Technical field ontology construction method based on patent data

Abstract

The invention discloses a technical field ontology construction method based on patent data, and the method comprises the steps: firstly obtaining a patent data set, and carrying out the preprocessing; then, respectively linking a part name, a large class name and a small class name corresponding to the IPC main classification number with entities in a universal knowledge base to obtain a first-layer candidate technology concept set, a second-layer candidate technology concept set and a third-layer candidate technology concept set; performing manual screening on the first-layer candidate technical concept and the second-layer candidate technical concept, performing automatic screening on the third-layer candidate technical concept by using a GAT model to obtain a first technical concept layer, a second technical concept layer and a third technical concept layer, and establishing a hierarchical relationship among the layers to obtain a static layer; then, constructing dynamic layers based on patent names, including fourth, fifth and sixth technical concept layers and hierarchical relationships thereof; and finally, establishing a hierarchical relationship between the third technical concept layer and the fourth technical concept layer, realizing integration of the static layer and the dynamic layer, and obtaining a directed acyclic graph structure, namely the technical field ontology. According to the method, the technical concept can be automatically extracted from the patent data, the multi-level technical field ontology is constructed, and the construction efficiency is improved.

Description

Technical Field

[0001] The present invention belongs to the field of patent data processing technology, and in particular relates to a method for constructing a technical field ontology based on patent data. Background Art

[0002] With the rapid development of science and technology, technological innovation is showing a trend of interdisciplinary and multi-field integration. Traditional technology classification systems, such as the International Patent Classification (IPC), are static and have limited coverage, making it difficult to effectively reflect the dynamic changes in current technological development. This makes it difficult to effectively utilize massive amounts of patent data for technology trend analysis, innovation monitoring, and R&D strategy decision-making.

[0003] At present, the construction of ontologies in the field of technology mainly relies on human experts, structured data and machine learning. Manual construction methods are time-consuming and labor-intensive, and are difficult to adapt to the rapid changes in the field of technology. Construction methods based on structured data, such as using existing classification systems, are difficult to capture emerging technologies and interdisciplinary technology concepts. Construction methods based on machine learning still have shortcomings when processing large-scale, unstructured patent data: 1) When processing patent names, existing technical concept recognition models have low recognition accuracy due to lack of domain knowledge and targeted optimization, making it difficult to accurately extract technical concepts; 2) Existing clustering methods are difficult to effectively capture the hierarchical relationship between technical concepts, and it is difficult to construct an ontology structure with semantic meaning; 3) Existing methods are difficult to process large-scale patent data and are difficult to adapt to the rapid changes in the field of technology.

[0004] Therefore, the present invention proposes a method for constructing a technical field ontology based on patent data, which can automatically, efficiently and accurately extract technical concepts from large-scale patent names and construct a multi-level and scalable technical field ontology to meet application needs in technology trend analysis, innovation monitoring, R&D strategic decision-making and other aspects. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the technical problem to be solved by the present invention is to provide a method for constructing a technical field ontology based on patent data.

[0006] The technical solution adopted by the present invention to solve the above technical problems is:

[0007] A method for constructing a technical field ontology based on patent data, characterized in that the method comprises the following steps:

[0008] Step 1: Obtain patent dataset and preprocess it;

[0009] Step 2: Construct a static layer based on the IPC main classification number and the general knowledge base, including the first, second, and third technical concept layers and their hierarchical relationships;

[0010] The section names, major category names, and minor category names corresponding to the IPC main classification numbers are linked to entities in the general knowledge base to obtain the first, second, and third-level candidate technical concept sets; the first and second-level candidate technical concepts are manually screened and their hierarchical relationships are determined, and the third-level candidate technical concepts are automatically screened using the GAT model to obtain the first, second, and third technical concept layers, each of which contains multiple technical concepts;

[0011] For the hierarchical relationship between the third technical concept layer and the second technical concept layer, each second-level technical concept with the same major category symbol as the third-level technical concept IPC main classification number is used as a candidate parent node, and the similarity between the third-level technical concept and the candidate parent node is calculated, and the second-level technical concept with the highest similarity is used as the parent node;

[0012] Step 3: Construct a dynamic layer based on the patent name, including the fourth, fifth, and sixth technical concept layers and their hierarchical relationships;

