Inventor’s note. This article explains, for a general audience, the concept and operation of an AI-based patent-analysis technology that I invented and filed for patent protection. Semantic similarity produced by the system should be understood as an analytical signal for prioritizing patents or products for closer review, not as a final legal determination of infringement.
Training on Claims and Descriptions to Detect Patent Risk
Patent searching is far more difficult than it first appears. When a product specification and a patent claim use the same terminology, comparison is relatively straightforward. Real disputes, however, are rarely so convenient. The same technical concept can be expressed with completely different terms, reordered features, or a mixture of broader and narrower concepts. Pure keyword search can easily miss those relationships.
The technology I invented and filed for patent protection starts from that problem. Its purpose is to move beyond simple word matching and train an AI model to evaluate how closely the technical meaning of a patent claim corresponds to descriptions found in other documents. The key idea is to use structural relationships that already exist inside patent documents as training signals. Instead of requiring human experts to read tens of thousands of document pairs and manually label each one, the patent documents themselves provide a substantial part of the supervision.
To understand the approach, it is useful to begin with the structure of a patent specification. The claims define the legal boundary of the right, while the description explains the invention and how it can be implemented. Article 42(4)(1) of the Korean Patent Act requires claims to be supported by the description. The requirement is intended to prevent an applicant from monopolizing subject matter that was not actually disclosed and to ensure that a person skilled in the art can understand the relationship between the claimed invention and the specification.
From a machine-learning perspective, that legal and documentary structure can be converted into a useful data relationship. A claim and a description extracted from the same patent will generally have a high degree of semantic relevance. By contrast, a claim taken from one patent and a randomly selected description passage from another patent will, statistically, be less likely to be relevant. This makes it possible to assign label 1 to same-patent claim–description pairs and label 0 to randomly paired passages from different patents.
One distinction is critical. Label 1 should not be equated with a legal finding of infringement, and label 0 should not be treated as a legal finding of non-infringement. These are proxy labels for teaching the model semantic correspondence, not judgments issued by a court. Two different patents can be technically related, and a claim paired with its own description says nothing by itself about infringement by a third party. The inventive idea lies in converting high-probability positive and negative relationships embedded in patent-document structure into large-scale training data.
This design reduces the bottleneck of manual labeling. In conventional supervised learning, experts must review document pairs and decide whether they are relevant. In the patent field, accurately reading even a single document can be time-consuming, so labeling costs rise quickly as the dataset grows. By automatically pairing a claim with its own description and randomly pairing it with descriptions from other patents, software can generate training pairs at scale.
Negative samples play an equally important role. If a model sees only strongly related pairs, it may overfit and begin treating a few overlapping words as sufficient evidence of relevance. Training with claim–description pairs drawn from different patents teaches the model that lexical overlap alone is not enough. Positive samples teach what meaningful correspondence looks like; negative samples help define where that correspondence ends.
The next problem is document length. Standard BERT has a maximum input length of 512 tokens, while patent descriptions routinely contain thousands of words. Feeding a claim together with an entire description into BERT is therefore impractical.
In the implementation described in my invention, the long description is divided into smaller passages of roughly 310 tokens. The system then evaluates the semantic relationship between the claim and each passage and prioritizes passages with higher relevance. Instead of simply truncating the patent from the beginning, the model first identifies portions that are more likely to explain the claimed subject matter.
Vector operations such as the dot product can be used during this process. In intuitive terms, the claim is represented as a semantic vector and each description passage is represented as another vector. The system then scores how closely those vectors point in the same semantic direction. A higher score suggests a stronger semantic relationship between the claim and the passage.
The selected text is then arranged in a structure that allows BERT to distinguish the two inputs. A representative sequence is [CLS] + claim + [SEP] + description + [SEP]. The [SEP] token marks the boundary between the two text segments, segment embeddings indicate which tokens belong to which segment, and the final representation of [CLS] can serve as a pooled vector for classification.
This architecture matters because a model does not naturally perceive document boundaries in the way a human reader does. People can use headings, line breaks, and layout to understand that two passages come from different sources. A language model needs explicit structural signals that tell it where comparison target A ends and comparison target B begins.
The training stage applies an alternating-batch strategy to label 1 and label 0 datasets. Rather than indiscriminately mixing the two groups, the system computes their losses separately and updates parameters in a direction that reduces both. The objective is to reduce the risk that the model becomes overly influenced by whichever class is more numerous or easier to learn.
