A hierarchical structured-based technical feature contrast analysis method

Abstract

The application discloses a layered structured technical feature contrast analysis method. The method generates a structured description of the user input technical document, constructs a candidate set through multiple rounds of semantic and keyword iterative retrieval; extracts technical features from the source document and the comparison file and maps them into vectors, filters core feature nodes through two density clustering, and constructs a three-layer technical feature hierarchical structure diagram; calculates the similarity of each layer using the cross-query average distance algorithm and weighted ranking, and generates a technical difference evaluation suggestion accordingly. The application converts patents into multi-layer graphs by constructing a technical feature hierarchical structure diagram, thereby accurately identifying and highlighting core technical features.

Description

Technical Field

[0001] This application relates to the field of technical document comparison and analysis technology, and in particular to a method for technical feature comparison and analysis based on hierarchical structure. Background Technology

[0002] Large language models are trained on a wide range of knowledge from various industries, thus possessing a level of understanding beyond that of ordinary people in each field. If trained on a corpus specific to a particular domain, they can reach or even surpass the professional level of experts. Therefore, large language models can be used to assist technical personnel in comparative analysis of technical documents.

[0003] Following the emergence of large-scale patent models, agent technology centered around them has rapidly developed. Agent technology comprises three key components: large language models, fixed or dynamic workflows, and various external functions that the agent can invoke. Based on this, an intelligent patent retrieval and query workflow based on agent workflows can be designed. However, simply invoking large models using simple prompts within this workflow is insufficient to solve all problems. For instance, given the large number of retrieved results, how to more accurately rank them—that is, how to prioritize comparison documents with technical features more similar to the source document—requires detailed algorithm design.

[0004] To rank the technical features between the comparison files and the source documents in the search results, existing technologies mainly employ the following three methods:

[0005] The first category is Boolean search ranking methods based on keyword matching. Traditional retrieval systems typically employ term frequency-inverse document frequency algorithms such as TF-IDF and BM25 to rank documents according to the degree of keyword matching between the comparison documents and the source documents. However, this method can only capture co-occurrence relationships at the lexical level and cannot understand the semantic relationships between technical features. When the source document and the comparison documents use different terms to describe the same technical feature (such as "voltage detection" and "potential measurement"), it is easy to miss or misrank entries.

[0006] The second category is semantic retrieval ranking methods based on vector space models. Recent semantic retrieval systems based on word embeddings or pre-trained language models (such as BERT) encode the source document and comparison documents as separate vectors, ranking them using cosine similarity. While this method addresses the synonym problem to some extent, compressing the entire document into a single global vector loses the structural relationships and hierarchical information between technical features, resulting in core and peripheral technical features being treated equally in similarity calculations. Furthermore, this method cannot answer the question of "in which specific technical features are the two documents similar," leading to a lack of interpretability in the ranking results.

[0007] The third category is structural comparison methods based on technical feature trees. Some scholars have proposed extracting feature trees from claims for structural comparison. However, existing feature tree construction mostly relies on manual annotation or simple rule extraction, resulting in a fixed tree structure and a lack of adaptability. In terms of feature tree similarity calculation, existing methods mostly use graph edit distance or maximum common subgraph algorithms. These algorithms have NP-hard computational complexity, which cannot meet the performance requirements of real-time sorting of hundreds of comparison documents.

[0008] In summary, existing technologies still have significant shortcomings in automatically identifying core technical features, constructing hierarchical feature structures, and efficiently calculating feature map similarity. There is an urgent need for a file contrast analysis method that can adaptively construct hierarchical structures of technical features and achieve efficient similarity assessment. Summary of the Invention

[0009] Based on this, this application provides a hierarchical structured technical feature contrast analysis method, which can transform technical documents into multi-layered graphs containing core and refinement layers by constructing a "technical feature hierarchical structure graph", thereby accurately identifying and highlighting core technical features; at the same time, it adopts the "cross-query average distance" algorithm to replace the traditional graph similarity calculation, and quickly realizes the similarity calculation of the algorithm documents.

[0010] Firstly, a method for technical feature contrast analysis based on hierarchical structure is provided, the method comprising:

[0011] S1. Input the technical document input by the user into the large language model to generate a structured description that includes technical topics, technical problems, technical solutions, technical effects, technical features and technical summaries;

[0012] S2. Based on the structured description, perform an iterative retrieval that combines semantic retrieval with multi-round keyword retrieval, and construct a candidate set of retrieval results after merging and deduplication;

[0013] S3. Extract technical features from each comparison file and source document in the candidate set and map them into vectors. Filter core feature nodes through two density clustering operations to construct a three-layer technical feature hierarchy structure diagram containing a core layer, a refinement layer, and an edge layer. Extract the three-layer node sets of the source document and comparison file respectively. Calculate the symmetric mean of the bidirectional nearest neighbor distance for the node sets at the same layer using the cross-query average distance algorithm and convert it into a similarity score. Calculate the comprehensive similarity score by weighting the results according to preset weights. Based on this, sort the comparison files in descending order.

