A multi-dimensional feature extraction and intelligent recommendation system of scientific and technological achievements based on fusion of text semantic mining

Abstract

The application discloses a kind of fusion text semantic mining's scientific and technological achievement multidimensional feature extraction and intelligent recommendation system, this system includes: data acquisition analysis module, space domain construction module, knowledge graph construction module and intelligent recommendation module.System obtains multi-source heterogeneous industrial text data, extracts the basic attribute of three kinds of nodes, key quantitative parameter and time stamp;Build technical performance boundary domain, subject capacity coverage domain and industry demand distribution domain;Analysis space position relationship establishes connection, generates timing sensitive graph in combination with time decay factor.The system can respond to query instruction to execute node matching, path analysis or technical quantification mode, respectively realize achievement positioning based on space domain, technology prediction based on trajectory fitting and technology quantification based on volume and stability.The application realizes multidimensional technical feature depth mining and accurate recommendation.

Description

Technical Field

[0001] This invention relates to the fields of text semantic mining and intelligent recommendation, and in particular to a multi-dimensional feature extraction and intelligent recommendation system that integrates text semantic mining with scientific and technological achievements. Background Technology

[0002] With the advent of the big data era, massive amounts of multi-source and heterogeneous industrial text data are growing explosively. How to use computer systems to extract key technical features from this data and make intelligent recommendations is crucial for industrial planning and technological innovation.

[0003] However, existing technology recommendation and analysis systems have significant shortcomings. First, traditional systems often rely on keyword matching or simple statistical analysis modules, making it difficult to transform discrete textual information into quantitative features with spatial geometric attributes. This hinders the accurate construction of technological effectiveness boundaries and the scope of subject capabilities, resulting in insufficient feature extraction depth and low accuracy. Second, existing knowledge graph systems often neglect the impact of time on information value. The weights of connections are typically stored statically, lacking a dynamic adjustment mechanism that incorporates time decay, leading to poor timeliness of recommendation results and an inability to accurately reflect the activity level of technological iteration. Furthermore, existing recommendation systems are mostly limited to static retrieval and matching of known information, lacking the ability to fit and compute technological leap trajectories. They struggle to extrapolate and predict future technological evolution directions and critical nodes based on historical data, failing to provide forward-looking quantitative evidence for industrial decision-making.

[0004] Therefore, there is an urgent need to develop a feature extraction and intelligent recommendation system that can deeply integrate text semantic mining and has the capabilities of multi-dimensional spatial quantization, temporal dynamic perception, and trend prediction, in order to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the shortcomings of existing technologies, embodiments of the present invention provide a multi-dimensional feature extraction and intelligent recommendation system for scientific and technological achievements that integrates text semantic mining, comprising: Data acquisition and parsing module: used to acquire multi-source heterogeneous industrial text data, and use semantic analysis and entity extraction technology to extract the basic attributes, key quantitative parameters and timestamps of subject nodes, result nodes and industry nodes; Spatial domain construction module: used to construct geometric spatial domains for three types of nodes to map discrete nodes into multi-dimensional feature spaces; the geometric spatial domains include the technical effectiveness boundary domain of the achievement node, the main capability coverage domain of the main node, and the industry demand distribution domain of the industry node; Knowledge graph construction module: used to analyze the spatial relationships between different levels of geometric spatial domains to establish the connection edges between main nodes, result nodes and industry nodes and quantify the initial weights to construct the initial network topology graph; establish a multi-level dynamic weight adjustment mechanism, and combine the time decay factor to correct the connection edge weights in the initial network topology graph to generate a target industry knowledge graph with time sensitivity. The intelligent recommendation module is used to respond to query commands containing query information, perform feature parsing on the query information to generate target parameter vectors and / or target semantic vectors, perform graph reasoning and technology quantification in the target industry knowledge graph, and execute any of the following modes according to the type of query command: If the query command corresponds to the node matching mode: map the target parameter vector to the multi-dimensional feature space, locate the several related result nodes corresponding to the technical effectiveness boundary domain in which it falls, and identify several related subject nodes connected to the related result nodes based on the attribution connection between the result nodes and the subject nodes. If the query command corresponds to the path analysis mode: match the target subject node based on the target semantic vector, identify several target result nodes based on the attribution connection between the subject node and the result node, fit the temporal geometric center point of all target result nodes to generate the technology leap trajectory, calculate the intersection point of the technology leap trajectory and the boundary of the industry demand distribution domain corresponding to the target subject node, and linearly extrapolate along the tangent direction of the intersection point to generate the predicted result node and timestamp. If the query command corresponds to the technology quantification mode: the target subject node is determined based on the target semantic vector, and the technology quantification index is obtained by weighting the geometric volume of the subject capability coverage domain of the target subject node and the temporal stability index of the achievement nodes contained in the domain.

[0006] According to a preferred embodiment, the data acquisition and parsing module utilizes semantic analysis and entity extraction techniques to extract basic attributes, key quantitative parameters, and timestamps, including: The named entity recognition model extracts organizational entities as subject nodes, technology product or patent entities as achievement nodes, and industry domain entities as industry nodes from multi-source heterogeneous industrial text data. The names and descriptions of the main nodes, achievement nodes, and industry nodes are used as basic attributes. A pre-trained language model is used to vectorize these basic attributes to generate corresponding semantic vectors. Extract the semantic attribution relationship between the main nodes and the result nodes, and classify the result nodes under the corresponding industry node names according to the semantic category; By using regular expressions combined with dependency parsing, numerical technical indicators are extracted from the descriptive text of the result nodes, a standard parameter indicator library is established, semantic mapping technology is used to align the extracted numerical technical indicators to the corresponding categories in the standard parameter indicator library, and numerical normalization is performed. Correlation analysis and dimensionality reduction were performed on the normalized numerical technical indicators. The retained principal components were selected as key quantitative parameters to ensure that the dimensions of the key quantitative parameters were consistent with the orthogonal basis of the multidimensional feature space. Parse the release time, application time, or effective time in multi-source heterogeneous industrial text data, and convert them into standard UNIX timestamp format as the time attribute of the node.