According to the experimental results recorded in the source manuscript, cross-entropy loss decreased from 0.89 to 0.19 over 15 epochs, while accuracy reached approximately 80%. In a comparison designed to test semantically similar content expressed with different terminology, all 17 text pairs were classified as label 1. These figures should be understood as results obtained under the specific experimental conditions described in the manuscript, not as a guarantee of identical performance across all technologies and patent disputes.
The significance of the result lies in the type of problem being addressed. One of the hardest situations in patent searching occurs when the same technical content is described with different vocabulary. Suppose one document refers to an “adhesive fixing mechanism,” while another describes a functionally similar structure using completely different terminology. A keyword engine may assign low relevance simply because the documents share few literal words. A semantic model has the potential to compare how the features function and relate to one another in context.
This naturally brings the doctrine of equivalents to mind. Patent infringement analysis does not necessarily end merely because an accused product or method differs from the literal wording of a claim. Korean Supreme Court precedent recognizes that, under specified conditions, a modified element may still fall within the scope of the patent under the doctrine of equivalents when the problem-solving principle, substantially identical effect, ease of substitution, and other requirements are satisfied and no exclusion applies.
Semantic similarity produced by an AI model, however, is not the same thing as legal equivalence. A high similarity score does not automatically establish infringement, and a low score does not eliminate legal risk. A legal analysis still requires consideration of every claim element, prosecution history, prior art, conscious exclusion, and the specific requirements of the doctrine of equivalents.
The practical value of the invention is therefore not that an AI system replaces a judge. Its more important role is to narrow the universe of documents that human experts need to examine. From thousands or millions of patents and product documents, the system can identify candidates that appear semantically close to a target claim and allow patent attorneys, lawyers, engineers, and R&D teams to focus their detailed review where it matters most.
During product development, the technology can help screen competitor patents and product specifications for early risk signals. Patent owners can compare market-facing technical descriptions with their own claims to prioritize monitoring. The same underlying architecture can also support prior-art searching by finding technically relevant documents that use different vocabulary from the claim being searched.
In patent AI, the most expensive resource is not necessarily the GPU or the number of model parameters. The harder problem is often obtaining high-quality labeled data. Expert labels are valuable but expensive and slow. Public patent data is abundant, but raw documents do not automatically provide task-specific labels. This invention is an attempt to bridge that gap by exploiting the legal and documentary structure already embedded in patent specifications.
I would not reduce the idea to the slogan that it “turns law into mathematics.” A more precise description is that it identifies machine-readable learning signals within the structures and rules accumulated in legal documents. The relationship between claims and descriptions, the selection of relevant passages, the construction of positive and negative samples, and the explicit separation of paired text together transform semantic patent comparison into a computable problem.
There are clear areas for further development. Patent terminology and drafting conventions differ across technical fields, making domain-specific training important. Random negative sampling can occasionally pair a claim with a genuinely relevant passage from another patent. Closing the gap between semantic relevance and legal infringement will require richer information, including claim-element analysis, prosecution history, prior art, and expert-labeled decisions. BERT’s sequence-length limitation can also be addressed with longer-context models or hierarchical architectures.
Even so, the starting insight remains powerful. A patent is not merely a block of text. It contains a distinctive relationship between claims that define legal scope and a description that explains the underlying technology. Once that relationship is converted into training data, AI can move beyond finding documents that use the same words and begin identifying documents that may be technically related for deeper review.
The proper role of AI in patent practice is not to take the final legal judgment away from human experts. It is to find meaningful connections across a scale of documents that humans cannot efficiently examine one by one, and then place the most relevant candidates in front of the people responsible for the legal and technical conclusion. My filed invention begins at precisely that boundary: not simply searching the words of patents, but attempting to read the technical meaning they are intended to protect.
References
Reference
Chinsu Lee, Method for Generating a Patent Analysis Model Using an Artificial Neural Network and Text-Pair Embedding, Patent Analysis Method, and Computing Device, Korean Patent Application No. 10-2024-0075102, filed June 10, 2024.
- Application No.: 10-2024-0075102
- Filing Date: June 10, 2024
- Earlier Application: 10-2023-0093439, filed July 18, 2023
- Inventor / Applicant: Chinsu Lee
- Title of Invention: Method for Generating a Patent Analysis Model Using an Artificial Neural Network and Text-Pair Embedding, Patent Analysis Method, and Computing Device
No comments:
Post a Comment