[0014] S4. Select high similarity comparison files from the sorting results, filter and classify node features similar to the source documents, and combine them with the difference assessment suggestions of the structured description generation technology.

[0015] Optionally, after generating the structured description in S1 that includes the technical topic, technical problem, technical solution, technical effect, technical features, and technical summary, an integrity check is also included:

[0016] The integrity check includes checking whether there are empty or obviously unreasonable items in the six-tuple. If so, it automatically triggers a supplementary call, uses prompt words to retry, and provides the missing input material fragments back to the large language model.

[0017] Optionally, the iterative retrieval described in S2 includes:

[0018] Based on the structured description, core words are extracted and synonyms are expanded to generate initial search terms, and the first round of semantic retrieval is performed;

[0019] Extract high-frequency content words from the current search results and identify key technical features that are not present by comparing them with the technical features of the source documents. Update the keywords accordingly and then execute the next round of keyword search.

[0020] Repeat the update and retrieval process until the preset round is reached, then merge the results from each round, remove duplicates, and form a candidate set.

[0021] Optionally, the two density clustering operations in S3 include:

[0022] Perform the first density-based spatial clustering on the feature set, dividing the features into several clusters and isolated points that are not assigned to any cluster;

[0023] Within each cluster, the node closest to the center is selected as the representative node to obtain a set of representative feature nodes. The isolated point is not included in this set.

[0024] A second density-based spatial clustering is performed on the representative feature node set to filter out the core feature node set.

[0025] Optionally, the construction of the three-layer technical feature hierarchy diagram described in S3 includes:

[0026] Place the core feature node set into the first layer;

[0027] The remaining nodes are assigned to the second layer with a preset probability and to the third layer with a complementary probability, and nodes in the upper layer must appear in the lower layer;

[0028] Within each layer, for each node, an approximate nearest neighbor search is used to find the m other nodes in that layer whose vectors are most similar to its own, and bidirectional edges are established.

[0029] Optionally, the cross-query average distance algorithm described in S3 includes:

[0030] For any two sets of nodes at the same level, X and Y, randomly sample N nodes from X, find the node with the closest cosine distance in the vector index of Y and record the distance, calculate the average to obtain the forward distance Dist(X→Y); symmetrically sample nodes from Y and calculate the reverse distance Dist(Y→X) in the index of X.

[0031] Calculate the symmetrical average distance:

[0032] ;

[0033] And convert it to a similarity score:

[0034] .

[0035] Optionally, the weighted summation according to preset weights in S3 includes: calculating the similarity scores between the source document and the comparison document in the first, second, and third layer node sets, respectively. According to the formula:

[0036]

[0037] Calculate the overall similarity score, where the weights are... satisfy .

[0038] Optionally, S4 specifically includes:

[0039] Select the top N highly similar comparison files from the sorting results, traverse the nodes in their three-layer technical feature hierarchy structure diagram, filter out similar features whose distance from the corresponding layer node of the source document is less than a preset threshold, and classify them into similar feature sets according to the core layer and the refinement layer.

[0040] The structured description and similar feature set are input into a large language model, and structured technical difference assessment opinions are generated through dedicated prompt words.

[0041] Secondly, a hierarchical structured technical feature contrast analysis system is provided, the system comprising:

[0042] The input module is used to input the technical documents entered by the user into the large language model and generate a structured description that includes technical topics, technical problems, technical solutions, technical effects, technical features and technical summaries.

[0043] The module is used to perform iterative retrieval combining semantic retrieval and multi-round keyword retrieval based on the structured description, and to construct a candidate set of retrieval results after merging and deduplication;

[0044] The sorting module is used to extract technical features from each comparison file and source document in the candidate set and map them into vectors. It filters core feature nodes through two density clustering operations and constructs a three-layer technical feature hierarchy structure diagram containing a core layer, a refinement layer, and an edge layer. It extracts the three-layer node sets of the source document and comparison file respectively. It calculates the symmetric mean of the bidirectional nearest neighbor distance for the node sets at the same layer using the cross-query average distance algorithm and converts it into a similarity score. It then calculates the comprehensive similarity score by weighting the nodes according to a preset weight and sorts the comparison files in descending order accordingly.

[0045] The generation module is used to select highly similar comparison files from the sorting results, filter and classify node features similar to the source documents, and combine the structured description to generate difference assessment suggestions.

[0046] Thirdly, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the hierarchical structure-based technical feature contrast analysis method described in any of the first aspects above.

[0047] The beneficial effects of the technical solutions provided in this application include at least the following:

[0048] (1) By adaptively constructing a three-layer technical feature hierarchy structure diagram through two density clusterings, core features are placed at the top level and refined features are gradually delegated down, which effectively solves the problem of inaccurate ranking caused by treating core technical features and marginal features equally in traditional methods.

