A method for automatically analyzing and extracting technical features of a patent application file

Abstract

The application provides a patent application file technical feature automatic analysis and extraction method, comprising the following steps: analyzing an original mark symbol list from a patent application file drawing, scanning the drawing content, obtaining all symbol numbers and position coordinate sets, and the set is recorded as an initial mark set; based on the number characteristics and the spatial distribution relationship of the position coordinates in the initial mark set, hierarchical clustering analysis is carried out, and a potential hierarchical tree structure among the mark symbols is constructed; the potential hierarchical tree structure among the symbols is traversed, the sub-level membership with high nesting depth and the independent components with low nesting depth are identified according to the depth, the branch condition and the number semantics of the nodes, and the subordinate identification is determined; the sub-level membership symbols are filtered from the subordinate identification, the core symbol number set of the independent components is obtained, the subsidiary marks are merged to the parent components, the symbol quantity statistical virtual high risk is evaluated, and the symbol quantity count is determined.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for automatically parsing and extracting technical features from patent application documents. Background Technology

[0002] In the field of automated patent application document processing, accurate extraction and analysis of technical features plays a crucial role in technical understanding, examination efficiency, and innovation assessment. This area directly impacts patent quality control and the reliability of technical intelligence analysis. Existing methods often struggle to effectively distinguish the actual meaning of symbols when processing reference numerals, leading to discrepancies in component counts. These methods typically ignore the inherent structural relationships between symbols, tending to treat all symbols equally, thus failing to accurately reflect the true composition of the technical solution. Reference numerals in the drawings point to different technical parts; a larger number may seem to correspond to more independent parts, but this is not the case. There is a hierarchical relationship between reference numerals; some point to the overall part, while others point to subdivided structures within that part. These subdivided structure markers are subordinate to higher-level markers and do not represent entirely new independent parts. If this hierarchical relationship is ignored, subdivided markers are directly counted as independent parts, resulting in an inflated count. For example, in a patent drawing of a mechanical device, one marker points to the entire transmission system, while multiple markers below it point to gears, shafts, and connectors within the system. If only the total number of markings is considered, the markings for gears, shafts, and connectors will all be treated as independent parts, thus breaking the transmission system down into multiple unrelated components and significantly overestimating the overall number of technical components. This overestimation distorts the assessment of the complexity of the technical solution, affecting subsequent feature extraction and comparative analysis. Therefore, accurately identifying the hierarchical relationships between markings when analyzing figure markings, avoiding misjudging subordinate markings as independent parts, and thus achieving accurate statistics on the number of technical components, becomes a key issue in improving the accuracy of automatic extraction of patent technical features. Summary of the Invention

[0003] This invention provides a method for automatically parsing and extracting technical features from patent application documents, mainly including:

[0004] Parse the original list of symbols from the drawings of the patent application document, scan the contents of the drawings, and obtain the set of all symbol numbers and position coordinates. This set is denoted as the initial symbol set.

[0005] Based on the spatial distribution relationship of the numbering features and location coordinates in the initial tag set, hierarchical clustering analysis is performed to construct a potential hierarchical tree structure among the tag symbols;

[0006] Traverse the potential hierarchical tree structure between symbols, and identify the child members with high nesting depth and the independent components with low nesting depth based on the node depth, branching situation and numbering semantics, and determine the subordinate identifier.

[0007] Filter child-level subordinate symbols from subordinate identifiers to obtain the core symbol number set of independent components. By merging subordinate tags into parent components, assess the risk of inflated symbol count and determine the symbol count.

[0008] Based on the number of symbols, obtain the complete description information of the parent component, locate and extract the functional description text of the corresponding component, correct the hierarchical tree structure, and determine the unique technical function of each independent component.

[0009] For the unique technical functions of each independent component, duplicate dependency identifiers are checked, multiple child-level memberships are analyzed and merged into a single parent component to obtain an optimized component list. The technical feature descriptions corresponding to each independent component are extracted to form a structured technical feature list as the final output of patent analysis.

[0010] Furthermore, the step of parsing the original list of marked symbols from the drawings of the patent application document, scanning the contents of the drawings, and obtaining a set of all symbol numbers and position coordinates, denoted as the initial mark set, includes:

[0011] The patent drawings are preprocessed, and edge detection is used to identify the contour features of the marker symbols. The coordinates of the bounding box and the position of the center point of the symbol region are determined based on the contour area and aspect ratio.

[0012] Based on the bounding box coordinates, the marked region is segmented, and the text content of the symbol number within the bounding box is extracted using a text recognition method to obtain the symbol number string representation and the corresponding center point coordinates.

[0013] A data mapping relationship is established based on the symbol number string representation and the center point coordinates. Symbols with a distance less than a threshold are grouped and labeled by calculating the distance between adjacent symbols to form an initial label set.

[0014] Furthermore, based on the spatial distribution relationship of the numbering features and location coordinates in the initial marker set, hierarchical clustering analysis is performed to construct a potential hierarchical tree structure among the marker symbols, including:

[0015] The symbol numbers in the initial tag set are parsed to extract the number sequence and letter sequence. The prefix and suffix patterns of the numbers are identified by regular expression matching. The number of symbols that share the same prefix is ​​counted to obtain the number structure features of each symbol.

[0016] The similarity between symbols is calculated based on the numbering structure features. Symbols with shared prefixes are assigned high similarity values, and symbols with consecutively increasing numbers are given a higher similarity weight. At the same time, spatial proximity is calculated based on the Euclidean distance of the symbols in the attached figure. By fusing the numbering similarity value and spatial proximity, a comprehensive correlation matrix between symbols is obtained.

