review-mining-stp（审查挖矿 stp）

Review Mining STP

Overview

This skill converts review text into Segmentation -> Targeting -> Positioning -> Strategy outputs through a strict workflow contract.

• review scoring workflow: reads raw reviews, infers scored items, assigns theory tags, and preserves verbatim review text.
• scripts: accept scored artifacts only and perform statistical analysis plus report assembly.

The review scoring workflow is an upstream workflow boundary, not a requirement to use any specific API, service, or orchestration tool.

The scripts are tools, not the main workflow. They do not read raw reviews, decide how to score them, or define the scoring process.

When To Use

Use this skill when:

• you need STP outputs from reviews, comments, or feedback text
• you need a repeatable scored-artifact contract before running statistics
• you need segmentation, targeting, or positioning outputs with explicit methods and theory labels
• you need report sections backed by verbatim review quotes instead of unsupported interpretation

Do not use this skill when:

• the task is only qualitative summarization with no scoring and no downstream statistics
• the user only wants raw review tagging with no STP analysis
• the user expects the CLI to ingest raw reviews directly

Theory Framework

Attribute extraction and theory annotation should draw from these four theory families. Additional families may be added when the corpus clearly warrants it, but every attribute must map to at least one family.

1. Product Positioning Theory (product_positioning)

Subtheories:

• attributes — physical or verifiable product properties (e.g. ANSI certification, lens material, weight)
• functions — what the product does in use (e.g. anti-fog, side coverage, UV blocking)
• benefits — perceived value or outcome the customer gains (e.g. confidence, style, value for money)
• usage_context_service_experience — context of use, service touchpoints, post-purchase experience

2. Maslow's Hierarchy of Needs (maslow)

Subtheories:

• physiological — sensory comfort, physical ease, visual clarity during use
• safety — protection from harm, certification compliance, structural durability
• belongingness — fitting into a community, sports group, or professional identity
• esteem — status signalling, brand prestige, professional image display
• self_actualization — enabling personal performance goals, empowerment, achievement

3. Purchase Motivation Theory (purchase_motivation)

Subtheories:

• functional — driven by performance, fit, ergonomics, multi-scenario utility
• security — driven by safety standards, brand trust, durability assurance, after-sales protection
• relational — driven by customer service quality, gifting intent, repeat purchase loyalty

4. Word-of-Mouth Motivation Theory (wom_motivation)

Subtheories:

• altruistic — sharing to genuinely help other buyers (tips, warnings, balanced reviews)
• social_identity — sharing to signal group membership (sports community, professional role)
• self_enhancement — sharing to display expertise or superior knowledge
• emotional_expression — sharing driven by strong positive or negative emotion

Attribute Extraction Rules

When extracting attributes from the full corpus:

• Extract at least 30 attributes whenever the corpus supports it.
• Every attribute must be mappable to at least one of the four theory families above.
• Each attribute must carry theory_annotations listing all applicable family + subtheory pairs.
• Attribute themes are dynamically inferred from the corpus — do not hardcode theme names.
• Freeze the attribute catalog before formal scoring begins.
• Attributes must cover all four theory families. If any family has zero coverage, flag it in attribute_extraction_summary.theory_gap.

Attribute Discovery Pass — How To Execute

The discovery pass is a dedicated read-through of the full corpus before any scoring begins. Its sole output is the frozen attribute catalog. Execute it in four stages:

Stage 1 — Read all reviews and collect raw signals

Read every review in the corpus. For each review, note any concern, praise, complaint, or observation the reviewer expresses about the product or their experience. Do not score yet. Collect these as raw signals in a working list. Signals can be short phrases or paraphrases — they do not need to be final attribute names yet.

Examples of raw signals:

• "fogged up immediately with mask on"
• "arms hooked in my hair every time I removed them"
• "military-grade, Z87 certified"
• "bought three pairs over two years"
• "bought these as a gift for my son"

Stage 2 — Cluster signals into candidate attributes

Group raw signals that represent the same underlying customer concern or product dimension. Each cluster becomes one candidate attribute. Apply these rules:

• One attribute per distinct construct. Do not merge two different concerns just because they co-occur (e.g. "fogging" and "scratching" are separate attributes even if one reviewer mentions both).
• Split an attribute if reviewers clearly treat it as two separate things (e.g. "nose pad comfort" and "nose pad staying in place" may warrant two attributes if complaints differ).
• Name each attribute with a short noun phrase that a non-specialist can understand. Use the language reviewers actually use, not academic terminology.
• Count how many reviews contributed signals to each cluster — this becomes mention_count.

Stage 3 — Map every attribute to the four theory frameworks

For each candidate attribute, assign theory_annotations by working through all four theory families in order:

product_positioning — ask: does this attribute describe a physical property (attributes), a functional capability (functions), a perceived benefit or outcome (benefits), or a usage context / service touchpoint (usage_context_service_experience)? Assign all that apply.

maslow — ask: which need does concern about this attribute reflect?