[0007] According to a preferred embodiment, the spatial domain construction module constructs geometric spatial domains for three types of nodes, including: Construct a multidimensional feature space, where each dimension corresponds to a key quantization parameter; For each achievement node, its key quantitative parameters are mapped to coordinate points in a multidimensional feature space. A high-dimensional hypersphere is constructed with the coordinate point as the center and the preset technical error tolerance as the radius as the boundary domain of the technical effectiveness. For the main node, obtain the coordinates of all the result nodes associated with the main node, calculate the minimum convex hull containing the coordinates of all result nodes, and expand the geometric boundary of the minimum convex hull based on the technical error tolerance to construct the main capability coverage domain. For an industry node, the coordinates of all the result nodes of that industry node are aggregated, and its spatial probability density distribution is calculated. The spatial region with a probability density higher than a preset threshold is taken as the industry demand distribution domain.

[0008] According to a preferred embodiment, the knowledge graph construction module parses the spatial relationships between different levels of geometric spatial domains, establishes connecting edges, and quantifies initial weights, including: The spatial inclusion relationship between the technical effectiveness boundary domain of the result node and the main capability coverage domain of the main node is calculated. If the center coordinates of the result node are located inside or on the boundary of the main capability coverage domain of the main node, a belonging connection edge is established and the initial weight is set to a preset weight value. Calculate the overlap volume between the main capability coverage area of ​​the main node and the industry demand distribution area of ​​the industry node. If the overlap volume is greater than zero, establish an adaptive connection edge, and set the initial weight to be proportional to the ratio of the overlap volume to the total volume of the main capability coverage area. Calculate the Euclidean distance between the technical effectiveness boundary domains of different achievement nodes. If the distance is less than a preset threshold, establish similar connection edges with initial weights inversely proportional to the Euclidean distance.

[0009] According to a preferred embodiment, the weight dynamic adjustment mechanism includes: Calculate the time difference between the current system time and the node timestamp; The attenuation coefficient is calculated using a preset time decay model; the attenuation coefficient is negatively correlated with the time difference. The initial weights of the connecting edges are adjusted using a decay coefficient.

[0010] According to a preferred embodiment, in a node matching mode, locating the technical performance boundary domain into which a node falls includes: Calculate the geometric distance between the target parameter vector and the center of the technical effectiveness boundary domain of all result nodes in the multidimensional feature space; Determine whether the geometric distance is less than the radius of the corresponding technical performance boundary region: If the value is less than the radius, the corresponding result node is identified as an associated result node, and the matching confidence level is set to the preset confidence level. If all geometric distances are not less than the radius, then select the few result nodes with the closest geometric distance as associated result nodes, and calculate the matching confidence. The matching confidence decreases non-linearly with the increase of geometric distance.

[0011] According to a preferred embodiment, in the path analysis mode, fitting the technical transition trajectory includes: Obtain the target semantic vector, calculate the cosine similarity between the target semantic vector and the semantic vector of each subject node in the graph, and determine the subject node with the highest cosine similarity as the target subject node; Sort all target result nodes associated with the target subject node in chronological order of their timestamps. Set a sliding time window and calculate the geometric centroid of the result node coordinates within each time window; By using regression analysis algorithms to fit vector functions to the geometric centroids of all time windows, a position function with respect to time is obtained, which is the trajectory of technological leaps. The first derivative of the calculated technological leap trajectory is used as the speed vector of technological evolution, and the second derivative is used as the acceleration vector of technological evolution.

[0012] According to a preferred embodiment, in the path analysis mode, generating the predicted result node and timestamp includes: Construct implicit functions or parametric equations that describe the boundaries of the industry demand distribution domain; By combining the position function of the technology leap trajectory with the boundary equation of the industry demand distribution domain, the coordinates of the intersection point of the trajectory and the boundary are obtained. If there is no intersection point, the point on the technology leap trajectory that is closest to the boundary equation is used as the intersection point coordinates. The coordinates of this intersection point are determined as the critical technical point that meets industry needs; Obtain the time solution corresponding to the intersection coordinates, and use it as the timestamp to meet industry requirements; Calculate the tangent vector at the intersection point, extend it along the tangent direction by a preset step size, and obtain the coordinates of the predicted node.

[0013] According to a preferred embodiment, in the technology quantification mode, the calculation process of the technology quantification index includes: Obtain the target semantic vector, calculate the cosine similarity between the target semantic vector and the semantic vector of each subject node in the graph, and determine the subject node with the highest cosine similarity as the target subject node; Calculate the spatial volume of the target main node's main capability coverage area; Calculate the time interval variance of the timestamp sequence of all result nodes within the main capability coverage domain of the target main node, and construct a time series stability index reflecting the uniformity of the result release frequency based on the time interval variance. The smaller the variance, the higher the time series stability index. The spatial volume and the temporal stability index are assigned preset weight coefficients, and the technology quantification index is calculated by weighted summation.

