Patent data analysis method based on dynamic context window

Abstract

The invention discloses a patent data analysis method based on a dynamic context window, which belongs to the technical field of text data analysis, and comprises the following steps: sorting patent documents; decomposing each patent document into a plurality of semantic blocks, and outputting a block set after algorithm processing; selecting the block sets processed by the two algorithms, and calculating and storing the technical association probability between the two segments; soft clustering is carried out on all semantic blocks according to the technology association probability, and a technology theme membership degree vector is obtained; constructing retrieval numbers for the semantic blocks on the basis of technical topic membership degree vectors; establishing a quick retrieval tree; and according to the size of the context window of each semantic block, retrieving and expanding the context of the corresponding semantic block by adopting the quick retrieval tree, and finally outputting an expanded result. According to the patent data analysis method based on the dynamic context window, the problem that the analysis effect is poor due to the fact that an existing patent data analysis mode is limited by an upper window and a lower window is solved.

Description

Technical Field

[0001] This invention relates to the field of text data analysis technology, and in particular to a patented data analysis method based on a dynamic context window. Background Technology

[0002] In today's information explosion era, the quantity and complexity of patent documents are constantly increasing. Therefore, a method based on context windows has emerged to analyze massive amounts of patent data. However, existing analysis methods have the following shortcomings: 1. Incomplete Capture of Technical Information: Patent documents typically include detailed technical specifications, claims, and background descriptions, all closely interconnected. A limited context window may fail to capture this information completely, leading to incomplete analysis, especially when identifying the novelty and inventiveness of a patent. 2. Impact on Patent Infringement Analysis: When conducting patent infringement analysis, analysts need to compare patent claims with other documents and products in detail. If the window length is insufficient, incomplete comparison information will result, affecting the accuracy of the analysis and the correctness of legal judgments. 3. Reduced Effectiveness of Patent Portfolio Analysis: Companies often need to analyze patent portfolios consisting of a series of related patents to formulate R&D strategies or assess competitors' technological strength. A limited context window makes it difficult to cover all relevant patents, resulting in insufficient strategic basis. 4. Constraints on Identifying Technology Trends: In technology trend analysis, researchers need to extract the direction and key areas of technological development from a large amount of patent data. Limited context windows may lead to overly coarse analysis granularity, making it impossible to predict and identify industry trends in a timely and accurate manner. 5. Reduced semantic understanding and reasoning ability: For complex patent documents, the limited window leads to the truncation of deep semantic information and implicit reasoning chains within the document, affecting the reasoning ability of analysis tools, such as the identification of inventiveness and the assessment of patent value.

[0003] In summary, existing methods for analyzing patent data based on context windows suffer from limitations in the upper and lower windows, resulting in poor analysis performance. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention provides a patent data analysis method based on dynamic context windows to solve the above-mentioned problems.

[0005] The technical solution adopted by this invention to solve its technical problem is: a patent data analysis method based on a dynamic context window, comprising the following steps: S1: Obtain the technical features, technical fields, and innovative points of the analysis target, extract field features by combining the patent classification number, and obtain the query volume q; sort the patent document d based on the similarity between the patent document d and the query volume q. S2: Apply a recursive block decomposition algorithm to each patent document d in sorted order, decomposing it into multiple semantic blocks (chunks) with hierarchical relationships, denoted as... The process continues until a preset minimum chunk is reached, and each semantic chunk is assigned a unique number ChunkID, which is recorded in the metadata of that semantic chunk. The number is used to establish contextual reference relationships between chunks. Then, a hierarchical reduction algorithm is applied to output the chunk set after the algorithm is processed, so as to maintain the document structure of each patent document d. S3: Select two sets of blocks processed by different algorithms, and then arbitrarily select one semantic block from each of the two sets of blocks processed by the algorithms to form a second semantic block pair. The two different semantic blocks that form the second semantic block pair are respectively regarded as fragments s1 and s2. Calculate and store the technical association probability between the two fragments. Then, based on the technical association probability, all semantic blocks are soft-clustered to obtain a technical topic membership vector containing the probability that each semantic block belongs to each preset technical topic; S4: Construct retrieval numbers for semantic blocks based on technical topic membership vectors; establish a fast retrieval tree based on the constructed retrieval numbers; S5: Select semantic blocks in order of importance, then determine the context window size of each semantic block according to the dynamic adjustment algorithm, and use a fast search tree to retrieve and expand the context of the corresponding semantic block according to the context window size of each semantic block, and finally output the expanded result.

