Intellectual property information analysis and management method based on multi-dimensional data

Abstract

The invention discloses an intellectual property information analysis and management method based on multi-dimensional data, and the method comprises the following steps: obtaining patent original data, constructing a patent record, and carrying out the preprocessing of an intellectual property data set; dividing the patent text data into a plurality of coding units according to a preset field range, and constructing vectorization representation for each patent; generating an adjusted sub-query set for the retrieval input based on a query decomposition algorithm; obtaining a sub-query mark sequence and a comprehensive candidate result set; and calculating a comprehensive sorting score, and generating a retrieval result list. According to the method, by introducing a patent retrieval framework combining multi-dimensional query decomposition and ConstBERT fixed multi-vector document representation, executable sub-query configuration can be stably generated and fine-grained semantic matching can be realized in a scene in which a long retrieval type, multiple limits and exclusion conditions coexist; and the accuracy, controllability, traceability and large-scale application performance of patent retrieval are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of information management, and in particular to a knowledge property right information analysis management method based on multi-dimensional data. BACKGROUND

[0002] The existing knowledge property right retrieval system is usually based on keyword matching and Boolean retrieval of title, abstract, claim, etc. fields, and combined with IPC / CPC, applicant, time, legal status, etc. metadata for filtering and sorting. In recent years, there is also a vector retrieval scheme based on deep semantic representation. The BERT model is used to vectorize the query and patent text to improve the semantic recall ability. Some systems further use a multi-vector interaction retrieval framework to enhance the matching ability of long text and fine-grained semantics.

[0003] However, the above-mentioned prior art still has deficiencies when facing complex patent retrieval scenarios. On the one hand, long retrieval and natural language requirements often contain mandatory elements, optional elements and exclusion elements, accompanied by field limitations and multi-dimensional constraints. The existing system relies on manual rewriting, expansion and weight adjustment, and it is difficult to form a structured and reusable subquery configuration. The retrieval stability and controllability are insufficient. On the other hand, the multi-vector retrieval usually needs to store a large number of variable number of vector representations on the document side, resulting in index volume and calculation overhead increasing with the length of the text, making it difficult to balance the response efficiency of large-scale library. In addition, the retrieval results often lack evidence-level association with dimension labels and original text positions, making it difficult to support traceable analysis management and report output of the retrieval project. SUMMARY

[0004] To solve the technical problems in the background art, the present application proposes a knowledge property right information analysis management method based on multi-dimensional data.

[0005] The knowledge property right information analysis management method based on multi-dimensional data proposed by the present application comprises the following steps: S1, obtaining patent original data, constructing a patent record for each patent and generating a patent identifier, constructing a knowledge property right data set and preprocessing; S2, dividing the patent text data in the patent record into a plurality of coding units according to a pre-set field range, constructing a vector representation for each patent based on the plurality of coding units and using the ConstBERT algorithm; S3, receiving a retrieval input and performing constraint analysis to obtain a basic constraint set, generating an adjusted subquery set for the retrieval input based on a query decomposition algorithm, and encapsulating the adjusted subquery set as a retrieval execution configuration; S4, obtaining a subquery mark sequence for each subquery in the adjusted subquery set, and obtaining a comprehensive candidate result set through the ConstBERT algorithm; S5, calculating a comprehensive ranking score for each candidate patent in the comprehensive candidate result set, and generating a search result list according to the comprehensive ranking score.

[0006] Preferably, patent original data is acquired, a patent record is constructed for each patent, and a patent identifier is generated, as follows: Acquiring patent original data, constructing a patent record for each patent , the patent record being composed of patent text data and patent metadata , the patent text data including title text, abstract text, claim text, and specification text, wherein, the title text, the abstract text, the claim text, and the specification text, the patent metadata including classification number field, applicant field, inventor field, application date field, publication date field, homologous identification field, citation relationship field, and legal status field, , the classification number field, the applicant field, the inventor field, the application date field, the publication date field, the homologous identification field, the citation relationship field, and the legal status field; a patent identifier is generated for each patent record , such that wherein, the publication number or announcement number, the type code, represents a string concatenation operation, and the patent identifier is stored in association with the patent text data and the patent metadata , forming a knowledge property data set , , wherein, the patent record corresponding to the th patent, and the total number of acquired patent records.

