A deep learning based patentability determination system and method

Abstract

The invention relates to a system and method for determining patentability by semantically comparing patent documents with each other. In particular, the invention compares a patent document received from a user through an interface with a plurality of relevant documents in the prior art. As a result of this comparison, the invention identifies the prior art patent documents most similar to the examined patent document and determines X, Y, or A documents indicating the degree of similarity of the patents, thereby providing a patentability determination system and method.

Description

[0001] DESCRIPTION

[0002] A DEEP LEARNING BASED PATENTABILITY DETERMINATION SYSTEM

[0003] AND METHOD

[0004] Technical Field of the Invention:

[0005] The invention relates to a system and method for determining patentability by semantically comparing patent documents with each other.

[0006] In particular, the invention concerns a patentability determination system and method which, through an interface, receives a patent document from the user and compares it with a plurality of relevant documents in the prior art. As a result of this comparison, the invention identifies the closest patent documents in the prior art with respect to the investigated patent document, thereby enabling the determination of documents X, Y, or A, which represent the degree of similarity of the patents.

[0007] Prior art:

[0008] In order for an invention to obtain patent protection, it must satisfy the criteria of patentability. The patentability criteria are novelty, inventive step, and industrial applicability. Novelty refers to an invention that is not included in the prior art. Prior art encompasses everything that has been made available to the public anywhere in the world before the filing date, whether by written or oral disclosure, by use, or by any other means of disclosure. Inventive step refers to an invention that, having regard to the prior art, is not obvious to a person skilled in the relevant technical field. Industrial applicability refers to an invention that can be produced or used in any field of industry, including agriculture.

[0009] In patent applications, it is a very challenging problem to decide whether the claimed invention satisfies the criteria of novelty and inventive step. Making this determination requires expertise in the relevant field. Since inventors and R&D personnel are often unable to assess the patentability of their inventions, they spend considerable time and effort in filing patent applications for inventions that ultimately cannot be patented. In the prior art, patent examiners at patent offices prepare reports on patentability. These search and examination reports concern whether the criteria of patentability have been met. Search reports contain certain universal codes. Documents cited with the X code are those that, alone, demonstrate that the claimed invention does not meet the novelty or inventive step criteria. Documents cited with the Y code are those which, in combination, show that the claimed invention lacks inventive step. Documents cited with the A code are those which, although in the same technical field and related, cannot prevent the grant of a patent for the claimed invention. Even the process of identifying the X, Y, and A documents through prior art searches by patent examiners requires considerable time. For this reason, applications may remain pending at the search and examination stages for long periods in patent offices. One of the main reasons why patent procedures take a long time worldwide is the difficulty of preparing reports on patentability.

[0010] In the prior art, there exist some inventions and patent applications developed to facilitate patent searches or to assist in assessing similarities. However, such inventions generally perform broad comparisons (e.g., differences in use between two inventions) for the convenience of inventors. Some of these applications are set out below.

[0011] In the prior art, the invention disclosed in application no. TR2023 / 018197, entitled “Dynamic Patent Management System”, is described in the abstract section as follows: This invention relates to a dynamic management system which enables the format control of notifications prepared by inventors regarding their inventions, evaluates whether the subject matter of the invention provides solutions to unresolved issues, calculates the degree of differentiation from existing patents, estimates the probability of patentability using artificial intelligence, and provides information to the inventor and the patent attorney.

[0012] In the prior art, the invention disclosed in application no. KR20210053539 relates to an apparatus and method for determining the novelty of a patent by receiving a structured patent document defined in natural language. According to this invention, when a patent to be evaluated is entered, preprocessing is performed taking into account the features of the patent document. As a result of the preprocessing, similarity with previously granted patent documents is measured, and novelty is determined using the measured similarity score together with learning algorithms such as deep learning. Finally, if lack of novelty is detected, a list of patents infringing novelty and a low probability value of novelty are presented. After the result is notified to the user, the outcome of the determination is used as training data to improve the performance of the novelty determination model by means of feedback. This invention is a method that performs novelty determination and similarity comparison among patent texts. However, in application KR20210053539 it is not disclosed how the training data is obtained. Furthermore, analysis of the claims and descriptions together with a deep learning algorithm for rendering a patentability decision is not present in application KR20210053539. Instead, the application only compares patent texts using conventional methods, although it also performs semantic comparison of patent texts. There is, however, a significant difference between semantic similarity of texts and patentability assessment. Documents assessed as category A in patent search reports may be in the same technical field and semantically very similar to the invention under examination. Many semantic similarity methods may find the similarity of X-coded and A-coded documents to the invention at the same level, or even consider A documents more similar than X documents. However, once patentability judgment comes into play, it is necessary to decide which documents fall into categories X, Y, or A. The system and method subject to the present invention, described in detail below, is capable of making this patentability judgment.