[0017] Based on the comprehensive correlation matrix, perform agglomerative hierarchical clustering operation, and use the minimum distance criterion to merge symbol pairs with correlation exceeding the threshold layer by layer. During the merging, record the hierarchical depth and the number of symbols contained, and construct a potential hierarchical tree structure among the labeled symbols.

[0018] Furthermore, the construction of the potential hierarchical tree structure among the marked symbols includes: forming symbol groups based on the recorded hierarchical depth and the number of symbols contained, determining the parent-child relationship between nodes according to the hierarchical depth recorded in the clustering process, determining the parent component and the child membership, and completing the construction of the potential hierarchical tree structure.

[0019] Furthermore, the potential hierarchical tree structure among the traversal symbols, based on node depth, branching, and numbering semantics, identifies child members with high nesting depth and independent components with low nesting depth, determining the subordinate identifier, including:

[0020] The depth-first traversal algorithm is used to visit all nodes in the hierarchical tree structure. The depth value of each node from the root node and the number of direct child nodes are recorded. The numerical hierarchy relationship in the symbol number of the node is parsed to obtain node feature data containing depth value, number of child nodes and number hierarchy.

[0021] Based on the node feature data, the component type attribute is determined, and a membership mapping is constructed according to the component type attribute to determine the subordinate identifier.

[0022] Furthermore, the component type attribute includes child-level membership tags and independent component tags. For child-level membership tags, the symbol number of its direct parent node is recorded. For independent component tags, the set of node numbers in its subtree is retained, and each symbol is marked as an independent component or a subordinate part.

[0023] Furthermore, the step of filtering child-level subordinate symbols from the subordinate identifiers to obtain the core symbol number set of independent components, merging subordinate tags into the parent component, assessing the risk of inflated symbol counts, and determining the symbol count includes:

[0024] Traverse the list of subordinate identifiers, identify symbols marked as child-level members, remove the child-level members from the initial identifier set, merge the child-level members into the corresponding parent component, and retain the independent component symbols to form a core symbol number set;

[0025] Based on the core symbol number set, the number of independent components is counted, the difference between the total number of the initial tag set and the number of independent components is calculated, and the number of core symbol number sets is determined as the symbol count.

[0026] Furthermore, the process of obtaining complete description information of the parent component based on the number of symbols, locating and extracting the functional description text of the corresponding component, correcting the hierarchical tree structure, and determining the unique technical functions of each independent component includes:

[0027] The text paragraphs corresponding to each parent component in the patent specification are queried by counting the number of symbols. The description area containing the component mark is located by the number index. The component name, structural features and function description are extracted to form the description text. The description text is segmented and action words and function words are identified to obtain the set of functional keywords for each component.

[0028] Based on the set of functional keywords, the text similarity between components is calculated. The cosine similarity algorithm is used to measure the cosine value of the angle between the keyword vectors. If the similarity between two components exceeds the threshold and they are on different branches in the tree structure, it is determined that there is a difference in the affiliation relationship. The parent-child relationship is determined by comparing the functional descriptions, and a list of node pairs that need to be corrected is obtained.

[0029] Adjust the hierarchical tree structure of the list of nodes that need to be corrected, reconnect the child nodes with different affiliations to their corresponding parent nodes, update the depth values ​​and path information of the affected nodes, and verify the uniqueness of the corrected path by traversing to obtain the accurate hierarchical relationship.

[0030] Based on the accurate hierarchical relationship analysis, the keyword distribution characteristics of each independent component are analyzed, the functional verbs and nouns unique to that component are identified, shared general words are excluded, and the unique technical function descriptions of each independent component are determined.

[0031] Furthermore, for the unique technical functions of each independent component, duplicate dependency identifiers are checked, multiple child-level memberships are analyzed and merged into a single parent component to obtain an optimized component list. The technical feature descriptions corresponding to each independent component are extracted to form a structured technical feature list as the final output of the patent analysis, including:

[0032] Traverse the unique technical function description of each independent component, use the edit distance algorithm to calculate the similarity of the function description text, if the similarity of two child symbol function descriptions exceeds the threshold, they are judged as duplicates, and the affiliation relationship is determined by comparing the parent component to obtain the duplicate affiliation identifier pairs that need to be merged;

[0033] Based on the duplicate subordinate identifiers that need to be merged, perform the membership fusion, retain the child symbol with the smaller number as the representative, delete other duplicate child records, merge the duplicate child description information into the retained symbol, update the child symbol count of the parent component, and obtain the optimized component list and the number of independent components.

[0034] Based on the optimized component list, the complete technical feature description corresponding to each independent component is extracted from the patent specification and arranged in a tree-like hierarchical order and component number to form a structured technical feature list.

[0035] Furthermore, the complete technical feature description includes three aspects: function, structural composition, and connection relationship.

[0036] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0037] This invention discloses an automatic method for parsing and extracting technical features from patent application documents. Addressing the challenges of complex symbol markings, ambiguous hierarchical relationships, and confusion between independent components and their sub-level affiliations in patent drawings, this method solves the problems of misjudgment of symbol attribution and inflated statistical counts by constructing a hierarchical tree structure and cross-validation mechanism. First, an initial set of markers is obtained by scanning the drawings. Hierarchical clustering is performed based on numbering features and spatial distribution to construct potential hierarchical relationships between symbols. Then, the tree structure is traversed to identify nesting depth and independent components. Sub-level symbols are filtered using subordinate identifiers, and subordinate markers are merged into parent components. Finally, functional description text is used to cross-validate hierarchical attribution, optimize the component list, and extract structured technical feature descriptions. This invention significantly improves the accuracy of component identification and the completeness of technical feature extraction, providing an efficient and reliable solution for patent parsing. Attached Figure Description

[0038] Figure 1 This is a flowchart of a method for automatically parsing and extracting technical features from a patent application document according to the present invention. Detailed Implementation

[0039] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] like Figure 1 This embodiment of the method for automatically parsing and extracting technical features from a patent application document may specifically include:

[0041] Step S101: Parse the original list of symbols from the drawings of the patent application document, scan the contents of the drawings, and obtain the set of all symbol numbers and position coordinates. This set is denoted as the initial symbol set.