[0014] The present invention has the following beneficial effects: 1. This invention maps discrete text data into a multi-dimensional feature space with spatial attributes, constructs a geometric technical performance and subject capability domain, breaks through the limitations of traditional keyword matching, realizes in-depth quantitative mining of technical features and industrial logic, and significantly improves the accuracy of feature extraction.

[0015] 2. This invention establishes a dynamic weight adjustment mechanism based on geometric spatial relationships and time decay factors, accurately linking technology supply and industry demand, ensuring that the knowledge graph can reflect the temporal changes of technology iteration in real time, and effectively enhancing the timeliness and reliability of the recommendation results.

[0016] 3. This invention provides multiple intelligent modes such as path analysis and technology quantification. It generates technology transition trajectories and predicts evolution directions by fitting the temporal geometric center. At the same time, based on the strength of spatial volume quantification technology, it provides a forward-looking scientific basis for industrial planning and technology decision-making. Attached Figure Description

[0017] Figure 1 This is a structural block diagram of a scientific and technological achievement multidimensional feature extraction and intelligent recommendation system that integrates text semantic mining, as an exemplary embodiment. Detailed Implementation

[0018] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0019] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The singular forms “a,” “the,” and “the” used in this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0020] It should be understood that although the terms first, second, third, etc., may be used in this invention to describe various information, this information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first information may also be referred to as second information without departing from the scope of this invention, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0021] See Figure 1 The present invention discloses a multi-dimensional feature extraction and intelligent recommendation system for scientific and technological achievements that integrates text semantic mining, comprising: Data acquisition and parsing module: used to acquire multi-source heterogeneous industrial text data, and use semantic analysis and entity extraction technology to extract the basic attributes, key quantitative parameters and timestamps of subject nodes, result nodes and industry nodes.

[0022] Preferably, the data acquisition and parsing module utilizes semantic analysis and entity extraction techniques to extract basic attributes, key quantitative parameters, and timestamps, including: The named entity recognition model extracts organizational entities as subject nodes, technology product or patent entities as achievement nodes, and industry domain entities as industry nodes from multi-source heterogeneous industrial text data. The names and descriptions of the main nodes, achievement nodes, and industry nodes are used as basic attributes. A pre-trained language model is used to vectorize these basic attributes to generate corresponding semantic vectors. Extract the semantic attribution relationship between the main nodes and the result nodes, and classify the result nodes under the corresponding industry node names according to the semantic category; By using regular expressions combined with dependency parsing, numerical technical indicators are extracted from the descriptive text of the result nodes, a standard parameter indicator library is established, semantic mapping technology is used to align the extracted numerical technical indicators to the corresponding categories in the standard parameter indicator library, and numerical normalization is performed. Correlation analysis and dimensionality reduction were performed on the normalized numerical technical indicators. The retained principal components were selected as key quantitative parameters to ensure that the dimensions of the key quantitative parameters were consistent with the orthogonal basis of the multidimensional feature space. Parse the release time, application time, or effective time in multi-source heterogeneous industrial text data, and convert them into standard UNIX timestamp format as the time attribute of the node.

[0023] In this embodiment, the multi-source heterogeneous industrial text data includes, but is not limited to, patent specifications, academic papers, industry analysis reports, and corporate news.

[0024] Key nodes: These refer to the implementers of technology research and development or application, with a focus on extracting entities such as "applicant", "inventor", "author institution", and "company name".

[0025] Outcome nodes: These refer to specific technological outputs, such as the model being trained to identify "invention name", "product model", and "core algorithm name".

[0026] Industry nodes: These refer to the macro-level sectors to which the technology belongs. The system uses IPC classification numbers or preset national economic industry classification standards, combined with text classification algorithms, to map text to specific industry nodes such as "integrated circuit manufacturing" and "industrial internet".

[0027] In order for computers to understand the semantic information of nodes, the system uses pre-trained language models (such as BERT) to transform the text description of nodes into high-dimensional dense vectors.

[0028] The textual input model, consisting of a brief introduction to the main nodes, a summary of the outcome nodes, and an industry definition for the industry nodes, is used to extract the output of the last hidden layer of the model as a semantic vector. This vector contains contextual information and reflects the semantic similarity between "neural networks" and "deep learning."

[0029] Standard parameter index library and alignment: Since different texts describe the same index differently, the system has a pre-built standard parameter index library. Word2Vec is used to calculate the cosine similarity between the extracted attribute nouns and the words in the standard library. When the similarity exceeds a threshold, semantic mapping and alignment are completed.

[0030] Spatial domain construction module: used to construct geometric spatial domains for three types of nodes to map discrete nodes into multi-dimensional feature spaces; the geometric spatial domains include the technical effectiveness boundary domain of the achievement node, the main capability coverage domain of the main node, and the industry demand distribution domain of the industry node.

[0031] Preferably, the spatial domain construction module constructs geometric spatial domains for the three types of nodes, including: Construct a multidimensional feature space, where each dimension corresponds to a key quantization parameter; For each achievement node, its key quantitative parameters are mapped to coordinate points in a multidimensional feature space. A high-dimensional hypersphere is constructed with the coordinate point as the center and the preset technical error tolerance as the radius as the boundary domain of the technical effectiveness. For the main node, obtain the coordinates of all the result nodes associated with the main node, calculate the minimum convex hull containing the coordinates of all result nodes, and expand the geometric boundary of the minimum convex hull based on the technical error tolerance to construct the main capability coverage domain. For an industry node, the coordinates of all the result nodes of that industry node are aggregated, and its spatial probability density distribution is calculated. The spatial region with a probability density higher than a preset threshold is taken as the industry demand distribution domain.