[0006] The beneficial effects of the present invention are as follows: the patent data analysis method based on dynamic context window generates pseudo-documents and combines semantic segmentation and soft clustering techniques to achieve dynamic context window management, which solves the efficiency problem in processing multiple patent documents and improves the quality and reliability of search results. Attached Figure Description

[0007] Figure 1 This is a flowchart of a patent data analysis method based on a dynamic context window, as described in one embodiment of the present invention. Detailed Implementation

[0008] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that these descriptions are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0009] like Figure 1 As shown, a patent data analysis method based on a dynamic context window includes the following steps: S1: Obtain the technical features, technical fields, and innovative points of the analysis target, and extract domain features by combining the patent classification number (IPC / CPC) to obtain the query volume q; rank the patent document d based on the similarity between the patent document d and the query volume q; In step S1, a dual-perspective retrieval paradigm is adopted by combining coarse-grained and fine-grained levels to obtain text similarity and technical relevance, and the similarity between patent document d and query volume q is obtained by fine-tuning the data construction pipeline. Step S1 includes: S11: Calculate the text similarity between the query volume q and the patent document d: Text similarity ; and All are fusion weight parameters, among which , ; The preferred value range is 0.3-0.4. When the document structure is complete and the overall integrity is strong, the value can be appropriately increased. When processing short documents, the value of can be slightly reduced. The possible values ​​of ; The preferred value range is 0.6-0.7. The more detailed the technical feature description, the better. The higher the value, the better it can be when precise matching is required. The possible values ​​of ; coarse-grained text similarity ; Fine-grained text similarity ; For the query volume q, the text coarse-grained level, For the text of patent document d, coarse-grained For the i-th fine-grained unit of query quantity q, the text fine-grainedness is... q1 represents the text fineness of the i-th fine-grained unit in patent document d; N is the number of fine-grained units, qi is the i-th fine-grained unit in query q, and di is the i-th fine-grained unit in patent document d; when i=1, q1 is the first fine-grained unit in query q, representing specific technical parameters, and d1 is the first fine-grained unit in patent document d, representing specific technical features of independent claims; when i=2, q2 is the second fine-grained unit in query q, representing specific technical implementation methods, and d2 is the second fine-grained unit in patent document d, representing additional features of dependent claims; when i=3, q3 is the third fine-grained unit in query q, representing key technical components, and d3 is the third fine-grained unit in patent document d, representing specific steps of the technical solution; sim() is the similarity function, and cos() represents cosine similarity, which is an index used to measure the degree of similarity between two vectors. ,in Represents the vector dot product. and The Euclidean norm (length) of a vector. It is the angle between two vectors; In this embodiment, the text similarity in the text representation refers to the textual description of technical features, technical fields, and innovative points. S12: Calculate the technical similarity between query volume q and patent document d: Technical similarity ; and All are fusion weight parameters, among which , ; The preferred value range is 0.4-0.5, and its influencing factors are the overall classification of the technical field and the IPC / CPC classification level; The preferred value range is 0.5-0.6, and its influencing factors are the depth of description of technical features and the specific implementation details of innovative points; coarse-grained technical similarity ; Fine-grained technical similarity ; For query volume q, the technical coarse-grained level. For the technology of patent document d, coarse-grained The technical fineness is the i-th fine-grained unit of the query quantity q. is the technical fineness of the i-th fine-grained unit in patent document