[0013] The invention disclosed in application no. CN110134925A relates to a method for calculating patent text similarity based on a vector space model. In this method, SAG triplets are extracted from the patent text, similarity of words in the SAG triplet group is calculated, and similarity between SAG structures is computed using a vector space model and fusion-based methods. Application CN110134925A semantically compares patent application texts. There is, however, a significant difference between semantic similarity of texts and patentability assessment. Documents assessed as category A in patent search reports may be in the same technical field and semantically very similar to the invention under examination. Many semantic similarity methods may find the similarity of X-coded and A-coded documents to the invention at the same level, or even consider A documents more similar than X documents. However, once patentability judgment comes into play, it is necessary to decide which documents fall into categories X, Y, or A. The system and method subject to the present invention, described in detail below, is capable of making this patentability judgment.

[0014] In the prior art, the invention disclosed in application no. WO2023195769A1 relates to a method for retrieving similar patent documents using a neural network model. According to this invention, an embedding vector-based method is employed to calculate similarity between patent documents, comprising the steps of: obtaining first embedding vectors and second embedding vectors of the patent documents; and calculating similarity between the patent documents on the basis of the first and second embedding vectors, wherein the first and second embedding vectors are embedded by using respective first and second embedding models, and preferably, the first embedding model generates the first embedding vector on the basis that the number of tokens exceeds the maximum token number of the second embedding model. In application WO2023195769A1 , patent application texts are semantically compared. However, there is a significant difference between semantic similarity of texts and patentability assessment. Documents categorized as A in patent search reports may belong to the same technical field as the invention under examination and may be semantically very similar to it. Many semantic similarity methods may find that the similarity of X-coded and A-coded documents to the invention is at the same level, or may even consider A documents more similar than X documents. However, once patentability judgment comes into play, it is necessary to decide which documents fall into categories X, Y, or A. The system and method subject to the present invention, described in detail below, is capable of making this patentability judgment.

[0015] In the prior art, the invention disclosed in application no. KR102437118B1 relates to a patent search application device and a patent search method using the same. This invention comprises a patent search application device through a thesis linkage search, including an interface unit for receiving search information from a client terminal and a first target thesis search unit for searching first target thesis materials from the search information. The invention enables thesis-focused searching of patent documents. However, in this application, no evaluation or analysis relating to patentability criteria is disclosed.

[0016] In the prior art, the invention disclosed in application no. JP2019040402A relates to a patentability compliance prediction apparatus and a patentability compliance prediction program. This application includes compliance prediction processing tools for patentability requirements, including an inventive step prediction processing unit and a compliance prediction server for patentability requirements. However, it is not disclosed how the training data is obtained in this application. Moreover, no categorization into X, Y, and A documents is performed.

[0017] In the prior art, the invention disclosed in application no. JP2019101944A relates to an intellectual property system, an intellectual property assistance method, and an intellectual property assistance program capable of supporting the creation of intellectual property. This invention describes a search engine that conducts searches of intellectual property documents based on the input data. However, in this application, no evaluation or analysis relating to patentability criteria is disclosed.

[0018] In the prior art, the invention disclosed in application no. KR20210046178A relates to a device and method for determining patentability by performing claim analysis using a dual-layer artificial neural network. However, it is not disclosed how the training data for the neural networks is obtained in this application. Moreover, no evaluation or analysis relating to patentability criteria is disclosed.

[0019] In the prior art, there are no methods or systems capable of solving patentability determination by using a deep learning model specifically designed for patentability, and performing both novelty and inventive step assessments.

[0020] Instead, existing methods are limited to comparing patent texts using conventional techniques and / or conducting novelty comparisons only on specific sections of specific documents. However, there is a fundamental difference between semantic similarity of texts and patentability assessment. Documents assessed as A in patent search reports may be semantically very similar to the invention subject to the application, but they do not prevent patentability. Many semantic similarity methods may find X and A documents equally similar, or even consider A documents more similar than X documents. However, patentability judgment requires deciding which documents fall into the X, Y, or A categories. The inventions disclosed in the prior art cannot achieve such categorization. As a result, due to the drawbacks described above and the insufficiency of the existing solutions on this subject, it has become necessary to provide an improvement in the relevant technical field.

