Patent Taxonomy Builder

Patent Taxonomy Builder — Help

Getting Started

Upload Your Patent Data

Drag and drop an Excel (.xlsx, .xls) or CSV file. Supports exports from Quartet, Lens.org, and PatSnap.

Review Auto-Generated Categories

The tool scans CPC codes to build initial categories. Review them in the sidebar — edit, delete, or keep as-is.

Create Your Own Categories

Click + New Category to define categories using CPC codes and keywords with precise matching logic.

Run & Export

Click Run Taxonomy to classify all patents. Export as XLSX for use in the Patent Landscape Dashboard.

Creating a Category

Click + New Category in the sidebar. Each category has a name and two groups of matching conditions:

✓ Must Match ALL

Every item here is required. A patent must match all of them. Use for precise targeting.

Example: CPC G06N + keyword neural network = only patents about neural networks classified under G06N.

| Match ANY

A patent matching any single item here qualifies. Use to cast a wider net.

Example: CPC G06K7/10 or CPC G06F16/33 = patents from either classification.

Using Both Together

A patent qualifies if it matches all required items, OR if it matches any optional item. This lets you define a core condition with broader alternatives.

CPC Code Matching

Prefix matching: G06N matches G06N, G06N3/08, G06N20/00, etc.
Space-insensitive: G06K7/10 matches G06K 7/10 in the data.
Up to 5 CPC codes per category.

Keyword Matching

Matched in patent titles and abstracts (case-insensitive).
Add + for prefix matching: medicat+ matches medication, medicate, medicated.
Up to 10 keywords per category.

Sidebar Actions

✎ — Edit a category (reopen the modal to change its conditions).
× — Delete a category.
Auto-Generate from Data — Rebuild categories from CPC data (replaces current ones).

Priority

Categories you create always take priority over auto-generated ones when assigning patents.

Patent Table

After running the taxonomy, each patent shows:

Category pill — primary assigned category. Click × to unassign.
Confidence % — how many conditions matched. Green (≥70%), amber (≥40%), red (<40%).
Overlapping categories — dashed-border pills for secondary matches. Click to swap to primary.

Bulk Operations

Use checkboxes to select multiple patents.
Assign to… — move selected patents to a category.
Clear — remove category assignments from selected patents.

Search

Use the search bar to filter patents by title, assignee, CPC code, or any other field.

Detail Panel

Click any patent row to open a detail panel showing all fields and category information.

Export XLSX

Click Export XLSX in the top bar. The file includes all original columns plus:

Category — primary category name
Confidence % — match confidence
Overlapping Categories — other matching categories

The exported file can be uploaded directly to the Patent Landscape Dashboard for visual analysis with charts and maps.

Scoring Algorithms

①

Simple Match

Fast • Transparent • Best for clear-cut CPC categories

The simplest approach. Counts how many of your conditions a patent matches and divides by the total number of conditions.

            confidence = items_matched ÷ total_items
          

How it works:

Each CPC code or keyword is worth 1 point
A match is binary — it either matches or it doesn’t
All conditions are weighted equally
No distinction between title and abstract matches

Example: Category has 3 conditions (CPC G06N, keyword “neural”, keyword “learning”). A patent matches 2 of 3 → confidence = 67%

Best for: Categories defined primarily by CPC codes, where keyword matching is secondary. Quick exploratory classification.

②

TF-IDF + CPC Depth

Balanced • Smart weighting • Default choice

Combines Term Frequency–Inverse Document Frequency (TF-IDF) for keywords with CPC classification depth analysis. This is the default algorithm.

            keyword_score = (1 + log(tf)) × log(N / df)

            cpc_score = code_depth + exact_match_bonus

            confidence = weighted_sum / max_possible

Three scoring factors:

TF-IDF for keywords: Rare, discriminating terms (e.g., “CRISPR”, “LiDAR”) score higher than common ones (“method”, “system”). Term frequency is sublinear: 1 + log(tf) prevents repeated words from inflating scores.
CPC depth scoring: More specific CPC codes score higher. G06N3/08 (7 chars) scores more than G06N (4 chars). Exact matches get a +2 bonus.
Title boost: Keywords found in the patent title score ~2× higher than abstract-only matches, since titles describe the core invention.