d; N is the number of fine-grained units, qi is the i-th fine-grained unit in query volume q, di is the i-th fine-grained unit in patent document d, sim() is the similarity function; cos() represents cosine similarity. S13: Calculate the overall relevance between query volume q and patent document d. ; and All are fusion weight parameters, among which , ; S14: After constructing the pipeline based on the overall relevance, fine-tune the data and calculate the similarity between each patent document d and the query volume q: First, we build evaluation data for the fine-tuning pipeline: ; q represents the query volume q (query): a description of the query target, usually input by the user, containing descriptive information such as technical features, technical field, and innovation points, and is the core content used to calculate the relevance of patent documents; d represents the patent document d (document): a specific document in the patent database, usually the patent text to be evaluated, including information such as patent claims, abstract, specification, and classification number (IPC / CPC); y represents the ground truth relevance: a label indicating the ground truth relevance between the query volume q and the patent document d. y represents the overall relevance obtained in step S13. ; Then, the predicted relevance score is calculated as the similarity between each patent document d and the query volume q: ; ; ; ; ; Let be the fine-grained unit in patent document d that is most similar to the i-th fine-grained unit of query quantity q. This refers to the fine-grained text of the fine-grained unit in patent document d. This represents the technical fineness of the fine-grained unit in patent document d; sim() is the similarity function; cos() represents the cosine similarity. Calculate the mean square error The mean squared error (MSE) is used as the similarity between patent document d and query volume q; N0 represents the evaluation dataset. The number of samples, i.e., the fine-grained partitioning of query-document pairs; An automatic weight adjustment mechanism is used to update the fusion weight parameters. ; ; ; ; ; ; ; ; ; ; ; ; ; in The partial derivative of the loss function, , For relative to the fusion weight parameters The change The fusion weight parameters in the current iteration step The value, For the updated fusion weight parameters The value, For relative to the fusion weight parameters The change The fusion weight parameters in the current iteration step The value, For the updated fusion weight parameters The value, For relative to the fusion weight parameters The change The fusion weight parameters in the current iteration step The value, For the updated fusion weight parameters The value, For relative to the fusion weight parameters The change The fusion weight parameters in the current iteration step The value, For the updated fusion weight parameters The value, For relative to the fusion weight parameters The change The fusion weight parameters in the current iteration step The value, For the updated fusion weight parameters The value, For relative to the fusion weight parameters The change The fusion weight parameters in the current iteration step The value, For the updated fusion weight parameters The value, The learning rate; It is the mean function, used to calculate The average value; This is the standard deviation adjustment factor, used to control the degree of influence of the fluctuation range; This is the standard deviation function, used to calculate... Standard deviation; This is a threshold value used to determine the correlation. This is the mean adjustment coefficient, used to control the degree of influence of the mean; threshold. As a performance monitoring metric, it is used to determine termination conditions. The specific process is as follows: Step 1, Initialization: Randomly initialize the weight parameters. , , , , and Set the learning rate and the maximum number of iterations, MaxEpochs; Step 2, Iterative training (For each epoch from 1 to MaxEpochs): Take a batch of data from the evaluation data D_eval; Calculate the Loss (i.e., MSE): based on the current weight parameters. , , , , and The mean square error (MSE) is calculated as the loss. c. Calculate the gradient: Calculate the loss with respect to each weight parameter using the backpropagation algorithm. , , , , and The partial derivative of ) , , , , and ); d. Update weights: Update the weights according to the gradient descent formula, for example... .