[0049] (2) The cross-query average distance algorithm replaces the traditional graph edit distance or the maximum common subgraph and other NP-hard problems. It transforms the graph structure similarity into the vector set distribution similarity, which achieves efficient calculation while ensuring the accuracy of the evaluation. It is suitable for real-time sorting of hundreds of comparison documents.

[0050] (3) By clearly distinguishing the similarity between the core layer and the refinement layer through the hierarchical structure, it can intuitively answer which specific technical features are similar between the two technical documents, and supports flexible adjustment of the weight of each layer according to the application scenario, providing a fine-grained decision basis for the assessment of technical differences.

[0051] (4) Combining six-tuple structured description with multi-round semantic / keyword iterative retrieval strategy, it automatically constructs a high-quality candidate set and accurately sorts it, helping technical personnel to get rid of the heavy and repetitive search of similar technical documents and focus on innovative core judgments. Attached Figure Description

[0052] To more clearly illustrate the embodiments of this application or the technical solutions in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0053] Figure 1 A flowchart illustrating the steps of a hierarchical structured technical feature contrast analysis method provided in this application embodiment;

[0054] Figure 2 A hierarchical architecture diagram of the technical feature contrast provided for embodiments of this application;

[0055] Figure 3 This is a system block diagram of a hierarchical structured technical feature contrast analysis method provided in an embodiment of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0057] In the description of this application, the terms "comprising," "having," and any variations thereof are intended to cover non-exclusive inclusion, such as a process, method, system, product, or apparatus that includes a series of steps or units, not necessarily limited to those steps or units that are expressly listed, but may also include other steps or units that are not expressly listed but are inherent to these processes, methods, products, or apparatuses, or steps or units added based on further optimizations conceived in this application.

[0058] Please refer to Figure 1 The diagram illustrates a flowchart of a hierarchical structured technical feature contrast analysis method provided in this application embodiment. This method may include the following steps:

[0059] S1. Input the technical document input by the user into the large language model to generate a structured description that includes technical topics, technical problems, technical solutions, technical effects, technical features and technical summaries.

[0060] This step primarily generates a structured description based on user input. Specifically, the core objective of this step is to parse the user-input technical document into a standardized six-tuple structured description using a large language model, providing a unified semantic foundation for subsequent retrieval and ranking.

[0061] 1.1 Input type judgment and adaptation

[0062] The system first identifies the type of input material. When targeting technical personnel (including technical document analysts or editors):

[0063] The input material is the original source document, including abstracts, instructions, etc. The system will send the above content as a whole or in segments into the large model for processing.

[0064] 1.2 Structured Extraction Based on Large Models

[0065] The system employs large language models (such as the GPT series and DeepSeek) as its core parsing engine. Through designed prompt word templates, it guides the model to extract and generate six-tuples from the input text. To improve processing efficiency and stability, multiple tuple items can be called concurrently; that is, multiple independent prompt word requests are sent simultaneously to extract different tuple items, and the results are finally merged.

[0066] For each item in the six-tuple, a corresponding prompt word template was designed, as detailed in Table 1:

[0067] Table 1

[0068]

[0069] 1.3 Output and Verification of Structured Description

[0070] After receiving the six items returned by the large model, the system performs the following processing:

[0071] Standardized format: Output the six-tuple in a fixed format to facilitate parsing in subsequent steps. The format is plain text in JSON format.

[0072] Completeness check: Checks for empty or obviously unreasonable items in the six-tuple (e.g., the "Technical Issues" field outputs "None" or an empty string). If a field is missing, the system automatically triggers a supplementary call, retrying with a prompt word specifically for that field, and providing the missing input material fragment back to the large model.

[0073] Manual Confirmation Interface: The generated six-tuple is displayed to the user (technical document analyst or editor) on the system interface, allowing the user to manually edit and correct it. The corrected six-tuple will be used as input for subsequent steps, and the correction record can serve as feedback data for prompt word optimization.

[0074] S2. Iterative retrieval combining semantic retrieval and multi-round keyword retrieval based on structured description, and constructing a candidate set of retrieval results after merging and deduplication.

[0075] This step primarily focuses on constructing a candidate set of search results based on multi-round searches. The core objective is to build a comprehensive and high-quality candidate set of comparison documents based on the six-tuples generated in step 1 through multi-round, multi-mode iterative searches, providing ample material for subsequent accurate ranking. Single search modes (semantic only or keyword only) are prone to missed detections. This step employs a hybrid strategy of "semantic search + multi-round keyword search," utilizing feedback from each round of search results to optimize search terms, achieving a closed-loop improvement in the search process.

[0076] 2.1 Generate initial search terms

[0077] The system extracts key information from the six-tuples output in step 1 and automatically generates an initial search term set, including:

[0078] Core word extraction: Extract noun phrases from <technical topic> and <technical feature>, and remove stop words (such as "method", "system", "a kind").

[0079] Synonym expansion: Call the built-in thesaurus or large model to generate 1-2 synonyms or near-synonyms for each core word (e.g., "sorting" is expanded to "sorting, rearrangement, sorting algorithm").