[0032] In one embodiment, a technical performance boundary domain (high-dimensional hypersphere) is constructed for the achievement node. In traditional databases, the achievement node is often regarded as an isolated point, but in actual engineering applications, a technical indicator often has a certain range of applicability or measurement error. First, the key quantitative parameter value of the achievement node is used as a coordinate vector to locate a centroid point in space, and a "technical error tolerance" is set, which can be set according to industry standards.

[0033] A high-dimensional hypersphere is constructed with the centroid as the center and the technical error tolerance as the radius. This hypersphere represents the effective range of the technology. As long as the external demand falls within this sphere, the technology is considered to fully meet the demand. This modeling method is more robust than simple point-to-point matching.

[0034] The capabilities of a principal node are determined by all the achievement nodes it possesses. This invention utilizes the "convex hull" algorithm from computational geometry to define the principal's technological landscape. The system obtains the coordinate set of all achievement nodes under the principal's name. A simple convex hull only represents "achievements already made." To characterize the principal's "potential capabilities," the system extends the convex hull surface outward in the normal direction based on a technical error tolerance.

[0035] An industry node is not composed of a single entity, but rather represents a macro-trend embodied by a collection of thousands of achievement nodes. The system does not use rigid geometric shapes to define an industry, but instead employs a kernel density estimation algorithm. All achievement coordinate points under the industry are considered observation samples, and the probability density value of each point is calculated in space. Areas with concentrated achievements (technological hotspots) have high density values; areas with low density values ​​have low density values. A preset density threshold is used to extract all areas in space with density values ​​greater than the threshold, forming an irregular geometric shape, i.e., the "industry demand distribution domain." The shape of this domain reflects the current "mainstream track" of the industry. The larger the domain volume, the higher the technological diversity of the industry; the more compact the domain, the more concentrated the technological routes.

[0036] Knowledge graph construction module: used to analyze the spatial relationships between different levels of geometric spatial domains to establish the connection edges between main nodes, result nodes and industry nodes and quantify the initial weights to construct the initial network topology graph; establish a multi-level dynamic weight adjustment mechanism, and combine the time decay factor to correct the connection edge weights in the initial network topology graph to generate a target industry knowledge graph with time sensitivity.

[0037] Preferably, the knowledge graph construction module parses the spatial relationships between different levels of geometric spatial domains, establishes connecting edges, and quantifies initial weights, including: The spatial inclusion relationship between the technical effectiveness boundary domain of the result node and the main capability coverage domain of the main node is calculated. If the center coordinates of the result node are located inside or on the boundary of the main capability coverage domain of the main node, a belonging connection edge is established and the initial weight is set to a preset weight value. Calculate the overlap volume between the main capability coverage area of ​​the main node and the industry demand distribution area of ​​the industry node. If the overlap volume is greater than zero, establish an adaptive connection edge, and set the initial weight to be proportional to the ratio of the overlap volume to the total volume of the main capability coverage area. Calculate the Euclidean distance between the technical effectiveness boundary domains of different achievement nodes. If the distance is less than a preset threshold, establish similar connection edges with initial weights inversely proportional to the Euclidean distance.

[0038] In this embodiment, the main function of the knowledge graph construction module is to construct logical connections between nodes by calculating their positional relationships in a multi-dimensional space. The system no longer relies solely on text labels, but instead determines the strength of associations through "geometric relationships."

[0039] The purpose of establishing the belonging connection edge is to verify and establish the subordinate relationship between the technological achievement and the R&D entity. The system determines whether the center coordinate point of the achievement node falls within the "entity capability coverage domain" constructed by the entity node.

[0040] The "core capability coverage domain" acts like a fence defining the scope of a company's technology. If the coordinates of a technological achievement fall within this fence, it means that the technology aligns with the company's technological DNA. Based on this, the system establishes attribution connections and sets the initial weight of these connections to a preset value. This not only confirms attribution but also helps the system identify abnormal data that, although registered under the company's name, exhibit technological characteristics that clearly deviate from the company's core business.

[0041] The purpose of establishing adaptive connection edges is to assess a company's position and suitability within the current industry. The system calculates the spatial overlap between the "subject capability coverage domain" of the subject node and the "industry demand distribution domain" of the industry node.

[0042] The "Industry Demand Distribution Domain" represents the mainstream technology demand areas in the market. If a company's technology coverage domain overlaps significantly with the market demand domain, it indicates that the company's products are exactly what the market urgently needs, and therefore it is given a higher weight (high suitability). If the overlap is small, it indicates that the company may be on the periphery of the industry or in a niche market.

[0043] The establishment of similar connection edges is to find similar technologies among massive amounts of technological achievements for technology comparison or plagiarism detection. The system measures the straight-line distance (Euclidean distance) between the center points of two achievement nodes (technical effectiveness boundary domains) in multi-dimensional space.

[0044] In a multidimensional feature space, the closer the coordinates of two technologies, the more similar their parameters are. The system sets a distance threshold; only nodes with a distance less than this threshold are connected by a similarity edge. The closer the distance, the higher the initial weight of the connection edge, indicating a higher degree of technological similarity.