[0010] (Key Steps) Performance Evaluation and Threshold Application: Use all predicted scores from the current batch Calculate dynamic threshold .

[0011] e. will and Comparison: If >= If the result is positive, the model determines it to be relevant (positive class); otherwise, it is irrelevant (negative class), and this result is used as the prediction result. f. Compare the real label y with Comparison: If y>= If the predicted result is positive, it is considered relevant (positive class); otherwise, it is considered irrelevant (negative class), and this result is taken as the actual judgment result. The accuracy, precision, or F1 score of the current batch is calculated based on the predicted judgment result and the actual judgment result. Check termination conditions: Condition 1: Reaching the maximum number of iterations: Training will automatically stop after reaching the preset maximum number of iterations, MaxEpochs. Condition 2: Performance no longer improves: If, in several consecutive iterations, the evaluated performance metrics (such as accuracy, precision, or F1 score) no longer improve significantly, or even begin to decline, training is terminated early to prevent model overfitting. Condition 3: Loss function convergence: If the value of Loss is already very small (less than the preset threshold) and hardly decreases in multiple iterations, it means that the model has reached a stable state and can be stopped.

[0012] S2: Apply a recursive block decomposition algorithm to each patent document d in sorted order, decomposing it into multiple semantic blocks (chunks) with hierarchical relationships, denoted as... The process continues until a preset minimum segmentation is reached, and each semantic segment is assigned a unique number ChunkID, which is recorded in the metadata of that semantic segment. The number is used to establish contextual reference relationships between segments. Then, a hierarchical reduction algorithm is applied to output the processed segmentation set to maintain the document structure of each patent document d; thus, the semantic segmentation processing of patent documents is achieved. In step S2, for each patent document d, an initial block set corresponding to each patent document d is first obtained using a recursive block partitioning algorithm, wherein the recursive block partitioning algorithm includes: ; ; in ; ; ; ; ; Where D represents the document that needs to be recursively divided into blocks, which can be patent document d, or sub-documents resulting from multiple block divisions of patent document d. The block size is the maximum semantic block size. Indicates for all , For the first Each semantic block is the result of semantic segmentation of the sub-document; for , This indicates the semantic chunks generated. All constraints must be met; specifically... This means for each semantic block Its size (e.g., number of characters or words) cannot exceed the preset maximum value. This ensures that the blocks are not too large, meeting the requirement of "smallest semantic unit"; and The `Split()` function is a temporary sub-document executed each time the recursive block splitting algorithm is run. `Split(D)` is a function that performs block splitting on the document `D` that needs to be recursively split. `∪` is the union operation of sets. `Split()` is the core of the recursive block splitting algorithm; it shows that the process of splitting a document `D` that needs to be recursively split can be decomposed into: first, finding an optimal splitting point to divide `D` in two, resulting in... and Then, respectively for and Calling its own Split function, which is to... As D in Split(D), or as As D in Split(D), it continues to be divided into smaller parts; finally, all the smallest blocks produced by all subprocesses are merged (the union is taken) to obtain the final block result of the patent document d, and the smallest block at this point is the semantic block. ; middle, This represents the best split point for document D that needs to be recursively divided into chunks. This indicates that an index ix is ​​found such that the value of the subsequent expression Score() is maximized; This refers to the portion from the beginning of document D to position ix. This represents the portion from position ix to the end of document D; This formula describes how to find the optimal position to split a document D that needs to be recursively divided into two parts. The algorithm iterates through all possible positions ix in document D, imaginarily splitting document D into two parts at each position. and The process involves two parts; then, a scoring function called Score is invoked to evaluate the quality of the split; finally, the split position ix that gives the highest score to the Score function is selected as the best split point B(D) for this recursion. The scoring function Score is obtained through... In this calculation, Coherence represents "coherence" or "cohesion"; cos is the cosine similarity function; this is part of the scoring function Score, used to measure the two parts after splitting. and The degree of semantic association; it is calculated by computing the vector representations of two text blocks. and Coherence is achieved through the cosine similarity between them; a high Coherence score (close to 1) means and If the semantics are very similar, it means that the segmentation point may destroy a complete semantic unit, and therefore it is a bad segmentation; conversely, a low Coherence score (close to 0 or negative) means that the semantics of the two are very different, indicating that the segmentation point may be located between two different technical topics, and is a good segmentation point. and They respectively represent the following: Functions that convert to a mathematical vector and convert... This function converts text into a mathematical vector; it's a process of text vectorization or embedding; in the context of the patent, this function will convert a piece of text (such as...) into a mathematical vector. or The semantic information of text is compressed into a fixed-dimensional numerical vector, so that computers can determine the semantic relationship between texts by calculating the mathematical relationship between the vectors (such as cosine similarity). In Chinese, Balance means "balance". express Length, express Length; It is another part of the scoring function Score, used to evaluate the two parts after splitting. and Are they approximately equal in length? The absolute value of the difference between the lengths of the two parts was calculated, by... Normalize this difference; when the two parts are exactly the same length (difference is 0), the Balance score is 1 (most balanced); when one part is much longer than the other, the Balance score is close to 0 (most unbalanced); the purpose of introducing this term is to avoid generating an invalid split with an extremely long and an extremely short length, and to encourage the algorithm to split from the middle as much as possible; α and β are used to calculate The weighting parameters; α and β control semantic coherence and length balance in the scoring, respectively. The importance of α and β can be adjusted to control the semantic segmentation strategy: if α is set much larger than β, the algorithm will be more inclined to find the least semantically related segmentation point, even if the lengths of the two segments differ greatly; if β is set much larger than α, the algorithm will be more inclined to segment evenly from the middle, even if this may interrupt a complete sentence or paragraph; combining the two aims to find a segmentation point that is relatively optimal in both semantics and structure.

