A machine learning-based intellectual property precise retrieval analysis method

Abstract

The application discloses a kind of based on machine learning's intellectual property precision retrieval analysis method, it is related to artificial intelligence and information retrieval technical field, including: generating semantic vector and dimension spectrum vector to document library literature and query literature respectively;Recall candidate literature based on semantic vector and calculate semantic similarity;Based on dimension spectrum vector, calculate dimension spectrum distance, and determine the intensity of deviation of dimension semantics in combination with semantic similarity;Based on the statistical distribution of deviation intensity, calculate inhibition coefficient, calculate ranking score in combination with semantic similarity and deviation intensity and output result.The application can improve the precision and technical matching of retrieval result.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and information retrieval technology, and in particular to a method for precise retrieval and analysis of intellectual property rights based on machine learning. Background Technology

[0002] In the field of intellectual property analysis and examination, facing a massive amount of patent literature, standard documents, and scientific journals, the key to assessing technological innovation, mitigating infringement risks, and establishing research and development directions lies in quickly and accurately extracting existing technologies highly relevant to the subject of analysis from this vast amount of data. With the exponential growth in the number of technical documents, traditional retrieval methods are facing enormous challenges, especially in parameter-sensitive fields such as precision manufacturing, materials science, and chemical processes. The core differences in technical solutions often lie not only in the logical structure of the textual description but also in the specific range of physical parameters and engineering specifications, such as temperature ranges, force magnitudes, and size levels. Therefore, balancing the understanding of textual meaning with the comparison of numerical physical properties during the retrieval process has become a crucial issue for improving retrieval accuracy.

[0003] Existing retrieval technologies are mainly divided into two categories: keyword-based Boolean retrieval and deep learning-based semantic retrieval. However, both have significant limitations in addressing the aforementioned problems. Keyword-based retrieval relies on manually constructing complex search queries, making it difficult to exhaustively list synonyms and understand context. Current mainstream deep learning-based semantic retrieval methods typically convert text into high-dimensional vectors for similarity calculation. While this approach can capture the literal meaning and contextual association of text well, it often ignores the physical constraints implied by numerical values ​​and units of measurement within the text. For example, when describing heating devices, a semantic model might determine that the metallurgical field heating to 2000 degrees Celsius and the biological culture field heating to 40 degrees Celsius have extremely high textual similarity because they are highly similar in linguistic structure. However, in actual engineering applications, these are completely different technical branches. Because existing technologies lack effective mechanisms for extracting and comparing physical dimension information in documents, retrieval systems easily recall a large number of semantically similar but technically unrelated documents, severely reducing the accuracy and reference value of the retrieval results. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies that rely on textual semantic similarity for retrieval, ignoring the specific constraints of physical parameters and engineering specifications in technical documents. This results in the inability to effectively distinguish technical solutions that have similar textual descriptions but differ greatly in actual engineering scale, operating environment, or physical dimensions, leading to low accuracy and poor technical matching of retrieval results. Therefore, this invention proposes a machine learning-based method for precise intellectual property retrieval and analysis.

[0005] To address the problems existing in the prior art, the present invention adopts the following technical solution:

[0006] A machine learning-based method for precise intellectual property retrieval and analysis includes:

[0007] S1. Perform semantic vectorization and numerical unit extraction on each document in the intellectual property document library to generate semantic vectors and dimensional spectrum vectors respectively, and construct a semantic index.

[0008] S2. Obtain the document to be queried, and generate the query semantic vector and query dimension spectrum vector of the document;

[0009] S3. Based on the query semantic vector, perform a nearest neighbor search in the semantic index to recall a set of candidate documents, and calculate the semantic similarity between each candidate document in the set and the query document.

[0010] S4. Calculate the dimensional spectral distance based on the query dimensional spectral vector and the dimensional spectral vector of each candidate document in the candidate document set, and determine the dimensional semantic deviation intensity by semantic similarity and dimensional spectral distance.

[0011] S5. Calculate the inhibition coefficient based on the dimensional semantic deviation strength of the candidate document set. Calculate the ranking score of each candidate document based on semantic similarity, dimensional semantic deviation strength and inhibition coefficient. Sort the candidate document set according to the ranking score and output the structured ranking result.