Example: Keyword “quantum” appears in 12 of 8,700 patents (IDF = 6.6). A patent with “quantum” in the title scores 1 × 6.6 × 2 = 13.2. Keyword “method” appears in 4,000 patents (IDF = 0.78) → only 0.78.

Best for: Categories that mix CPC codes with keywords. The smart weighting helps when some keywords are generic and others are highly specific.

③

BM25 (Best Matching 25)

Industry-standard • Length-normalized • Best for keyword-heavy categories

The ranking function used by search engines like Elasticsearch and Apache Lucene. An evolution of TF-IDF that better handles varying document lengths and term saturation.

            score = IDF × (tf × (k₁ + 1)) / (tf + k₁ × (1 - b + b × dl/avgdl))

            k₁ = 1.5 (term saturation), b = 0.75 (length normalization)

Key advantages over TF-IDF:

Term saturation: The 5th occurrence of a keyword adds less score than the 1st. Controlled by k₁ = 1.5. Prevents keyword-stuffed patents from dominating.
Document length normalization: A short patent (200 words) mentioning “neural” once scores similarly to a long patent (2,000 words) mentioning it 10 times. Controlled by b = 0.75.
Better IDF formula: Uses log((N - df + 0.5) / (df + 0.5) + 1) which handles edge cases more gracefully than standard IDF.

Example: Two patents match keyword “autonomous”. Patent A: 150 words, 2 mentions. Patent B: 3,000 words, 8 mentions. TF-IDF ranks B higher (more mentions). BM25 ranks them similarly because B’s length dilutes the signal.

Best for: Large datasets with abstracts of varying lengths. Categories defined mainly by keywords rather than CPC codes. When you want the most robust text-based matching.

Quick Comparison

	Simple	TF-IDF	BM25
Speed	⚡ Fastest	⚡ Fast	⚡ Fast
Keyword weighting	Equal	By rarity	By rarity + saturation
Doc length handling	None	None	Normalized
Title vs Abstract	Same	Title 2×	Title 2×
CPC depth scoring	Prefix only	Depth-weighted	Depth-weighted
Best for	CPC-heavy	Mixed CPC + KW	Keyword-heavy

How It Works

Patent Taxonomy Builder

Classify thousands of patents into technology categories in four steps

🔍

1. Discover

Upload your dataset. The system extracts the top keywords and CPC classes automatically.

📋

2. Review

Review auto-generated categories. Edit names, add or remove CPC codes and keywords, set AND/OR logic.

⚡

3. Classify

Run the taxonomy. Every patent is scored and assigned to its best-matching category.

Navigate Freely

You can move between all steps at any time. Click the step indicators at the top, or use the “← Discovery” button in the taxonomy builder to go back and refine your categories. Changes are preserved when you navigate.

Step 1: Data Discovery

After uploading your patent file, the system analyzes the entire dataset and extracts two sets of building blocks:

📝 Top Keywords (40–50)

Extracted using TF-IDF (Term Frequency × Inverse Document Frequency) from patent titles and abstracts. This ranks words by how distinctive they are — common words like “method” or “system” are filtered out, while domain-specific terms score high.

Bigrams included: Multi-word phrases like “neural network” or “power supply” are detected as single units
Truncation (+): Keywords automatically get a + suffix for prefix matching. comput+ captures computer, computing, computational, etc.
Patent jargon filtered: Words like comprising, plurality, embodiment, apparatus are removed
Ranked by relevance: Each keyword shows a score bar and patent count

🏷️ Top CPC Classes (10)

The 10 most common CPC (Cooperative Patent Classification) classes in your dataset, shown at the 4-character level (e.g., G06F, H04L). Each class shows its official description and patent count.

Click to expand: See sub-classes found in your data (e.g., G06F16 — Information Retrieval)
300+ descriptions: Built-in lookup table covers sections, classes, subclasses, and common groups
Sorted by count: Most common classes appear first

Step 2: Review & Edit Categories

The system auto-generates 7–8 categories by matching keywords to CPC classes. Each category is shown as an editable card.