[0021] Objective of the Invention:

[0022] The primary objective of the invention is to receive a patent text prepared by a researcher, to compare it in terms of patentability with previously published patent documents in the relevant field, and to determine the documents coded as X, Y, and A, presenting the result to the researcher in the form of a report. To perform this comparison, the invention employs a deep learning-based model. This model has been specifically developed with a focus on patentability determination. A semantic similarity model designed for another field would be insufficient to solve the problem of patentability determination, since patentability determination is a process with unique challenges. Deciding on X, Y, and A documents requires a command of the technical literature in patent documents and the ability to make novelty and inventive step assessments specific to the patent system. In order to carry out patentability determination, patent search reports prepared by patent examiners are utilized as the training dataset.

[0023] The fundamental aim of the invention is to develop a system capable of automatically determining whether the invention subject to a patent application satisfies the criteria of novelty and inventive step. In this way, inventors and R&D staff can avoid wasting time and effort on patent applications for inventions that are not patentable. For this purpose, the invention compares the patent document received from the user with relevant documents in the prior art. As a result of this comparison, it identifies the closest patent documents in the prior art to the examined patent document, determines whether they correspond to X, Y, or A documents indicating similarity ratios, and provides the patentability evaluation using a deep learning model trained on real expert assessments, presenting the results as a report to the user.

[0024] Another objective of the invention is to ensure that the matrices to be analyzed by the deep learning model are derived from patent texts. In accordance with the principles of patent law that “the scope of protection of the invention is determined by the claims” and “the basis of the claims is the description,” for each claim, the corresponding description and the comparative description are combined and submitted to the deep learning algorithm. The algorithm analyzes the claim and descriptions together to render a patentability decision.

[0025] A further objective of the invention is that the invention entitled “Deep Learning-Based Patentability Determination System” provides a solution to the difficulties faced by both inventors and experts in determining patentability, thereby contributing significantly to the R&D and innovation ecosystem of the country.

[0026] Another aim of the invention is that the deep learning model is specifically designed for patentability, both during the training phase and in the determination of inputs and outputs. The model is designed to directly simulate the patent examiner’s process of determining patentability. The problem of patentability determination, being a highly complex evaluation, can only be overcome by a model specifically developed for this field.

[0027] Another objective of the invention is to eliminate the difficulties of preparing patentability reports, which is one of the reasons why patentability assessments are challenging and why patent procedures take a long time. Thus, it provides a solution to the problems experienced by both inventors and experts in patentability determination, making a significant contribution to the country’s R&D and innovation ecosystem.

[0028] Another objective of the invention is that the deep learning model is specifically designed for patentability, both during the training phase and in the determination of inputs and outputs. The model is designed to directly simulate the process of a patent examiner in determining patentability. The very challenging problem of patentability determination can only be overcome with a model developed specifically for this field. The invention enables the deep learning model to be trained in accordance with expert opinions in the technical field where patentability determination is to be made and, as a result, to perform categorization into X, Y, and A.

[0029] Another objective of the invention is to employ a deep learning model specifically developed for patentability determination. This model evaluates the semantic similarities of patent documents and renders patentability decisions accordingly. This specialization allows the model to deliver more accurate results in patentability assessment.

[0030] A further objective of the invention is to generate its training dataset directly from published patent search reports. This enables the model to be trained on real-world patent decisions, thereby improving the accuracy of its results.

[0031] Another aim of the invention is to ensure that the model, similar to patent examiners, categorizes patent documents into X, Y, and A, and is capable of imitating the decisionmaking process of human experts, producing results with comparable accuracy.

[0032] A further objective of the invention is to work with historical documents provided by the user. The user can upload documents obtained from online patent search engines in PDF format, and the invention extracts relevant data from these documents to perform the evaluation.

[0033] Another aim of the invention is to operate independently of language. In this way, the model can produce results in whichever language it is trained, ensuring that the invention can also be effectively applied to international patent documents.

[0034] Description of the Drawings:

[0035] FIGURE - 1 : The drawing illustrating the diagram of the system according to the invention.

[0036] Reference Numerals:

[0037] 110. Interface

[0038] 120. Cloud Service

[0039] 130. Computer

[0040] 140. User Database

[0041] 150. Cloud Database Detailed description of the invention:

[0042] The invention is a system that determines patentability by semantically comparing patent documents with one another. Through an interface (110), the invention receives a patent document from the user and compares it with a plurality of relevant documents in the prior art. As a result of this comparison, it identifies the closest prior art patent documents with respect to the investigated patent document, and determines whether these correspond to documents coded as X, Y, or A, representing the degree of similarity. The invention employs a deep learning model trained on data derived from actual expert opinions contained in patent search reports in the relevant technical field, thereby performing patentability assessment. The model used in the invention is designed to most closely approximate the patentability assessments of real experts, thus achieving significant success in solving the difficult problem of patentability evaluation. The invention presents the output to the user in the form of a report.

[0043] In general terms, the system for determining patentability by semantically comparing patent documents comprises an interface (110), a cloud service (120), a computer (130), a user database (140), and a cloud database (150). The method according to the invention is executed on the computer (130).