[0012] Preferably, each document in the intellectual property document database undergoes semantic vectorization and numerical unit extraction to generate semantic vectors and dimensional spectrum vectors, including:

[0013] The text sequence of each document in the intellectual property document library is obtained, the text sequence is input into a machine learning vectorizer to obtain the corresponding semantic vector, and a semantic index is constructed based on the semantic vector;

[0014] The text sequence is parsed to obtain numerical unit entries, each numerical unit entry including a numerical value and a unit identifier corresponding to the numerical value;

[0015] Map the unit identifiers in the numerical unit entries to unit dimension exponent vectors;

[0016] Calculate the weight of the numerical unit entry based on the value in each entry;

[0017] The dimensional index vectors of each document are weighted and summed according to the weights of the numerical unit entries, and the weighted summation results are normalized to obtain the dimensional spectrum vector of each document.

[0018] Preferably, the unit dimension index vector is a seven-dimensional vector, with the seven dimensions corresponding to the dimension indices of length, mass, time, current, thermodynamic temperature, amount of substance, and luminous intensity, respectively.

[0019] Preferably, calculating the semantic similarity between each candidate document in the candidate document set and the query document includes:

[0020] The semantic vector of the query and the semantic vector of each candidate document are normalized to obtain the normalized semantic vector of the query and the normalized semantic vector of the candidates.

[0021] The semantic similarity is obtained by performing a vector inner product operation on the query normalized semantic vector and the candidate normalized semantic vector.

[0022] Preferably, the dimensional spectral distance is calculated based on the query dimensional spectral vector and the dimensional spectral vectors of each candidate document in the candidate document set, and the dimensional semantic deviation strength is determined by semantic similarity and dimensional spectral distance, including:

[0023] Calculate the Euclidean distance between the query dimensional spectrum vector and the dimensional spectrum vector of the candidate documents, and use the Euclidean distance as the dimensional spectrum distance.

[0024] The dimensional semantic deviation strength is obtained by multiplying the semantic similarity by the dimensional spectral distance.

[0025] Preferably, the suppression coefficient is calculated based on the dimensional semantic deviation strength of the candidate document set, including:

[0026] Calculate the median of the semantic deviation intensity of all dimensions within the candidate document set;

[0027] Calculate the absolute difference between the semantic deviation intensity of each dimension and the median, and determine the median of the absolute difference as the median absolute deviation;

[0028] The reciprocal of the sum of the median of the semantic deviation intensity of all dimensions and the absolute deviation of the median is used as the suppression coefficient.

[0029] Preferably, the ranking score of each candidate document is calculated based on semantic similarity, dimensional semantic deviation strength, and suppression coefficient, including:

[0030] Calculate the product of the suppression coefficient and the dimensional semantic deviation intensity, and take the negative of the product as the exponent to calculate the value of the exponential function with the natural constant as the base and the exponent as the power.

[0031] The ranking score of the candidate documents is obtained by multiplying the semantic similarity by the value of the exponential function.

[0032] Preferably, the structured ranking result includes document identifier, ranking score, semantic similarity, dimensional spectral distance, and dimensional semantic deviation intensity.

[0033] Compared with the prior art, the beneficial effects of the present invention are:

[0034] 1. This invention effectively solves the problem that existing technologies cannot distinguish essential differences in technical solutions due to neglecting physical parameters and engineering specifications by constructing a seven-dimensional dimensional spectral vector based on the International System of Units (SI). This invention not only generates semantic vectors representing the meaning of text, but also innovatively extracts numerical unit entries from the literature, mapping them to dimensional index vectors corresponding to length, mass, time, current, thermodynamic temperature, amount of substance, and luminescence intensity, and generates dimensional spectral vectors based on numerical weight normalization. This dual-channel feature extraction mechanism enables the retrieval system to establish an additional computable physical dimensional space in addition to the high-dimensional semantic space, thereby accurately identifying and distinguishing technical solutions with similar textual descriptions but drastically different actual engineering application scenarios, significantly improving the matching degree and reference value of retrieval results in terms of physical attributes.

[0035] 2. In this invention, by introducing a ranking algorithm based on dimensional semantic deviation strength and an adaptive suppression coefficient, precise suppression and dynamic weight reduction of interfering documents that are semantically similar but dimensionally incompatible are achieved. The dimensional semantic deviation strength is constructed by multiplying semantic similarity and dimensional spectral distance, and the suppression coefficient is calculated based on the statistical characteristics of the candidate set. Then, an exponential function is used to nonlinearly correct the ranking score. This mechanism can automatically identify semantic trap documents that have high keyword matching but huge differences in physical dimensions, and significantly reduce their ranking weight through an exponential decay factor. At the same time, the adaptive suppression coefficient ensures that the algorithm remains robust in candidate sets of different sizes and distributions. Finally, a structured ranking result that takes into account both text relevance and technical feasibility is output, which greatly improves the accuracy and efficiency of intellectual property retrieval. Attached Figure Description

[0036] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings:

[0037] Figure 1 This is a flowchart illustrating a machine learning-based method for precise intellectual property retrieval and analysis according to the present invention.