[0013] The patent name is input into the PatentNER model to identify technical concepts, and the identified technical concepts are post-processed; the technical concepts retained after post-processing are input into the SCIBERT model for vector embedding to obtain the technical concept embedding vector; the UMAP algorithm is used to reduce the dimension of the technical concept embedding vector to obtain the reduced-dimensional technical concept embedding vector; the DBSCAN algorithm is applied to the reduced-dimensional technical concept embedding vector space for density clustering to generate multiple clusters, with the size of each cluster not exceeding a preset threshold; the Ward algorithm is used to perform hierarchical clustering within each cluster to construct a tree-like hierarchical structure; each technical concept is assigned to the preset three technical concept layers according to its depth in the tree-like hierarchical structure, thereby obtaining the fourth, fifth, and sixth technical concept layers;

[0014] Each technical concept in the fifth and sixth technical concept layers is used as a technical concept of the parent node to be assigned, and the technical concept in the previous technical concept layer with the same IPC main classification number as the technical concept of the parent node to be assigned is used as a candidate parent node. The cosine similarity between the technical concept of the parent node to be assigned and each candidate parent node is calculated, and the cosine similarity is sorted in descending order. The more similar candidate parent nodes are selected using the elbow method to obtain the optimal parent node set; based on the optimal parent node set, the technical concept of the parent node to be assigned is connected with the parent node to complete the construction of the hierarchical relationship;

[0015] Step 4: Using the same method as that used to construct the hierarchical relationship between the fifth and sixth technical concept layers, establish a hierarchical relationship between the third technical concept layer and the fourth technical concept layer, realize the integration of the static layer and the dynamic layer, and thus obtain a directed acyclic graph structure, namely the technical field ontology.

[0016] Furthermore, in the second step, the SCIBERT model is used to encode the third-layer technical concepts and their candidate parent nodes to obtain the embedding vectors of each word; based on the word embedding vectors, the similarity between the third-layer technical concepts and the candidate parent nodes is calculated using formula (1);

[0017]

[0018] Where, sim(e 3i ,e 2j ) represents the i-th third-layer technical concept e 3i The jth second-level technical concept e with the same IPC classification number major category symbol 2j The similarity between Indicates e 3i The embedding vector of the fth word in , Indicates e 2j The embedding vector of the lth word in ω f represents the weight of the fth word, n represents e 3i The number of words contained, m means e 2f The number of words included, cos(·) represents the cosine similarity.

[0019] Furthermore, the PatentNER model includes a SCIBERT model, a BiLSTM model, and a CustomCRF layer. The CustomCRF layer is used to predict the labels of each word in the patent name. There are three possible predicted labels for each word: "B-FOT" represents the first word of the technical concept, "I-FOT" represents the remaining words of the technical concept except the first word, and "O" represents a non-technical concept. The CustomCRF layer generates labels according to the following constraints:

[0020] a) The B-FOT tag must be followed by the I-FOT tag;

[0021] b) Based on the part-of-speech tagging of patent names, the probability of words representing non-technical concepts being predicted as technical concepts is reduced;

[0022] c) Increase the probability that multiple adjacent words after a word labeled B-FOT are predicted to be labeled I-FOT.

[0023] Furthermore, in the third step, post-processing of technical concepts includes: removing technical concepts that begin or end with invalid parts of speech; replacing hyphens in technical concepts with spaces; and removing duplicate technical concepts that already exist in the static layer.

[0024] Furthermore, in the first step, preprocessing includes: Unicode character standardization, special character processing, duplicate data removal, and data format conversion.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] 1. The present invention can automatically extract technical concepts from patent data and construct a multi-level technical field ontology without manual intervention, greatly improving construction efficiency. A static layer is constructed based on the IPC main classification number and the general knowledge base. Each technical concept in the static layer is associated with a certain part of the IPC classification number to ensure the consistency of patent classification. A dynamic layer is constructed based on the patent name. The technical concepts in the dynamic layer may be associated with multiple parts of the IPC classification number. Therefore, the technical concepts in the dynamic layer can capture interdisciplinary characteristics.

[0027] 2. The PatentNER model employed in this paper combines the advantages of the SCIBERT model, the BiLSTM model, and the CustomCRF layer. The CustomCRF layer proposes label constraint rules applicable to the patent domain, enabling the model to accurately identify technical concepts and improving their accuracy. During the dynamic layer construction process, a combination of DBSCAN and Ward algorithms is used to cluster technical concepts, effectively capturing the hierarchical relationships between technical concepts and facilitating the construction of a semantically meaningful technical domain ontology.

