💻 computer science

A Dual-Channel Patent Vectorization Method Based on Claim Structure Analysis

This paper proposes a dual-channel patent vectorization method that leverages claim structure analysis to separately encode system-level and detail-level semantics, which are then fused via a gated mechanism with decorrelation correction to significantly improve patent matching accuracy over existing full-text baselines.

Original authors: Jiangtao Li, Wenlong Feng, Mengxing Huang, Siling Feng, Fu Gao

Published 2026-06-26

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jiangtao Li, Wenlong Feng, Mengxing Huang, Siling Feng, Fu Gao

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Problem: The "One-Size-Fits-All" Patent Soup

Imagine you are trying to find a specific recipe in a massive cookbook. The problem with current computer methods for reading patents is that they treat every patent like a giant, unorganized bowl of soup. They dump the entire text—the main idea, the specific ingredients, the cooking times, and the safety warnings—into a single blender.

The researchers argue that this approach misses the nuance. In a patent, the Independent Claims are like the main recipe (e.g., "A cake made of flour and eggs"). The Dependent Claims are the specific details and restrictions (e.g., "The flour must be organic," or "The oven temperature must be exactly 350°F").

When computers mash these together into one single "vector" (a digital fingerprint of the text), the specific details get lost in the noise of the main idea. It's like trying to find a recipe for a "gluten-free cake" but the computer only sees the general concept of "cake," making it hard to distinguish between a standard cake and a specialized one.

The Solution: A Dual-Channel Kitchen

To fix this, the authors (Jiangtao Li and team from Hainan University) built a new system called a Dual-Channel Patent Vectorization Method. Instead of blending everything together, they set up two separate "channels" or conveyor belts to process the patent text.

Channel 1: The "Big Picture" Conveyor (The Overall Channel)

This channel only looks at the Independent Claims. Think of this as the "Executive Summary" or the "Main Dish." It captures the core system-level idea of the invention. It answers the question: What is this invention generally?

Channel 2: The "Fine Print" Conveyor (The Local Channel)

This channel only looks at the Dependent Claims. Think of this as the "Ingredients List" or the "Special Instructions." It captures the specific restrictions, extra features, and technical details. It answers the question: What makes this specific version unique?

The Secret Sauce: The "Smart Mixer"

Just having two separate channels isn't enough; you still need to combine them to get the final result. But if you just pour them together equally, you might get a messy mix. The authors designed a Dual-Channel Gated Concatenation Fusion Mechanism.

Think of this as a Smart Mixer with a Volume Knob.

Cleaning the Signal (Decorrelation): Before mixing, the system checks if the two channels are repeating the same information. If the "Big Picture" and the "Fine Print" are saying the same thing, the system subtracts the duplicate parts from the "Fine Print" channel. This ensures the second channel is only adding new details, not just echoing the first one.
Reading the Room (Fusion Tendency): The system looks at the specific patent to decide how much weight to give each channel.
- If a patent has very few key words that define it (high "Keyword Concentration"), the system knows the "Big Picture" is the most important part. It turns the volume up on Channel 1.
- If a patent is full of unique, rare technical terms (high "Term Specificity"), the system knows the "Fine Print" is crucial. It turns the volume up on Channel 2.
The Final Blend (Gated Concatenation): Using a mathematical "gate" (like a traffic light), the system dynamically decides the perfect ratio of Main Idea to Specific Details for that specific patent. It then combines them into a single, high-quality digital fingerprint.

The Results: A Better Search Engine

The researchers tested this new method against older methods (like standard word counting or simple text averaging) using a real patent database.

The Analogy: Imagine searching for a specific type of car.
- Old Methods: Might return a generic "Sedan" because they missed the specific detail that the car has "turbocharged V6 engines."
- New Method: Because it kept the "Main Idea" (Sedan) and the "Details" (Turbo V6) separate but connected, it found the exact car you wanted much more accurately.

The Findings:

The new method was better at finding the right patents (higher Precision).
It ranked the best matches higher on the list (higher MAP and NDCG scores).
It did this by respecting the legal structure of the patent (Independent vs. Dependent claims) rather than treating the text as a flat block of words.

The Trade-Off

The paper notes that this "Smart Mixer" takes a bit more computing power and time to run than the simple "blender" methods. However, the authors argue that the extra time is worth it because the results are much more accurate and useful for finding similar technologies.

In short: The paper proposes that to understand a patent, you shouldn't just read the whole thing at once. You should separate the "Main Idea" from the "Specific Rules," clean up the duplicates, and then mix them together in a smart way that adapts to the specific document. This leads to a much clearer picture of what the invention actually is.