[0080] 2.2 First Round: Semantic Retrieval

[0081] Semantic retrieval does not rely on precise keyword matching. Instead, it converts the overall semantics of the source document into a vector and searches for semantically similar comparative documents in a document library. The search input is primarily "technical solution," concatenating "technical problem" and "technical effect" as the source text for semantic vectorization. The search scope can be selected from the full-text library or limited to the title, abstract, or claims; the default recommendation is the full-text library to broaden recall. The top K most similar documents are selected as the first round of search results, and the similarity score and matched paragraphs for each document are recorded. The first round of search results are added to the candidate set, while highly relevant paragraphs in the results list (such as the abstracts or claims of the top 10 documents) are temporarily stored for subsequent keyword updates.

[0082] 2.3 Keyword Update Strategy (Executed after each round of retrieval)

[0083] After each round of retrieval (including semantic retrieval and keyword retrieval), the system performs a keyword update to optimize the input for the next round of retrieval. From the top-ranked (e.g., top 10) similar documents in the current round of search results, high-frequency content words (nouns, verbs) in the titles, abstracts, and claims are extracted, filtered for duplicates and stop words, and added as new keywords. The system compares the <technical features> of the current round of search results with those of the source documents to identify key technical features that did not appear or appeared with low frequency in the current round of results, and these are converted into new keywords. The updated keywords are merged with the original keywords and deduplicated to form the keyword set for the next round.

[0084] 2.4 Rounds 2 to 4: Keyword Search (three rounds in total)

[0085] Based on the updated keywords, three rounds of keyword retrieval (Boolean search) are performed consecutively. After each round, the keywords are updated as input for the next round. "OR" is used to connect synonyms, and "AND" is used to connect core concepts from different dimensions. Priority is given to searching the "title, abstract, and claims" fields; if the number of retrieved results is too small (e.g., less than 50), the search is automatically expanded to full-text search.

[0086] The three rounds of keyword retrieval were performed sequentially, with the results of each round being counted independently in the candidate set. To balance recall efficiency and effectiveness, the three rounds of retrieval used a progressive keyword list: the second round was based on the keywords updated from the first round of semantic retrieval, the third round was based on the keywords updated from the second round, and the fourth round was based on the keywords updated from the third round.

[0087] 2.5 Candidate set merging and deduplication

[0088] The search results from all rounds are merged to form the final candidate set of search results. During the merging process, duplicates are removed based on the file numbers of the comparison documents. If the size of the merged and deduplicated candidate set is too large (e.g., more than N articles), the system can select the top N articles in descending order based on the average similarity across rounds (semantic retrieval uses a similarity score, while keyword retrieval can use a score calculated based on the frequency of hit words).

[0089] S3. Extract technical features from each comparison file and source document in the candidate set and map them into vectors. Filter core feature nodes through two density clustering operations and construct a three-layer technical feature hierarchy structure diagram containing a core layer, a refinement layer, and an edge layer. Extract the three-layer node sets of the source document and comparison file respectively. Calculate the symmetric mean of the bidirectional nearest neighbor distance for the node sets at the same layer using the cross-query average distance algorithm and convert it into a similarity score. Calculate the comprehensive similarity score by weighting the results according to preset weights. Based on this, sort the comparison files in descending order.

[0090] This step mainly implements the ranking based on the contrast of technical features. The core objective of this step is to construct a multi-layer structure graph that reflects the differences in importance and the hierarchy of relationships between features for each comparison file (as well as the source document itself) in the candidate set of search results, so as to provide a structured data foundation for subsequent feature contrast ranking.

[0091] 3.1 Construction of the Technical Feature Comparison Chart

[0092] The core objective of this step is to construct a multi-layered structure diagram that reflects the differences in importance and the hierarchy of relationships between features for each comparison document (as well as the source document itself) in the candidate set of search results, providing a structured data foundation for subsequent feature contrast ranking.

[0093] Traditional methods typically construct the technical features of technical documents into a single-layer feature list, tree structure, or tiled graph structure. These methods have two inherent drawbacks: first, all features are treated equally, failing to distinguish between core and peripheral technical features, leading to interference from numerous non-critical features during subsequent similarity comparisons; second, the lack of a hierarchical organization makes it difficult to reflect the primary and secondary relationships and abstract levels between features, resulting in coarse and inefficient structural similarity measurements.

[0094] To address the aforementioned problems, this invention proposes a method for constructing a hierarchical structure graph of technical features based on density clustering and hierarchical organization. This method uses two density clustering operations to select the core feature node set and, drawing on the concept of hierarchical navigation, constructs a three-layer structure. This allows the structure graph to adaptively highlight core technical features while retaining the association information of fine-grained features, thereby significantly improving the accuracy and robustness of subsequent graph similarity comparisons.

[0095] 3.1.1 Technical Feature Extraction and Vectorization

[0096] First, technical features are extracted from the comparison files. The large model is invoked, and an initial feature list is generated according to the "technical feature" extraction rules defined in step 1, with each feature being a short sentence. The final output feature set is then determined. , The total number of features.