[0007] Preferably, the knowledge property data set is constructed and preprocessed, as follows: Perform field integrity verification and field standardization on the intellectual property dataset; Perform text segmentation on the patent text data to obtain a set of text segments. It generates paragraph position identifiers for each text segment, performs clause segmentation on the claims text, and generates a set of claims clauses. ,in, For the first The first patent The text of each claim clause is generated, and a clause position identifier is generated for each claim clause. The paragraph position identifiers and clause position identifiers are combined into a position identifier set. For the set of claims Perform dependency annotation and construct a dependency set. ,in, , For the first The parent number of each claim, when When it is an independent claim, ,when When it is a dependent claim The independent claim clause and dependent claim clause are obtained based on the dependent claim set, according to the number of the claim to which it belongs.

[0008] Preferably, the patent text data in the patent record is divided into multiple encoding units according to a preset field range. Based on these multiple encoding units, a vectorized representation is constructed for each patent using the ConstBERT algorithm, as follows: Based on the preset field range The patent text data is divided to construct a set of coding units, which includes title coding units formed by title text, abstract coding units formed by abstract text, independent claim coding units formed by independent claim clauses, dependent claim coding units formed by dependent claim clauses, and specification coding units formed by specification text. Perform word segmentation and serialization processing on each coding unit in the coding unit set to generate a token sequence. , , For the first One tag, To mark the length of the sequence; The labeled sequence is input into the text encoder of ConstBERT to obtain the label-level representation matrix. , ,in, Output dimensions for the ConstBERT encoder. Indicates the first a mark of ConstBERT is a multi-vector retrieval model, primarily used for semantic retrieval. Its overall architecture consists of a query encoder, a document encoder, and a scoring module. And the label-level representation matrix Perform fixed multi-vector generation processing to obtain a fixed vector set. , , A fixed number of vectors are allocated to this coding unit, and the following conditions are met: , A fixed number of document vectors are generated for each patent; By concatenating the fixed vector sets of each coding unit, we obtain the first... Document vector set of patents And generate a vector identifier for each document vector in the document vector set. The vector identifier is derived from the patent identifier of the patent. The index of the document vector in the document vector set Together, we determine and establish the mapping relationship between vector identifiers and field ranges, as well as the mapping relationship between vector identifiers and the set of position identifiers; A vector index library is constructed based on the document vector set, vector identifier, mapping relationship between vector identifier and field range, and mapping relationship between vector identifier and location identifier set of patents. An association index is established between the vector index library and patent metadata.