[0038] Figure 2 This is a schematic diagram of the generation of dimensional spectral vectors according to the present invention. Detailed Implementation

[0039] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0040] Example: This example provides a machine learning-based method for precise intellectual property retrieval and analysis. See [link to example]. Figure 1 Specifically, including:

[0041] S1. Perform semantic vectorization and numerical unit extraction on each document in the intellectual property document library to generate semantic vectors and dimensional spectrum vectors respectively, and construct a semantic index.

[0042] S2. Obtain the document to be queried, and generate the query semantic vector and query dimension spectrum vector of the document;

[0043] S3. Based on the query semantic vector, perform a nearest neighbor search in the semantic index to recall a set of candidate documents, and calculate the semantic similarity between each candidate document in the set and the query document.

[0044] S4. Calculate the dimensional spectral distance based on the query dimensional spectral vector and the dimensional spectral vector of each candidate document in the candidate document set, and determine the dimensional semantic deviation intensity by semantic similarity and dimensional spectral distance.

[0045] S5. Calculate the inhibition coefficient based on the dimensional semantic deviation strength of the candidate document set. Calculate the ranking score of each candidate document based on semantic similarity, dimensional semantic deviation strength and inhibition coefficient. Sort the candidate document set according to the ranking score and output the structured ranking result.

[0046] In embodiments of the present invention, each document in the intellectual property document library undergoes semantic vectorization and numerical unit extraction to generate semantic vectors and dimensional spectrum vectors, respectively, and a semantic index is constructed, including:

[0047] The text sequence of each document in the intellectual property document library is obtained, the text sequence is input into a machine learning vectorizer to obtain the corresponding semantic vector, and a semantic index is constructed based on the semantic vector;

[0048] Specifically, to ensure that the text data input to the machine learning vectorizer accurately represents the core technical information of the documents, key text fragments of each document need to be extracted from the intellectual property document library to form a text sequence. The intellectual property document library refers to a database storing a massive amount of intellectual property-related documents such as patents, trademarks, and copyrights. The text sequence specifically selects one or more combinations of the core paragraphs of the document's title, abstract, claims, and specification. The purpose of selecting these text fragments is to ensure they centrally carry the core information of the document's technical solution, scope of protection, and technical features, providing a high-quality data foundation for the accurate generation of subsequent semantic vectors. After obtaining the text sequence, preprocessing is required. Preprocessing includes removing invalid characters such as special symbols, meaningless spaces, and page number identifiers; standardizing text formatting, such as converting full-width characters to half-width characters; standardizing capitalization; and standardizing technical terminology, such as unifying different expressions of the same technical feature into standardized terminology, ensuring the consistency and standardization of the text data.

[0049] Specifically, the preprocessed text sequence is input into a pre-defined machine learning vectorizer to generate corresponding semantic vectors. This machine learning vectorizer is a text vectorization model built upon a pre-trained language model, specifically the Sentence-BERT model (a sentence vector generation model optimized based on the BERT pre-trained model). This model effectively captures the deep semantic information of the text and adapts to the semantic representation needs of complex technical terms and long texts in intellectual property literature. The specific implementation process is as follows: First, the Sentence-BERT model parameters are initialized using general domain pre-trained weights. Then, the model is fine-tuned using intellectual property literature corpus labeled with technical semantic associations. During fine-tuning, the model parameters are optimized using a contrastive learning loss function, enabling the model to learn the specific semantic features of intellectual property texts. After fine-tuning, the model is input into the preprocessed text sequence, and the pooling layer (using an average pooling strategy) outputs a 768-dimensional real-valued vector. This real-valued vector is the semantic vector that represents the semantic information of the text. The purpose of generating semantic vectors is to transform technical information in text form into a vector form that can be directly computed by a computer, providing data support for subsequent nearest neighbor retrieval. A semantic index is constructed based on the generated semantic vectors of all documents. The semantic index refers to a data structure used for efficient storage and retrieval of semantic vectors. Specifically, it adopts the Hierarchical Navigable Small Worlds (HNSW) index structure in the Approximate Nearest Neighbor (ANN) index. The construction process is as follows: the semantic vectors are input into the HNSW index construction module, the index level parameter is set to 16 and the number of candidate nodes is 200, and a multi-level navigation graph structure is generated through the index construction algorithm. This structure can realize fast nearest neighbor matching of semantic vectors, improve the efficiency of semantic nearest neighbor retrieval in subsequent query processes, and enable the system to quickly recall a set of candidate documents with semantic similarity to the query document in a massive amount of intellectual property documents, meeting the real-time requirements of accurate retrieval.