[0045] Preferably, the dynamic weight adjustment mechanism includes: calculating the time difference between the current system time and the node timestamp; calculating the decay coefficient using a preset time decay model; the decay coefficient is negatively correlated with the time difference; and using the decay coefficient to perform weighted correction on the initial weights of the connecting edges.

[0046] In this embodiment, the dynamic weight adjustment mechanism adds a "time filter" to the knowledge graph to ensure that the system recommends the most cutting-edge technologies to users, rather than outdated technologies.

[0047] The system automatically acquires the current time and reads the timestamp marked on each technological achievement node to determine how long ago the achievement was published. If the time difference is small (the technology is new), the calculated decay coefficient is large (close to 1, indicating full retention); if the time difference is large (the technology is old), the calculated decay coefficient is small (approaching 0, indicating value loss). This rule can be flexibly adjusted. For industries with extremely rapid updates (such as chips and AI), the system will be set to make the coefficient decrease faster; while for traditional basic industries (such as building structures), the coefficient decreases more slowly.

[0048] The construction of the time decay model determines the rate at which technology "depreciates." In this embodiment, an exponential decay function is preferably used as the preset time decay model.

[0049] Through this mechanism, when users search or the system makes recommendations, the latest technological achievements will be ranked first due to their high weight, ensuring that the recommendation results are highly timely and avoiding recommending outdated technologies to users.

[0050] The intelligent recommendation module is used to respond to query commands containing query information, perform feature parsing on the query information to generate target parameter vectors and / or target semantic vectors, perform graph reasoning and technology quantification in the target industry knowledge graph, and execute any of the following modes according to the type of query command: If the query command corresponds to the node matching mode: map the target parameter vector to the multi-dimensional feature space, locate the several associated result nodes corresponding to the technical effectiveness boundary domain in which it falls, and identify several associated subject nodes connected to the associated result nodes based on the attribution connection between the result nodes and the subject nodes.

[0051] Preferably, in the node matching mode, locating the technical performance boundary domain into which it falls includes: Calculate the geometric distance between the target parameter vector and the center of the technical effectiveness boundary domain of all result nodes in the multidimensional feature space; Determine whether the geometric distance is less than the radius of the corresponding technical performance boundary region: If the value is less than the radius, the corresponding result node is identified as an associated result node, and the matching confidence level is set to the preset confidence level. If all geometric distances are not less than the radius, then select the few result nodes with the closest geometric distance as associated result nodes, and calculate the matching confidence. The matching confidence decreases non-linearly with the increase of geometric distance.

[0052] In this embodiment, when a user selects the "node matching mode" for querying (for example, when a user inputs a set of specific technical parameters to find a corresponding solution), the intelligent recommendation module does not perform text search, but performs spatial geometric search.

[0053] The calculation of geometric distance involves the system spatially comparing the user's "requirements" with the "outcomes" in the database. First, the system converts the user's query information into a coordinate point in a multi-dimensional feature space, called the "target point." The system calculates the straight-line distance between this "target point" and each outcome node in the space (i.e., the center of the technical effectiveness boundary region). The shorter the distance, the closer the user's required parameters are to the core parameters of the technology. The system determines whether the user's "target point" falls within the "technical effectiveness boundary region" of the outcome node. If the geometric distance is less than the radius, it means the user's needs fall entirely within the coverage of the technology. The system classifies this as an "exact match," assigns a very high match confidence level (e.g., 100% or a preset high score), and directly recommends it to the user.

[0054] If none of the technologies in the system covers the user's "target point" (i.e., all distances are greater than the radius), the system will not return an empty result. Instead, it will automatically find the nearest possible result nodes. Although these technologies are "nearest," they are not entirely satisfactory. Therefore, when calculating the confidence score, the score will decrease rapidly as the distance increases.

[0055] In this embodiment, the confidence level decay after exceeding the boundary can be calculated using a Gaussian radial basis function. The system presets a decay bandwidth parameter, and when the geometric distance exceeds the radius of the technical performance boundary domain, the confidence level decreases exponentially with the square of the distance, to simulate the physical characteristic that the technical relevance decreases sharply as the difference increases.

[0056] If the query command corresponds to the path analysis mode: match the target subject node based on the target semantic vector, identify several target result nodes based on the attribution connection between the subject node and the result node, fit the temporal geometric center point of all target result nodes to generate the technology leap trajectory, calculate the intersection point of the technology leap trajectory and the boundary of the industry demand distribution domain corresponding to the target subject node, and linearly extrapolate along the tangent direction of the intersection point to generate the predicted result node and timestamp.

[0057] Preferably, in the path analysis mode, the fitted technical transition trajectory includes: Obtain the target semantic vector, calculate the cosine similarity between the target semantic vector and the semantic vector of each subject node in the graph, and determine the subject node with the highest cosine similarity as the target subject node; Sort all target result nodes associated with the target subject node in chronological order of their timestamps. Set a sliding time window and calculate the geometric centroid of the result node coordinates within each time window; By using regression analysis algorithms to fit vector functions to the geometric centroids of all time windows, a position function with respect to time is obtained, which is the trajectory of technological leaps. The first derivative of the calculated technological leap trajectory is used as the speed vector of technological evolution, and the second derivative is used as the acceleration vector of technological evolution.