[0009] Preferably, the retrieval input is received and constraint parsing is performed to obtain a basic constraint set. An adjusted subquery set is generated from the retrieval input based on a query decomposition algorithm, and the adjusted subquery set is encapsulated as a retrieval execution configuration, as follows: Receive search input And determine the set of input types to be retrieved. ,in, For the search content entered by the user, , These represent Boolean search input, natural language input, and mixed input, respectively. To retrieve the number of input types; For search input Perform constraint resolution to obtain the basic constraint set. The basic constraint set includes classification number constraints, time constraints, legal status constraints, and inventor entity constraints. ,in, For classification number constraints, Due to time constraints, Constrained by legal status, For inventor entity constraints, the time constraint is expressed as: ,in, The start date of the search. The search period ends on the specified date. Enter the search results Transform into a structured retrieval intent representation The structured search intent representation includes a set of required constraint descriptions, a set of optional constraint descriptions, a set of excluded constraint descriptions, a set of parameter range descriptions, a set of application scenario descriptions, and a set of field range descriptions. ,in, This is a set of mandatory constraint descriptions. This is a set of optional constraint descriptions. To exclude the constraint description set, Describes a set of parameters. A collection of application scenario descriptions. Describe a collection for a range of fields; The query decomposition algorithm is used to decompose the structured search intent representation into dimensions, generating a set of subqueries. ,in, For the first Each subquery generates a set of dimension labels based on the set of subqueries. Field range collection Weight parameter set Extended vocabulary collection and a collection of negative words Each subquery in the subquery set corresponds to each dimension label in the dimension label set, each field range in the field range set, each weight parameter in the weight parameter set, each extended word set in the extended word set, and each negative word set in the negative word set. For the first Dimension labels for each subquery ; Pair query set Apply a set of preset performance constraints Adjustments are made to obtain the adjusted subquery set. The preset performance constraint set includes a maximum constraint on the number of subqueries and rules for the effectiveness of negative keywords. , This is a constraint on the maximum number of subqueries. This is a rule for the effectiveness of negation words and is used to determine the total set of negation words. Within the scope of application at the subquery level and the search input level; Adjusted subquery set Dimensional label set Field range collection Weight parameter set Extended vocabulary collection A collection of negative words and the basic constraint set Encapsulated as a retrieval execution configuration : .

[0010] Preferably, a preset set of performance constraints is applied to the subquery set for adjustment, resulting in the adjusted subquery set, as follows: When the number of subqueries exceeds the upper limit constraint, the weight parameter is used. Eliminate low-priority subqueries; Based on the rules for negation effectiveness, negations in the subquery are divided into globally effective negations and locally effective negations within the subquery. The number of negations is then pruned, and excess negations are added to the exclusion constraint description set. middle.

[0011] Preferably, the negative words in the subquery are divided into globally effective negative words and locally effective negative words in the subquery according to the rules for the effectiveness of negative words, as follows: When negative words originate from structured search intent representation Exclusion constraint description set When this happens, the negation word is defined as a globally effective negation word; When the negation word originates from the negation word set of a certain subquery in the subquery set. When a negation word is defined as a negation word that is locally effective within a subquery.

[0012] Preferably, a subquery tag sequence is obtained based on each subquery in the adjusted subquery set, and a comprehensive candidate result set is obtained through the ConstBERT algorithm, as follows: A set of candidate patent identifiers is obtained by filtering patent metadata based on a set of fundamental constraints. The subquery text is obtained by concatenating each subquery in the adjusted subquery set with the corresponding extended word set. The subquery text is segmented and serialized to obtain a subquery tag sequence, and then the subquery tag sequence is input into the ConstBERT query encoder to generate the corresponding query vector set. For each query vector in the query vector set, perform a vector nearest neighbor search in the vector index, and simultaneously apply a set of candidate patent identifiers during the search process. and field range The constraints are used to obtain a set of vector recall results that satisfy the constraints. The candidate result set corresponding to the subquery is obtained by merging the vector recall result sets of the same subquery, and the candidate result sets corresponding to all subqueries are further merged to obtain the comprehensive candidate result set. The correspondence record between each candidate patent and the source subquery in the comprehensive candidate result set is established.

[0013] Preferably, a comprehensive ranking score is calculated for each candidate patent in the comprehensive candidate result set, and a search result list is generated based on the comprehensive ranking score, as follows: For each subquery in the adjusted subquery set, read the corresponding query vector set, field range, and negative word set, and then, based on the comprehensive candidate result set, extract the document vector set of the candidate patents. Select document vectors whose field ranges match the field ranges of the subquery to form a subset of matching vectors. ; Subquery-based query vector set Matching a subset of vectors Perform multi-vector similarity calculation to obtain the similarity sub-scores of the subquery to the candidate patents. : ; in, For the first The first subquery A query vector, For the first The number of query vectors for each subquery. Indicates cosine similarity; Similarity sub-scores obtained for the same candidate patent under multiple sub-queries According to the weight parameters corresponding to each subquery Weighted fusion is performed to obtain the comprehensive ranking score of the candidate patents. : ; The comprehensive ranking scores of each candidate patent in the comprehensive candidate result set are sorted to generate a ranking result list. The ranking result list is then deduplicated to generate a search result list. The basic constraint set, search execution configuration, and search result list are written into the search item record.