[0050] The text sequence is parsed to obtain numerical unit entries, each numerical unit entry including a numerical value and a unit identifier corresponding to the numerical value;

[0051] Specifically, to accurately extract key elements carrying physical scale information from text sequences, targeted parsing of the preprocessed text sequences is required. Here, the parsing employs a named entity recognition model based on a fusion of rules and deep learning. This model automatically identifies numerical values ​​and corresponding unit identifiers related to technical features in the text by combining a pre-set technical terminology dictionary with a BERT fine-tuning model (covering common numerical association expressions in intellectual property literature, such as thickness, pressure, frequency, etc.), forming numerical unit entries. Among them, numerical unit entries refer to the smallest data unit that can characterize the physical properties of technical parameters, while unit identifiers refer to the physical quantity measurement symbols or names associated with the numerical values ​​(such as mm, ℃, etc.). Through this parsing process, the core physical quantity information that determines technical comparability can be filtered out from complex text information, laying a data foundation for subsequent dimensional analysis.

[0052] Map the unit identifiers in the numerical unit entries to unit dimension exponent vectors;

[0053] Specifically, after obtaining the numerical unit entries, the unit identifiers need to be mapped to unit dimension index vectors. These unit dimension index vectors are seven-dimensional vectors constructed based on the seven fundamental dimensions of the International System of Units (SI): length, mass, time, electric current, thermodynamic temperature, amount of substance, and luminous intensity. The numerical value of each dimension corresponds to the exponent of the unit identifier on that fundamental dimension (e.g., the unit mm corresponds to a length dimension exponent of 1, and the exponents of the other dimensions are 0). The mapping process is achieved by constructing a pre-defined unit dimension index mapping table. This table covers all units commonly used in intellectual property literature, as well as specialized units, unifying different forms of units into a standardized dimensional representation, ensuring dimensional comparability between different documents. The unit dimension index vectors are seven-dimensional because the dimensional basis is selected according to the fundamental dimensions of the SI. A commonly used and reproducible set of independent basis dimensions includes: length, mass, time, electric current, thermodynamic temperature, amount of substance, and luminous intensity. Mapping units to exponential vectors of these seven basic dimensions is essentially performing a fixed-dimensional dimensional decomposition, enabling any engineering unit or derived unit to be represented as a power combination of these seven basic dimensions. This allows for the direct calculation of dimensional differences and dimensional spectral distances between different units within the same coordinate system. Furthermore, the independence of these seven dimensions covers frequently occurring physical quantities in intellectual property literature, such as size, mass, time constant, electrical parameters, temperature conditions, stoichiometry, and optical intensity. This ensures a unified representation space for dimensional spectral vectors across cross-disciplinary literature and guarantees the comparability and interpretability of dimensional spectral vectors from different documents.

[0054] Calculate the weight of the numerical unit entry based on the value in each entry;

[0055] The dimensional index vectors of units are weighted and summed according to the weights of the numerical unit entries, and the weighted summation results are normalized to obtain the dimensional spectrum vector of each document.

[0056] Specifically, the weight of each item is calculated based on the numerical value in each unit item. The item value is converted to a real number, and its absolute value is taken. The weight is obtained by adding one to the absolute value and then taking the logarithm. The addition of one ensures that items with zero values ​​can still participate in subsequent calculations, and the logarithm is used to suppress the dominant effect of extremely large values ​​on the aggregation result and to keep the weight monotonically changing with the numerical amplitude. This makes the dimensional spectrum vector more sensitive to typical process parameter ranges and more robust to abnormal extreme values. When the unit dimensional index vector is weighted, summed, and normalized based on the weights, the unit dimensional index vector of each item is multiplied by the corresponding weight and accumulated in seven dimensions to obtain a weighted sum vector. The weighted sum vector is divided by the sum of all weights to obtain the normalized result. The normalized result is defined as the dimensional spectrum vector of each document, ensuring that the dimensional spectrum vector is comparable between documents with different numbers of items and different text lengths, and can reflect the strength of the difference in physical dimensional composition between the query document and the candidate document when calculating the dimensional spectrum distance.