[0042] Image preprocessing is performed on the patent drawings. Clear image data is obtained through grayscale conversion and binarization. The Canny edge detection algorithm is used to identify the contour features of the marked symbols in the drawings. The validity of the symbol regions is determined based on the contour area and aspect ratio, obtaining the bounding box coordinates of each marked symbol and its center point position in the drawing coordinate system. Image segmentation is performed on each marked region based on the bounding box coordinates. A deep learning-based text recognition method is used to extract the symbol number text content within the bounding box. If the recognition confidence is below 0.8, histogram equalization is used to enhance image contrast before re-recognition, obtaining the string representation of the symbol number and its corresponding center point coordinates. A data mapping relationship is established based on the symbol number string representation and center point coordinates. The spatial distribution characteristics of the symbols are determined by calculating the Euclidean distance between adjacent symbols. Symbols with a distance less than a preset threshold T are grouped and marked, where T is dynamically determined based on the average symbol density of the drawings, typically 0.1 times the width W of the drawings. This determines the unique identifier and precise position of each symbol number in the drawings, forming an initial mark set containing the symbol number, center coordinates, and group identifier.

[0043] In one embodiment, when resolving the symbols in the drawings of a patent application document, image data suitable for subsequent analysis is first obtained through image preprocessing.

[0044] Specifically, the black-and-white image is binarized using an adaptive thresholding method, with the threshold automatically adjusted based on local pixel intensity. When identifying the outline of the marker symbol, the Canny edge detection algorithm smooths image noise using a Gaussian filter, calculates the image gradient intensity and direction, performs non-maximum suppression to preserve edge details, and finally uses a dual threshold to detect connecting edges. The outline feature judgment is based on a comparison between the outline area and a preset area threshold; outlines with excessively small areas are considered noise points and filtered out, while outlines with an aspect ratio between 0.3 and 3.0 are identified as valid marker regions.

[0045] Preferably, the deep learning-based character recognition method employs a convolutional neural network architecture, comprising three main parts: a feature extraction layer, a sequence modeling layer, and a transcription layer. The feature extraction layer extracts visual features from the image through multi-layer convolutional operations, forming a feature map. The sequence modeling layer uses a bidirectional long short-term memory network to capture the contextual relationships of the feature sequences. The transcription layer converts the feature sequences into symbol-numbered text through a sequential classification algorithm. Recognition confidence is obtained by calculating the geometric mean of the predicted character probabilities. When the confidence is below 0.8, histogram equalization is used to enhance image contrast, expand the grayscale dynamic range, highlight the difference between the marked symbols and the background, and the image is re-input into the recognition network for secondary recognition.

[0046] In one possible implementation, the Euclidean distance is calculated based on the coordinates of the center points of two marker symbols, obtaining the spatial distance value through square root operation. The preset threshold is dynamically determined according to the average symbol density of the attached drawing, typically set to one-tenth of the width of the attached drawing. Symbols with a distance less than this threshold are determined to be different markers belonging to the same technical component.

[0047] For example, when processing a mechanical patent drawing containing a transmission device, the marker "100" points to the entire transmission system, while the markers "101", "102", and "103" point to the main shaft, gear, and bearing within the system, respectively. Through spatial distance analysis, these sub-markers are all less than a threshold distance from the main marker "100", and are grouped and marked as the same component series. A hierarchical association record is established in the initial marker set to avoid misjudging internal structure markers as independent components.

[0048] Step S102: Based on the spatial distribution relationship of the numbering features and location coordinates in the initial label set, hierarchical clustering analysis is performed to construct a potential hierarchical tree structure among the label symbols.

[0049] The initial set of symbols is parsed to extract the numerical and alphanumeric sequences. Regular expressions are used to identify prefix and suffix patterns. If a symbol contains consecutive digits, its increasing order is determined. The number of symbols sharing the same prefix is ​​counted to obtain the symbol structure features. Based on these features, the similarity between symbols is calculated. Symbols sharing prefixes are assigned higher similarity values, and symbols with consecutively increasing numbers receive increased similarity weights. Spatial proximity is calculated based on the Euclidean distance between the symbols in the attached diagram. A comprehensive association matrix is ​​obtained by linearly combining the symbol similarity value and spatial proximity, where the symbol similarity value has a weight of 0.6 and the spatial proximity has a weight of 0.4. Agglomerated hierarchical clustering is performed based on this matrix. Using the minimum distance criterion, symbol pairs with an association score exceeding 0.7 are merged layer by layer from the bottom layer. The current layer depth and the number of symbols in that layer are recorded at each merge. Clustering terminates when all association scores are below a threshold, resulting in multi-level symbol groups. A tree-like hierarchical structure is constructed based on the symbol groups. The parent-child relationship between nodes is determined according to the level depth recorded during the clustering process. Symbol groups with smaller level depths are used as parent components, and symbols with larger level depths that are included are used as child members. The parent-child mapping in each merging step is identified by traversing the branch order of the clustering tree. All child symbols under each parent component are identified by traversing the tree structure, thus completing the construction of the potential hierarchical tree structure.