[0014] The intellectual property information analysis and management method based on multi-dimensional data proposed in this invention has the following beneficial technical effects: 1. This application introduces the ConstBERT fixed multi-vector document representation mechanism, which constructs encoding units for patent text according to preset field ranges and generates a fixed number of document vectors. It establishes a mapping relationship between document vectors and field labels and text position identifiers. This maintains the fine-grained semantic matching capability of multi-vectors while changing the document-side index size from variable to constant level. This significantly reduces the vector storage and retrieval computation overhead in large-scale patent database scenarios, improves retrieval response efficiency and system scalability, and enables consistent management and reuse of matching paths under different field ranges.

[0015] 2. This application combines a query decomposition algorithm to convert the search input into a structured search intent and output a set of subqueries with dimension labels, field ranges, weight parameters, extended terms, and negative terms. At the same time, it applies performance constraints such as the number of subqueries and the rules for the effectiveness of negative terms to the decomposition results, forming an executable search configuration. On this basis, multi-vector recall and weighted fusion ranking are performed, and evidence location and project record solidification are achieved based on the hit document vector and text location identifier. This allows complex search intents to be stably implemented in the form of multi-dimensional constraints, reducing reliance on manual rewriting and parameter tuning, improving search accuracy, controllability, and result traceability, and thus supporting the standardized output of intellectual property search analysis and management. Attached Figure Description

[0016] Figure 1 This is a flowchart of the intellectual property information analysis and management method based on multi-dimensional data according to the present invention. Detailed Implementation

[0017] Embodiments of the present invention are described in detail below. Examples of these embodiments are illustrated in the accompanying drawings, wherein the same or similar symbols denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0018] like Figure 1 The intellectual property information analysis and management method based on multi-dimensional data shown includes the following steps: S1. Obtain the original patent data, construct a patent record and generate a patent identifier for each patent, construct an intellectual property dataset and perform preprocessing. In an optional embodiment, the original patent data is obtained, a patent record is constructed for each patent, and a patent identifier is generated, as follows: Obtain the original patent data and construct a patent record for each patent. The patent record consists of patent text data. With patent metadata Composition, the patent text data This includes the title text, abstract text, claims text, and specification text. ,in, For the title text, This is the abstract text. For the text of the claims, The patent metadata is the text of the specification. This includes fields for classification number, applicant, inventor, application date, publication date, family identifier, citation relationship, and legal status. , For the category number field, For the applicant field, For the inventor field, For the application date field, For the date of publication, For family identification field, For citation relation fields, For legal status fields; For each patent record Generate patent logo , making ,in, This is a public number or announcement number. For category codes, This represents a string concatenation operation and includes the patent identifier. With patent text data and patent metadata Linked storage to form an intellectual property dataset , , ,in, For the first The patent records corresponding to each patent, The total number of patent records obtained; In an optional embodiment, an intellectual property dataset is constructed and preprocessed as follows: Perform field integrity verification and field standardization on the intellectual property dataset; Perform text segmentation on the patent text data to obtain a set of text segments. It generates paragraph position identifiers for each text segment, performs clause segmentation on the claims text, and generates a set of claims clauses. ,in, For the first The first patent The text of each claim clause is generated, and a clause position identifier is generated for each claim clause. The paragraph position identifiers and clause position identifiers are combined into a position identifier set. For the set of claims Perform dependency annotation and construct a dependency set. ,in, , For the first The parent number of each claim, when When it is an independent claim, ,when When it is a dependent claim The independent claim clause and dependent claim clause are obtained based on the dependent claim set, according to the number of the claim to which it belongs.