[0044] The interface (110) enables the user to upload documents, perform searches, and view reported results. User query information is activated through the interface, and after activation, the relevant commands are sent via the cloud service (120) or directly to the computer (130) to execute the operations.

[0045] The cloud service (120) provides data exchange among the interface (110), the computer (130), the user database (140), and the cloud database (150). In this way, the system can operate locally or in an internet-based manner.

[0046] The computer (130), operating locally or via the internet, performs the creation of the training dataset, vectorization of the dataset, model construction, and comparison of the user-provided patent document (received via the interface (110) and user database (140)) with earlier patent documents in the relevant field. The computer (130) identifies the features upon which the deep learning architecture will base the patentability decision, produces a value in the outputs representing the similarities X, Y, and A, and generates a similarity report. The computer (130) also ensures the execution of the method according to the invention.

[0047] The user database (140) stores the documents of the user and, when the system is used locally, stores the downloaded pre-trained model. The cloud database (150) stores the information obtained from patent databases and the models generated by the computer (130) using this information, as well as the documents uploaded by the user when the system is used in an internet-based manner.

[0048] The system according to the invention receives as input a patent text prepared by the user, executes a pre-trained deep learning model via a processor on the computer (130), and performs a patentability comparison with other patent documents stored in the cloud database (150) or user database (140). Here, the computer (130) may operate locally or as a remote server through the cloud service (120). After deciding on the X, Y, and A categories, the system provides a corresponding report to the user. The system according to the invention employs a deep learning model specifically developed for patentability determination. This model is designed to directly simulate the process of patent examiners in determining patentability and has been developed exclusively for this field.

[0049] The system according to the invention further comprises a method executed on the computer (130). The method for determining patentability by semantically comparing patent documents comprises the following steps:

[0050] • Creation of the dataset o Patent documents relevant to the technical field in which patentability will be assessed are retrieved using essential keywords and / or IPC / CPC classifications. Published EP and WO applications containing search reports are included. o Using this information, the description text of the invention under examination is obtained from its publication number, the claim text is obtained from the claim number to be evaluated, and the description set (full text of the description including claims) of the cited patent is obtained from its publication number. These are combined with the similarity category (X, Y, or A) to prepare the training dataset.

[0051] • Vectorization of the texts in the dataset to represent their meaning o Here, the BERT model, a state-of-the-art and highly effective text representation method based on Transformers, is used for the vectorization process. o In another embodiment of the invention, in order to better reflect the technical character of the invention, word groups representing subject- verb-object (SVO) triplets, which also capture the relationships among words, are determined using natural language processing methods instead of keywords. The vectorization process is then carried out based on the identified triplets. Here, SVO triplets are extracted from the texts by means of natural language processing methods, and these triplets are used to better represent the meaning of the text and the relationships it contains. o In the vectorization process, when dependent claims are vectorized, the vector of the independent claim on which the dependent claim relies is combined with the vector of the dependent claim text. o Here, three vectors are obtained: a vector of the description of the investigated patent, a vector of the claim under focus, and a vector of the full text of the cited patent. o In another embodiment of the invention, when constructing vectors, in order to obtain a model closest to an expert’s assessment, the description of the investigated patent and the focused claim are not treated as separate vectors but are combined into a single vector representing the scope of protection of the claim. For this purpose, the keywords contained in the claim are determined using natural language processing methods. The claim is treated as a keyword pool. The paragraphs of the description in which the keywords derived from the claim occur most intensively are identified, and instead of vectorizing the entire description, only the most relevant paragraphs are vectorized. In this way, a single vector representing the scope of protection of the claim is obtained. Alternatively, rather than using keywords to select the paragraphs to be vectorized, the claim is first vectorized in its entirety, and then all paragraphs of the description are vectorized separately. To identify the paragraphs in the description that are closest in meaning to the claim, cosine similarity is used to find the paragraph vectors nearest to the claim vector. The composite or average of these paragraph vectors is then taken to produce a single vector representing the scope of protection of the claim. In this case, the matrix formation is performed using a first vector representing the investigated document and a second vector representing the cited document.

[0052] • Combination of vectors into a single matrix o Here, the matrix formation is carried out by arranging the vectors vertically. If the combined three vectors correspond to patentability category X, the resulting matrix represents an X patentability evaluation. For rows in the dataset where the patentability category is Y, the constructed matrix represents Y similarity, whereas for rows where the patentability category is A, the constructed matrix represents A similarity.