Technical Summary: A Dual-Channel Patent Vectorization Method Based on Claim Structure Analysis

Problem Statement
Existing patent vectorization methods predominantly rely on unified full-text representation strategies. While approaches such as TF-IDF, Word2Vec, and Doc2Vec have advanced semantic modeling, they often treat patent texts as flat inputs. This approach fails to explicitly leverage the hierarchical structure inherent in patent claims. Consequently, these methods struggle to simultaneously capture the overall system-level semantics (typically found in independent claims) and local technical details (found in dependent claims). The mixing of these distinct semantic levels within a single vector space can dilute discriminative ability, leading to reduced patent matching accuracy and lower quality in vector representations.

Methodology
The paper proposes a Dual-Channel Patent Vectorization Method that explicitly models the structural hierarchy of patent claims. The workflow consists of three primary stages:

Claim Structure Segmentation and Dual-Channel Construction:
- Segmentation: Patent claims are segmented based on numbering and citation relationships. Claims that do not cite others and summarize the core solution are classified as Independent Claims; those citing previous claims to add features or restrictions are classified as Dependent Claims.
- Channel Definition:
  - Overall Channel: Constructed from concatenated independent claims to represent system-level semantics. This channel is modeled using Doc2Vec to capture the core technical solution.
  - Local Channel: Constructed from concatenated dependent claims to represent fine-grained technical details (parameters, additional features). This channel is modeled using TF-IDF weighted word vector aggregation to highlight discriminative terms.
- Local Decorrelation Correction: To prevent redundancy between the two channels (since dependent claims often repeat core terms), a decorrelation step is applied. The local channel vector is adjusted by subtracting a proportion of the overall channel vector, ensuring the local channel focuses on supplementary, non-redundant details.
Dual-Channel Gated Concatenation Fusion:
- Feature Extraction: Two statistical features are extracted from the complete claim text: Keyword Concentration (measuring the weight of top keywords) and Term Specificity (measuring the rarity and discriminative power of terms).
- Fusion Tendency: These features are normalized and combined into a patent-sample-level fusion tendency ( $\alpha_i$ ), controlled by a balance parameter ( $\eta$ ).
- Gated Mapping: A sigmoid function converts the fusion tendency into a gate weight ( $gate_i$ ). This allows the model to dynamically adjust the contribution ratio of the overall and local channels based on the specific textual characteristics of each patent sample.
- Final Fusion: The final patent vector is generated via gated concatenation: $[(1 - gate_i)v_{o,i}; gate_i v'_{l,i}]$ , where $v_{o,i}$ is the overall vector and $v'_{l,i}$ is the decorrelated local vector.

Key Contributions

Claim-Structure-Oriented Framework: The paper introduces a representation framework that separates independent and dependent claims into an overall channel and a local channel, respectively. This explicitly distinguishes core solution semantics from detailed limitations, enhancing the vector's ability to capture claim hierarchy.
Gated Concatenation Mechanism: A novel fusion mechanism is designed to address semantic complementarity. By using local decorrelation to reduce redundancy and a gated mapping strategy driven by keyword concentration and term specificity, the method dynamically balances the influence of overall and local semantics for each patent sample.
Empirical Validation: The method is rigorously evaluated on a real-world patent retrieval dataset (DAPFAM), demonstrating improvements over standard full-text baselines.

Experimental Results
Experiments were conducted using the DAPFAM dataset, comparing the proposed method against full-text TF-IDF, full-text Avg Word2Vec, and full-text Doc2Vec baselines.

Ranking Quality: The proposed method achieved superior performance across all metrics: Precision@10 (P@10), Mean Average Precision@10 (MAP@10), and Normalized Discounted Cumulative Gain@10 (NDCG@10). Specifically, it reached a P@10 of 0.3316, outperforming the best baseline (Doc2Vec) at 0.3172.
Ablation Studies: Results confirmed that the dual-channel structure is the foundation of performance gains. While direct concatenation improved results, the addition of decorrelation and the full gated mechanism yielded the highest performance, validating the necessity of reducing redundancy and dynamically weighting channels.
Efficiency: The proposed method incurs higher computational costs (approx. 7.26 ms/doc for online vectorization) compared to lightweight baselines like TF-IDF (0.89 ms/doc) or Avg Word2Vec (0.51 ms/doc). However, the authors note this overhead remains within an acceptable range, representing a reasonable trade-off for the significant gains in retrieval accuracy.

Significance and Claims
The paper claims that its primary significance lies in shifting patent vectorization from a "flat" full-text modeling approach to a hierarchical, structure-aware modeling approach. By acknowledging that independent and dependent claims serve different semantic functions, the method provides a more targeted vector representation. This improves the quality of patent vector representation, thereby supporting more accurate similar patent retrieval, technology opportunity discovery, and patent value analysis. The authors position this work as a step toward systematic evaluation and task adaptation in patent NLP, offering a modeling approach that better distinguishes between core technical solutions and local technical details.