• physiological: physical sensation, visual comfort, weight, pressure on face
• safety: protection level, certification, structural durability, after-sales security
• belongingness: fitting into a community, professional group, sport team
• esteem: brand status, professional image, visible identity
• self_actualization: enabling personal goals, performance achievement, empowerment

purchase_motivation — ask: what drives someone to care about this at the point of purchase?

• functional: performance, fit, ergonomics, multi-use utility
• security: standards compliance, brand trust, warranty, durability assurance
• relational: customer service, gifting intent, loyalty and repeat purchase

wom_motivation — ask: why would a reviewer write about this attribute?

• altruistic: to warn or help other buyers
• social_identity: to signal membership in a group (sport, profession, military)
• self_enhancement: to display expertise or superior product knowledge
• emotional_expression: because strong feeling (delight or frustration) compels them to write

An attribute may carry multiple families and multiple subtheories. There is no maximum. However, every attribute must carry at least one family, and the full catalog must cover all four families.

Stage 4 — Freeze and validate the catalog

Before scoring begins:

1. Confirm the attribute count meets the target_minimum (at least 30 when corpus supports it).
2. Confirm all four theory families appear at least once across the catalog. If any family is absent, revisit Stage 2 — missing coverage usually means signals were merged incorrectly or a whole class of reviewer concerns was overlooked.
3. Assign a stable attribute_key to each attribute (snake_case, e.g. anti_fog_performance, hinge_durability). Keys must not change after freezing.
4. Write the plain_language_definition for each attribute — one sentence describing what the attribute measures and what a high vs low signal looks like in a review.
5. Select one example_review_id and example_quote per attribute from the raw corpus. The quote must be verbatim.
6. Record the frozen catalog in attribute_catalog.csv and review_foundation.json -> dimension_catalog before any scoring row is written.

No attribute may be added, removed, or renamed after the catalog is frozen. If a gap is found during scoring, record it in attribute_extraction_summary.shortfall_reason and complete scoring with the frozen catalog as-is.

Workflow Contract

Review Scoring Workflow

The review scoring workflow is the main process. It is responsible for:

• reading every review one by one
• extracting at least 30 important attributes from the full review set whenever the corpus supports it
• freezing the attribute catalog before formal scoring begins
• inferring scored items from the full review set
• assigning each item to dynamic themes inferred from the full review set
• attaching theory metadata at both family and subtheory level — exclusively from the four permitted families
• keeping a paired salience and product-quality scoring plan for every inferred attribute
• preserving the original review_text so later report evidence can quote the real source text

Theme names and theme count are not fixed. They come from the corpus, not from a hardcoded taxonomy.

Two Scoring Axes

This skill uses two distinct scoring axes applied to different units of analysis:

Axis A — Customer × Attribute: Salience (0–7)

Applied per review (or per customer). Measures how prominently the attribute features in a given review.

• 0: the attribute is absent from the review — no relevant content at all
• 1–3: the attribute is mentioned slightly or indirectly
• 4: the review is neutral, ambiguous, or tangential on this attribute
• 5–6: the review clearly and explicitly addresses this attribute
• 7: the review strongly emphasises this attribute as a central concern

Dependency rule: when salience = 0, the attribute is treated as absent for this review and must not be included in any per-review analysis. Only reviews with salience ≥ 1 are counted as mentioning the attribute.

Axis B — Product × Attribute: Quality Score (0–10)

Applied per product (aggregated across all reviews for that product). Measures the overall quality or performance of the product on this attribute as judged by its reviewers.

• 0: extremely poor — near-universal negative evaluation
• 1–3: below average — more negative than positive mentions
• 4: below average leaning negative — complaints outweigh praise
• 5: mixed or neutral — roughly equal positive and negative signals
• 6–7: above average — more positive than negative, with some complaints
• 8–9: strong — predominantly positive, only minor issues noted
• 10: exceptional — near-universal praise with no notable complaints

This score is a product-level aggregate, not a per-review score. It is computed or estimated after all reviews for a product have been read, and represents the overall quality positioning of the product on that attribute.

Column naming:

• Per-review salience: <attribute_key>_salience
• Product-level quality: <attribute_key>_quality

Scoring Workflow Steps

1. Read each review one by one.
2. Run an attribute-discovery pass across the full corpus.
3. Freeze the attribute catalog with definitions, theory annotations (from the four permitted families only), and paired score-column names.
4. Score every review against the frozen attribute catalog using Axis A (salience 0–7).
5. After reading all reviews per product, compute Axis B (quality 0–10) per product per attribute.
6. Convert qualitative review text into quantitative data on both axes.
7. Use the scored output for downstream statistical analysis and research models.

