After studying this material, you should be able to:

1. Explain what taxonomy construction is and why it matters in NLP
2. Describe the core components of the TaxCE framework
3. Understand the progressive pipeline from raw text to hierarchy
4. Define and apply the three EEG evaluation metrics
5. Compare TaxCE against alternative approaches and articulate its advantages

---

Step-by-Step Study Material

Step 1: The Problem — Why Taxonomy Construction Is Hard

What is a Taxonomy?

A taxonomy is a hierarchical classification system that organizes concepts from broad categories down to specific ones.

Example:
Product Feedback
├── Delivery Experience
│   ├── Shipping Speed
│   └── Packaging Quality
└── Product Quality
    ├── Durability
    └── Design

Why Does This Matter?

Organizations receive feedback at massive scale through:

• Customer reviews
• Call transcripts
• Survey responses

The core challenge: How do you automatically organize thousands of unstructured text entries into a meaningful, navigable hierarchy?

What's Wrong With Existing Approaches?

💡 Key Insight: Existing methods fail on at least one of these three dimensions. TaxCE addresses all three simultaneously.

---

Step 2: Introducing TaxCE — The Big Picture

What is TaxCE?

TaxCE is a fully automated framework that:

1. Takes raw unstructured text as input
2. Produces a multi-level hierarchical taxonomy as output
3. Evaluates and refines itself using built-in metrics

Think of it as a pipeline with two major phases:

Phase 1: CONSTRUCTION
Raw Text → Segments → Semantic Units → Topics → Hierarchy

Phase 2: EVALUATION & REFINEMENT
Hierarchy → EEG Metrics → Diagnose Issues → Apply Fixes → Repeat

---

Step 3: Phase 1 — Construction Pipeline (Step by Step)

Stage 1: Progressive Condensation

The framework processes raw text through three layers of condensation:

Layer A: Actionable Segments

• Raw text is broken into meaningful, self-contained pieces
• Noise and irrelevant content is filtered out
• Think of this as extracting "what actually matters" from each document

Raw text: "The delivery was super fast! Loved it. 
           My cat knocked over my coffee this morning."

Actionable segment: "Delivery was fast" ✓
Irrelevant noise: "Cat knocked over coffee" ✗

Layer B: Deduplicated Semantic Units

• Similar segments are grouped and merged
• Removes redundancy across thousands of documents
• Preserves unique meaning while eliminating repetition

"Delivery was fast"
"Shipping arrived quickly"     →  [Fast Delivery]
"Package came ahead of time"

Layer C: Granular Topics with Definitions

• Semantic units are elevated into named topics
• Each topic receives a formal definition
• This ensures clarity and prevents ambiguity later

Topic: "Shipping Speed"
Definition: "Customer experiences related to how quickly 
             orders are delivered after placement"

💡 Why progressive? Each stage reduces complexity while preserving meaning. You can't jump straight from raw text to topics without losing important nuance.

---

Stage 2: Bottom-Up Hierarchy Construction

Once granular topics exist, TaxCE builds the hierarchy from the bottom up.

What does "bottom-up" mean?

• Start with the most specific topics (leaves)
• Group related topics into mid-level categories
• Continue grouping until you reach broad top-level themes

Bottom (specific):     "Next-day delivery", "Same-day shipping"
                                    ↓ group
Middle (general):           "Shipping Speed"
                                    ↓ group
Top (broad):              "Delivery Experience"

Why bottom-up instead of top-down?

• Top-down risk: You impose categories that may not reflect the actual data
• Bottom-up advantage: Categories emerge from the corpus itself — this is called corpus-groundedness

💡 Corpus-groundedness means every node in the taxonomy is justified by actual evidence in the text data, not by assumptions.

---

Step 4: Phase 2 — The EEG Evaluation Metrics

This is one of TaxCE's most important contributions. The framework introduces three metrics to evaluate taxonomy quality.

Metric 1: Exclusivity

Question it answers: Do topics overlap too much?

• High exclusivity = topics are distinct from each other
• Low exclusivity = topics are redundant or duplicated

BAD (low exclusivity):
- "Slow shipping"
- "Late delivery"       ← These mean the same thing!
- "Delayed packages"

GOOD (high exclusivity):
- "Shipping Speed"
- "Packaging Quality"   ← Clearly different topics
- "Tracking Accuracy"

---

Metric 2: Exhaustivity

Question it answers: Does the taxonomy cover everything in the corpus?

• High exhaustivity = all important topics from the text are represented
• Low exhaustivity = significant topics are missing (especially long-tail ones)

If 15% of reviews mention "return policy" but your 
taxonomy has no node for it → LOW exhaustivity

---

Metric 3: Granularity

Question it answers: Is the taxonomy detailed enough to be actionable?

• High granularity = topics are specific enough to act on
• Low granularity = topics are too vague to be useful

LOW granularity:  "Product Issues"  ← What do you do with this?

HIGH granularity: "Battery Life Under Cold Conditions"  
                  ← Now you know exactly what to fix

---

The EEG Summary Table

💡 Together, EEG measures whether a taxonomy is useful in practice — not just structurally valid.

---

Step 5: Metrics-in-the-Loop Refinement

TaxCE doesn't just evaluate — it uses evaluation to improve itself.

How the Loop Works:

┌─────────────────────────────────────────┐
│                                         │
│  Taxonomy → Compute EEG → Diagnose      │
│      ↑                        ↓         │
│  Apply Fix ← Targeted Correction        │
│                                         │
└──────── Repeat until convergence ───────┘

Step-by-Step:

1. Compute EEG scores on current taxonomy
2. Diagnose deficiency — which metric is failing and why?
3. Apply targeted correction:
   • Low exclusivity? → Merge overlapping topics
   • Low exhaustivity? → Add missing topics
   • Low granularity? → Split vague topics into subtopics
4. Recompute EEG on updated taxonomy
5. Repeat until scores converge (stop improving)

💡 "Metrics-in-the-loop" is a powerful design pattern: instead of evaluating after the fact, evaluation drives the construction process.

---

Step 6: How TaxCE Compares to Alternatives

Baseline Categories

TaxCE's Improvements Over Best Baseline

The largest gain is in exhaustivity (+20.5pp), confirming that existing methods systematically miss important topics — especially long-tail ones.

---

Step 7: Synthesis — Putting It All Together

The Complete TaxCE Story

PROBLEM: Unstructured feedback at massive scale
         ↓
INPUT:   Raw text (reviews, transcripts, surveys)
         ↓
STAGE 1: Progressive Condensation
         Raw Text → Segments → Semantic Units → Topics+Definitions
         ↓
STAGE 2: Bottom-Up Hierarchy Construction
         (corpus-grounded, multi-level)
         ↓
STAGE 3: EEG Evaluation
         Measure Exclusivity, Exhaustivity, Granularity
         ↓
STAGE 4: Targeted Refinement Loop
         Diagnose → Fix → Re-evaluate → Repeat
         ↓
OUTPUT:  High-quality, actionable, navigable taxonomy

---

Quick Review: Key Concepts Checklist

• I can explain why taxonomy construction is challenging at scale
• I can describe the three stages of progressive condensation
• I understand why bottom-up construction ensures corpus-groundedness
• I can define Exclusivity, Exhaustivity, and Granularity with examples
• I understand how metrics-in-the-loop refinement works
• I can compare TaxCE to classical, neural, and LLM-based baselines

---

Key Vocabulary