[0013] In this embodiment, the hierarchical reduction algorithm uses the Hierarchical Reduction algorithm, which includes: Hierarchical definition: Hierarchy ; Block set ; Hierarchical attributes of each semantic block: ; Reduction rules: Merging conditions for adjacent semantic blocks: ; Overall similarity calculation: ; in In Indicates the first semantic blocks , Indicates the first semantic blocks , This condition means Must equal or In other words, semantic chunking and semantic chunking The blocks are adjacent in the sequence, which ensures that the algorithm will only consider merging physically adjacent semantic blocks, and will not skip steps in merging. Semantic blocks and semantic chunking They must be at the same level in the current tree hierarchy. This is a key prerequisite for merging. The algorithm is built up layer by layer, and it will not merge a semantic block at level 3 and a semantic block at level 4. The acquisition process: Step 1, Initial Level (Level 0): All semantic blocks obtained from the initial block set generated by the recursive block division algorithm are initialized to the lowest level, i.e., level 0. Step 2, hierarchical aggregation to generate a new hierarchy: when two adjacent blocks belonging to level 0 are semantically segmented... and When they are merged based on the merging condition (sufficient similarity), they will form a new parent node P. The level of this newly generated parent node P will be defined as a higher level, namely level1. Step 3, layer by layer upward: The algorithm continues this process, constantly merging adjacent nodes of level 1 into nodes of level 2, and so on, building the entire hierarchical structure tree from bottom to top; for content_sim (content / text similarity): measures the similarity between two semantic chunks. and The degree of similarity in literal content and semantics is usually achieved by calculating the cosine similarity of their text vectors (e.g., using Sentence-BERT). A high content_sim means that the two semantic chunks are talking about the same thing; structure_sim (structural similarity): measures the similarity between two semantic chunks. and The similarity of their roles in the document structure, for example, they may both belong to "specific embodiments of claim 2" or be steps in a certain technical process. This similarity can be calculated based on their parent nodes, depth, and other attributes in the initial tree structure; tech_sim (technical similarity): measures the similarity between two semantic blocks. and The similarity of the core technical concepts involved is usually achieved by comparing their technical classification numbers (IPC / CPC) and the set of extracted technical keywords. A high tech_sim means that they are discussing the same or very related technical points.

[0014] In step S2, the hierarchical reduction algorithm is a hierarchical reduction algorithm, which includes: Initial hierarchy: Obtain the initial set of blocks from the recursive block partitioning algorithm. For this initial set of blocks, the initial set of blocks is used as the leaf node for hierarchical construction to build the initial tree hierarchy. Hierarchical aggregation: Based on the similarity between the patent document d obtained in step S1 and the query volume q, the similarity between adjacent blocks is identified, and similar blocks are merged layer by layer to form a parent node until the preset simplification level is met; among them, adjacent semantic blocks with text similarity ≥ 0.70, technical similarity ≥ 0.80, structural similarity ≥ 0.65 and comprehensive similarity ≥ 0.75 are considered similar blocks; Semantic clustering: The alignment results of the local alignment algorithm (output of locally optimal aligned semantic block pairs) guide the merging of highly similar blocks to preserve semantic integrity; The guidance method is to ensure that the alignment path is continuous, the matching pattern is consistent, the boundary points are complete, and the semantic unit correlation degree is greater than 0.7; The local alignment algorithm is the Smith-Waterman algorithm, which specifically involves obtaining the initial block set corresponding to each patent document d from a recursive block partitioning algorithm. When performing local alignment using a local alignment algorithm, it is possible to select two semantic blocks from different initial block sets that require detailed comparison. Align the first semantic blocks that make up the structure. Each pair of first semantic blocks is calculated using a scoring function. The similarity of word vectors is evaluated by a score. A scoring function is used to reflect the similarity strength and matching position of each pair of first semantic blocks. After calculating the local alignment similarity strength using a core formula, a similarity matrix is ​​constructed as the basis for alignment. A local alignment algorithm is used to find the first semantic block pair with the highest local score in the similarity matrix and backtracks to obtain the alignment path, which is taken as the optimal alignment path. All first semantic block pairs in the alignment path are identified as the locally optimal aligned first semantic block pairs to reflect the local sequence matching obtained by the alignment operation in the similarity matrix. The alignment score of each first semantic block pair in the optimal alignment path is output to measure the matching quality and similarity between semantic blocks in each pair of first semantic blocks.