Category Card Structure

Name: Click to edit. Auto-generated from CPC class description.
✓ Must Match ALL (AND): Every item in this section must match. Typically used for CPC codes that define the technology field.
| Match ANY (OR): Any single match qualifies. Typically used for keywords that refine within the field.
× Delete: Remove the entire category

Adding Items

CPC codes: Use the dropdown selector, choose CPC, and type a code. A lookup dropdown shows matching classes with descriptions. Click to add.
Sub-classes: When a broad CPC code (e.g., G06F) is in the AND section, click “Show sub-classes” to see and add specific sub-classes.
Keywords: Use the dropdown selector, choose Keyword, type your term. Use + for prefix matching (e.g., sensor+).
Maximum 3 CPC classes per category (across AND and OR sections). No limit on keywords.

AND Logic Example

AND: CPC G06N → patent must be in Machine Learning/AI classification
OR: KW neural+ | KW deep learn+ | KW reinforc+
Result: patent must have CPC G06N and mention at least one of the keywords.

OR Logic Example

OR: CPC H04L | CPC H04W | KW wireless+ | KW network+
Result: patent matches if it has any of these CPC codes or keywords.

Step 3: Classify & Refine

Click “Apply & Classify” to run the taxonomy. Every patent is scored against each category and assigned to its best match.

Scoring Algorithms

Choose from three algorithms via the dropdown in the toolbar:

Simple Match: Items matched ÷ total items. Fast and transparent.
TF-IDF + CPC Depth: Keywords weighted by rarity, CPC codes by specificity, title matches boosted 2×. Default choice.
BM25: Industry-standard search ranking. Handles document length normalization and term saturation. Best for keyword-heavy categories.

In the Taxonomy Builder

Sidebar tree: Shows categories with patent counts. Click to filter the table.
Sub-categories: Click + on any category to add a sub-category that further narrows it.
Edit categories: Click the pencil icon to modify AND/OR rules.
Re-run: Click “▶ Run Taxonomy” after changes to re-classify all patents.
Manual assignment: Select patents and use “Assign to…” to override.
Sort columns: Click any column header to sort by category, confidence, patent number, etc.

Exporting

Export XLSX: Full dataset with category columns marked “X”. Compatible with the Patent Dashboard for visualisation.
Export Selected: Only checked patents, same format.
Patent hyperlinks: Patent numbers are clickable links to Google Patents or Espacenet.
Save JSON: Save taxonomy + data locally for later editing.

← Discovery Button

Click “← Discovery” in the toolbar to go back to Step 2. Your categories sync back so you can add CPC codes, change keywords, and re-apply without starting over.

Tips & Best Practices

Category Design

Start broad, then narrow: Begin with 6–8 top-level categories. Add sub-categories for areas that need finer breakdown.
CPC in AND, keywords in OR: Use CPC codes to define the technology field (mandatory match), and keywords to capture variations within it (any match).
Avoid overlap: If two categories share many patents, make their criteria more distinct or merge them.
Check “Uncategorized”: Patents that don’t match any category may reveal a missing category.

Keyword Techniques

Use + for stemming: detect+ matches detect, detection, detecting, detector, detectable.
Be specific: neural network+ is better than just neural+ which might match unrelated uses.
Check false positives: After classifying, click a category and scan the patent titles. If irrelevant patents appear, tighten the keywords.
4–8 keywords per category usually gives good coverage without noise.

CPC Classification Tips

Use sub-classes when available: G06N3 (Neural Networks) is more precise than G06N (all AI/ML).
Check sub-class expansion: Click “Show sub-classes” to see what’s actually in your data.
Max 3 per category: If you need more CPC codes, consider splitting into two categories.

Workflow for the Dashboard

1. Build your taxonomy here → 2. Export as XLSX → 3. Upload to the Patent Dashboard → 4. Open the “Categories” tab to see distribution charts and trends by year. Each category becomes a filterable column.

Data Discovery

Suggested Categories