TLDR: Patentformer is an AI-powered platform developed by Samsung Semiconductor, Inc. to assist patent attorneys in drafting high-quality patent specifications. It addresses the limitations of general Large Language Models (LLMs) by utilizing a fine-tuned LLM trained on over a million patent data samples and a specialized input processing methodology. The system takes patent claims and drawing descriptions, processes them, and generates detailed, legally compliant specifications. A user study demonstrated high performance in legal and linguistic quality, with future research focusing on integrating multimodal capabilities to enhance image understanding and automate figure generation.
Patent drafting is a complex and time-consuming process that demands extensive experience and specialized expertise from patent attorneys. These professionals must possess a deep understanding of both legal principles and the technical intricacies of an invention to create patent applications that adhere to a formal legal writing style. The cost of drafting a moderately complex patent can exceed $10,000, and the process typically involves attorneys reviewing invention disclosure documents and interviewing inventors to grasp technical details before drafting claims, drawings, and the detailed description, known as the specification.
The specification is the most substantial part of a patent document, requiring significant effort to draft as it provides a detailed description of the invention in accordance with the claims and drawings. While Transformer-based Large Language Models (LLMs) like GPT-4 have shown remarkable capabilities in natural language generation, they face significant challenges in producing high-quality patent specifications. Patent documents differ fundamentally from general text due to their complex legal and technical nature, domain-specific language, intricate relationships between claims and drawing descriptions, and extensive technical details. Additionally, patent specifications are lengthy, often exceeding the fixed token limits of most LLMs, and most pre-trained LLMs are not trained on patent data, leading to difficulties in adapting to the precise writing style and legal requirements.
Introducing Patentformer: An AI-Powered Solution
To address these challenges, researchers from Samsung Semiconductor, Inc. have developed and demonstrated Patentformer, an AI-powered automated patent drafting platform. Patentformer is designed to support patent attorneys by rapidly producing high-quality patent applications that meet legal writing standards. The system takes a patent claim and any associated drawing text as input, then pre-processes and enhances this input to improve readability and structure for better comprehension by its underlying LLM. This LLM has been specifically fine-tuned on publicly available patent data to learn the stylistic and structural conventions of patent writing, enabling it to generate high-quality patent specifications that align with legal and technical standards.
The Patentformer system is deployed as an interactive patent drafting assistant, providing users with an intuitive interface to streamline the patent writing process. Key contributions of this work include the development and deployment of the Patentformer platform, a specialized training data construction methodology that transforms plain text into an enriched representation (using a dataset of over one million samples processed from USPTO patents), and a user study to quantitatively evaluate its effectiveness.
How Patentformer Works
The Patentformer system operates in three main steps. First, users upload claim text and corresponding drawing figures. The platform then preprocesses this input, generating structured claim features from the text and automatically identifying key components and their numbers within the drawings. Users also have the option to modify the processed text and manually add image descriptions through the interactive interface.
Second, the platform presents a Mapping Interface, allowing users to define relationships between various claims and drawing features. While users can manually establish these connections, Patentformer also implements a simple strategy to automatically match components to claim features based on cosine similarity and BLEU scores. This mapping ensures that the generated patent specification accurately aligns claims with their respective visual components.
Finally, each mapped claim, along with its associated components, is processed into a structured text input tuple and passed to the fine-tuned LLM. The LLM generates an enriched version of the patent specification for each claim. A final post-processing step refines the output to conform to standard patent writing conventions before it is presented to the user.
Also Read:
- PatentVision: Integrating Visuals and Text for Advanced Patent Drafting
- A New Framework for AI-Powered Customer Support Efficiency
Evaluation and Future Directions
The effectiveness of Patentformer was evaluated through a user study involving an expert patent writer. The user assessed the system’s quality across four dimensions: linguistic, legal, figure descriptive, and technical quality. The results showed high scores for legal and linguistic quality, indicating that Patentformer effectively learns from legal patent documents and generates well-written text. However, the model’s ability to understand images directly was limited, leading to lower scores in figure descriptive and technical qualities. This is expected, as the current model relies on text descriptions of images rather than direct image input, and it is a text-to-text model without multimodality.
The researchers acknowledge these limitations and highlight future work focusing on developing multimodal models, such as Large Vision-Language Models (LVLMs), to integrate both textual and visual information. These advancements would enable automatic interpretation of patent figures, reducing the need for manual input and further improving specification generation. Additionally, leveraging image generative models like Stable Diffusion could automate the creation of patent figures, further streamlining the patent drafting process. For more in-depth information, you can refer to the original research paper: Patentformer: A demonstration of AI-assisted automated patent drafting.
The Patentformer system represents a significant step towards automating and improving the patent drafting process, offering a valuable tool for legal professionals navigating the complexities of intellectual property.
