INVENTOR'S TECHNOLOGY COLUMN · AI AND PATENT ANALYTICS
Training AI on Patent Claims and Detailed Descriptions to Identify Patent Risk
The same technology can be described in entirely different words. This article explains an AI-based patent-analysis technique that turns the native structure of patent documents into training signals and identifies passages in a lengthy disclosure that are most closely related to a claim.
1. Why Keyword Search Misses Technical Relationships
Patent searching is not simply a matter of finding identical words. Comparing a product description with a claim is relatively straightforward when both use the same terminology. The harder cases arise when two documents describe the same technology in different language. A drafter may reorder components, move between broader and narrower concepts, or describe the same function in different terms. A keyword search can easily miss those connections.
The technology I invented and filed for patent protection begins with that problem. Its objective is to move beyond word matching and train an AI model to evaluate how closely the technical meaning of a claim corresponds to a passage in another document. The central idea is to use structural relationships already present within patent documents as training data. Rather than requiring experts to read tens of thousands of documents and manually create an answer key, the system is designed so that the patent documents themselves supply part of the training signal.
2. Turning Patent Documents into Training Data
The starting point is the structure of a patent document. The claims define the legal boundaries of the invention, while the detailed description explains what the invention is and how it may be implemented. Article 42(4)(i) of the Korean Patent Act requires the claims to be supported by the detailed description. That requirement helps prevent an applicant from monopolizing subject matter that was not disclosed and enables a person skilled in the art to understand the relationship between the claimed invention and the specification.
Viewed through a machine-learning lens, that legal structure creates an opportunity. A claim and a passage from the detailed description of the same patent will ordinarily have a strong relationship. By contrast, a claim and a randomly selected passage from a different patent are statistically more likely to be unrelated. The system can therefore assign label 1 to a claim–description pair drawn from the same patent and label 0 to a pair assembled from different patents.
The immediate benefit is a substantial reduction in manual labeling. In a conventional supervised-learning project, experts might need to read each document pair and decide whether it is related. Patents take time to read carefully, so the cost rises quickly as the dataset grows. Automatically pairing claims with passages from the same patent—and combining claims with randomly selected passages from other patents—can remove much of that bottleneck.
Negative samples are not mere filler. If a model sees only highly related pairs, it may learn to treat a few overlapping words as sufficient evidence of a meaningful relationship. Training it on claim–description pairs drawn from different patents forces the model to distinguish superficial word overlap from technical correspondence. Positive samples show the model what a relationship looks like; negative samples help define its boundary.
3. Working Around the 512-Token Limit
Document length presents the next obstacle. In its basic form, BERT accepts no more than 512 tokens in a single input. A patent's detailed description, however, often runs to thousands of words. Feeding a claim and the entire description into the model at once is therefore impractical.
My proposed approach first divides the detailed description into chunks of approximately 310 tokens. It then calculates the semantic relationship between the claim and each chunk and selects the highest-scoring passages as input candidates. This differs from mechanically taking text from the beginning of the document. The limited input window is allocated first to passages most likely to bear directly on the claim.
A dot product between vector representations can be used to rank the chunks. The underlying intuition is straightforward. The claim is represented as one semantic vector and each description chunk as another. The operation scores how closely the two vectors point in the same direction. A higher score indicates a greater likelihood that the claim and the passage are semantically related.
- Segment: Divide the detailed description into chunks of roughly 310 tokens.
- Score: Calculate the semantic relationship between the claim and each chunk.
- Select: Choose the highest-scoring chunks as candidates for the BERT input.
4. How BERT Reads the Relationship Between Two Texts
The selected text is formatted so that BERT can distinguish the two inputs. A representative sequence is [CLS] + claim + [SEP] + detailed description + [SEP]. The [SEP] tokens mark the boundary between the texts, while segment embeddings identify which tokens belong to which input. The final representation of the [CLS] token can then serve as a summary vector for classifying the relationship between the two texts.
This structure prevents the claim and the description from collapsing into a single undifferentiated passage. Human readers naturally use headings, paragraphs, and line breaks to identify document boundaries. A model has no comparable intuition unless the input structure expressly tells it, in effect, “Text A ends here, and Text B begins here.”
During training, label-1 and label-0 examples are handled through an alternating-batch strategy. Rather than mixing both classes indiscriminately from the outset, the system calculates the loss for each group and updates the model parameters in a direction that reduces both losses. The purpose is to keep the model's decision rule from drifting toward one class merely because that class is larger or easier to learn.