[0028] 3. The present invention can process large-scale patent data, adapt to rapid changes in the technology field, and has good scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is the overall flow chart of the present invention;

[0030] Figure 2 It is a flow chart of constructing a static layer according to the present invention;

[0031] Figure 3 It is a flow chart of constructing a dynamic layer according to the present invention;

[0032] Figure 4 It is a schematic structural diagram of the PatentNER model of the present invention;

[0033] Figure 5 It is a flow chart of each technical concept layer of the dynamic layer obtained by the present invention. DETAILED DESCRIPTION

[0034] Specific embodiments are given below in conjunction with the accompanying drawings. The specific embodiments are only used to introduce the technical solutions of the present invention in detail and are not intended to limit the scope of protection of the present application.

[0035] The present invention provides a method for constructing a technical field ontology based on patent data, comprising the following steps:

[0036] Step 1: Obtain the patent dataset and perform preprocessing including Unicode character standardization, special character processing, duplicate data removal, and data format conversion;

[0037] Unicode character standardization: Unicode character standardization is performed on patent names to unify the representation of special characters, including (1) converting full-width characters to half-width characters to ensure consistency in character encoding; (2) unifying the formats of various types of quotation marks into standard English quotation marks; (3) normalizing whitespace characters, such as replacing tabs and multiple consecutive spaces with a single space; and (4) unifying punctuation marks into a standard form, such as unifying different forms of hyphens and dashes into a standard form.

[0038] Special character processing to ensure the cleanliness and consistency of patent names; (1) Use regular expressions to remove HTML and XML tags, such as 、、 etc.; (2) remove control characters and unprintable characters, such as characters with ASCII values 0-31.

[0039] Deduplication: For duplicate data, only the data with the latest date or the most complete fields are retained to ensure the uniqueness and accuracy of the data. Duplicate data includes: (1) the same patent that appears multiple times, which is identified by the application number; (2) different patents with exactly the same patent name, which are compared by calculating the SHA-256 hash value of the patent name; (3) data with essentially the same content but different patent names, which are compared by calculating the Jaccard similarity of the patent name. The threshold of the Jaccard similarity is set to 0.95.

[0040] Data format conversion: Convert the data format of patent names into a unified structured format, such as JSON or CSV format.

[0041] Step 2: Construct a static layer based on the IPC main classification number and the general knowledge base, including the first, second, and third technical concept layers and their hierarchical relationships;

[0042] 1) Link the section names, major category names, and minor category names corresponding to the IPC main classification numbers to entities in a general knowledge base (e.g., Wikipedia) to transform the abstract IPC classification system into structured, concrete technical concepts, and obtain the first, second, and third-level candidate technical concept sets;

[0043] Specifically, the IPC main classification number is divided into three levels: department (such as "A" for human necessities), major category (such as "A61" for medicine or veterinary medicine; hygiene) and minor category (such as "A61B" for diagnosis; surgery; identification); entity linking is performed on the department name, which is input as a query into the entity linking model and matched with the entities in the general knowledge base. Multiple similar entities are selected as the first-level candidate technical concepts to obtain the first-level candidate technical concept set; similarly, the second-level candidate technical concept set and the third-level candidate technical concept set are obtained. The candidate technical concept contains information such as name, text and URL; this embodiment selects the BLINK model as the entity linking model.

[0044] 2) Screening candidate technical concepts at each layer to obtain various technical concept layers, each layer containing multiple technical concepts;

[0045] Since the first-layer candidate technical concepts and the second-layer candidate technical concepts represent the highest level and a wide range of technical fields, their accuracy is crucial to the quality of the entire technical field ontology. Therefore, manual methods are used for precise screening to obtain the first and second technical concept layers;

[0046] The GAT model is used to automatically screen the third-layer candidate technical concepts, resulting in the third-layer technical concept layer. The GAT model's input is a graph structure composed of third-layer candidate technical concepts as nodes and hyperlink relationships between entities in the general knowledge base as edges. The output is a determination of whether the third-layer candidate technical concepts are technical concepts. If so, they are considered third-layer technical concepts; otherwise, they are eliminated. If a page for an entity in the general knowledge base contains a link to the page for another entity, then there is an association between the two entities, and this association serves as an edge connecting the two entities in the graph structure. The training data for the GAT model includes positive and negative samples, with positive samples being technical concepts and negative samples being non-technical concepts.

[0047] 3) Construct hierarchical relationships between various technical concept layers;

[0048] The hierarchical relationship between the first and second technical concept layers has been determined during the manual screening process, so it is only necessary to construct the parent-child relationship between the third and second technical concept layers. Specifically, for each third-level technical concept, the second-level technical concepts with the same major category symbol as its IPC main classification number are selected as candidate parent nodes. The similarity between the third-level technical concept and the candidate parent nodes is calculated, and the second-level technical concept with the highest similarity is selected as the parent node.