[0050] In one implementation, when identifying the hierarchical relationships of the marking symbols in the patent drawings, a tree-like hierarchical structure is constructed by analyzing numbering features and spatial distribution. The extraction of the numbering structure features is based on string pattern recognition to identify potential affiliation relationships between marking symbols. The numbering string parsing employs a regular expression matching method. For purely numeric numbers such as "100", "101", and "102", the hundreds digit is extracted as a prefix feature; for alphanumeric mixed numbers such as "A1", "A2", and "B1", the letter part is separated as a category identifier, and the numeric part as a sequence identifier. Statistical analysis reveals that symbols sharing the same prefix usually belong to different parts of the same technical component. The group of symbols with the prefix "10" may represent a complete transmission device, while "101" to "109" point to the internal structure of that device.

[0051] Preferably, the calculation of the number similarity value adopts a multi-dimensional feature fusion method, constructing a feature vector containing three dimensions: prefix matching degree, digit continuity, and letter consistency. Prefix matching degree is obtained by calculating the ratio of the longest common prefix length of two numbers to the total length of the number; digit continuity is judged based on the difference in the digit parts of the numbers, with adjacent numbers having a difference of 1 assigned a continuity weight of 0.9, a difference of 2 assigned a weight of 0.7, and a difference exceeding 5 assigned a weight below 0.3; letter consistency is applied to numbers containing letters, assigning a value of 1 for identical letters and 0 for different letters. The three feature values ​​are weighted and summed to obtain the final number similarity value, where prefix matching degree has a weight of 0.5, digit continuity has a weight of 0.3, and letter consistency has a weight of 0.2. This multi-dimensional fusion method can accurately reflect the inherent correlation between numbers, avoiding errors caused by single feature judgment.

[0052] Specifically, spatial proximity is calculated based on the center point position of the marker symbols in the attached coordinate system, using the Euclidean distance formula to calculate the pixel distance between any two symbols. This distance value needs to be normalized according to the overall size of the attached image, by dividing the distance value by the length of the attached image's diagonal, mapping the spatial proximity to a standard range of 0 to 1; the closer the distance, the closer the proximity is to 1.

[0053] In one embodiment, the comprehensive correlation matrix is ​​constructed through a linear combination, where the correlation value is equal to 0.6 times the number similarity value plus 0.4 times the spatial proximity. This weighting reflects the dominant role of number features in hierarchical recognition while also taking into account the auxiliary role of spatial location.

[0054] Understandably, the agglomerative hierarchical clustering operation starts with each symbol as an independent cluster and gradually merges the cluster pairs with the highest correlation. The single-link clustering criterion is defined as the maximum correlation value of any symbol pair in two clusters. When the correlation between symbol a in cluster A and symbol b in cluster B reaches 0.8, even if the correlation of other symbol pairs is low, the two clusters are still considered to be merged. Each merge operation records the current iteration number as the hierarchical depth. The symbol pair depth for the first merge is 1, for the second it is 2, and so on. The merge process continues until the maximum correlation of all remaining cluster pairs is lower than a threshold of 0.7. This threshold is determined based on statistical analysis of a large number of patent figures. Being below the merging threshold avoids over-merging and effectively distinguishes independent components from subordinate parts. After clustering, the symbol groups form multiple sets of different sizes. Groups containing more than 5 symbols usually represent complex technical components, while groups containing only 2 to 3 symbols are mostly simple components. Furthermore, the construction of the tree-like hierarchical structure is based on the merging order and hierarchical depth information recorded during the clustering process. Symbol groups with smaller depth values ​​are located at higher-level nodes in the tree, representing complete technical components; symbols with larger depth values ​​are located at lower-level nodes, representing the internal structure or details of the components. A depth-first traversal algorithm is used to visit all nodes in the tree, identifying all child nodes under each parent node.

[0055] For example, when processing a patent drawing containing a hydraulic control device, the marker "200" refers to the entire hydraulic system, with "201" to "205" pointing to the cylinder, piston, control valve, pipeline, and oil tank, respectively. Through number similarity analysis, these markers all share the same hundreds digit "2," with a similarity value exceeding 0.75. Spatial analysis shows they are closely distributed in the drawing, with a mean proximity of 0.82. Clustering operations group them into the same group, constructing a tree structure where "200" is the root node and the remaining markers are leaf nodes, clearly demonstrating the hierarchical structure of the hydraulic system. By constructing this hierarchical tree structure, the affiliation of each marker in the patent drawing can be accurately identified, distinguishing between independent components and subordinate parts.

[0056] Step S103: Traverse the potential hierarchical tree structure between symbols, and identify the child members with high nesting depth and the independent components with low nesting depth based on the node depth, branching situation and numbering semantics, and determine the subordinate identifier.

[0057] A depth-first traversal algorithm is used to visit all nodes in the hierarchical tree structure. The depth value of each node from the root node and the number of its direct child nodes are recorded. The hierarchical relationship of the numbers in the node's corresponding symbol number is parsed. When the number contains a decimal separator, it is marked as a multi-level number. The nesting level of the node in the tree is determined based on the path length from the node to the root node, resulting in node feature data containing depth value, number of child nodes, and number level. Component type determination is performed based on this node feature data. These thresholds are set based on statistical experience with tree structures, with a depth threshold of 3 and a child node threshold of 3. When the node depth exceeds 3 and the number of child nodes is less than 2, it is determined to be a child-level membership. If the node depth is less than or equal to 2 and the number of child nodes is greater than or equal to 3, it is determined to be an independent component. The determination result is verified by comparing the numbering continuity of nodes at the same level. Specifically, the continuity judgment is to check whether the difference in numbers between adjacent nodes is less than or equal to 1 to confirm the increasing pattern, thus obtaining the component type attribute of each node. Based on the component type attribute, a membership mapping is constructed. For child membership tags, the symbol number and complete path of its direct parent node are recorded. For independent component tags, the set of numbers of all nodes in its subtree is retained. The transitivity of the membership relationship is verified by traversing the complete path from each node to the root node. The subordinate identifier is determined and each symbol is marked as an independent component or a subordinate part.