[0019] S2. Divide the patent text data in the patent record into multiple coding units according to the preset field range, and construct a vectorized representation for each patent based on the multiple coding units and the ConstBERT algorithm. In an optional embodiment, the patent text data in the patent record is divided into multiple encoding units according to a preset field range. Based on these multiple encoding units, a vectorized representation is constructed for each patent using the ConstBERT algorithm, as follows: Based on the preset field range The patent text data is divided to construct a set of coding units, which includes title coding units formed by title text, abstract coding units formed by abstract text, independent claim coding units formed by independent claim clauses, dependent claim coding units formed by dependent claim clauses, and specification coding units formed by specification text. Perform word segmentation and serialization processing on each coding unit in the coding unit set to generate a token sequence. , , For the first One tag, To mark the length of the sequence; The labeled sequence is input into the text encoder of ConstBERT to obtain the label-level representation matrix. , ,in, Output dimensions for the ConstBERT encoder. Indicates the first a mark of ConstBERT is a multi-vector retrieval model, primarily used for semantic retrieval. Its overall architecture consists of a query encoder, a document encoder, and a scoring module. And the label-level representation matrix Perform fixed multi-vector generation processing to obtain a fixed vector set. , , A fixed number of vectors are allocated to this coding unit, and the following conditions are met: , A fixed number of document vectors are generated for each patent; By concatenating the fixed vector sets of each coding unit, we obtain the first... Document vector set of patents And generate a vector identifier for each document vector in the document vector set. The vector identifier is derived from the patent identifier of the patent. The index of the document vector in the document vector set Together, we determine and establish the mapping relationship between vector identifiers and field ranges, as well as the mapping relationship between vector identifiers and the set of position identifiers; A vector index library is constructed based on the document vector set, vector identifier, mapping relationship between vector identifier and field range, and mapping relationship between vector identifier and location identifier set of patents. An association index is established between the vector index library and patent metadata.

[0020] S3. Receive the retrieval input and perform constraint parsing to obtain the basic constraint set. Generate an adjusted subquery set based on the query decomposition algorithm and encapsulate the adjusted subquery set into a retrieval execution configuration. In an optional embodiment, the retrieval input is received and constraint parsing is performed to obtain a basic constraint set. An adjusted subquery set is generated from the retrieval input based on a query decomposition algorithm, and the adjusted subquery set is encapsulated as a retrieval execution configuration, as follows: Receive search input And determine the set of input types to be retrieved. ,in, For the search content entered by the user, , These represent Boolean search input, natural language input, and mixed input, respectively. To retrieve the number of input types; For search input Perform constraint resolution to obtain the basic constraint set. The basic constraint set includes classification number constraints, time constraints, legal status constraints, and inventor entity constraints. ,in, For classification number constraints, Due to time constraints, Constrained by legal status, For inventor entity constraints, the time constraint is expressed as: ,in, The start date of the search. The search period ends on the specified date. Enter the search results Transform into a structured retrieval intent representation The structured search intent representation includes a set of required constraint descriptions, a set of optional constraint descriptions, a set of excluded constraint descriptions, a set of parameter range descriptions, a set of application scenario descriptions, and a set of field range descriptions. ,in, This is a set of mandatory constraint descriptions. This is a set of optional constraint descriptions. To exclude the constraint description set, Describes a set of parameters. A collection of application scenario descriptions. Describe a collection for a range of fields; The query decomposition algorithm is used to decompose the structured search intent representation into dimensions, generating a set of subqueries. ,in, For the first Each subquery generates a set of dimension labels based on the set of subqueries. Field range collection Weight parameter set Extended vocabulary collection and a collection of negative words Each subquery in the subquery set corresponds to each dimension label in the dimension label set, each field range in the field range set, each weight parameter in the weight parameter set, each extended word set in the extended word set, and each negative word set in the negative word set. For the first Dimension labels for each subquery ; Pair query set Apply a set of preset performance constraints Adjustments are made to obtain the adjusted subquery set. The preset performance constraint set includes a maximum constraint on the number of subqueries and rules for the effectiveness of negative keywords. , This is a constraint on the maximum number of subqueries. This is a rule for the effectiveness of negation words and is used to determine the total set of negation words. Within the scope of application at the subquery level and the search input level; Adjusted subquery set Dimensional label set Field range collection Weight parameter set Extended vocabulary collection A collection of negative words and the basic constraint set Encapsulated as a retrieval execution configuration : ; In an optional embodiment, a preset set of performance constraints is applied to the subquery set for adjustment, resulting in an adjusted subquery set, as follows: When the number of subqueries exceeds the upper limit constraint, the weight parameter is used. Eliminate low-priority subqueries; Based on the rules for negation effectiveness, negations in the subquery are divided into globally effective negations and locally effective negations within the subquery. The number of negations is then pruned, and excess negations are added to the exclusion constraint description set. middle; In an optional embodiment, the negative words in the subquery are divided into globally effective negative words and subquery locally effective negative words according to the negative word effectiveness rules, as follows: When negative words originate from structured search intent representation Exclusion constraint description set When this happens, the negation word is defined as a globally effective negation word; When the negation word originates from the negation word set of a certain subquery in the subquery set. When a negation word is defined as a negation word that is locally effective within a subquery.