[0049] For example, if the IPC main classification number of the third-level technical concept is A21C, the second-level technical concept with the major classification number A21 is selected as its candidate parent node; the third-level technical concept and its candidate parent node are encoded using the SCIBERT model to obtain the embedding vector of each word; based on the word embedding vector, the similarity between the third-level technical concept and the candidate parent node is calculated using formula (1);

[0050]

[0051] Where, sim(e 3i ,e 2j ) represents the i-th third-layer technical concept e 3i The jth second-level technical concept e with the same IPC classification number major category symbol 2j The similarity between Indicates e 3i The embedding vector of the fth word in ω f represents the weight of the fth word, n represents e 3i The number of words contained, m means e 2f The number of words contained, Indicates e 2j The embedding vector of the lth word in , cos(·) represents the cosine similarity.

[0052] Step 3: Construct a dynamic layer based on the patent name, including the fourth, fifth, and sixth technical concept layers and their hierarchical relationships;

[0053] 1) Input the patent name into the PatentNER model and identify the technical concept based on the patent name;

[0054] like Figure 4 As shown in the figure, the PatentNER model is a technical concept recognition model designed by the present invention for patent names, including the SCIBERT model, the BiLSTM model and the CustomCRF layer; the patent name consists of words w0~w k These words are vector-embedded in the SCIBERT model to capture their semantic information, and word embedding vectors v0~v k ; where k represents the number of words in the name; all word embedding vectors are modeled in the BiLSTM network to obtain vectors y0~y k ; Vector y0~y k After the CustomCRF layer, predict the labels of each word;

[0055] There are three possible predicted labels for each word: B-FOT (Begin-Field Of Technology), I-FOT (Inside-Field Of Technology), and O. B-FOT represents the first word of a technical concept, I-FOT represents the remaining words of a technical concept except the first word, and O represents a non-technical concept. The CustomCRF layer introduces the following label constraints applicable to the patent field:

[0056] a) The B-FOT tag must be followed by the I-FOT tag;

[0057] b) Based on the part-of-speech tagging of patent names, reduce the probability that words representing non-technical concepts (such as prepositions, articles, pronouns, etc.) are predicted as technical concepts;

[0058] c) Increase the probability of predicting multiple adjacent words with the I-FOT label after the word with the B-FOT label, to ensure the continuity of the labels of technical concepts and avoid isolated labels;

[0059] The PatentNER model is trained using the multi-objective loss function of formula (2);

[0060]

[0061] Where, L total Represents the total loss, L CRF represents the conditional random field loss, L focal represents the focal loss, L custom represents the constraint rule loss, represents the L2 regularization term of the model parameter θ, and α, β, γ and λ are hyperparameters;

[0062] Conditional random field loss is used to measure the difference between the predicted label sequence and the true label sequence, and is calculated using the negative log-likelihood loss function. Focus loss is used to deal with the problem of class imbalance and is calculated using the following formula:

[0063] L focal =-α t (1-p t ) τ logp t (3)

[0064] Where p t is the probability of predicting the label to be category t, α t is the weight of category t, τ is the focus parameter;

[0065] The constraint rule loss is used to penalize the label sequence that violates the label constraint rule and is calculated as follows:

[0066]

[0067] Where N is the number of training samples, R is the set of label constraint rules, and s r (x g ) is the label constraint rule r in the training sample x g The penalty value on

[0068] The PatentNER model innovatively integrates the semantic understanding capabilities of the SCIBERT model, the sequence modeling capabilities of the BiLSTM model, and the label dependency capture capabilities of the CustomCRF layer, and significantly improves the accuracy of technical concept recognition through label constraint rules.

[0069] 2) Post-processing the technical concepts identified by the PatentNER model, including: (1) eliminating technical concepts that begin or end with invalid parts of speech (such as prepositions and determiners); (2) replacing hyphens in technical concepts with spaces; (3) eliminating duplicate technical concepts that already exist in the static layer;