[0053] • Training of the deep learning model using the matrices and their associated patentability labels o Here, the deep learning model under training processes the vectorized text data and decides on X, Y, and A similarities. The model analyzes the similarities and relationships among the texts to perform a patentability assessment and obtains the outputs. The outputs of the deep learning model comprise three results, each representing one of X, Y, and A similarities. If the input matrix corresponds to X similarity, the output is set to “1 -0-0”; if it corresponds to Y similarity, the output is set to “0-1 -0”; and if it corresponds to A similarity, the output is set to “0-0-1 .” o In one embodiment of the invention, the trained deep learning model is a CNN-based model. o In another embodiment of the invention, the trained deep learning model is an Autoencoder-based model. o In yet another embodiment of the invention, the trained deep learning model is a hybrid CNN + Autoencoder model.

[0054] • Receiving a user-created patent document and comparing it with earlier patent documents in the relevant field o Here, for the comparison process, the user-provided document and claims together with earlier patent documents are vectorized, converted into a matrix, and provided as input to the trained deep learning architecture.

[0055] • Determining, by using the coefficients established during training of the deep learning architecture, the features of the matrix on which the patentability decision will be based

[0056] • Production of output values representing X, Y, and A similarities o If the value of the X output is the largest, the similarity is determined as X; if the value of the Y output is the largest, the similarity is determined as Y; and if the value of the A output is the largest, the similarity is determined as A.

[0057] • Generating a similarity report by using the X, Y, and A documents determined based on the produced values o The report contains information similar to patent search reports, including the claim number of the investigated invention, the publication number of the cited patent document, and the similarity category.

[0058] • Delivery of the similarity report to the user o The report is delivered to the user via the interface (110) or sent by e- mail through the cloud server (120).

[0059] The system and method according to the invention utilize patent search reports prepared by patent examiners as the training dataset in order to address the difficult problem of patentability determination. The fact that the training of patentability is carried out on decisions made by patent examiners, which are publicly accessible, provides a significant advantage for the invention. Another superiority of the invention lies in the fact that the training dataset is created specifically for the technical field in which patentability determination is to be made. Indeed, patent examiners themselves are only able to provide patentability assessments within their own fields of expertise. An additional originality of the invention is that the matrices to be analyzed by the deep learning model are derived directly from patent texts. In accordance with the principles of patent law that “the scope of protection of the invention is determined by the claims” and “the basis of the claims is the description,” for each claim, the description to which the claim belongs and the description used for comparison are combined and submitted to the deep learning algorithm. The deep learning algorithm analyzes the claim to be assessed for patentability together with the corresponding descriptions.

[0060] The method of the invention receives a patent text prepared by a researcher and compares it with previously published patent documents in the relevant field with respect to patentability. Patentability comparison is different from semantic similarity comparison in other domains. Through this comparison, it is necessary to determine documents categorized as X, Y, or A. The system and method of the invention provide the researcher with a report containing this information. In order to perform such comparison, a deep learning-based model is employed. This model has been developed specifically and exclusively for patentability determination. A semantic similarity model from another domain would be insufficient to solve the problem of determining patentability, since patentability assessment is a process with its own inherent difficulties. The determination of X, Y, and A documents requires mastery of the technical literature in patent documents and the ability to conduct novelty and inventive step evaluations specific to the patent system.

[0061] It is critical for accurate decision-making that the dataset created for training the deep learning model is specifically prepared for patentability. The system and method of the invention generate the training data themselves from published patent search reports. Numerous previously published patent search reports are available in publicly accessible patent search engines. These reports include documents coded as X, Y, and A, which indicate the patentability decisions of patent examiners. Using these search reports, the dataset for training the deep learning model is constructed. An advantage of the method of the invention is that the training dataset is created specifically for the technical field in which patentability determination is to be made. Indeed, patent search examiners are also able to provide patentability assessments only within their own fields of expertise. For example, a patent examiner specialized in the field of electrical engineering cannot decide on the patentability of inventions in the field of automotive technology. In particular, the assessment of the inventive step criterion requires domain-specific expertise. Accordingly, the model of the invention follows the same principle, whereby the deep learning model must be trained with a dataset constructed for the specific technical field in which the patentability decision is to be made. To achieve this, relevant documents are retrieved from patent search engines by using basic keywords and / or IPC / CPC classes relating to the technical field in which the patentability determination is to be conducted. In order to obtain search reports containing X, Y, and A coded documents, which are necessary for preparing the dataset, it is required to perform a publication number query and to access EP or WO publications, since not every patent publication includes a search report, whereas EP and WO publications do. Moreover, among EP and WO publications, search reports are included in A1 and A3 publications. Accordingly, a search conducted in patent search engines with the criteria “keyword and / or IPC class” + “EP or WO” + “A1 or A3” enables access to the published search reports in the relevant technical field. The system and method of the invention store the retrieved search reports in PDF format and, by means of OCR or similar scanning technology, extract the publication number of the investigated patent, the claim number subject to the patentability assessment, the publication number of the cited patent, and the similarity category (X, Y, or A) contained in the report. Based on this information, the description text of the investigated invention is obtained using the publication number of the investigated patent; the claim text is obtained using the claim number subject to patentability assessment; and the full description of the cited patent is obtained using the publication number of the cited patent. These pieces of information are then combined with the similarity category (X, Y, or A) to prepare the training dataset. The training dataset comprises four main columns: the first is the description of the investigated patent; the second is the text of the claim of the investigated invention that is subject to patentability assessment; the third is the full description of the cited patent; and the fourth is the patentability category. These four entries reflect the patentability assessment made by the examiner who prepared the search report. If the similarity category is X, then when the description and the relevant claim of the investigated patent document are compared with the full text of the cited patent document, the patentability decision is understood to be X. In this way, within the method of the invention, a dataset containing decision labels of X, Y, or A — enabling supervised learning — can be constructed independently. For training the deep learning model using this dataset, the texts in the dataset must be vectorized so as to represent their semantic meaning. For the vectorization process, the BERT model, a state-of-the-art and highly effective Transformer-based text representation method, is employed. In one embodiment of the invention, the description text of the investigated patent, the text of the focused claim, and the full text of the cited patent are vectorized using the BERT model.