[0097] Each technical feature The mapping is performed as a high-dimensional vector representation for subsequent density clustering and distance calculation. A word embedding model fine-tuned with a corpus of technical documents is used, and all feature vectors are normalized to unit length. Cosine distance is used uniformly for subsequent distance calculations.

[0098] 3.1.2 Two-stage density clustering

[0099] For feature set All vectors are subjected to density-based spatial clustering to identify densely distributed feature regions. The clustering result divides the features into several clusters. This includes isolated points (noise points) that are not assigned to any cluster. Each cluster represents a set of semantically highly similar or closely related technical features. For example, in battery management technical documents, "detecting cell voltage," "detecting cell temperature," and "detecting cell current" may cluster into the same cluster, all belonging to the relatively coarse-grained concept of "cell status sensing."

[0100] In each cluster Within each sub-cluster, the feature node closest to the sub-cluster center is selected as the representative node of that sub-cluster. This process is repeated for each cluster to obtain a set of representative feature nodes. That is, for each cluster Take its center node:

[0101]

[0102] Obtain a representative feature set Isolated points that were not assigned to any cluster in the first clustering were considered non-core features and not included. .

[0103] right Density-based spatial clustering is performed to obtain the core feature node set. The core feature node set selected through two density clustering processes. While the number of features is far less than the total number of original features, it semantically covers representative features across all major technical dimensions. This ensures that the top-level nodes in the subsequently constructed structure graph can highly summarize the core inventive points of the technical document, avoiding the problem in traditional methods where core features are buried among a large number of fragmented features.

[0104] 3.1.3 Constructing a three-tiered technical feature hierarchy diagram

[0105] This invention employs a three-layer structure to organize all technical features. The design concept is: the most core and most generalized feature nodes are placed on the top layer. The lower layers are progressively refined, placing more specific and peripheral features; the connections between nodes are determined by semantic similarity. This results in the construction of a three-tiered structure graph.

[0106]

[0107] in , Represents the node set of each layer. This represents the edges between nodes. The specific construction steps are as follows:

[0108] (1) Determine the hierarchical affiliation of each node

[0109] First layer (top layer): will All feature nodes are placed in the first layer, i.e. ;

[0110] The second and third layers: For the remaining feature nodes, a randomization method is used to determine whether they belong to the second or third layer. Specifically, for each remaining node, a pre-set probability is used... (For example It is assigned to the second layer with probability. The nodes are assigned to the third layer. This random layering method ensures that the structure graph maintains a reasonable distribution of nodes in each layer in a statistical sense, while avoiding the introduction of additional complex optimization processes.

[0111] Nodes in higher-level nodes will always appear in lower-level nodes. For the second and third levels, this can be represented as:

[0112]

[0113] in, This represents the entire feature point set, and RandSubSet() represents a random selection function.

[0114] like Figure 2 The hierarchical architecture diagram of technical feature contrast is given, from top to bottom: L1 (top layer features, sparsest, most representative features), L2 (middle layer), and L3 (bottom layer features, densest, containing all features).

[0115] (2) Constructing edge connections within the layer

[0116] Within each layer, for each node, an approximate nearest neighbor search is used to find the node most similar to itself in the vector space within that layer. Establish bidirectional edges with other nodes. This connection method allows similar features within the same layer to be correlated, forming local semantic clusters.

[0117] That is, for each layer For each node The approximate nearest neighbor search is used to find the m nodes in the layer that are most similar to its vector. Establish a bidirectional edge:

[0118]

[0119] final:

[0120]

[0121] (3) Construction Results

[0122] Ultimately, each technical document (source document or comparison document) is represented as a three-tiered technical feature hierarchy diagram:

[0123] The first layer consists of a set of core feature nodes (obtained through two density clustering processes). These nodes are few in number and have strong semantic generalization, representing the main technical dimensions of the technical document.

[0124] The second layer consists of medium-granularity feature nodes, which refine and supplement the core features to a certain extent.

[0125] The third layer consists of fine-grained or edge feature nodes, which contain more specific implementation details or minor technical points.

[0126] Edges between nodes include both horizontal associations between semantically similar features at the same level and vertical anchoring between lower-level features and upper-level core features.

[0127] Combining the formulas above, we can obtain the overall construction formula for the three-layer structure diagram:

[0128]

[0129] The advantages of this construction method compared to traditional feature trees / feature maps are shown in Table 2:

[0130] Table 2

[0131]

[0132] 3.2 Similarity Assessment of Technical Feature Hierarchical Structure Diagram

[0133] The core objective of this step is to calculate the structural similarity score between the source document and the comparison document based on the three-layer technical feature hierarchy diagram constructed in step 3.1, and then sort the comparison documents in the candidate set in reverse order according to this score. The higher the similarity, the closer the technical features of the comparison document are to those of the source document, and the higher it should be ranked.