[0070] 3) The technical concepts retained after post-processing are input into the SCIBERT model for vector embedding to obtain the technical concept embedding vector; the UMAP algorithm is used to reduce the dimensionality of the technical concept embedding vector, while retaining the semantic relationship of the technical concepts and reducing the complexity of subsequent clustering calculations, to obtain the reduced dimensionality technical concept embedding vector; the DBSCAN algorithm is applied to the space of the reduced dimensionality technical concept embedding vector for density clustering, identifying clusters of arbitrary shapes and generating multiple clusters; for clusters whose size exceeds a preset threshold, the DBSCAN algorithm is recursively applied for subdivision until the size of all clusters does not exceed the preset threshold; the Ward algorithm is used for hierarchical clustering within each cluster to construct a tree-like hierarchical structure; according to the depth of each technical concept in the tree-like hierarchical structure, it is assigned to the preset three technical concept layers through the hierarchical mapping function, thereby forming the fourth, fifth, and sixth technical concept layers;

[0071] 4) Construct the hierarchical relationship between the fifth and sixth technical concept layers;

[0072] Each technical concept in the fifth and sixth technical concept layers is taken as the technical concept of the parent node to be assigned, and the technical concept in the previous technical concept layer with the same IPC main classification number as the technical concept of the parent node to be assigned is taken as the candidate parent node. The FAISS vector database is used to calculate the cosine similarity between the technical concept of the parent node to be assigned and each candidate parent node. The cosine similarity is sorted in descending order, and the elbow rule is applied to analyze the changing trend of the cosine similarity to obtain the optimal parent node set, which contains all candidate parent nodes whose cosine similarity is above the elbow point; according to the optimal parent node set, the technical concept of the parent node to be assigned is connected with the parent node to complete the construction of the hierarchical relationship.

[0073] The dynamic layer contains newly discovered technical concepts according to patent names and is linked together in a structured multi-level form.

[0074] Step 4: Integrate the static layer and the dynamic layer to form the technical domain ontology;

[0075] The key to integration lies in establishing a hierarchical relationship between the third technical concept layer and the fourth technical concept layer, which is similar to the hierarchical relationship between the fifth and sixth technical concept layers, thereby achieving an effective connection between the third technical concept layer and the fourth technical concept layer, and obtaining a unified, coherent directed acyclic graph structure containing six levels, namely the technical field ontology.

[0076] Example

[0077] The method for constructing a technical field ontology based on patent data in this embodiment includes the following steps:

[0078] Step 1: Obtain patent datasets from GooglePatentsPublicData and preprocess the patent data, including Unicode character standardization, special character processing, duplicate data removal, and data format conversion;

[0079] Step 2: Construct a static layer based on the IPC main classification number and the general knowledge base, including the first, second, and third technical concept layers and their hierarchical relationships;

[0080] Among them, the first technical concept layer contains 35 technical concepts, the second technical concept layer contains 280 technical concepts, the third technical concept layer contains 11,519 technical concepts, and the static layer contains a total of 11,834 technical concepts; these technical concepts cover the eight main sections (AH) of the IPC classification number, and are significantly representative in sections A (necessities of human life), B (operations; transportation) and G (physics). For example, section A contains 2,365 technical concepts, section B contains 3,142 technical concepts, and section G contains 1,750 technical concepts; each technical concept in the static layer is associated with a section of the IPC classification number, ensuring consistency with the existing patent classification system.

[0081] Step 3: Construct a dynamic layer based on the patent name, including the fourth, fifth, and sixth technical concept layers and their hierarchical relationships;

[0082] The PatentNER model identified 468,583 technical concepts from 88,878,307 patent titles. These technical concepts constitute the dynamic layer. The fourth technical concept layer contains 100,576 technical concepts, the fifth technical concept layer contains 158,023 technical concepts, and the sixth technical concept layer contains 209,984 technical concepts. The number of technical concepts in each section of the IPC classification code in the dynamic layer has increased significantly. For example, Section C (Chemistry; Metallurgy) in the dynamic layer contains 244,402 technical concepts, far exceeding the 1,503 technical concepts contained in the static layer.

[0083] The dynamic layer also exhibits significant cross-disciplinary characteristics, where only 37.6% of technical concepts are associated with a single section of the IPC classification number, and the remaining 62.4% of technical concepts are associated with multiple sections of the IPC classification number, which shows that the dynamic layer is able to capture the interdisciplinary characteristics of modern technology.

[0084] Step 4: Integrate the static layer and the dynamic layer to form a technical field ontology consisting of six levels.

[0085] Any matters not described in the present invention are applicable to the prior art.