[0057] In an embodiment of the present invention, obtaining the query document to be queried and generating the query semantic vector and query dimension spectrum vector of the query document specifically involves: when obtaining the query document to be queried, the object to be retrieved is taken as the query document and processed query input data is formed in the electronic computing device. The query document can be text submitted by the user, uploaded document files, or document content passed in by an external interface. The system assigns a query identifier to the query document and extracts text content with the same field scope as the intellectual property document database to generate a query text sequence. The query text sequence is a character sequence obtained by concatenating the title, abstract, and claims text of the query document in sequence and retaining field boundary separators to ensure that the input form of subsequent semantic vectorization processing is consistent with the document database; a query is generated based on the query text sequence. When querying the semantic vector, the query text sequence is input into a machine learning vectorizer that performs the same semantic vectorization processing as the document library, and the output is the query semantic vector. When generating the query dimensional spectrum vector based on the query text sequence, numerical units are extracted from the query text sequence to obtain a set of query numerical unit entries. The unit identifiers are mapped to unit dimensional index vectors through a unit dimensional index mapping table. Then, the corresponding weights are calculated according to the values ​​in the query numerical unit entries. The unit dimensional index vectors are weighted and summed according to the weights and normalized to obtain the query dimensional spectrum vector. This ensures that the query semantic vector and the query dimensional spectrum vector are consistent with the semantic vector and dimensional spectrum vector in the intellectual property document library in terms of generation rules, and guarantees that the subsequent calculation of dimensional spectrum distance and dimensional semantic deviation strength is comparable and reproducible.

[0058] In embodiments of the present invention, a nearest neighbor retrieval is performed in the semantic index based on the query semantic vector to recall a set of candidate documents, and the semantic similarity between each candidate document in the candidate document set and the query document is calculated, including:

[0059] Based on the query semantic vector, a nearest neighbor search is performed in the semantic index to recall a set of candidate documents;

[0060] The semantic vector of the query and the semantic vector of each candidate document are normalized to obtain the normalized semantic vector of the query and the normalized semantic vector of the candidates.

[0061] The semantic similarity is obtained by performing a vector dot product operation on the query normalized semantic vector and the candidate normalized semantic vector;

[0062] Specifically, when performing nearest neighbor retrieval in the semantic index based on the query semantic vector to recall the candidate document set, the query semantic vector generated from the queried documents is denoted as... The semantic vectors and document identifiers of each document in the document library are read from the semantic index, which is a data structure that uses the document semantic vectors as index keys and supports nearest neighbor retrieval. To ensure the stability of subsequent similarity measurements and consistency with vector dot product calculations, vector normalization is performed on the query semantic vectors to obtain query normalized semantic vectors. ,in And for each document semantic vector stored in the semantic index Candidate normalized semantic vectors are obtained in the same way. The L2 norm used in normalization is employed to eliminate vector magnitude differences caused by text length and redundancy, ensuring that vector comparisons primarily reflect directional consistency; in the nearest neighbor retrieval process, As input for retrieval, the document library vector set is retrieved using a similarity metric consistent with semantic indexing, and the top K most similar document identifiers are returned as a candidate document set. Where K is the preset recall number, the preset recall number is set based on the fact that semantic retrieval, as a candidate recall stage, needs to strike a balance between coverage and computational cost. Furthermore, when the intellectual property literature database reaches the millions level, returning the top thousand candidate sets can significantly reduce the overall complexity of subsequent fine-grained ranking calculations while maintaining coverage of highly similar documents. Therefore, K can be set to one thousand. After obtaining the candidate document set... Then, the dot product of the query normalized semantic vector and each candidate normalized semantic vector is calculated as the semantic similarity. Candidate documents... The semantic similarity is denoted as In the formula This indicates that the query normalized semantic vector is at the th... The components of the dimension, The candidate normalized semantic vector is represented in the th... The reason for using the vector inner product for the components of the dimension is that after the vectors are normalized, the inner product of the two is equivalent to the cosine similarity, which can directly represent the degree of closeness of the angle between the two texts in the high-dimensional semantic space and reduce the computational complexity.

[0063] The above processing enables the recall of candidate document sets to be based on semantic similarity, and provides reproducible semantic similarity input for subsequent dimensional spectral distance calculation and dimensional semantic deviation strength construction, thereby explicitly representing the degree of textual semantic similarity of candidate documents in numerical form.

[0064] In an embodiment of the present invention, the dimensional spectral distance is calculated based on the query dimensional spectral vector and the dimensional spectral vectors of each candidate document in the candidate document set, and the dimensional semantic deviation strength is determined by semantic similarity and dimensional spectral distance, including:

[0065] Calculate the Euclidean distance between the query dimensional spectrum vector and the dimensional spectrum vector of the candidate documents, and use the Euclidean distance as the dimensional spectrum distance.

[0066] The dimensional semantic deviation strength is obtained by multiplying the semantic similarity and the dimensional spectral distance.