[0134] Traditional methods typically employ algorithms such as graph edit distance or maximum common subgraph to calculate the similarity between two graph structures. While these methods are theoretically complete, their high computational complexity (often NP-hard) makes them unsuitable for real-time ranking of hundreds of comparative documents in technical document retrieval scenarios. This invention introduces a similarity evaluation method based on cross-query average distance. This method fully leverages the vectorized nature of nodes in a three-layer graph structure, transforming the graph structure similarity problem into a distributional similarity problem between vector sets, achieving efficient computation while ensuring evaluation accuracy.

[0135] 3.2.1 Overall Computational Framework

[0136] For the structure diagram of the source document Structure diagram of the comparison file Similarity assessment is divided into the following three levels:

[0137] First-layer (core layer) similarity calculation: Only the first-layer core feature node set of the two images is compared to quickly determine the degree of similarity between the two technical documents on the core invention points.

[0138] Similarity calculation between the second and third layers (refinement layers): The node sets of the second and third layers are compared respectively to evaluate the degree of similarity at the refinement and edge feature levels.

[0139] Comprehensive similarity calculation: The three levels of similarity are weighted and summed according to preset weights to obtain the final comprehensive similarity score.

[0140] The advantage of hierarchical calculation is that different weights can be assigned to different needs during subsequent ranking. For example, technical difference assessment focuses more on the similarity of core technical features, so the weight of the first layer can be appropriately increased; while infringement risk screening may need to take into account the matching of fine-grained features.

[0141] 3.2.2 Calculation of Average Distance for Cross-Query of a Single-Layer Node Set

[0142] For any two node sets X (from the source document) and Y (from the comparison document), the similarity between them is calculated using the cross-query average distance. The core idea of ​​this method is: if the contents of the two sets are similar, then the distance to the nearest neighbor node found in set Y starting from a node in set X should be small; conversely, if the two sets are significantly different, then the distance will be large.

[0143] (1) Calculate the forward distance Dist(X→Y)

[0144] Randomly sample N nodes from set X as the query set (N is min(|X|, 10000); if the set size is small, use all nodes). For each query node, find the node closest to it in the index of Y and record the distance between them. Calculate the average of all query distances to obtain Dist(X→Y):

[0145]

[0146] in, This represents a set of nodes randomly sampled from X. Represents the cosine distance between two nodes

[0147] (2) Calculate the reverse distance Dist(Y→X)

[0148] Symmetrically, sample M nodes from Y, query the nearest neighbor distance in the index of X and calculate the average, to obtain Dist(Y→X).

[0149] (3) Normalization

[0150] Take the average of the distances in both directions as the final metric:

[0151]

[0152] The smaller the value, the more similar the characteristics of the two sets.

[0153] Convert distance into a similarity score to conform to the intuitive principle that "the higher the score, the more similar the similarity":

[0154]

[0155] The converted score falls within the interval (0,1], and the smaller the distance, the closer the score is to 1.

[0156] 3.2.3 Weighted Comprehensive Similarity of Three-Layer Structure

[0157] Calculate the source documents separately Compared with the comparison piece The similarity scores are calculated between the node sets in the first, second, and third layers. The final comprehensive similarity score is a weighted sum of the similarities across the three layers:

[0158]

[0159] Among them, weight satisfy The weight configuration can be flexibly adjusted according to the application scenario. Substituting the formula for calculating the average distance of cross-queries in a single-layer node, we can obtain:

[0160] in, for A subset of query nodes randomly sampled from the k-th layer node set, y for The system randomly samples a subset of query nodes from the k-th layer node set. Users (technical document analysts or technical document editors) can manually adjust the weight configuration on the interface to adapt to different search needs.

[0161] 3.3 Sort in descending order based on similarity score

[0162] After calculating the overall similarity score between the source document and each comparison file in the candidate set, the system sorts all comparison files according to... The scores are sorted in descending order, with higher scores indicating a closer overall structural similarity between the comparison document and the source document in terms of technical features. The top-ranked comparison documents are prioritized for step 4, where technical difference assessment opinions are generated. Simultaneously, the system retains three levels of detailed similarity scores for each comparison document. This is for users to view and to help them understand the specific source of the similarity.

[0163] S4. Select highly similar comparison files from the sorting results, filter and classify node features similar to the source documents, and combine them with structured description generation technology to provide difference assessment suggestions.

[0164] This step primarily generates recommendations for assessing technological differences.

[0165] Based on the top-ranked comparison files, the system automatically generates draft assessment opinions to assist technical document analysts or editors in judging technical differences.

[0166] Extract similar features from highly similar comparison files: Select the top N comparison files by similarity from the ranking results in step 3. For each comparison file, traverse all nodes in its three-layer technical feature hierarchy diagram and filter out feature nodes whose distance to the corresponding layer node of the source document is less than a preset threshold (i.e., similarity greater than the mean). Categorize and organize the selected similar feature nodes according to their respective layers (core layer / refinement layer) to form a similar feature set, and record the source of the comparison file corresponding to each similar feature.