[0062] In one embodiment of the invention, since patent texts are complex, instead of vectorizing the entire text as it is, keywords are identified by natural language processing methods and these keywords are vectorized.

[0063] In another embodiment of the invention, in order to better reflect the technical character of the invention, instead of keywords, Subject-Action-Object (SAO) triplets, which also represent the relationships among words, are identified by natural language processing methods. The vectorization process is then performed based on the identified triplets. During vectorization, when dependent claims are vectorized, the vector of the independent claim on which the dependent claim relies is combined with the vector of the dependent claim text. As a result of the vectorization process, three vectors are obtained: a vector of the description of the investigated patent, a vector of the focused claim, and a vector of the full text of the cited patent. The reason for using these three vectors during model training is to replicate as closely as possible the manner in which an actual expert performs a patentability assessment. In practice, a patent examiner, when conducting a patentability assessment, first reads the relevant claim, then reviews the description to clearly define the scope of protection of the claim, thereby clarifying the technical scope for which protection is sought. The examiner also reviews the full text of comparable prior art patents and compares it with the claimed technical scope. For this reason, in the training stage, the three vectors are employed in the same manner as an expert’s assessment. During training of the deep learning model, the existing text vectors are combined into a single matrix. The matrix formation is carried out by arranging the vectors one below another. If the combined three vectors correspond, for example, to patentability category X, the resulting matrix represents an X patentability evaluation. For rows in the dataset where the patentability category is Y, the constructed matrix represents Y similarity, whereas for rows where the category is A, the constructed matrix represents A similarity. Converting the complex decisionmaking process of patentability assessment into a digital form in this manner, based on expert opinions, constitutes an important advantage of the system and method of the invention over the prior art. In another embodiment of the invention, when constructing the vectors for matrix formation, in order to replicate as closely as possible an expert assessment, the description vector and the claim vector of the investigated patent are not treated as separate vectors but are combined into a single vector representing the scope of protection of the claim. For this process, the keywords contained in the claim are identified using natural language processing techniques. The claim is included in the process as a pool of keywords. The paragraphs in the description in which these keywords occur most frequently are identified, and instead of vectorizing the entire description, only the paragraphs most relevant to the claim are vectorized. In this way, a single vector representing the scope of protection of the claim is obtained. For comparison, instead of vectorizing the entire cited description, the paragraphs most relevant to the claim are similarly identified and vectorized. In determining the paragraphs to be vectorized, either keywords or Subject-Action- Object (SAO) triplets may be used. Instead of relying on keywords, in one embodiment, the claim is first vectorized in its entirety. Subsequently, all paragraphs of the description are vectorized individually. To identify the paragraphs in the description that are most semantically relevant to the meaning of the focused claim, the cosine similarity method is applied to determine the paragraph vectors closest to the claim vector. The combination or average of these paragraph vectors is then taken to construct a single vector representing the scope of protection of the claim. Using the same method, the paragraphs of the cited description most relevant to the claim are also identified and vectorized. In this case, the matrix formation is carried out using a first vector representing the investigated document and a second vector representing the cited document. The matrices thus obtained, together with their associated patentability labels, are used to train the deep learning model. The matrix is provided as input to the deep learning model. The output of the deep learning model consists of three outputs, each representing one of the similarities X, Y, and A. If the input matrix corresponds to X similarity, the output is set as “1 -0-0”; if it corresponds to Y similarity, the output is set as “0-1 -0”; and if it corresponds to A similarity, the output is set as “0- 0-1 .” In this way, the deep learning model is trained by using each row of the dataset constructed from the search reports. As a result of the training, the deep learning model identifies the features within the matrices that enable a decision on patentability.