[0021] S4. Obtain the subquery tag sequence based on each subquery in the adjusted subquery set, and obtain the comprehensive candidate result set through the ConstBERT algorithm; In an optional embodiment, a subquery tag sequence is obtained based on each subquery in the adjusted subquery set, and a comprehensive candidate result set is obtained through the ConstBERT algorithm, as follows: A set of candidate patent identifiers is obtained by filtering patent metadata based on a set of fundamental constraints. The subquery text is obtained by concatenating each subquery in the adjusted subquery set with the corresponding extended word set. The subquery text is segmented and serialized to obtain a subquery tag sequence, and then the subquery tag sequence is input into the ConstBERT query encoder to generate the corresponding query vector set. For each query vector in the query vector set, perform a vector nearest neighbor search in the vector index, and simultaneously apply a set of candidate patent identifiers during the search process. and field range The constraints are used to obtain a set of vector recall results that satisfy the constraints. The candidate result set corresponding to the subquery is obtained by merging the vector recall result sets of the same subquery, and the candidate result sets corresponding to all subqueries are further merged to obtain the comprehensive candidate result set. The correspondence record between each candidate patent and the source subquery in the comprehensive candidate result set is established.

[0022] S5. Calculate the comprehensive ranking score for each candidate patent in the comprehensive candidate result set, and generate a search result list based on the comprehensive ranking score; In an optional embodiment, a comprehensive ranking score is calculated for each candidate patent in the comprehensive candidate result set, and a search result list is generated based on the comprehensive ranking score, as follows: For each subquery in the adjusted subquery set, read the corresponding query vector set, field range, and negative word set, and then, based on the comprehensive candidate result set, extract the document vector set of the candidate patents. Select document vectors whose field ranges match the field ranges of the subquery to form a subset of matching vectors. ; Subquery-based query vector set Matching a subset of vectors Perform multi-vector similarity calculation to obtain the similarity sub-scores of the subquery to the candidate patents. : ; in, For the first The first subquery A query vector, For the first The number of query vectors for each subquery. Indicates cosine similarity; Similarity sub-scores obtained for the same candidate patent under multiple sub-queries According to the weight parameters corresponding to each subquery Weighted fusion is performed to obtain the comprehensive ranking score of the candidate patents. : ; The comprehensive ranking scores of each candidate patent in the comprehensive candidate result set are sorted to generate a ranking result list. The ranking result list is then deduplicated to generate a search result list. The basic constraint set, search execution configuration, and search result list are written into the search item record.

