review-salience-xlsx（审查 salience XLSX）

Review Salience → PCA → K-means Customer Segmentation

Full pipeline from a review corpus to customer segments:

1. Score every review on every attribute (salience 0–7) → review × attribute matrix
2. Reduce attributes to latent dimensions via PCA
3. Segment reviews into customer groups via iterative K-means

Each stage is independently useful. Run only the stages the user needs.

For the Excel output format, read references/xlsx-format.md. For PCA and K-means implementation details, read references/pca-kmeans.md. For a worked example (safety-eyewear, 923 reviews, 30 attributes, 4 clusters), read references/worked-example.md.

Salience measures *how prominently* a reviewer mentions an attribute — not sentiment. Each cell is an integer 0–7:

For the Excel output format, read references/xlsx-format.md. For PCA and K-means implementation details, read references/pca-kmeans.md. For a worked example (safety-eyewear corpus, 923 reviews, 30 attributes, 4 clusters), read references/worked-example.md.

Concepts

Attribute catalog

A frozen, ordered list of attributes produced upstream (e.g. by the review-scoring-docx skill). Each attribute has an id (zero-padded, e.g. 01) and a label. The catalog must not change after scoring begins.

Scorer

The component that reads one review and returns N integers. This skill is scorer-agnostic: Claude reads and scores by default, but the architecture supports swapping in any external scorer without changing the rest of the pipeline. See Scorer contract below.

Salience matrix

A table with one row per review and one column per attribute, plus metadata columns (review_id, product, review_text). Column names follow the pattern s01, s02 … sN. This is the input to both PCA and K-means.

PC scores matrix

Produced by PCA on the standardised salience matrix. Shape: (n_reviews, n_components). Each column is a latent dimension (e.g. "整體使用價值感", "場景創新適應力"). This matrix is the direct input to K-means clustering.

Customer segments

K-means groups applied to the PC scores matrix. Each review (= each customer voice) is assigned to exactly one segment. The iterative pruning rule ensures no segment is smaller than 5% of the total corpus.

Workflow

Step 0 — Locate required skills

Check your available skills for an xlsx skill (covers .xlsx or spreadsheet creation) and a docx skill if Word output is needed. Read their SKILL.md files before writing any output code.

Step 1 — Ingest reviews

Load all reviews for each product. Never truncate. Accept all languages.

import csv

def load_reviews(filepath):
    candidates = ['body', 'Body', 'review', 'text', 'content']
    with open(filepath, encoding='utf-8', errors='replace') as f:
        reader = csv.DictReader(f)
        col = next((c for c in candidates if c in reader.fieldnames), None)
        if col is None:
            raise ValueError(f"No text column found. Headers: {reader.fieldnames}")
        return [r[col].strip() for r in reader
                if r[col].strip() and len(r[col].strip()) > 15]

Print a per-product count summary before proceeding.

Step 2 — Confirm the attribute catalog

Either receive the catalog from upstream or rediscover it by reading the corpus. Freeze as an ordered list before scoring starts:

ATTRS = [
    ("01", "attribute label 1"),
    ("02", "attribute label 2"),
    # ...
]

Step 3 — Score reviews ← scorer swap point

This is the scorer contract boundary. Anything that satisfies the contract below can replace Claude's built-in scoring.

Scorer contract

Input: a single review string (any language) Output: a list of integers, one per attribute, in catalog order, each 0–7

Built-in scorer (Claude reads semantically): Read each review, understand its meaning, assign scores. No keyword matching. See references/worked-example.md for calibration examples.

External scorer (n8n / API / other): See references/external-scorer.md for integration patterns.

scores = {}
for pid, reviews in all_reviews.items():
    scores[pid] = []
    for review_text in reviews:
        scores[pid].append(score_one_review(review_text, ATTRS))

Save progress incrementally if corpus > 200 reviews.

Step 4 — Build salience Excel

See references/xlsx-format.md for exact sheet layout, colour scheme, and openpyxl patterns. Two sheets: full matrix + product summary.

Step 5 — PCA (if requested)

Read references/pca-kmeans.md → Section A before writing any PCA code.

Key steps:

1. Standardise: X_std = StandardScaler().fit_transform(X)
2. Fit PCA: use Kaiser criterion (eigenvalue > 1) to choose n_components
3. Compute factor loadings: loadings = components_.T × sqrt(eigenvalues_)
4. Name each PC by its dominant loadings (|≥ 0.30|)
5. Save PC scores matrix for clustering: PC = pca.fit_transform(X_std)

Step 6 — K-means segmentation (if requested)

Read references/pca-kmeans.md → Section B before writing any clustering code.

Key steps:

1. Scan K=2–9 for Elbow + Silhouette
2. Pick starting K (favour interpretability over peak silhouette if gap is small)
3. Run iterative pruning: remove clusters < 5% of corpus, decrement K, refit
4. After convergence: call km.predict(ALL_reviews) — never discard pruned reviews
5. Name clusters by PC centroid pattern and top attribute means

Step 7 — Output

• Save .xlsx to /mnt/user-data/outputs/
• Save .docx to /mnt/user-data/outputs/ (if Word output was requested)
• Call present_files for every output file
• Write a prose summary: total reviewed, cluster sizes + names, top finding per cluster

Hard rules

1. Never truncate reviews. Every valid review (> 15 chars) must be scored and clustered.
2. All languages count. Do not filter by language.
3. Catalog is frozen before scoring. No mid-run additions or reordering.
4. Integer scores only. Each salience cell is a whole number 0–7.
5. Scorer is swappable. The pipeline must not assume Claude is the scorer.
6. Never discard pruned reviews. After iterative pruning converges, use km.predict() on the full corpus to assign every review — including those removed during pruning — to the nearest final centroid.
7. Use absolute column widths. Percentage widths break in Google Sheets.
8. Present with present_files. Never ask the user to navigate to the file.