[0058] Preferably, in the path analysis mode, generating the predicted result nodes and timestamps includes: Construct implicit functions or parametric equations that describe the boundaries of the industry demand distribution domain; By combining the position function of the technology leap trajectory with the boundary equation of the industry demand distribution domain, the coordinates of the intersection point of the trajectory and the boundary are obtained. If there is no intersection point, the point on the technology leap trajectory that is closest to the boundary equation is used as the intersection point coordinates. The coordinates of this intersection point are determined as the critical technical point that meets industry needs; Obtain the time solution corresponding to the intersection coordinates, and use it as the timestamp to meet industry requirements; Calculate the tangent vector at the intersection point, extend it along the tangent direction by a preset step size, and obtain the coordinates of the predicted node.

[0059] In this embodiment, when the user selects the "path analysis mode," the intelligent recommendation module is no longer limited to static retrieval but instead models the dynamic process of technological development. This mode consists of two stages: the first stage reviews historical trajectories, and the second stage predicts future trends.

[0060] The system finds the enterprise or organization (target subject) in the database that best matches the user's description (target semantic vector). The system then arranges all the achievement nodes under that enterprise in chronological order. This is analogous to unfolding the enterprise's development history along a timeline.

[0061] Individual technological achievements may be accidental. The system sets a "sliding time window" (e.g., 1 or 2 years) and calculates an "average position" (geometric centroid) for all achievements within the window. This centroid represents the company's "core technological focus" during that specific time period. As the time window slides, a series of continuously moving centroid points are generated.

[0062] Trajectory Connection: Using regression analysis algorithms, the discrete centroids mentioned above are connected into a smooth curve, which is the "technology leap trajectory". It intuitively demonstrates the changes in the direction of enterprise technology research and development.

[0063] In this embodiment, a polynomial regression algorithm supported by the least squares method is employed. Considering that technological evolution typically exhibits nonlinear acceleration or deceleration characteristics, the system constructs a second- or third-order polynomial model to fit the geometric centroid sequence. This algorithm can smooth out local time jitter and generate a continuous and differentiable curve, thus facilitating subsequent calculations of the evolution speed (first derivative) and acceleration (second derivative) using differentiation techniques.

[0064] First derivative (velocity): Represents the speed of technological iteration. A large value indicates that the company has achieved a significant technological leap in a short period of time. Second derivative (acceleration): Represents the trend of R&D investment. If the acceleration is positive, it indicates that the company is accelerating its efforts in this field; if it is negative, it indicates weak innovation or a transformation.

[0065] After fitting the historical trajectory, the system uses geometric extrapolation to answer the question "What will happen in the future?". The system continuously extends the "technology leap trajectory" drawn in the previous stage, while simultaneously constructing the geometric boundary of the "industry demand distribution domain" (equivalent to a market demand threshold). The system calculates when the trajectory will cross this boundary. This intersection (critical technology point) has significant commercial implications—it represents the company's existing technological roadmap and at what point it is expected to truly meet mainstream market demand or reach industry entry standards.

[0066] If the trajectory deviates from the direction and never intersects the boundary, the system calculates the point on the trajectory closest to the boundary, which represents the "theoretical best level" that the enterprise can achieve under the existing route.

[0067] The system calculates the time required to reach the intersection point based on the length of the trajectory extension and the previous "evolution speed." The system assumes that technological development has inertia. At the intersection point, the system extends forward one step (preset step size) along the tangent of the trajectory. The coordinates of the endpoint of this extension represent the system's predicted "next-generation technological achievement."

[0068] If the query command corresponds to the technology quantification mode: the target subject node is determined based on the target semantic vector, and the technology quantification index is obtained by weighting the geometric volume of the subject capability coverage domain of the target subject node and the temporal stability index of the achievement nodes contained in the domain.

[0069] Preferably, in the technology quantification model, the calculation process of the technology quantification index includes: Obtain the target semantic vector, calculate the cosine similarity between the target semantic vector and the semantic vector of each subject node in the graph, and determine the subject node with the highest cosine similarity as the target subject node; Calculate the spatial volume of the target main node's main capability coverage area; Calculate the time interval variance of the timestamp sequence of all result nodes within the main capability coverage domain of the target main node, and construct a time series stability index reflecting the uniformity of the result release frequency based on the time interval variance. The smaller the variance, the higher the time series stability index. The spatial volume and the temporal stability index are assigned preset weight coefficients, and the technology quantification index is calculated by weighted summation.

[0070] In this embodiment, when the user selects the "Technology Quantification Mode," the intelligent recommendation module no longer performs retrieval or prediction, but instead evaluates the technological capabilities of a specific entity. The system evaluates from two dimensions: "total scale" and "R&D pace."

[0071] The spatial volume of a subject's capability coverage area represents the "size" of the subject's technological landscape. A large volume means that the subject's technological achievements are widely distributed and span a large area in the feature space. For example, if a subject has patents for both "hardware architecture" and "software algorithms," the resulting geometric volume is large. This indicates that the subject has broad technological coverage and deep comprehensive technological accumulation. A small volume means that the achievements are highly concentrated on a very small sub-point, indicating a singular technology.

[0072] For high-dimensional convex polyhedra, the system uses algorithms based on simplex subdivision or Monte Carlo integration to accurately calculate their volume values.

[0073] The time series stability index is used to judge the R&D continuity of an entity. The system extracts the timestamps of all the entity's achievements, calculates the time difference between two adjacent achievement releases, and calculates the variance of these interval values.