[0067] Specifically, when calculating the dimensional distance between the query dimensional spectrum vector and the dimensional spectrum vector of the candidate document, the query dimensional spectrum vector of the query document is denoted as... Let the dimensional spectral vector corresponding to any candidate document in the candidate document set be denoted as ,in and All are seven-dimensional real vectors, with the seven dimensions corresponding to the dimensional index aggregation results of length, mass, time, current, thermodynamic temperature, amount of substance, and luminescence intensity, respectively. The dimensional spectral vectors are obtained from numerical unit entries through unit mapping, weighting, and normalization; therefore, the components of each dimension are comparable on the same scale. Based on the above... and Calculate Euclidean distance as dimensional spectral distance Its calculation formula is In the formula, Indicate candidate documents Dimensional spectral distance, This indicates the component of the query dimension spectral vector in the k-th dimension. The component of the candidate dimensional spectral vector in the k-th dimension is represented. The Euclidean distance synthesizes the differences across the seven dimensional dimensions using the same metric system, characterizing the overall deviation between the query and candidate documents in terms of their physical dimensional composition. Furthermore, this distance satisfies non-negativity and symmetry, facilitating its direct use as a deviation measure in subsequent calculations; it is also crucial for obtaining semantic similarity. Dimensional spectral distance Then, the product of the two is defined as the dimensional semantic deviation strength. Its calculation formula is In the formula To query the dot product of the normalized semantic vector and the candidate normalized semantic vector, which is used to characterize the semantic similarity of the text, The product form is used to characterize the degree of difference in dimensional spectrum because it only produces a large deviation intensity when the semantic proximity is high and the dimensional spectrum difference is large. This highlights the potential dimensional aliasing in semantic nearest neighbor retrieval, attracting candidates and suppressing results that are semantically close but dimensionally incompatible. When the candidate and the query are not semantically close... A smaller value naturally reduces the deviation intensity, avoiding over-amplification of low-relevance candidates, while when the candidate and query are close in dimensionality... A smaller deviation reduces the intensity of the deviation, making the system more inclined to retain candidate documents that simultaneously satisfy semantic similarity and similar dimensional spectra, thereby providing an interpretable deviation input for subsequent suppression coefficient calculation and ranking score calculation.

[0068] In embodiments of the present invention, a suppression coefficient is calculated based on the dimensional semantic deviation strength of the candidate document set; a ranking score is calculated for each candidate document based on semantic similarity, dimensional semantic deviation strength, and suppression coefficient; the candidate document set is ranked according to the ranking scores; and a structured ranking result is output, including:

[0069] Calculate the median of the semantic deviation intensity of all dimensions within the candidate document set;

[0070] Calculate the absolute difference between the semantic deviation intensity of each dimension and the median, and determine the median of the absolute difference as the median absolute deviation;

[0071] The suppression coefficient is calculated based on the median of the semantic deviation intensity across all dimensions and the median absolute deviation.

[0072] Specifically, when calculating the suppression coefficient based on the candidate document set, the dimensional semantic deviation intensity corresponding to each candidate document in the candidate document set is regarded as a sortable real number sequence. Median statistics are then performed on this real number sequence to obtain the median of all dimensional semantic deviation intensities. The median is calculated by sorting the real number sequence from smallest to largest and taking the value in the middle of the sequence as the median. When the number of candidate documents is even, the arithmetic mean of the two middle values ​​is taken as the median. After obtaining the median, the absolute difference between each dimensional semantic deviation intensity in the candidate document set and the median is calculated. All the obtained absolute differences are formed into an absolute difference sequence and sorted from smallest to largest. The median of the absolute difference sequence is taken as the median absolute deviation. The median absolute deviation is used to characterize the dispersion of the dimensional semantic deviation intensity in the candidate set and is insensitive to a small number of extreme deviation values. The suppression coefficient is calculated based on the median and the median absolute deviation. The formula for calculating the suppression coefficient is:

[0073]

[0074] In the formula, This is the suppression coefficient, used to adjust the attenuation of candidate document scores by the intensity of dimensional semantic deviation in subsequent ranking score calculations; The median of the semantic deviation intensity across all dimensions is used to characterize the central position of the deviation level of the candidate set; The median absolute deviation is used to characterize the dispersion of the candidate set's deviation level. The suppression coefficient calculated above can adaptively change with the overall deviation level and dispersion of the candidate set. When the candidate set's deviation level is high, the suppression coefficient decreases to avoid excessive decay, while when the candidate set's deviation level is low, the suppression coefficient increases to enhance the suppression effect on a small number of highly deviation candidates. This provides stable and interpretable parameters for the subsequent ranking output based on the ranking score.