If upstream information is incomplete, the review scoring workflow may produce MissingDataOutput.

Scripts

The scripts start only after scoring is already complete. Their responsibilities are:

• schema and numeric sanity checks for scored artifacts
• reliability checks
• factor or theme reduction
• clustering for segmentation
• ANOVA / regression for continuous targeting outcomes
• chi-square / logistic regression for binary targeting outcomes
• perceptual-map generation
• ideal-point distance analysis
• pairwise competition-distance analysis
• report assembly with explicit methods, theory labels, and evidence quotes

The scripts do not:

• infer items from raw review text
• define the scoring rubric
• enforce the wording of the scoring process
• change the attribute catalog during statistical execution
• rewrite review quotes
• auto-backfill missing scored artifacts

If prerequisites are missing, the scripts return MissingPrerequisiteOutput.

Canonical Input Artifacts

review_scoring_table.csv

Required columns:

• review_id
• unit_id
• brand
• product
• review_text

All inferred attributes must appear as salience columns (Axis A, per review):

• <attribute_key>_salience

Each salience column must follow these rules:

• integer scores only, range 0–7
• 0 means the attribute is absent from this review

Optional metadata columns may include:

• profile_*
• channel
• rating

The table is per-review. If no stable person-level identity exists, unit_id may default to review_id.

product_quality_scorecard.csv

New artifact replacing the per-review valence axis. Required columns:

• product
• brand
• n_reviews — number of reviews used to compute the scores
• One <attribute_key>_quality column per attribute (Axis B, 0–10)

This is the product × attribute quality matrix. Each cell is the product-level quality score for that attribute, estimated from all reviews of that product.

review_foundation.json

Required keys for the scripts:

• dimension_catalog
• theme_mapping
• attribute_extraction_summary
• people_insights
• product_triggers
• context_scenarios
• system1_system2_split
• maslow_keywords

Optional audit metadata:

• scoring_rubric

Each dimension_catalog item must include:

• column
• label
• theme
• attribute_group
• salience_column
• quality_column
• stat_roles
• plain_language_definition
• theory_annotations

attribute_group must use one of:

• attribute_function
• benefit_use
• brand_personality
• brand_image

theory_annotations must map each scored item to at least one theory family plus subtheory. The four default families for this skill are:

• product_positioning (subtheories: attributes, functions, benefits, usage_context_service_experience)
• maslow (subtheories: physiological, safety, belongingness, esteem, self_actualization)
• purchase_motivation (subtheories: functional, security, relational)
• wom_motivation (subtheories: altruistic, social_identity, self_enhancement, emotional_expression)

Additional theory families may be used when the corpus clearly calls for them. Any added family must be documented in review_foundation.json with its name, rationale, and subtheory list.

attribute_extraction_summary must record:

• target_minimum
• actual_count
• shortfall_reason
• theory_gap — list any of the four theory families with zero attribute coverage

attribute_catalog.csv

Required columns:

• attribute_key
• label
• theme
• attribute_group
• definition
• source_type
• mention_count
• salience_column
• quality_column
• example_review_id
• example_quote
• theory_families — comma-separated list of applicable theory families from the four permitted
• theory_subtheories — comma-separated list of applicable subtheories

The catalog is the script-facing bridge from upstream attribute extraction into downstream statistics and report evidence.

Auto-Discovered Context Files

• analysis_context.json

- analysis_goal - comparison_axes - scope_limits

• brands.json
• ideal_point.json

Run Modes

• full: starts from canonical scored artifacts and emits the three statistical intermediates
• full canonical input requires review_scoring_table.csv + product_quality_scorecard.csv + review_foundation.json + attribute_catalog.csv + analysis_context.json + brands.json + ideal_point.json
• segmentation: uses review_foundation.json + segmentation_variables.csv
• targeting: uses segment_profiles.json + targeting_dataset.csv
• positioning: uses positioning_scorecard.csv + brands.json + ideal_point.json
• custom: runs only requested downstream modules

Generated intermediate artifacts in full mode:

• segmentation_variables.csv
• targeting_dataset.csv
• positioning_scorecard.csv

Statistical Rules

Segmentation

• standardize salience columns (Axis A) across reviews to identify customer concern patterns
• use factor_analysis -> K-means
• rerun when any cluster falls below the >5% guardrail
• record cluster_threshold, reruns_performed, and final_k
• retain System 1 / System 2, Maslow, cluster share, and consumer-portrait outputs

Targeting

• resolve current and potential targeting variables from dimension_catalog.stat_roles
• allow analysis_context.comparison_axes to override the default comparison axes
• model salience columns (Axis A) as customer-side drivers
• model quality columns (Axis B) as product-side performance indicators
• use ANOVA / regression for continuous outcomes
• use chi-square / logistic regression for binary outcomes
• emit pairwise_comparison_table when ANOVA significance justifies post-hoc comparison
• emit priority_segments, secondary_segments, and deprioritized_segments