[0064] In one embodiment of the invention, the trained deep learning model is a CNN-based model. In another embodiment, the trained deep learning model is an Autoencoderbased model. In yet another embodiment, the trained deep learning model is a hybrid model combining CNN and Autoencoder architectures. As a result of training, a deep learning model is obtained that has learned the features enabling a determination of patentability in the relevant technical field. In operation, the system and method of the invention take a user-generated patent document and compare it with earlier patent documents in the relevant field. For the comparison process, in the same manner as described during training, the user-provided document and claims, together with the earlier patent documents, are vectorized, converted into a matrix, and provided as input to the trained deep learning architecture. The deep learning architecture applies the coefficients established during training to determine the features of the matrix on which the patentability decision will be based, and produces values in its outputs representing X, Y, and A similarities. If the value of the X output is the largest, the similarity is determined as X; if the value of the Y output is the largest, the similarity is determined as Y; and if the value of the A output is the largest, the similarity is determined as A. The system and method of the invention then generate a report using the determined X, Y, and A documents. The report includes, in a manner similar to patent search reports, the number of the focused claim of the investigated invention, the publication number of the cited patent document, and the similarity category information.

[0065] In one embodiment of the invention, the earlier documents to be compared are provided by the user as input through the interface (110). The user performs a simple query on online patent search engines, retrieves numerous documents, and uploads them to the system in PDF format via the interface (110). The method of the invention extracts the relevant data from these PDFs, presents them to the deep learning model, and reports the result. The invention is of great importance for R&D and innovation managers and personnel. With the system and method of the invention, it can be determined whether the R&D work carried out is patentable. In this way, unnecessary expenses for filing patent applications for inventions that cannot be patented are avoided, and insights can be obtained on possible improvements required to render an invention patentable. A search engine employing the model of the invention will attract significant demand from stakeholders in the field of technology. Moreover, at the global level, national offices engaged in patentability searches will benefit from a valuable tool to support the decision-making of patent examiners. This will contribute to faster preparation of patent search and examination reports and to shorter patent grant procedures. Furthermore, the use of the system and method of the invention will provide significant benefits to all private and legal entities offering preliminary patent search services. It may also be applied in the legal field, since patentability assessment is also at the core of decisions relating to disputes over patent invalidity or infringement.

Claims

CLAIMS1. A system for determining patentability by semantically comparing patent documents with each other, characterized in that the system comprises:• an interface (110) configured to enable the user to upload documents, perform searches, and view reported results, to activate user query information, and, upon such activation, to transmit relevant commands via a cloud service (120) or directly to a computer (130) for execution of operations by the computer (130);• a cloud service (120) configured to provide data exchange among the interface (110), the computer (130), a user database (140), and a cloud database (150), thereby enabling the system to operate locally or on an internet-based platform;• a computer (130) configured to operate locally or via the internet, to generate a training dataset, to vectorize the dataset, to construct and train a model, to receive a user-generated patent text / document via the interface (110) and compare it with earlier patent documents in the relevant field stored in the user database (140) and / or the cloud database (150), to run a deep learning architecture that identifies the features decisive for patentability, to produce values in its outputs representing X, Y, and A similarities, and to generate a similarity report;• a user database (140) configured to store the user’s documents and, in the case of local system operation, to store a previously downloaded trained model; and• a cloud database (150) configured to store information obtained from patent databases and the model generated by the computer (130) using such information, and, in the case of internet-based system operation, to store the documents uploaded by the user.

2. A method executed on a computer (130) for determining patentability by semantically comparing patent documents with each other, characterized in that the method comprises the steps of:- creating a dataset,- vectorizing the texts in the dataset so as to represent their semantic meaning,- combining the vectors into a single matrix,- training a deep learning model using the obtained matrices together with their associated patentability labels,- receiving a user-generated patent document and, for comparison with earlier patent documents in the relevant field, vectorizing the user- provided documents and claims together with the earlier patent documents, converting them into a matrix, and providing the matrix as input to the trained deep learning architecture,- determining, by means of the deep learning architecture, the features of the matrix on which the patentability decision will be based,- producing, in the outputs of the deep learning architecture, a value representing the X, Y, and A similarities,- generating a similarity report using the determined X, Y, and A documents, and- transmitting the similarity report to the user.