5. What the Reported Experiment Shows
In the experiment described in the source manuscript, cross-entropy loss decreased from 0.89 to 0.19 over 15 training epochs, and accuracy reached approximately 80%. The manuscript also reports that the model classified all 17 text pairs in a comparison set that used different terminology as label 1.
More important than the headline numbers is the model's response to changes in wording. One of the central difficulties in patent searching is that the same technology can be obscured by different language. Suppose one document refers to an “adhesive retention mechanism,” while another describes the same arrangement using entirely different terminology. A keyword search may push the documents apart because they share few words. A semantic model is intended to do the opposite: it asks what function each component performs in context and how the components relate to one another.
6. The Line Between Semantic Similarity and Infringement
This discussion naturally calls to mind the doctrine of equivalents. Patent infringement analysis does not necessarily end merely because an accused product departs from the literal wording of a claim. Under Korean Supreme Court precedent, an altered element may in appropriate circumstances be treated as equivalent when requirements concerning the invention's problem-solving principle, substantially identical operation and effect, and interchangeability are satisfied, absent a recognized bar to equivalence.
A legal analysis must separately address whether every claim limitation is satisfied, as well as prosecution history, the prior art, any deliberate exclusion or disclaimer, and each requirement of the doctrine of equivalents. The practical value of this invention therefore does not lie in having AI replace a judge or lawyer. Its more useful role is to narrow the set of documents that a human expert must read. Among thousands—or tens of thousands—of patents and product documents, the system can elevate candidates that correspond strongly to a claim, after which patent counsel, litigators, and R&D personnel can conduct a rigorous review.
7. Practical Applications and the Value of Data
The same semantic-comparison architecture supports several practical use cases. Patent-risk analysis, portfolio monitoring, and prior-art searching all converge on the task of finding technical relationships expressed in different language.
- Product development: Compare competitor patents with product specifications to identify potential risk candidates early.
- Rights monitoring: Compare descriptions of newly released products with existing claims to prioritize human review.
- Prior-art searching: Surface references that describe relevant subject matter using language different from the claim.
In patent AI, the most expensive resource is not always GPU capacity or model parameters. Often the harder problem is obtaining good training data. Expert labels can be accurate, but they are slow and costly to produce. Public patent documents are abundant, yet they are not ready-made answer keys. This invention seeks a middle path: use the legal and documentary structure already embedded in patents to reduce the cost of generating training data.
Describing the project simply as “turning law into mathematics” misses the point. More precisely, it identifies signals that AI can learn from the structure and rules accumulated in legal documents. The relationship between claims and the detailed description, the selection of relevant passages, the construction of positive and negative samples, and the formatting that distinguishes two text inputs together convert semantic correspondence in patent documents into a computationally tractable problem.
8. Remaining Challenges and the Path Forward
Substantial challenges remain. Terminology and sentence structure vary by technical field, which may require domain-specific training. Random negative sampling can also produce false negatives by pairing documents that are, in fact, technically related. Narrowing the gap between semantic similarity and legal infringement analysis will require claim-element-level analysis supplemented by prosecution history, prior art, and expert-labeled data. BERT's 512-token constraint may also be addressed through longer-context models or hierarchical architectures.
Even so, the starting proposition is clear: a patent is not merely an undifferentiated block of text. A special relationship exists between the claims that define the legal right and the detailed description that explains the technology. Once that relationship is converted into training data, AI can move beyond retrieving documents that use the same words and begin exploring why two documents may be technically related.
The proper role of AI in patent practice is not to displace the expert's ultimate judgment. It is to find, more quickly and across a wider field, the candidates that deserve that judgment. The system first detects semantic connections that may be easy to miss in a large corpus; lawyers and technical specialists then apply the governing legal and engineering standards. The invention I filed occupies that boundary. It is an effort to move beyond searching the words of a patent and toward reading the technical meaning the patent seeks to protect.
Reference
ChinSu Lee (이진수), “Method for Generating a Patent-Analysis Model Using an Artificial Neural Network and Text-Pair Embeddings, Patent-Analysis Method, and Computing Device” [English translation of title], Korean Patent Application No. 10-2024-0075102, filed June 10, 2024.
- Application number: 10-2024-0075102
- Filing date: June 10, 2024
- Priority application: 10-2023-0093439 (July 18, 2023)
- Inventor / applicant: ChinSu Lee (이진수)
- Original Korean title: 인공신경망 및 텍스트 쌍 임베딩을 이용한 특허 분석 모델의 생성 방법, 특허 분석 방법 및 컴퓨팅 장치