Positioning

• build the scorecard from stat_roles containing positioning
• use quality columns (Axis B) from product_quality_scorecard.csv as the primary product positioning features
• cross-reference with salience columns (Axis A) to weight attributes by customer concern level
• default to factor_analysis
• allow MDS when similarity-based input is explicitly requested
• include ideal-point distance and pairwise competition distance
• draw the perceptual map as a Python-generated figure from the coordinate table
• treat perceptual_map_figure + perceptual_map_coordinate_table + perceptual_map_method + perceptual_map_interpretation as the public positioning-map contract
• allow factor-analysis-only vector and projection diagnostics as optional internal outputs
• never fabricate attribute vectors for MDS
• emit dynamic_scorecard_summary with distance, gap, reliability, and validity sections

Report Contract

The final report must contain an Attribute Extraction Summary that shows:

• target_minimum
• actual_count
• shortfall_reason
• theory_gap — any of the four theory families with zero coverage
• discovered themes
• attribute-group counts
• theory family and subtheory coverage breakdown
• representative attributes with real example quotes

Each major report section must contain:

• What this section is doing
• Axis modeling summary — specify whether Axis A (salience), Axis B (quality), or both are used
• Statistical methods used
• Theories used — must name specific families and subtheories from the four permitted
• Theme coverage summary
• Theory coverage summary
• Plain-language explanation
• Evidence quotes

Each major report section must also contain a non-empty findings list.

Each finding must contain:

• finding_id
• finding_statement
• business_implication
• axes_used — salience, quality, or both
• methods_used
• theories_used
• themes_used
• subtheories_used
• reproducibility
• statistical_results
• plain_language_explanation
• evidence_quotes

Each reproducibility package must contain:

• input_artifacts
• input_columns
• filters
• preprocessing
• analysis_steps
• decision_rule

Each statistical_results package must contain:

• method_family
• test_or_model
• sample_size
• statistic
• degrees_of_freedom
• p_value
• effect_size
• coefficient
• confidence_interval
• result_direction
• axis_breakdown

Evidence-quote rules:

• quotes must come verbatim from review_scoring_table.csv.review_text
• each quote must include review_id
• each quote must explain why it matters
• each quote must link back to the scored items it supports
• when canonical review evidence is available, each major section should include 2-3 quotes
• when canonical review evidence is available, each finding should include at least 1 quote

The goal is to make the report readable for non-specialists while keeping every key claim traceable to real review text and reproducible from the emitted statistical artifacts.

The final report should visibly show:

• the dynamically inferred themes for this corpus
• which findings use which themes
• theory families plus subtheories — drawn exclusively from the four permitted families
• which subtheories are not_evidenced in the current dataset

Hard Rules

• Never blur review-scoring inputs with script-ready artifacts.
• Never let scripts consume raw reviews directly.
• Never hardcode a fixed item count into the validator or statistical pipeline.
• Always use product as the product field name.
• Axis A scoring (customer × attribute): always use salience 0–7.
• Axis B scoring (product × attribute): always use quality 0–10.
• Never use valence as a column name — it is replaced by quality in this skill.
• Always preserve verbatim review_text for evidence quoting.
• Always state the statistical method and theory used in each major report section.
• Always show how Axis A and Axis B were modeled in each major report section.
• Always show dynamic theme coverage and theory coverage in the report body.
• Always show the attribute-extraction summary and representative attributes in the report body.
• Always attach reproducibility steps and statistical results to each finding.
• Never fabricate evidence quotes or attribute vectors.
• Theory annotations default to the four built-in families: product_positioning, maslow, purchase_motivation, wom_motivation. Additional families may be introduced when the corpus clearly warrants it, provided they are documented with name, rationale, and subtheory list in review_foundation.json.
• Every attribute must be covered by at least one theory family. Attributes with no theory mapping must be flagged and reconsidered.

References

• references/01-router-and-gates.md
• references/02-segmentation.md
• references/03-targeting.md
• references/04-positioning.md
• references/05-output-contract-and-quality-rules.md
• references/06-end-to-end-examples.md
• references/07-traceability-evidence-matrix.md
• references/08-verification-scenarios.md

Scripts

• Install dependencies: python -m pip install -r requirements.txt
• Run analysis: python scripts/run_review_mining_stp.py --run-mode <mode> --input-dir <artifacts> --output-dir <o>
• Validate outputs: python scripts/validate_review_mining_stp.py --run-mode <mode> --output-dir <o>
• Script boundary: statistical analysis only; raw reviews must first be converted into scored artifacts during the review scoring workflow