3. The method according to claim 2, characterized in that the step of creating the dataset comprises:- retrieving relevant documents by using basic keywords and / or IPC / CPC classes relating to the technical field in which patentability is to be determined,- obtaining reports which include patent search reports among publications having EP or WO numbers,- accessing the description text of the investigated invention by using the publication number of the investigated patent,- accessing the claim text by using the claim number whose patentability is to be evaluated,- accessing the full description of the cited patent by using the publication number of the cited patent, and- preparing the training data by combining this information with the similarity category X, Y, or A.

4. The method according to claim 2, characterized in that in the step of vectorizing the texts in the dataset to represent their semantic meaning, the vectorization is performed by using a Transformer-based BERT model.

5. The method according to claim 2, characterized in that in the step of vectorizing the texts in the dataset to represent their semantic meaning, the vectorization is carried out based on Subject-Action-Object triplets determined by natural language processing methods.

6. The method according to claim 2 or claim 5, characterized in that in the step of vectorizing the texts in the dataset to represent their semantic meaning, when dependent claims are vectorized, the vector of the independent claim on which the dependent claim relies and the vector of the dependent claim text itself are vectorized and combined.

7. The method according to claim 2 or claim 5, characterized in that in the step of vectorizing the texts in the dataset to represent their semantic meaning, three vectors are obtained, namely a vector of the description of the investigated patent, a vector of the focused claim, and a vector of the entire text of the cited patent.

8. The method according to claim 2, characterized in that the step of vectorizing the texts in the dataset to represent their semantic meaning comprises the steps of: identifying the keywords contained in the claim by means of a natural language processing method,- including the claim as a keyword pool in the process,- detecting, within the description, the paragraphs in which the keywords generated from the claim are most densely present,- vectorizing, instead of the entire description, the paragraphs of the description that are most relevant to the claim, and- obtaining a single vector representing the scope of protection of the claim.

9. The method according to claim 2 or claim 8, characterized in that;- when determining the paragraphs to be vectorized, the entire claim is vectorized instead of identifying keywords,- each paragraph of the description is vectorized separately,- the paragraph vectors in the description that are closest to the claim vector are identified, by means of a cosine similarity method, in order to determine the paragraphs most relevant to the meaning of the focused claim,- a single vector representing the scope of protection of the claim is obtained by combining or averaging the vectors of the identified paragraphs, and- the matrix construction is performed by using a first vector representing the investigated document and a second vector representing the cited document.

10. The method according to claim 2, characterized in that in the step of combining the vectors into a single matrix, the matrix construction is carried out by arranging the vectors vertically.

11. The method according to claim 2, characterized in that in the step of combining the vectors into a single matrix:- the matrix constructed for the rows in the dataset where the patentability category is X represents X similarity,- the matrix constructed for the rows in the dataset where the patentability category is Y represents Y similarity, and- the matrix constructed for the rows in the dataset where the patentability category is A represents A similarity.

12. The method according to claim 2, characterized in that the step of training the deep learning model using the obtained matrices and their corresponding patentability labels comprises;- the deep learning model processing the vectorized text data to decide on X, Y, and A similarities,- the deep learning model analyzing the similarities and relationships among the texts to perform a patentability assessment, and- the deep learning model obtaining outputs.

13. The method according to claim 2 or claim 12, characterized in that in the step of obtaining the outputs of the deep learning model, the output of the deep learning model comprises three outputs each representing X, Y, and A similarities, and if the input matrix corresponds to X similarity, the output is set to “1 -0-0”, if the input matrix corresponds to Y similarity, the output is set to “0- 1 -0”, and if the input matrix corresponds to A similarity, the output is set to “0-0- 1”.

14. The method according to claim 2, characterized in that the trained deep learning model is a CNN-based model.

15. The method according to claim 2, characterized in that the trained deep learning model is an Autoencoder-based model.

16. The method according to claim 2, characterized in that the trained deep learning model is a hybrid CNN + Autoencoder-based model.

17. The method according to claim 2, characterized in that in the step of producing a value in the outputs representing X, Y, and A similarities:- if the value of the X output is the largest, the similarity is determined to be X,- if the value of the Y output is the largest, the similarity is determined to be Y, and- if the value of the A output is the largest, the similarity is determined to be A.

18. The method according to claim 2, characterized in that in the step of generating the similarity report using the X, Y, and A documents determined by the produced values, the generated report includes the focused claim number of the investigated invention, the publication number of the cited patent document, and the similarity category information.

19. The method according to claim 2, characterized in that in the step of transmitting the similarity report to the user, the report is transmitted to the user through the interface (110).

20. The method according to claim 2, characterized in that in the step of transmitting the similarity report to the user, the report is transmitted to the user through the cloud service (120).

Images