[0023] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intellectual property information analysis and management based on multi-dimensional data, characterized in that, Includes the following steps: S1. Obtain the original patent data, construct a patent record and generate a patent identifier for each patent, construct an intellectual property dataset and perform preprocessing. S2. Divide the patent text data in the patent record into multiple coding units according to the preset field range, and construct a vectorized representation for each patent based on the multiple coding units and the ConstBERT algorithm. S3. Receive the retrieval input and perform constraint parsing to obtain the basic constraint set. Generate an adjusted subquery set based on the query decomposition algorithm and encapsulate the adjusted subquery set into a retrieval execution configuration. S4. Obtain the subquery tag sequence based on each subquery in the adjusted subquery set, and obtain the comprehensive candidate result set through the ConstBERT algorithm; S5. Calculate the comprehensive ranking score for each candidate patent in the comprehensive candidate result set, and generate a search result list based on the comprehensive ranking score.

2. The intellectual property information analysis and management method based on multi-dimensional data according to claim 1, characterized in that, Obtain the original patent data, construct a patent record for each patent, and generate a patent identifier, as follows: Obtain the original patent data and construct a patent record for each patent. The patent record consists of patent text data and patent metadata. The patent text data includes title text, abstract text, claims text, and specification text. The patent metadata includes classification number field, applicant field, inventor field, application date field, publication date field, family identifier field, citation relationship field, and legal status field. A patent identifier is generated for each patent record, and the patent identifier is associated with and stored with the patent text data and patent metadata to form an intellectual property dataset.

3. The intellectual property information analysis and management method based on multi-dimensional data according to claim 1, characterized in that, The intellectual property dataset is constructed and preprocessed as follows: Perform field integrity verification and field standardization on the intellectual property dataset; The patent text data is segmented to obtain a set of text segments, and a paragraph position identifier is generated for each text segment. The claim text is segmented to generate a set of claim clauses, and a clause position identifier is generated for each claim clause. The paragraph position identifiers and clause position identifiers are combined to form a position identifier set. Perform dependency labeling on the set of claim clauses to construct a dependency set, and obtain independent claim clauses and dependent claim clauses based on the dependency set.

4. The intellectual property information analysis and management method based on multi-dimensional data according to claim 1, characterized in that, The patent text data in the patent record is divided into multiple encoding units according to preset field ranges. Based on these multiple encoding units, and using the ConstBERT algorithm, a vectorized representation is constructed for each patent, as follows: The patent text data is divided into a set of coding units based on a preset field range. The set of coding units includes a title coding unit formed by the title text, an abstract coding unit formed by the abstract text, an independent claim coding unit formed by the independent claim clauses, a dependent claim coding unit formed by the dependent claim clauses, and a specification coding unit formed by the specification text. Perform word segmentation and serialization on each coding unit in the coding unit set to generate a token sequence; The labeled sequence is input into the text encoder of ConstBERT to obtain the labeled representation matrix, and a fixed multi-vector generation process is performed on the labeled representation matrix to obtain a fixed vector set. The fixed vector sets of each coding unit are concatenated to obtain the document vector set of the patent, and a vector identifier is generated for each document vector in the document vector set. The vector identifier is determined by the patent identifier of the patent and the serial number of the document vector in the document vector set. At the same time, a mapping relationship between the vector identifier and the field range and a mapping relationship between the vector identifier and the location identifier set are established. A vector index library is constructed based on the document vector set, vector identifier, mapping relationship between vector identifier and field range, and mapping relationship between vector identifier and location identifier set of patents. An association index is established between the vector index library and patent metadata.