[0015] In this embodiment, the local alignment algorithm uses Smith-Waterman to find the optimal locally similar semantic blocks between two sequences. Here, the two sequences represent the initial block sets of two different patent documents d obtained through a recursive block partitioning algorithm. The optimal alignment scheme is calculated using a dynamic programming matrix. The local alignment algorithm includes: Core formula: Scoring function: during matching ; When mismatch ; Open space penalty: ; H(i, j) represents the highest possible score among all local alignment paths ending at the i-th unit (xi) of sequence X and the j-th unit (yj) of sequence Y, i.e., the alignment score, used to represent the similarity strength. It is not necessarily the global "best local similarity score," but rather the locally optimal solution up to the point (i, j). During the local alignment algorithm, the two selected initial block sets are labeled as sequence X and sequence Y, respectively. The i-th unit (xi) of sequence X is the i-th semantic block in the corresponding initial block set, and the j-th unit (yj) of sequence Y is the i-th semantic block in the corresponding initial block set.

[0016] in, A match is defined as two sequences xi and yj being equal, and a pre-set matching score will be assigned to them. , This indicates a mismatch when the sequences xi and yj are not equal, and a pre-defined mismatch score will be assigned to them. ; The gap penalty is the cost of introducing a "gap" in a sequence to allow other parts to match better when aligning two sequences. In the context of a patent document, this is like saying, "In order to align paragraphs 5, 6, and 7 of document 1 with paragraphs 8, 9, and 10 of document 2, we have to skip (i.e., introduce a gap) paragraph 4 of document 1." This "skipping" behavior needs to be penalized, and the penalty is the gap penalty g. In this embodiment, the gap penalty g is usually a pre-set, fixed negative number (e.g., g = -2), which is directly used in the core recursive formula of the dynamic programming matrix H(i,j). In this embodiment, based on the scoring function With the empty space penalty g, construct a local alignment similarity matrix H, where row index i corresponds to the i-th unit (xi) of sequence X and column index j corresponds to the j-th unit (yj) of sequence Y; calculate the local alignment similarity intensity H(i, j) for all (i, j), and combine all local alignment similarity intensities H(i, j) to obtain the complete similarity matrix H. Each element in the similarity matrix H is the local alignment similarity intensity H(i, j), and the maximum value max H(i, j) in the similarity matrix H is the local highest score of this local alignment.

[0017] After obtaining the similarity matrix H, first find the element (i, j) with the maximum value max H(i,j) in the similarity matrix H. The local alignment similarity intensity H(i, j) corresponding to this element is the local highest score of this local alignment. Starting from element (i, j), according to the source direction (i.e., the current value is from...) , or By backtracking back to the element with a score of 0, a continuous alignment path in the similarity matrix H can be obtained. The first semantic block pair covered by this path is the first semantic block pair of local optimal alignment, and this path is also the best alignment path.

[0018] By outputting the start and end positions of this optimal alignment path and the local maximum score max H(i,j) of this local alignment, we can simultaneously obtain: 1. The first semantic block pair with local optimal alignment; 2. Alignment score (used to quantify the similarity strength of this match, which is the local maximum score max H(i,j) of this local alignment).