[0058] In one implementation, the component affiliation of the marked symbols is identified by traversing a hierarchical tree structure. Node depth and the number of branches are used as the criteria to distinguish the boundaries between independent components and subordinate parts. Depth-first traversal starts from the root node, using a stack structure to store the nodes to be visited. For each visited node, its depth value is recorded and the number of child nodes is accumulated. For multi-level numbering systems containing decimal points, such as "100.1" and "100.2", the numbers before and after the decimal point are interpreted as the primary and secondary numbers, respectively, identifying them as subdivided markers with clear hierarchical relationships. The nested level number is obtained by calculating the number of edges traversed from the node to the root node, with the root node level being 0, direct child nodes being 1, and so on, increasing sequentially.

[0059] Preferably, the component type determination adopts a dual-threshold determination mechanism, with a depth threshold set to 3 and a branch threshold set to 3. These two thresholds are determined based on statistical analysis of a large number of patent drawings. When a node depth exceeds 3 and the number of child nodes is less than 2, it indicates that the node is located at a deeper level of the tree and has fewer branches, usually corresponding to the internal detailed structure of a component, and is determined as a child-level membership marker. Conversely, nodes with a depth of 2 or less and a number of child nodes greater than or equal to 3 are located at the upper level of the tree and have multiple branches, representing independent functional modules, and are determined as independent component markers. The continuity verification of the numbering of nodes at the same level is achieved by checking the increasing pattern of the numerical part of the number. Consecutive numbers such as "201", "202", and "203" are considered to belong to the same component series, enhancing the accuracy of the determination.

[0060] Specifically, the membership mapping table uses a key-value pair structure for storage, where the key is the child node number and the value contains the parent node number and the complete path string. Path transitivity verification is implemented by recursively traversing the ancestor nodes of each node to confirm that each level of parent-child relationship from the child node to the root node exists and is unique.

[0061] For example, when processing patent drawings containing complex mechanical structures, the designation "300" represents the entire gearbox assembly, with "301" to "306" pointing to the input shaft, output shaft, gear set, bearing, housing, and lubrication device, respectively. Based on component identification, "300" is recognized as an independent component because it has a depth of 1 and 6 child nodes; "301.1" and "301.2," pointing to the keyway and shoulder details of the input shaft, have a depth of 3 and no child nodes, and are therefore identified as sub-level members. The generated subordinate identifiers clearly distinguish the 6 main components and several detail markers, avoiding the error of classifying all markers as independent components.

[0062] Step S104: Filter child-level subordinate symbols from the subordinate identifiers to obtain the core symbol number set of independent components. By merging subordinate tags into the parent component, assess the risk of inflated symbol count and determine the symbol count.

[0063] Traverse the list of subordinate identifiers, identify symbols marked as child-level members, and remove these child-level members from the initial identifier set. Simultaneously, by querying the parent component number corresponding to each child-level symbol, merge the child-level symbols into their respective parent components, retaining only the independent component symbols to form a core symbol number set. Based on this core symbol number set, count the actual number of independent components, calculate the difference between the total number in the initial identifier set and the number of independent components, divide the difference by the initial total to obtain the inflated ratio. If the inflated ratio exceeds 0.3, a risk of statistical inflatation is identified, and the number of core symbol number sets is determined as the symbol count.

[0064] In one implementation, accurate identification of independent components is achieved by filtering child symbols in the subordinate identifiers, avoiding statistical inflated numbers in the attached figures. When traversing the subordinate identifier list, a hash table structure is used to store parent-child relationships, with the key being the child symbol number and the value being the corresponding parent number. After identifying a child-belonging symbol, the symbol record is deleted from the initial identifier set and simultaneously merged into the child symbol list of the parent component, thus merging subordinate symbols into the parent. The inflated ratio is calculated based on the difference between the initial total number of identifiers and the number of independent components. This difference reflects the number of child symbols misidentified as independent components; dividing the difference by the initial total yields the inflated ratio. A threshold of 0.3 is an empirical value determined based on the statistical regularities of the patent attached figures; exceeding this threshold indicates a significant statistical bias.

[0065] For example, in a patent drawing containing control circuitry, the initial symbol set contains 50 symbols. After filtering, 12 independent components and 38 sub-component symbols are identified. The inflated ratio is 0.76, far exceeding the threshold of 0.3, indicating that without filtering, the number of components would be overestimated by more than four times. The determined symbol count is 12, accurately reflecting the actual complexity of the technical solution.

[0066] Step S105: Based on the number of symbols, obtain the complete description information of the parent component, locate and extract the functional description text of the corresponding component, correct the hierarchical tree structure, and determine the unique technical function of each independent component.

