Sentiment is the most abused field in data models

1. What sentiment_score really measures

Sentiment score = emotional valence
Positive ↔ Neutral ↔ Negative

It answers:

“How does the person feel about this topic right now?”

Not:

• How important it is
• How much effort they’re putting in
• Whether they’ll act

You can have:

• High engagement + negative sentiment
• Low engagement + positive sentiment

That’s why this field matters.

---

2. Recommended scale (keep it simple)

Option A — Discrete scale (best for explainability)

Score	Label	Meaning
-2	Strongly Negative	Frustration, anger, resentment
-1	Negative	Dislike, concern, reluctance
0	Neutral / Mixed	Informational, balanced
+1	Positive	Enjoyment, optimism
+2	Strongly Positive	Excitement, pride, love

This is human-legible and stable over time.

---

Option B — Normalized (-1.0 → +1.0)

Use only if you’re piping NLP outputs directly.

Range	Interpretation
-1.0 to -0.6	Strong negative
-0.59 to -0.2	Mild negative
-0.19 to +0.19	Neutral
+0.2 to +0.59	Mild positive
+0.6 to +1.0	Strong positive

You can always bucket it later.

---

3. Examples (this is the important part)

A. Same topic, different sentiment

Topic: Work / Career

Utterance	sentiment_score	Why
“Work is fine.”	0	Flat, informational
“I like my job.”	+1	Mild positive
“I love what I’m building.”	+2	Strong positive
“I’m stuck at work.”	-1	Frustration
“I hate my job.”	-2	Strong negative

Engagement could be high in all of these.

---

B. Same sentiment, different engagement

Topic: Health & Gyms

Utterance	Sentiment	Engagement
“Gyms are great.”	+1	1
“I train 5x a week and track PRs.”	+1	5

Sentiment = feeling
Engagement = energy

---

C. Business context examples

Topic: Project Management

Utterance	sentiment_score
“We’re on track.”	+1
“This is harder than expected.”	-1
“This project is a mess.”	-2
“Challenging, but solvable.”	0

---

Topic: Negotiations & Proposals

Utterance	sentiment
“Let’s explore options.”	0
“I’m excited to move forward.”	+2
“This pricing doesn’t work.”	-1
“This is unacceptable.”	-2

---

D. Personal & emotional topics

Topic: Emotions / Love

Utterance	sentiment_score
“I’m okay.”	0
“I feel close to her.”	+2
“I’m confused.”	-1
“I’m heartbroken.”	-2

This is where sentiment is most valuable.

---

4. Common modeling mistakes (avoid these)

❌ Mistake 1: Using sentiment as importance

“This is a big issue” ≠ negative sentiment

❌ Mistake 2: Using sentiment as an outcome

“We closed the deal” ≠ positive sentiment

❌ Mistake 3: Averaging sentiment blindly

A -2 followed by +2 is not “neutral” — it’s volatility.

---

5. How sentiment works with your existing fields

Field	Question it answers
Topic	What are we talking about
Intent	Why now
Engagement score	How invested
Sentiment score	How it feels
Outcome	What happened

This separation is why your model is actually strong.

---

6. Design recommendation (this is subtle but important)

Store:

• sentiment_score
• sentiment_confidence (low / medium / high)

Why?

Some statements are emotionally clear. Some are ambiguous.

Example:

• “Interesting.” → sentiment unclear → low confidence

---

TL;DR

• Sentiment = emotional direction
• Use -2 to +2
• Keep it independent of engagement & outcome
• Anchor it to language, not interpretation

//todo

• Combine sentiment × engagement into a “heat map”
• Model sentiment drift over time
• Or detect hidden risk (high engagement + negative sentiment)

That’s where this turns into a real signal, not just labels.

---

1. What engagement_score really measures

Engagement score = how much energy + commitment is present in the interaction.

Not sentiment.
Not importance.
Not outcome.

It answers:

“How invested is the person in this conversation right now?”

Think attention × effort × continuity.

---

2. Signal strength (the inputs)

signal_strength is evidence, not judgment.

Common observable signals

Signal	Examples
Time	Long messages, long calls, repeat back-and-forth
Depth	Personal detail, specificity, vulnerability
Effort	Preparation, documents shared, follow-ups
Agency	Decisions made, commitments stated
Continuity	Refers to past conversations, future plans
Risk	Disagreement, asking for advice, asking for help

Each of these increases signal strength.

---

3. Example scales (practical, not academic)

Option A — Simple 1–5 scale (recommended)

This works shockingly well.

Score	Label	Observable signals
1	Passive	Polite replies, small talk, low effort
2	Light	Casual interest, short responses
3	Engaged	Asking questions, staying on topic
4	Invested	Sharing context, opinions, time
5	Committed	Decisions, plans, emotional or professional stake

---

Option B — Expanded 0–100 (only if you need math)

Use if you plan analytics.

Range	Meaning
0–20	Noise
21–40	Awareness
41–60	Interest
61–80	Engagement
81–100	Commitment

Internally, you can still bucket it to 1–5.

---

4. Mapping signal strength → engagement_score

You can model it explicitly.

Example signals → score

Signals present	engagement_score
Topic hopping, generic replies	1
Jokes, light banter	2
Asking follow-ups	3
Sharing constraints, goals	4
Making decisions / asking next steps	5

---

5. Real examples from your domain

Example 1: Salsa Dancing

• “Yeah I dance sometimes” → 1
• “I love salsa, where do you dance?” → 3
• “I’m training for a festival, practicing twice a week” → 5

Topic stayed the same.
Engagement changed drastically.

---

Example 2: Business — Marketing Campaigns

• “Send me the deck” → 2
• “What would success look like?” → 3
• “Let’s launch next month, budget is X” → 5

---

Example 3: Personal — Aspirational

• “I’ve been thinking about my career” → 3
• “I’m unhappy and need to change direction” → 4
• “I’m committing to a 90-day pivot” → 5

---

6. Where this fits in your model

Think in layers:

• Topic → What
• Intent → Why now
• Learning touchpoint → Where in the arc
• Engagement score → How much energy
• Signal strength → Why you believe that score

This keeps the model explainable, not magical.

---

7. Design advice (important)

• Don’t over-optimize this.
• Human-readable beats perfect.
• Engagement is situational, not identity.

A person can be:

• Engagement = 5 in Salsa
• Engagement = 1 in Investing

That’s not inconsistency — that’s truth.

---

TL;DR

• Signal strength = observable evidence
• Engagement score = normalized judgment of investment
• Use 1–5 unless you really need more
• Anchor every score to something you could point to

//todo

• Define decay over time
• Weight engagement by intent
• design a query like: “Which topics generate high engagement but low outcomes?”

lethal design in a good way.

Engagement is situational, but Energy and Effectiveness is Person Centric

ENERGY ASSESSMENT (Circle 0-10)

• Pre-interaction mood: Dread [0 1 2 3 4 5 6 7 8 9 10] Anticipation
• Post-interaction feeling: Drained [0 1 2 3 4 5 6 7 8 9 10] Energized
• Conversation flow: Strained [0 1 2 3 4 5 6 7 8 9 10] Natural
• Emotional safety: Unsafe [0 1 2 3 4 5 6 7 8 9 10] Safe

EFFECTIVENESS ASSESSMENT (Circle 0-10)

• Follow-through: Never [0 1 2 3 4 5 6 7 8 9 10] Always
• Value created: None [0 1 2 3 4 5 6 7 8 9 10] Transformative
• Proactivity: Passive [0 1 2 3 4 5 6 7 8 9 10] Initiator
• Efficiency: Time-waster [0 1 2 3 4 5 6 7 8 9 10] Multiplier