[0074] If the variance is small, it indicates that the company's release of results is very regular. This usually represents a mature and stable R&D team and continuous investment, making it a high-quality target. If the variance is large (for example, suddenly filing 50 patents in a single month, followed by three years of inactivity), it often means that the company may be rushing to meet project acceptance or IPO deadlines, indicating a lack of sustainability in its R&D.

[0075] The system sums the spatial volume and temporal stability indices according to preset weighting coefficients.

[0076] The resulting technology quantification index is a specific scalar value. Users can directly use this index to rank multiple competitors horizontally, thereby quickly identifying industry leaders who "have both deep technological accumulation and the ability to maintain continuous innovation."

[0077] This invention maps discrete text data into a multidimensional feature space with spatial attributes, constructs a geometric technical performance and subject capability domain, breaks through the limitations of traditional keyword matching, realizes in-depth quantitative mining of technical features and industrial logic, and significantly improves the accuracy of feature extraction.

[0078] This invention establishes a dynamic weight adjustment mechanism based on geometric spatial relationships and time decay factors, accurately linking technology supply and industry demand, ensuring that the knowledge graph can reflect the temporal changes of technology iteration in real time, and effectively enhancing the timeliness and reliability of recommendation results.

[0079] This invention provides multiple intelligent modes such as path analysis and technology quantification. It generates technology transition trajectories and predicts evolution directions by fitting the temporal geometric center. At the same time, based on the strength of spatial volume quantification technology, it provides a forward-looking scientific basis for industrial planning and technology decision-making.

[0080] The computer program instructions used to perform the operations of this invention may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing state information from the computer-readable program instructions. This electronic circuitry can execute the computer-readable program instructions to implement various aspects of the invention.

[0081] The present invention discloses a non-transitory computer-readable storage medium storing computer instructions, which, when executed by a processor, cause the processor to perform the above-described method.

[0082] Those skilled in the art will understand that all or part of the steps in the above methods can be implemented by a program instructing related hardware (e.g., processor, FPGA, ASIC, etc.), and the program can be stored in a readable storage medium, such as a read-only memory, a disk, or an optical disk. All or part of the steps in the above embodiments can also be implemented using one or more integrated circuits. Accordingly, each module in the above embodiments can be implemented in hardware, such as by using integrated circuits to implement its corresponding function, or it can be implemented as a software functional module, such as by a processor executing a program / instruction stored in memory to implement its corresponding function. The embodiments of the present invention are not limited to any particular combination of hardware and software.

[0083] Furthermore, the functional units in the various embodiments of this document can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0084] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this article, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this article. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0085] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-dimensional feature extraction and intelligent recommendation system for scientific and technological achievements that integrates text semantic mining, characterized in that, include: Data acquisition and parsing module: used to acquire multi-source heterogeneous industrial text data, and use semantic analysis and entity extraction technology to extract the basic attributes, key quantitative parameters and timestamps of subject nodes, result nodes and industry nodes; Spatial domain construction module: used to construct geometric spatial domains for three types of nodes to map discrete nodes into multi-dimensional feature spaces; the geometric spatial domains include the technical effectiveness boundary domain of the achievement node, the main capability coverage domain of the main node, and the industry demand distribution domain of the industry node; Knowledge graph construction module: used to analyze the spatial relationships between different levels of geometric spatial domains to establish the connection edges between main nodes, result nodes and industry nodes, quantify the initial weights, and construct the initial network topology graph; A multi-level dynamic weight adjustment mechanism is established, and the weights of the connecting edges in the initial network topology graph are corrected by combining the time decay factor to generate a target industry knowledge graph with time sensitivity. The intelligent recommendation module is used to respond to query commands containing query information, perform feature parsing on the query information to generate target parameter vectors and / or target semantic vectors, perform graph reasoning and technology quantification in the target industry knowledge graph, and execute any of the following modes according to the type of query command: If the query command corresponds to the node matching mode: map the target parameter vector to the multi-dimensional feature space, locate the several related result nodes corresponding to the technical effectiveness boundary domain in which it falls, and identify several related subject nodes connected to the related result nodes based on the attribution connection between the result nodes and the subject nodes. If the query command corresponds to the path analysis mode: match the target subject node based on the target semantic vector, identify several target result nodes based on the attribution connection between the subject node and the result node, fit the temporal geometric center point of all target result nodes to generate the technology leap trajectory, calculate the intersection point of the technology leap trajectory and the boundary of the industry demand distribution domain corresponding to the target subject node, and linearly extrapolate along the tangent direction of the intersection point to generate the predicted result node and timestamp. If the query command corresponds to the technology quantification mode: the target subject node is determined based on the target semantic vector, and the technology quantification index is obtained by weighting the geometric volume of the subject capability coverage domain of the target subject node and the temporal stability index of the achievement nodes contained in the domain.

2. The system according to claim 1, characterized in that, The data acquisition and parsing module utilizes semantic analysis and entity extraction techniques to extract basic attributes, key quantitative parameters, and timestamps, including: The named entity recognition model extracts organizational entities as subject nodes, technology product or patent entities as achievement nodes, and industry domain entities as industry nodes from multi-source heterogeneous industrial text data. The names and descriptions of the main nodes, achievement nodes, and industry nodes are used as basic attributes. A pre-trained language model is used to vectorize these basic attributes to generate corresponding semantic vectors. Extract the semantic attribution relationship between the main nodes and the result nodes, and classify the result nodes under the corresponding industry node names according to the semantic category; By using regular expressions combined with dependency parsing, numerical technical indicators are extracted from the descriptive text of the result nodes, a standard parameter indicator library is established, semantic mapping technology is used to align the extracted numerical technical indicators to the corresponding categories in the standard parameter indicator library, and numerical normalization is performed. Correlation analysis and dimensionality reduction were performed on the normalized numerical technical indicators. The retained principal components were selected as key quantitative parameters to ensure that the dimensions of the key quantitative parameters were consistent with the orthogonal basis of the multidimensional feature space. Parse the release time, application time, or effective time in multi-source heterogeneous industrial text data, and convert them into standard UNIX timestamp format as the time attribute of the node.