[0075] Calculate the product of the suppression coefficient and the dimensional semantic deviation intensity, and take the negative of the product as the exponent to calculate the value of the exponential function with the natural constant as the base and the exponent as the power.

[0076] The ranking score of the candidate documents is obtained by multiplying the semantic similarity by the value of the exponential function.

[0077] Specifically, when calculating the ranking score of candidate documents for the candidate document set, the dimensional semantic deviation strength and suppression coefficient corresponding to each candidate document are used as attenuation control quantities in the score calculation. The suppression coefficient is an adaptive parameter obtained based on the statistical analysis of the dimensional semantic deviation strength of the candidate document set, used to adjust the attenuation magnitude of the deviation strength on the ranking score. For any candidate document, the suppression coefficient is multiplied by the dimensional semantic deviation strength of that candidate document to obtain a product value. The negative of this product value is used as the exponent to calculate an exponential function with the natural constant as the base, yielding an exponential attenuation factor. This exponential attenuation factor is used to transform the monotonic relationship of stronger attenuation with greater deviation strength into a continuously differentiable and numerically stable attenuation curve. The formula for calculating the exponential function value is as follows:

[0078]

[0079] In the formula, The exponential decay factor corresponding to the candidate literature; The suppression coefficient is used to uniformly adjust the amplification or reduction of the semantic deviation intensity in the exponential domain, so that the attenuation amplitude under different candidate set sizes and different deviation levels remains controllable. The dimensional semantic deviation strength corresponding to the candidate document is obtained by multiplying the semantic similarity between the candidate document and the query document by the dimensional spectral distance. It is used to characterize the joint amplification effect of semantic proximity and dimensional deviation. The reason for using an exponential function with the natural constant as the base is that the exponential function rapidly reduces the output value in a monotonically decreasing manner when the independent variable increases. It can significantly reduce the score when the deviation strength is large to suppress candidates that are semantically close but dimensionally incompatible. At the same time, it keeps the output value close to one when the deviation strength is small to reduce the disturbance to high-confidence candidates. After obtaining the exponential decay factor, the semantic similarity is multiplied by the exponential decay factor to obtain the ranking score of the candidate document. The ranking score reflects both the proximity between the candidate and the query in the semantic space and the deviation penalty in the dimensional spectral space. Therefore, the ranking process prioritizes the output of candidate documents that are semantically similar and have a small dimensional spectral distance, and reduces the ranking position of candidate documents with large dimensional semantic deviation strength.

[0080] Specifically, when sorting the candidate document set according to the ranking score and outputting a structured ranking result, a candidate record is created for each candidate document in the candidate document set. The candidate record includes at least a document identifier and the ranking score, semantic similarity, dimensional distance, and dimensional semantic deviation strength corresponding to that document identifier. The document identifier is used to uniquely locate the candidate document, the ranking score is used as the ranking key, and the semantic similarity, dimensional distance, and dimensional semantic deviation strength are used as retrieval analysis and interpretation fields. After assembling the candidate records, all candidate records are sorted in descending order using the ranking score as the primary ranking key to obtain a ranking sequence. When there are candidate records with the same ranking score, semantic similarity is used as the secondary ranking key. The system sorts records in descending order to prioritize candidates with semantic similarity. When candidate records with the same semantic similarity appear, it re-sorts them in ascending order using the dimensional distance as the sorting key to prioritize candidates with closer dimensional spectra, thus forming a deterministic sorting output. After sorting, the system extracts the first few candidate records from the sorted sequence as the sorting result according to the output quantity parameter configured in the system. The sorting result is then encapsulated into a structured data object for output. The structured data object adopts a record structure with fixed fields. Each record contains a document identifier, sorting score, semantic similarity, dimensional distance, and dimensional semantic deviation strength. It may also optionally include basic metadata fields such as title and publication date associated with the document identifier for display on the output end.

[0081] This invention also provides a computer storage medium storing a computer program, which, when executed by a processor, implements the aforementioned machine learning-based method for precise intellectual property retrieval and analysis.

[0082] like Figure 2 As shown, Figure 2 This is a schematic diagram of dimensional spectrum vector generation. Taking a text sequence of intellectual property documents as input, the text sequence is processed to extract numerical units, resulting in a set of numerical unit entries. Each numerical unit entry contains a numerical value and a unit identifier. For the unit identifier, a unit mapping table is used to obtain a corresponding unit dimensional index vector. This unit dimensional index vector is a seven-dimensional vector, with dimensions corresponding to length, mass, time, current, thermodynamic temperature, amount of substance, and luminous intensity, respectively. For each numerical value, a weight corresponding to that value is calculated, and this weight, along with the unit dimensional index vector, is input into a weighted summation process to accumulate the dimensional contributions of each entry and obtain a weighted sum. After obtaining the weighted sum, it is normalized, and the normalized result is output as the dimensional spectrum vector of the document. This transforms the scattered numerical and unit information in the document into a vectorized representation that can be used for subsequent distance calculation and retrieval analysis.