Claims

1. A method for constructing a technical field ontology based on patent data, characterized in that: The method comprises the following steps: Step 1: Obtain patent dataset and preprocess it; Step 2: Construct a static layer based on the IPC main classification number and the general knowledge base, including the first, second, and third technical concept layers and their hierarchical relationships; The section names, major category names, and minor category names corresponding to the IPC main classification numbers are linked to entities in the general knowledge base to obtain the first, second, and third-level candidate technical concept sets; the first and second-level candidate technical concepts are manually screened and their hierarchical relationships are determined, and the third-level candidate technical concepts are automatically screened using the GAT model to obtain the first, second, and third technical concept layers, each of which contains multiple technical concepts; For the hierarchical relationship between the third technical concept layer and the second technical concept layer, each second-level technical concept with the same major category symbol as the third-level technical concept IPC main classification number is used as a candidate parent node, and the similarity between the third-level technical concept and the candidate parent node is calculated, and the second-level technical concept with the highest similarity is used as the parent node; Step 3: Construct a dynamic layer based on the patent name, including the fourth, fifth, and sixth technical concept layers and their hierarchical relationships; The patent name is input into the PatentNER model to identify technical concepts, and the identified technical concepts are post-processed; the technical concepts retained after post-processing are input into the SCIBERT model for vector embedding to obtain the technical concept embedding vector; the UMAP algorithm is used to reduce the dimension of the technical concept embedding vector to obtain the reduced-dimensional technical concept embedding vector; the DBSCAN algorithm is applied to the reduced-dimensional technical concept embedding vector space for density clustering to generate multiple clusters, with the size of each cluster not exceeding a preset threshold; the Ward algorithm is used to perform hierarchical clustering within each cluster to construct a tree-like hierarchical structure; each technical concept is assigned to the preset three technical concept layers according to its depth in the tree-like hierarchical structure, thereby obtaining the fourth, fifth, and sixth technical concept layers; Each technical concept in the fifth and sixth technical concept layers is used as a technical concept of the parent node to be assigned, and the technical concept in the previous technical concept layer with the same IPC main classification number as the technical concept of the parent node to be assigned is used as a candidate parent node. The cosine similarity between the technical concept of the parent node to be assigned and each candidate parent node is calculated, and the cosine similarity is sorted in descending order. The more similar candidate parent nodes are selected using the elbow method to obtain the optimal parent node set; based on the optimal parent node set, the technical concept of the parent node to be assigned is connected with the parent node to complete the construction of the hierarchical relationship; Step 4: Using the same method as that used to construct the hierarchical relationship between the fifth and sixth technical concept layers, establish a hierarchical relationship between the third technical concept layer and the fourth technical concept layer, realize the integration of the static layer and the dynamic layer, and thus obtain a directed acyclic graph structure, namely the technical field ontology.

2. The method for constructing a technical field ontology based on patent data according to claim 1 is characterized in that: In the second step, the SCIBERT model is used to encode the third-layer technical concepts and their candidate parent nodes to obtain the embedding vectors of each word. Based on the word embedding vectors, the similarity between the third-layer technical concepts and the candidate parent nodes is calculated using formula (1). Where, sim(e 3i ,e 2j ) represents the i-th third-layer technical concept e 3i The jth second-level technical concept e with the same IPC classification number major category symbol 2j The similarity between Indicates e 3i The embedding vector of the fth word in , Indicates e 2j The embedding vector of the lth word in ω f represents the weight of the fth word, n represents e 3i The number of words contained, m means e 2f The number of words included, cos(·) represents the cosine similarity.

3. The method for constructing a technical field ontology based on patent data according to claim 1 or 2, characterized in that: The PatentNER model includes a SCIBERT model, a BiLSTM model, and a CustomCRF layer. The CustomCRF layer is used to predict the labels of each word in the patent name. There are three possible predicted labels for each word: "B-FOT" represents the first word of the technical concept, "I-FOT" represents the remaining words of the technical concept except the first word, and "O" represents a non-technical concept. The CustomCRF layer generates labels based on the following constraints: a) The B-FOT tag must be followed by the I-FOT tag; b) Based on the part-of-speech tagging of patent names, the probability of words representing non-technical concepts being predicted as technical concepts is reduced; c) Increase the probability that multiple adjacent words after a word labeled B-FOT are predicted to be labeled I-FOT.

4. The method for constructing a technical field ontology based on patent data according to claim 3 is characterized in that: In the third step, post-processing of technical concepts includes: removing technical concepts that begin or end with invalid parts of speech; replacing hyphens in technical concepts with spaces; and removing duplicate technical concepts that already exist in the static layer.

5. The method for constructing a technical field ontology based on patent data according to claim 1, characterized in that: In the first step, preprocessing includes: Unicode character standardization, special character processing, duplicate data removal, and data format conversion.

Images