[0167] Generate technical difference assessment opinions: Input the six-tuple of the source document along with the aforementioned similarity feature set into the large language model. Use dedicated prompt words to guide the model in generating structured assessment opinions. The output draft opinions are provided for technical document analysts or editors to refer to and revise, and are not considered final conclusions.

[0168] In summary, this application addresses the following core issues:

[0169] Limitations of semantic understanding of technical features: Traditional keyword matching (such as TF-IDF, BM25) only stays at the level of word co-occurrence and cannot handle synonyms (such as "voltage detection" and "potential measurement") or near-synonyms, leading to missed detections or false negatives. This solution aims to break down word barriers and achieve matching of the essence of technology through deep semantic analysis.

[0170] Loss of structural information and submersion of core features: Existing vector semantic retrieval typically compresses the entire technical document into a single global vector or uses a flat feature list. This approach loses the hierarchical structure and logical connections between technical features, causing core technical features to be treated equally with marginal and non-critical features in similarity calculations, severely interfering with the accuracy of ranking.

[0171] The ranking results lack interpretability and are inefficient: Traditional comparison methods based on graph edit distance or maximum common subgraph can reflect structural similarity, but their computational complexity is NP-hard, making them unsuitable for real-time requirements. Furthermore, existing methods often only provide a similarity score, failing to intuitively answer "on which specific technical features are similar between the two technical documents," thus lacking useful support for decision-making.

[0172] To rank search results, a ranking algorithm based on feature contrast is designed. This algorithm constructs a "technical feature hierarchical structure graph," transforming technical documents into a multi-layered graph containing core and refined layers, thereby accurately identifying and highlighting core technical features. Simultaneously, it replaces traditional graph similarity calculation with a "cross-query average distance" algorithm, enabling rapid calculation of document similarity.

[0173] Simultaneously, an agent-based intelligent technical document retrieval system was designed. This system not only includes the aforementioned ranking algorithm but also integrates multi-round retrieval strategies (semantic + keyword iteration) and structured description generation. It efficiently assesses the technical differences of source documents, truly realizing the transformation from "manual needle-in-a-haystack" to "AI-precise push."

[0174] like Figure 3 This application also provides a hierarchical structured technical feature contrast analysis system, which may include:

[0175] The input module is used to input the technical documents entered by the user into the large language model and generate a structured description that includes technical topics, technical problems, technical solutions, technical effects, technical features and technical summaries.

[0176] The module is used to perform iterative retrieval combining semantic retrieval and multi-round keyword retrieval based on the structured description, and to construct a candidate set of retrieval results after merging and deduplication;

[0177] The sorting module is used to extract technical features from each comparison file and source document in the candidate set and map them into vectors. It filters core feature nodes through two density clustering operations and constructs a three-layer technical feature hierarchy structure diagram containing a core layer, a refinement layer, and an edge layer. It extracts the three-layer node sets of the source document and comparison file respectively. It calculates the symmetric mean of the bidirectional nearest neighbor distance for the node sets at the same layer using the cross-query average distance algorithm and converts it into a similarity score. It then calculates the comprehensive similarity score by weighting the nodes according to a preset weight and sorts the comparison files in descending order accordingly.

[0178] The generation module is used to select highly similar comparison files from the sorting results, filter and classify node features similar to the source documents, and combine the structured description to generate difference assessment suggestions.

[0179] Specific limitations regarding the hierarchical structured technical feature contrast analysis system can be found in the limitations of the hierarchical structured technical feature contrast analysis method described above, and will not be repeated here. Each module in the aforementioned hierarchical structured technical feature contrast analysis system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0180] In one embodiment, an electronic device, which may be a computer, is provided. The electronic device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device is used for technical feature contrast analysis data. The network interface of the computer device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a technical feature contrast analysis method.

[0181] In one embodiment of this application, a computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of the above-described hierarchical structure-based technical feature contrast analysis method.

[0182] In one embodiment of this application, a computer program product is provided, including a computer program / instructions, which, when executed by a processor, implements the steps of the above-described hierarchical structure-based technical feature contrast analysis method.

[0183] The computer-readable storage medium and computer program product provided in this embodiment are similar in implementation principle and technical effect to the above method embodiments, and will not be described again here.

[0184] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods.

[0185] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0186] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for comparing technical features based on hierarchical structure, characterized in that, The method includes: S1. Input the technical document input by the user into the large language model to generate a structured description that includes technical topics, technical problems, technical solutions, technical effects, technical features and technical summaries; S2. Based on the structured description, perform an iterative retrieval that combines semantic retrieval with multi-round keyword retrieval, and construct a candidate set of retrieval results after merging and deduplication; S3. Extract technical features from each comparison file and source document in the candidate set and map them into vectors. Filter core feature nodes through two density clustering operations to construct a three-layer technical feature hierarchy structure diagram containing a core layer, a refinement layer, and an edge layer. Extract the three-layer node sets of the source document and comparison file respectively. Calculate the symmetric mean of the bidirectional nearest neighbor distance for the node sets at the same layer using the cross-query average distance algorithm and convert it into a similarity score. Calculate the comprehensive similarity score by weighting the results according to preset weights. Based on this, sort the comparison files in descending order. S4. Select high similarity comparison files from the sorting results, filter and classify node features similar to the source documents, and combine them with the difference assessment suggestions of the structured description generation technology.