[0083] 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 machine learning-based method for precise intellectual property retrieval and analysis, characterized in that, Includes the following steps: S1. Perform semantic vectorization and numerical unit extraction on each document in the intellectual property document library to generate semantic vectors and dimensional spectrum vectors respectively, and construct a semantic index. The specific steps for semantic vectorization and numerical unit extraction of each document in the intellectual property document database to generate semantic vectors and dimensional spectrum vectors are as follows: The text sequence of each document in the intellectual property document library is obtained, the text sequence is input into a machine learning vectorizer to obtain the corresponding semantic vector, and a semantic index is constructed based on the semantic vector; The text sequence is parsed to obtain numerical unit entries, each numerical unit entry including a numerical value and a unit identifier corresponding to the numerical value; Map the unit identifiers in the numerical unit entries to unit dimension index vectors; Calculate the weight of the numerical unit entry based on the value in each entry; The dimensional index vectors of units are weighted and summed according to the weights of the numerical unit entries, and the weighted summation results are normalized to obtain the dimensional spectrum vector of each document. S2. Obtain the document to be queried, and generate the query semantic vector and query dimension spectrum vector of the document; S3. Based on the query semantic vector, perform a nearest neighbor search in the semantic index to recall a set of candidate documents, and calculate the semantic similarity between each candidate document in the set and the query document. S4. Calculate the dimensional spectral distance based on the query dimensional spectral vector and the dimensional spectral vector of each candidate document in the candidate document set, and determine the dimensional semantic deviation intensity by semantic similarity and dimensional spectral distance. The specific steps for calculating the dimensional spectral distance based on the query dimensional spectral vector and the dimensional spectral vectors of each candidate document in the candidate document set, and determining the dimensional semantic deviation strength based on semantic similarity and dimensional spectral distance, are as follows: Calculate the Euclidean distance between the query dimensional spectrum vector and the dimensional spectrum vector of the candidate documents, and use the Euclidean distance as the dimensional spectrum distance. The dimensional semantic deviation strength is obtained by multiplying the semantic similarity and the dimensional spectral distance. S5. Calculate the inhibition coefficient based on the dimensional semantic deviation strength of the candidate document set. Calculate the ranking score of each candidate document based on semantic similarity, dimensional semantic deviation strength and inhibition coefficient. Sort the candidate document set according to the ranking score and output the structured ranking result. The specific steps for calculating the suppression coefficient based on the dimensional semantic deviation strength of the candidate document set, and calculating the ranking score of each candidate document based on semantic similarity, dimensional semantic deviation strength, and suppression coefficient are as follows: Calculate the median of the semantic deviation intensity of all dimensions within the candidate document set; Calculate the absolute difference between the semantic deviation intensity of each dimension and the median, and determine the median of the absolute difference as the median absolute deviation; The reciprocal of the sum of the median of the semantic deviation intensity of all dimensions and the median absolute deviation is used as the suppression coefficient; Calculate the product of the suppression coefficient and the dimensional semantic deviation intensity, and take the negative of the product as the exponent to calculate the value of the exponential function with the natural constant as the base and the exponent as the power. The ranking score of the candidate documents is obtained by multiplying the semantic similarity by the value of the exponential function.

2. The method for precise intellectual property retrieval and analysis based on machine learning according to claim 1, characterized in that, The unit dimension index vector is a seven-dimensional vector, with the seven dimensions corresponding to the dimension indices of length, mass, time, current, thermodynamic temperature, amount of substance, and luminous intensity, respectively.

3. The method for precise intellectual property retrieval and analysis based on machine learning according to claim 1, characterized in that, Calculate the semantic similarity between each candidate document in the candidate document set and the query document, including: The semantic vector of the query and the semantic vector of each candidate document are normalized to obtain the normalized semantic vector of the query and the normalized semantic vector of the candidates. The semantic similarity is obtained by performing a vector inner product operation on the query normalized semantic vector and the candidate normalized semantic vector.

4. The method for precise intellectual property retrieval and analysis based on machine learning according to claim 1, characterized in that, The structured ranking results include document identifiers, ranking scores, semantic similarity, dimensional spectral distance, and dimensional semantic deviation strength.

5. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the machine learning-based precise intellectual property retrieval and analysis method as described in any one of claims 1 to 4.

Images