[0067] The patent specification retrieves text paragraphs corresponding to each parent component based on symbol count and core symbol number set. Using a constructed number index (a mapping table between component markers and text positions), the description region containing the component marker is located. Component names, structural features, and function descriptions are extracted to form description information records. The description text is segmented, and action words and function terms are identified to obtain a set of functional keywords for each component. Based on this set of functional keywords, text similarity between components is calculated. A cosine similarity algorithm is used to measure the cosine of the angle between keyword vectors. If the similarity between two components exceeds 0.7 and they are on different branches in the generated tree structure, an attribution error is identified. The correct parent-child relationship is determined by comparing their functional descriptions, resulting in a list of node pairs requiring correction. The hierarchical tree structure is adjusted based on this list, reconnecting incorrectly attributed child nodes to their correct parent nodes. The depth and path information of affected nodes are updated. A depth-first traversal is used to verify the uniqueness of the path from each corrected node to the root node, resulting in a cross-validated accurate hierarchical relationship. Based on the accurate hierarchical relationship analysis, the keyword distribution characteristics of each independent component are analyzed, the functional verbs and nouns unique to that component are identified, common words shared with other components are excluded, the unique technical function descriptions of each independent component are determined, and the accuracy of hierarchical classification is verified.

[0068] In one implementation, the accurate identification of the attribution relationship of marker symbols is achieved by extracting the functional description text from the patent specification and cross-validating it with a hierarchical tree structure. The cross-validation mechanism utilizes textual semantic information to verify and correct the initial hierarchical relationship constructed based on numbering features, improving the accuracy of component segmentation. The numbering index is constructed based on the paragraph numbering system of the patent specification, with each technical component having a corresponding detailed description paragraph in the specification. A mapping relationship between component numbers and specification paragraphs is established by matching citation patterns such as "see marker YYY in Figure X" using regular expressions. During the extraction process, complete sentences from the paragraphs are obtained, including descriptions of the component name, structural composition, working principle, and functional role, forming a structured record of descriptive information.

[0069] Specifically, the word segmentation process employs a dictionary-based mechanical segmentation method, combined with a specialized terminology dictionary for the patent field. Identified action words include verbs indicating function such as "transmission," "control," "adjustment," and "support"; functional nouns include nouns indicating components such as "gear," "bearing," and "valve." Part-of-speech tagging distinguishes different types of words, constructing a functional keyword vector for each component.

[0070] Preferably, the application process of the cosine similarity algorithm includes two stages: vectorization and similarity calculation. In the vectorization stage, the keyword set of each component is converted into a high-dimensional sparse vector. Each dimension of the vector corresponds to a word in the dictionary, and the value is the frequency of that word in the component description. A word frequency-inverse document frequency weighting method is used to adjust the word weights, reducing the influence of common words and highlighting the importance of feature words. Similarity calculation is achieved by calculating the cosine value of the angle between two vectors. The closer the cosine value is to 1, the more similar the functions of the two components are. The threshold of 0.7 is an empirical value determined based on statistical analysis of a large number of patent texts. This threshold can effectively distinguish between functionally similar components and functionally independent components. When the similarity between two components exceeds 0.7 but they are located on completely different branches in the tree structure, it indicates a possible misclassification, requiring further verification of their true parent-child relationship.

[0071] In one possible implementation, the determination of incorrect attribution combines two dimensions: textual similarity and structural position. If component A and component B have a functional similarity greater than 0.7, but A belongs to parent P1 while B belongs to parent P2, then the functional description relevance of A to P2 and B to P1 is compared. By calculating the cross-similarity matrix, the matching degree between each component and each potential parent is determined. The parent-child pairing with the highest relevance is considered the correct attribution relationship, and a list of node pairs that need adjustment is generated accordingly. Furthermore, the correction operation of the tree structure involves the synchronous update of multiple data structures. When an incorrect connection is broken, the node needs to be removed from the child node list of the original parent node, and the child node count of the original parent node needs to be updated. When a new connection is established, the node is added to the child node list of the new parent node, and the depth value of the node and all its descendant nodes is recalculated. The depth value update adopts a breadth-first traversal, starting from the moved node, updating the depth of its descendant nodes layer by layer to ensure that the depth information of the entire subtree remains consistent. The update of path information includes reconstructing the complete path string from each affected node to the root node for subsequent path uniqueness verification.

[0072] Understandably, path uniqueness verification is implemented using depth-first traversal, starting from each leaf node and traversing upwards to the root node, recording the sequence of node numbers traversed. If a node is found to be reachable from the root node via multiple paths, it indicates that the tree structure contains loops or multiple parent nodes, requiring further inspection and correction.

[0073] For example, the identification of unique functions employs a difference set operation method. Keywords from all components are collected to form a global vocabulary set. Then, vocabulary specific to each component is calculated—vocabulary that appears only in the description of that component and not in other components. These unique vocabulary terms typically represent the core functional features of the component. By semantically classifying these unique vocabulary terms, the component's affiliation within different functional categories such as hydraulic transmission, electrical control, and mechanical support is determined. In processing a complex patent for automated equipment, cross-validation of functional text revealed that "position sensor 405," originally belonging to "transmission mechanism 400," should actually belong to "control unit 500." The corrected hierarchical relationship accurately reflects the functional module division of the equipment, increasing the accuracy of technical feature extraction to over 95%.

[0074] Step S106: For the unique technical function of each independent component, check for duplicate subordinate identifiers, analyze multiple child-level memberships and merge them into a single parent component to obtain an optimized component list. Extract the technical feature descriptions corresponding to each independent component to form a structured technical feature list as the final output of patent analysis.

[0075] The algorithm iterates through the unique technical function descriptions of each independent component and uses the Levenshtein edit distance algorithm to calculate the similarity between pairs of function description texts. The input is the unique technical function description text pair, and the output is the similarity score. This algorithm is based on the minimum number of edit operations, unlike the cosine similarity used in S105, which is used for vector similarity. If the similarity of the function descriptions of two child symbols exceeds 0.9, they are considered functionally duplicated. The algorithm then compares their parent components to determine if they belong to the same category, obtaining duplicate subordinate identifier pairs that need to be merged. Based on these duplicate subordinate identifier pairs, a membership fusion is performed, retaining the child symbol with the smaller number as the representative, deleting other duplicate child records, merging the description information of all duplicate children under the retained symbol, updating the child symbol count of the parent component, and checking for any missing duplicate identifiers to obtain the optimized component list and the accurate number of independent components. Based on the optimized component list, a complete technical feature description corresponding to each independent component is extracted from the patent specification, including three aspects: function, structural composition, and connection relationship. These are arranged in a tree-like hierarchical order and component number to form a structured list containing component number, functional description, and technical feature, which serves as the final output of the automatic extraction of patent technical features.