5. The intellectual property information analysis and management method based on multi-dimensional data according to claim 1, characterized in that, The system receives the search input and performs constraint parsing to obtain the basic constraint set. Based on the query decomposition algorithm, it generates an adjusted subquery set from the search input and encapsulates this adjusted subquery set into a search execution configuration, as follows: Receive search input and determine the set of search input types; Constraint parsing is performed on the search input to obtain a basic constraint set, which includes classification number constraints, time constraints, legal status constraints, and inventor entity constraints. The search input is converted into a structured search intent representation, which includes a set of mandatory constraint descriptions, a set of optional constraint descriptions, a set of excluded constraint descriptions, a set of parameter range descriptions, a set of application scenario descriptions, and a set of field range descriptions. A query decomposition algorithm is used to decompose the structured retrieval intent representation into dimensions, generating a set of subqueries. Based on the set of subqueries, a set of dimension labels, a set of field ranges, a set of weight parameters, a set of extended words, and a set of negative words are generated. Each subquery in the set of subqueries corresponds to each dimension label in the set of dimension labels, each field range in the set of field ranges, each weight parameter in the set of weight parameters, each set of extended words in the set of extended words, and each set of negative words in the set of negative words. The subquery set is adjusted by applying a preset performance constraint set, which includes a maximum constraint on the number of subqueries and a rule for negation words to take effect. The adjusted subquery set, dimension label set, field range set, weight parameter set, extended term set, negative term set, and basic constraint set are encapsulated into a retrieval execution configuration.

6. The intellectual property information analysis and management method based on multi-dimensional data according to claim 1, characterized in that, The subquery set is adjusted by applying a preset set of performance constraints, resulting in the adjusted subquery set as follows: When the number of subqueries exceeds the upper limit constraint, low-priority subqueries are eliminated based on the weight parameter. Based on the rules for negation, the negations in the subquery are divided into globally effective negations and locally effective negations in the subquery. The number of negations is then pruned, and any excess negations are added to the exclusion constraint description set.

7. The intellectual property information analysis and management method based on multi-dimensional data according to claim 1, characterized in that, Based on the rules for negation effectiveness, negation words in the subquery are divided into globally effective negation words and subquery locally effective negation words, as follows: When a negative term originates from the set of exclusion constraints in a structured search intent representation, the negative term is defined as a globally effective negative term. When a negation word originates from the negation word set of a subquery within the subquery set, the negation word is defined as a locally effective negation word for the subquery.

8. The intellectual property information analysis and management method based on multi-dimensional data according to claim 1, characterized in that, Based on each subquery in the adjusted subquery set, a subquery tag sequence is obtained, and a comprehensive candidate result set is obtained through the ConstBERT algorithm, as follows: Based on the basic constraint set, the patent metadata is filtered to obtain the candidate patent identifier set. Each subquery in the adjusted subquery set and the corresponding extended word set of the subquery are concatenated to obtain the subquery text. The subquery text is segmented and serialized to obtain a subquery tag sequence, and then the subquery tag sequence is input into the ConstBERT query encoder to generate the corresponding query vector set. For each query vector in the query vector set, a vector nearest neighbor search is performed in the vector index library, and constraints on the candidate patent identifier set and field range are applied simultaneously during the search process to obtain a vector recall result set; The candidate result set corresponding to the subquery is obtained by merging the vector recall result sets of the same subquery, and the candidate result sets corresponding to all subqueries are further merged to obtain the comprehensive candidate result set. The correspondence record between each candidate patent and the source subquery in the comprehensive candidate result set is established.

9. The intellectual property information analysis and management method based on multi-dimensional data according to claim 1, characterized in that, For each candidate patent in the comprehensive candidate result set, a comprehensive ranking score is calculated, and a search result list is generated based on the comprehensive ranking score, as follows: For each subquery in the adjusted subquery set, read the query vector set, field range and negative word set corresponding to the subquery, and select document vectors from the document vector set of candidate patents whose field range is consistent with the field range of the subquery based on the comprehensive candidate result set to form a matching vector subset; Multi-vector similarity calculation is performed on the query vector set and matching vector subset based on the subquery to obtain the sub-score of the similarity between the subquery and the candidate patent; The similarity sub-scores obtained for the same candidate patent under multiple sub-queries are weighted and fused according to the weight parameters corresponding to each sub-query to obtain the comprehensive ranking score of the candidate patent. The comprehensive ranking scores of each candidate patent in the comprehensive candidate result set are sorted to generate a ranking result list. The ranking result list is then deduplicated to generate a search result list. The basic constraint set, search execution configuration, and search result list are written into the search item record.

Images