives clarity; board games and puns give specificity. Suggested anchors: 4, 4.
A: no concrete invitation or permission to pass → feasibility 0, exit 0.
B: time stated, budget/travel missing → feasibility 3; “Fine to pass” → exit 4.
Ratings need reasons. A proposed meeting still needs a feasible venue.
Revise under constraints · 6 min
Repair bio B’s missing logistics within 150 code points.
Keep the relationship possibility, at least one supported detail and an explicit way to pass.
Do not invent a venue, price quote, hobby or transport option.
Save the original and revision. Explain what information each change adds.
One possible revision
“Board games, bad puns, maybe a relationship? Meet Friday 5-7 pm: bus-friendly, $20 total. Fine to pass.”
103 code points, including punctuation and spaces.
This states constraints; it does not prove a venue is reachable or affordable.
The text bundle changed. An eventual A/B result could not isolate the effect of one word.
Pause for the full route · 10 minutes
For an uninterrupted full-pack session; skip this slide in Monday’s selection.
Keep your token trace and evidence quotes.
Next: a manager has found a winning bio by deleting the zone column.
Bring the denominator back.
Three separate teaching fixtures
Texts: our newly authored candidates have no measured response outcomes.
Exposure CSV: invented counts with A/B labels; it supplies no actual wording or assignment log.
Balanced example: separate invented 120/1000 and 140/1000 counts.
Matching A/B labels do not join these datasets. Do not claim the CSV tested our sentences.
Read the exposure rows
Synthetic CSV: zone, variant, exposures, positive responses.
North: A 24/80 · B 7/20.
South: A 2/20 · B 12/80.
One invented binary positive at most per exposure. These labels do not establish mutual selection or consent.
The pooled dashboard favours A
Is “deploy A because of its text” supported?
Both zones favour B
Reproduce the reversal · 7 min
Open bio-exposures.csv.
Check positive counts do not exceed exposures and all four zone/variant cells exist.
Calculate zone rates, pooled rates and B minus A for each comparison.
Identify the aggregation step that changes the ranking.
Reversal answers
Within North: 35% − 30% = +5 percentage points.
Within South: 15% − 10% = +5 points.
Pooled: 19% − 26% = −7 points.
Both arms have 100 total exposures, but their zone compositions differ. Equal totals do not imply comparable audiences.
Who contributed to each arm?
A gets more exposure in the higher-response zone.
The hidden weights are arithmetic
The weighted averages describe different mixtures.
This constructed reversal is often called Simpson’s paradox.
Give both the same mix · 5 min
Declare a hypothetical target: 50% North, 50% South.
Reweight each bio’s two observed zone rates to that target.
Compare the results with the pooled dashboard.
Does reweighting reconstruct an unrecorded random assignment?
Standardised descriptive rates
The difference is +5 points. Unmeasured selection could remain within zones; this is not a causal estimate.
Design the comparison you wanted
Hold the photo and case facts fixed. Freeze both candidate texts before allocation.
Assign variants randomly within each zone, retaining one assignment per simulated viewer.
Record the zone in the analysis, not just in the raw file.
A proposed allocation plan
Define one primary outcome
Example protocol: a binary positive response within 24 simulated hours of first exposure.
Denominator: every assigned simulated viewer, counted once.
A complete window with no positive is 0. Missing observation is not automatically 0.
Missing records or duplicate assignments fail this classroom run’s data check; repair before analysing.
Freeze the stopping rule too
Stop allocation at 500 viewers per variant per zone. Then complete every observation window.
Primary estimate: mean of the two zone differences, with 50/50 weights.
Declare +3 percentage points as an illustrative minimum worthwhile effect.
The 2,000-viewer cap is a teaching choice, with no power guarantee. The protocol has no outcomes yet.
Repair this protocol · 7 min
“Show A on Monday with one photo. Show B on Sunday with another. Keep checking until B wins.”
Find changed inputs, allocation bias and the moving stopping rule.
Rewrite the unit, controls, outcome, allocation, analysis weights and stopping rule.
Use a simulation proposal; no real profiles or app users are required or permitted.
Protocol debrief
Use one image and fixed texts; assign each unique simulated viewer once within zone.
Complete 24-hour windows; validate records; analyse once at the fixed cap.
Report each zone and the prespecified 50/50 difference.
Plan uncertainty for that stratified estimate. The next pooled-z exercise uses a separate, simpler model.
A separate balanced fixture
Observed difference: +2 percentage points.
Points and percent are different units
Absolute change: 14% − 12% = 2 percentage points.
Relative change: (0.14 − 0.12) / 0.12 ≈ 16.7%.
“B is 16.7 points better” is an arithmetic error.
Neither expression tells you whether the difference is distinguishable from sampling variation.
Build the equal-rate null model
Assume independent binary observations with a stable probability in each arm.
Under H₀: pA = pB, estimate their shared rate:
NIST’s two-proportion method gives the pooled standard error for this test.
Scale the difference by its uncertainty
A two-sided 5% normal test compares |z| with about 1.96.
Our sign convention is always B minus A.
Audit the calculation · 6 min
Reproduce 120/1000 versus 140/1000 in the calculator or workbook.
Explain the 0.13, 0.01504 and 1.33, with their units.
Reverse the arms. Which values should change sign?
Write a result sentence that does not announce equivalence or a causal winner.
What the result permits
1.33 < 1.96: no clear difference under this model and threshold.
Swapping the arms changes the signs of the difference and z; the pooled rate and SE stay fixed.
Non-rejection does not establish equivalence.
The +2-point estimate is below our example +3-point worthwhile effect, but that comparison alone does not rule out worthwhile effects.
Same rates, ten times the counts
Both differences are 2 points. Statistical detection does not make an effect larger or a confounded comparison causal.
Sparse counts need a different route
Try 2/8 versus 4/8. The course calculator withholds its normal approximation.
Its guard requires at least 10 in every positive/negative cell; this is a conservative course choice.
It implements no exact test, power calculation or equivalence test.
Increasing synthetic counts does not create real-world evidence.
Audit the report claims · 5 min
“Half the local singles are available, one in four is a match, and bio A is the winner.”
Use Week 2’s 300 active, 150 available, 40 reciprocal fictional profiles.
Use this week’s zone and pooled exposure comparisons.
For each claim: give the numerator, denominator, scope and an evidence-based correction.
Report debrief
150/300 = 50% of invented active profiles are available.
40/150 ≈ 26.7% of available profiles meet the reciprocal criteria; 40/300 ≈ 13.3% of active profiles do.
A leads only in the pooled exposure summary. B leads within both zones; neither establishes a text effect.
Report sampling uncertainty, model assumptions and the fixture’s lack of external validity separately.
Week 5 needs a supported candidate
Carry the bio revision, evidence quotes, frozen photo and denominator definitions forward.
A response label is not automatically a mutual-selection event. Week 5 defines the joint probability.
Exit check: can zero clichés, 26% pooled or z = 1.33 justify “this bio causes more matches”?
Name the missing evidence in each case.
The experiment now has a specification
Answers: zero clichés says nothing about truth; 26% hides the exposure mix; z = 1.33 neither detects a clear difference nor establishes causation.
Alex has a supported bio, but no causal winner or match forecast. Week 5 defines mutual selection.
Save the Bio experiment protocol. Report due 19 March, 5 pm.
Worked notes and five-file download · Tutorial · Report brief