[0076] In one implementation, by identifying and fusing duplicate dependent identifiers, the inflated component count caused by multiple references is eliminated, resulting in accurate independent component statistics and standardized technical feature output. The edit distance algorithm is used to quantify the degree of difference between two functional description texts. This algorithm calculates the minimum number of editing operations required to convert one string into another, including insertion, deletion, and replacement operations. For two description texts of length m and n, the edit distance value ranges from 0 to max(m,n). Dividing the edit distance by the length of the longer text yields the normalized distance; subtracting the normalized distance from 1 gives the similarity. When the similarity exceeds 0.9, it indicates that the two descriptions are almost identical.

[0077] Preferably, the identification and fusion of duplicate subordinate identifiers follow specific processing rules. Under the same parent component, if multiple child symbols are found to have the same or highly similar functional descriptions, their numbers are first compared, and the symbol with the smallest number is selected as the representative symbol and retained. Other duplicate symbols are marked as pending deletion. During the fusion process, the descriptive information of all duplicate symbols is collected, including functional descriptions, structural features, and connection relationships, and merged into the description record of the representative symbol. If there are slight differences between the descriptions, a union approach is used to retain all unique information to avoid information loss. When updating the child symbol list of the parent component, references to the symbols to be deleted are removed from the list, and the child symbol counter is updated simultaneously to ensure the consistency of statistical data.

[0078] Specifically, the traversal check is implemented recursively, starting from the root node and visiting each parent component to check for any unprocessed duplicates among its child symbols. The check includes horizontal comparisons between symbols at the same level and vertical verification across levels, ensuring that no potential duplicate identifiers are missed.

[0079] In one possible implementation, the structured technical feature list employs a three-tiered information organization. The first tier is a component number index; the second tier contains three fixed fields: function, structural composition, and connection relationships; and the third tier contains detailed descriptive text for each field. The arrangement follows a depth-first traversal sequence of a tree-like hierarchy, with components within the same tier arranged in ascending order of their numbers.

[0080] For example, when processing the accompanying drawings of a car patent containing an automatic transmission, it was found that the three identifiers "lock-up clutch 601", "hydraulic torque converter lock-up mechanism 601", and "direct drive device 601" actually refer to the same component. Using an edit distance algorithm, the functional description similarity was confirmed to be 0.92. After merging, "lock-up clutch 601" was retained as the unique identifier, and the number of independent components was optimized from the original 45 to 38.

[0081] The above description is only a preferred embodiment of the present invention. It should be noted that those skilled in the art can make several improvements and additions without departing from the principle of the present invention, and these improvements and additions should also be considered within the scope of protection of the present invention.

Claims

1. A method for automatically parsing and extracting technical features from patent application documents, characterized in that, The method includes: Parse the original list of symbols from the drawings of the patent application document, scan the contents of the drawings, and obtain the set of all symbol numbers and position coordinates. This set is denoted as the initial symbol set. Based on the spatial distribution relationship of the numbering features and location coordinates in the initial tag set, hierarchical clustering analysis is performed to construct a potential hierarchical tree structure among the tag symbols; Traverse the potential hierarchical tree structure between symbols, and identify the child members with high nesting depth and the independent components with low nesting depth based on the node depth, branching situation and numbering semantics, and determine the subordinate identifier. Filter child-level subordinate symbols from subordinate identifiers to obtain the core symbol number set of independent components. By merging subordinate tags into parent components, assess the risk of inflated symbol count and determine the symbol count. Based on the number of symbols, obtain the complete description information of the parent component, locate and extract the functional description text of the corresponding component, correct the hierarchical tree structure, and determine the unique technical function of each independent component. For the unique technical functions of each independent component, duplicate dependency identifiers are checked, multiple child-level memberships are analyzed and merged into a single parent component to obtain an optimized component list. The technical feature descriptions corresponding to each independent component are extracted to form a structured technical feature list as the final output of patent analysis.

2. The method for automatically parsing and extracting technical features from patent application documents according to claim 1, characterized in that, The process involves parsing the original list of symbols from the drawings of the patent application document, scanning the drawing content, and obtaining a set of all symbol numbers and position coordinates. This set is denoted as the initial symbol set, and includes: The patent drawings are preprocessed, and edge detection is used to identify the contour features of the marker symbols. The coordinates of the bounding box and the position of the center point of the symbol region are determined based on the contour area and aspect ratio. Based on the bounding box coordinates, the marked region is segmented, and the text content of the symbol number within the bounding box is extracted using a text recognition method to obtain the symbol number string representation and the corresponding center point coordinates. A data mapping relationship is established based on the symbol number string representation and the center point coordinates. Symbols with a distance less than a threshold are grouped and labeled by calculating the distance between adjacent symbols to form an initial label set.