3. The system according to claim 2, characterized in that, The spatial domain construction module constructs geometric spatial domains for three types of nodes, including: Construct a multidimensional feature space, where each dimension corresponds to a key quantization parameter; For each achievement node, its key quantitative parameters are mapped to coordinate points in a multidimensional feature space. A high-dimensional hypersphere is constructed with the coordinate point as the center and the preset technical error tolerance as the radius as the boundary domain of the technical effectiveness. For the main node, obtain the coordinates of all the result nodes associated with the main node, calculate the minimum convex hull containing the coordinates of all result nodes, and expand the geometric boundary of the minimum convex hull based on the technical error tolerance to construct the main capability coverage domain. For an industry node, the coordinates of all the result nodes of that industry node are aggregated, and its spatial probability density distribution is calculated. The spatial region with a probability density higher than a preset threshold is taken as the industry demand distribution domain.

4. The system according to claim 3, characterized in that, The knowledge graph construction module parses the spatial relationships between different levels of geometric spatial domains, establishes connecting edges, and quantifies initial weights, including: The spatial inclusion relationship between the technical effectiveness boundary domain of the result node and the main capability coverage domain of the main node is calculated. If the center coordinates of the result node are located inside or on the boundary of the main capability coverage domain of the main node, a belonging connection edge is established and the initial weight is set to a preset weight value. Calculate the overlap volume between the main capability coverage area of ​​the main node and the industry demand distribution area of ​​the industry node. If the overlap volume is greater than zero, establish an adaptive connection edge, and set the initial weight to be proportional to the ratio of the overlap volume to the total volume of the main capability coverage area. Calculate the Euclidean distance between the technical effectiveness boundary domains of different achievement nodes. If the distance is less than a preset threshold, establish similar connection edges with initial weights inversely proportional to the Euclidean distance.

5. The system according to claim 4, characterized in that, The dynamic weight adjustment mechanism includes: Calculate the time difference between the current system time and the node timestamp; The attenuation coefficient is calculated using a preset time decay model; the attenuation coefficient is negatively correlated with the time difference. The initial weights of the connecting edges are adjusted using a decay coefficient.

6. The system according to claim 5, characterized in that, In the node matching pattern, the technical performance boundary domain into which it falls includes: Calculate the geometric distance between the target parameter vector and the center of the technical effectiveness boundary domain of all result nodes in the multidimensional feature space; Determine whether the geometric distance is less than the radius of the corresponding technical performance boundary region: If the value is less than the radius, the corresponding result node is identified as an associated result node, and the matching confidence level is set to the preset confidence level. If all geometric distances are not less than the radius, then select the few result nodes with the closest geometric distance as associated result nodes, and calculate the matching confidence. The matching confidence decreases non-linearly with the increase of geometric distance.

7. The system according to claim 5, characterized in that, In the path analysis model, the fitted technical transition trajectory includes: Obtain the target semantic vector, calculate the cosine similarity between the target semantic vector and the semantic vector of each subject node in the graph, and determine the subject node with the highest cosine similarity as the target subject node; Sort all target result nodes associated with the target subject node in chronological order of their timestamps. Set a sliding time window and calculate the geometric centroid of the result node coordinates within each time window; By using regression analysis algorithms to fit vector functions to the geometric centroids of all time windows, a position function with respect to time is obtained, which is the trajectory of technological leaps. The first derivative of the calculated technological leap trajectory is used as the speed vector of technological evolution, and the second derivative is used as the acceleration vector of technological evolution.

8. The system according to claim 7, characterized in that, In the path analysis mode, the generated prediction result nodes and timestamps include: Construct implicit functions or parametric equations that describe the boundaries of the industry demand distribution domain; Solve the boundary equations of the location function of the technology leap trajectory and the industry demand distribution domain to obtain the coordinates of the intersection point of the trajectory and the boundary. If there is no intersection point, find the point on the technology leap trajectory that is closest to the boundary equation as the intersection point coordinates. The coordinates of this intersection point are determined as the critical technical point that meets industry needs; Obtain the time solution corresponding to the intersection coordinates, and use it as the timestamp to meet industry requirements; Calculate the tangent vector at the intersection point, extend it along the tangent direction by a preset step size, and obtain the coordinates of the predicted node.

9. The system according to claim 5, characterized in that, In the technology quantification model, the calculation process of the technology quantification index includes: Obtain the target semantic vector, calculate the cosine similarity between the target semantic vector and the semantic vector of each subject node in the graph, and determine the subject node with the highest cosine similarity as the target subject node; Calculate the spatial volume of the target main node's main capability coverage area; Calculate the time interval variance of the timestamp sequence of all result nodes within the main capability coverage domain of the target main node, and construct a time series stability index reflecting the uniformity of the result release frequency based on the time interval variance. The smaller the variance, the higher the time series stability index. The spatial volume and the temporal stability index are assigned preset weight coefficients, and the technology quantification index is calculated by weighted summation.

Images