2. The method for technical feature contrast analysis based on hierarchical structure according to claim 1, characterized in that, After generating a structured description in S1 that includes the technical topic, technical problem, technical solution, technical effect, technical features, and technical summary, it also includes an integrity check: The integrity check includes checking whether there are empty or obviously unreasonable items in the six-tuple. If so, it automatically triggers a supplementary call, uses prompt words to retry, and provides the missing input material fragments back to the large language model.

3. The method for technical feature contrast analysis based on hierarchical structure according to claim 1, characterized in that, The iterative retrieval described in S2 includes: Based on the structured description, core words are extracted and synonyms are expanded to generate initial search terms, and the first round of semantic retrieval is performed; Extract high-frequency content words from the current search results and identify key technical features that are not present by comparing them with the technical features of the source documents. Update the keywords accordingly and then execute the next round of keyword search. Repeat the update and retrieval process until the preset round is reached, then merge the results from each round, remove duplicates, and form a candidate set.

4. The method for technical feature contrast analysis based on hierarchical structure according to claim 1, characterized in that, The two density clustering operations mentioned in S3 include: Perform the first density-based spatial clustering on the feature set, dividing the features into several clusters and isolated points that are not assigned to any cluster; Within each cluster, the node closest to the center is selected as the representative node to obtain a set of representative feature nodes. The isolated point is not included in this set. A second density-based spatial clustering is performed on the representative feature node set to filter out the core feature node set.

5. The method for technical feature contrast analysis based on hierarchical structure according to claim 1, characterized in that, The construction of the three-layer technical feature hierarchy diagram described in S3 includes: Place the core feature node set into the first layer; The remaining nodes are assigned to the second layer with a preset probability and to the third layer with a complementary probability, and nodes in the upper layer must appear in the lower layer; Within each layer, for each node, an approximate nearest neighbor search is used to find the m other nodes in that layer whose vectors are most similar to its own, and bidirectional edges are established.

6. The method for technical feature contrast analysis based on hierarchical structure according to claim 1, characterized in that, The cross-query average distance algorithm described in S3 includes: For any two sets of nodes at the same level, X and Y, randomly sample N nodes from X, find the node with the closest cosine distance in the vector index of Y and record the distance, calculate the average to obtain the forward distance Dist(X→Y); symmetrically sample nodes from Y and calculate the reverse distance Dist(Y→X) in the index of X. Calculate the symmetrical average distance: ； And convert it to a similarity score: 。 7. The method for technical feature contrast analysis based on hierarchical structure according to claim 1, characterized in that, The weighted summation according to preset weights described in S3 includes: calculating the similarity scores between the source document and the comparison document between the node sets of the first, second, and third layers, respectively. According to the formula: Calculate the overall similarity score, where the weights are... satisfy .

8. The method for technical feature contrast analysis based on hierarchical structure according to claim 1, characterized in that, S4 specifically includes: Select the top N highly similar comparison files from the sorting results, traverse the nodes in their three-layer technical feature hierarchy structure diagram, filter out similar features whose distance from the corresponding layer node of the source document is less than a preset threshold, and classify them into similar feature sets according to the core layer and the refinement layer. The structured description and similar feature set are input into a large language model, and structured technical difference assessment opinions are generated through dedicated prompt words.

9. A technical feature contrast analysis system based on hierarchical structure, characterized in that, The system includes: The input module is used to input the technical documents entered by the user into the large language model and generate a structured description that includes technical topics, technical problems, technical solutions, technical effects, technical features and technical summaries. The construction module is used to perform iterative retrieval combining semantic retrieval and multi-round keyword retrieval based on the structured description, and construct a candidate set of retrieval results after merging and deduplication; The sorting module is used to extract technical features from each comparison file and source document in the candidate set and map them into vectors. It filters core feature nodes through two density clustering operations and constructs a three-layer technical feature hierarchy structure diagram containing a core layer, a refinement layer, and an edge layer. It extracts the three-layer node sets of the source document and comparison file respectively. It calculates the symmetric mean of the bidirectional nearest neighbor distance for the node sets at the same layer using the cross-query average distance algorithm and converts it into a similarity score. It then calculates the comprehensive similarity score by weighting the nodes according to a preset weight and sorts the comparison files in descending order accordingly. The generation module is used to select highly similar comparison files from the sorting results, filter and classify node features similar to the source documents, and combine the structured description generation technology to generate difference assessment suggestions.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the hierarchical structured technical feature contrast analysis method as described in any one of claims 1 to 7.

Images