3. The method for automatically parsing and extracting technical features from patent application documents according to claim 1, characterized in that, Based on the spatial distribution relationship of the numbering features and location coordinates in the initial tag set, hierarchical clustering analysis is performed to construct a potential hierarchical tree structure among the tag symbols, including: The symbol numbers in the initial tag set are parsed to extract the number sequence and letter sequence. The prefix and suffix patterns of the numbers are identified by regular expression matching. The number of symbols that share the same prefix is ​​counted to obtain the number structure features of each symbol. The similarity between symbols is calculated based on the numbering structure features. Symbols with shared prefixes are assigned high similarity values, and symbols with consecutively increasing numbers are given a higher similarity weight. At the same time, spatial proximity is calculated based on the Euclidean distance of the symbols in the attached figure. By fusing the numbering similarity value and spatial proximity, a comprehensive correlation matrix between symbols is obtained. Based on the comprehensive correlation matrix, perform agglomerative hierarchical clustering operation, and use the minimum distance criterion to merge symbol pairs with correlation exceeding the threshold layer by layer. During the merging, record the hierarchical depth and the number of symbols contained, and construct a potential hierarchical tree structure among the labeled symbols.

4. The method for automatically parsing and extracting technical features from patent application documents according to claim 3, characterized in that, The construction of the potential hierarchical tree structure among the marked symbols includes: forming symbol groups based on the recorded hierarchical depth and the number of symbols contained; determining the parent-child relationship between nodes according to the hierarchical depth recorded in the clustering process; determining the parent component and the child membership; and completing the construction of the potential hierarchical tree structure.

5. The method for automatically parsing and extracting technical features from patent application documents according to claim 1, characterized in that, The potential hierarchical tree structure among the traversal symbols, based on node depth, branching, and numbering semantics, identifies child members with high nesting depth and independent components with low nesting depth, determining the subordinate identifier, including: The depth-first traversal algorithm is used to visit all nodes in the hierarchical tree structure. The depth value of each node from the root node and the number of direct child nodes are recorded. The numerical hierarchy relationship in the symbol number of the node is parsed to obtain node feature data containing depth value, number of child nodes and number hierarchy. Based on the node feature data, the component type attribute is determined, and a membership mapping is constructed according to the component type attribute to determine the subordinate identifier.

6. The method for automatically parsing and extracting technical features from patent application documents according to claim 5, characterized in that, The component type attribute includes child-level membership tags and independent component tags. For child-level membership tags, the symbol number of its direct parent node is recorded. For independent component tags, the set of node numbers in its subtree is retained, and each symbol is marked as an independent component or a subordinate part.

7. The method for automatically parsing and extracting technical features from patent application documents according to claim 1, characterized in that, The process of filtering child-level subordinate symbols from subordinate identifiers to obtain the core symbol number set of independent components, merging subordinate tags into parent components, assessing the risk of inflated symbol counts, and determining the symbol count includes: Traverse the list of subordinate identifiers, identify symbols marked as child-level members, remove the child-level members from the initial identifier set, merge the child-level members into the corresponding parent component, and retain the independent component symbols to form a core symbol number set; Based on the core symbol number set, the number of independent components is counted, the difference between the total number of the initial tag set and the number of independent components is calculated, and the number of core symbol number sets is determined as the symbol count.

8. The method for automatically parsing and extracting technical features from patent application documents according to claim 1, characterized in that, The process of counting symbols to obtain the complete description information of the parent component, locating and extracting the functional description text of the corresponding component, correcting the hierarchical tree structure, and determining the unique technical functions of each independent component includes: The text paragraphs corresponding to each parent component in the patent specification are queried by counting the number of symbols. The description area containing the component mark is located by the number index. The component name, structural features and function description are extracted to form the description text. The description text is segmented and action words and function words are identified to obtain the set of functional keywords for each component. Based on the set of functional keywords, the text similarity between components is calculated. The cosine similarity algorithm is used to measure the cosine value of the angle between the keyword vectors. If the similarity between two components exceeds the threshold and they are on different branches in the tree structure, it is determined that there is a difference in the affiliation relationship. The parent-child relationship is determined by comparing the functional descriptions, and a list of node pairs that need to be corrected is obtained. Adjust the hierarchical tree structure of the list of nodes that need to be corrected, reconnect the child nodes with different affiliations to their corresponding parent nodes, update the depth values ​​and path information of the affected nodes, and verify the uniqueness of the corrected path by traversing to obtain the accurate hierarchical relationship. Based on the accurate hierarchical relationship analysis, the keyword distribution characteristics of each independent component are analyzed, the functional verbs and nouns unique to that component are identified, shared general words are excluded, and the unique technical function descriptions of each independent component are determined.

9. The method for automatically parsing and extracting technical features from patent application documents according to claim 1, characterized in that, For each independent component's unique technical function, duplicate dependency identifiers are checked, multiple child-level memberships are analyzed and merged into a single parent component to obtain an optimized component list. The technical feature descriptions corresponding to each independent component are extracted, forming a structured list of technical features as the final output of patent analysis, including: Traverse the unique technical function description of each independent component, use the edit distance algorithm to calculate the similarity of the function description text, if the similarity of two child symbol function descriptions exceeds the threshold, they are judged as duplicates, and the affiliation relationship is determined by comparing the parent component to obtain the duplicate affiliation identifier pairs that need to be merged; Based on the duplicate subordinate identifiers that need to be merged, perform the membership fusion, retain the child symbol with the smaller number as the representative, delete other duplicate child records, merge the duplicate child description information into the retained symbol, update the child symbol count of the parent component, and obtain the optimized component list and the number of independent components. Based on the optimized component list, the complete technical feature description corresponding to each independent component is extracted from the patent specification and arranged in a tree-like hierarchical order and component number to form a structured technical feature list.

10. The method for automatically parsing and extracting technical features from patent application documents according to claim 9, characterized in that, The complete technical feature description includes three aspects: function, structural composition, and connection relationship.

Images