[0019] In the guidance method: 1. The alignment path is continuous, which is directly reflected in the process of backtracking to generate the first semantic block pair with optimal alignment. When the local alignment algorithm Smith-Waterman constructs the similarity matrix H, the value of each cell depends on its three adjacent cells: left, top, and top-left. When the algorithm backtracks from the highest score point, it will work backward step by step along the calculation path to form a continuous path without breaks in the similarity matrix. 2. Consistent matching patterns are mainly reflected by analyzing the distribution of gaps in aligned semantic blocks. Consistent matching patterns: A high-quality aligned semantic block pair usually consists of a long series of continuous "match / non-match" (diagonal movement) with few or no "gap" (horizontal or vertical movement), indicating that the two semantic blocks are stably matched "end-to-end". Inconsistent matching patterns: If gaps frequently appear in the alignment path (e.g., match-gap-match-gap), it indicates that the two semantic blocks have performed a large number of "insertion" and "deletion" operations to align. In the textual context, this means that their internal structure or content order is very different, and the matching patterns are inconsistent. 3. Complete boundary points are reflected in the start and end points of the semantic blocks in the optimal local alignment. A core advantage of the Smith-Waterman local alignment algorithm is its ability to automatically find the optimal start and end boundaries within the aligned semantic blocks. The backtracking process stops when the algorithm backtracks to a unit with a value of 0 in the similarity matrix. The start point (the first non-zero unit found) and the end point (the globally highest-scoring unit) of this backtracking define the complete boundary of this optimal local match. 4. Semantic unit relevance > 0.7, reflected by the alignment score. This is essentially an application of a threshold to the final similarity score; after finding the locally optimal aligned semantic block pair, the algorithm obtains a local maximum score for that local alignment, i.e., max H(i,j); this score is used during matching. When mismatched The result, summed with the gap penalty along the entire path, comprehensively measures the path length, matching quality, and gap cost. For ease of comparison, this raw score usually needs to be normalized, for example, by dividing by the highest possible score (i.e., the alignment score between two semantic blocks themselves), converting it into a relevance score between 0 and 1. Output data: The output tree structure is a collection of chunks after the algorithm has processed the data. Each node represents a different document level, and the higher the level of the node, the larger the concept or paragraph it represents. Simplify document representation: Finally, by compressing redundancy and merging similar content, a high-level overview of the document is provided.

[0020] S3: Select two sets of blocks processed by different algorithms, and then arbitrarily select one semantic block from each of these two sets to form a second semantic block pair. The two different semantic blocks forming the second semantic block pair are respectively designated as fragments s1 and s2. Calculate and store the technical association probability between the two fragments. This process will continue until all second semantic block pairs formed by semantic blocks of the two sets of blocks processed by different algorithms have been calculated, thus obtaining a complete probability matrix describing the degree of mutual association between all semantic blocks. In the process of calculating the technical association probability, the technical feature weight DW(t) and the term t weight W0(t) defined later will play a core role in GMM model construction, feature matching, and domain relevance calculation, ensuring that technical importance runs through the entire analysis process. Then, based on the technology association probability, all semantic blocks are soft-clustered to obtain a technology topic membership vector containing the probability that each semantic block belongs to each preset technology topic. The result of soft clustering is to generate a technology topic membership vector for each semantic block. This technology topic membership vector represents the probability that this technology block belongs to each preset technology topic. This technology topic membership vector will be used as one of the core metadata of the semantic block and will be used in step S4 to build the index of the fast retrieval tree, thereby realizing fast positioning and retrieval based on technology topics. In step S3, the technical field feature weights ; The domain relevance score reflects the relevance of term t in a specific domain, and its value range is [0,1]. For technology category relevance, it reflects the technology category to which term t belongs, and the value range is [0,1]. This represents the relevance of the patent classification number, reflecting the importance of term t in international patent classifications (such as IPC / CPC), with a value range of [0,1]. For domain relevance, For relevance to technology categories, The classification number relevance is (in this embodiment, the domain relevance, technology category relevance, and classification number relevance are all obtained through experimental tuning on the validation set, and are therefore preset values). ; t is a term.

[0021] In step S3, the technical relevance between fragment s1 and fragment s2 is determined. ;in , The vector representation of the fragment. It is the overall representation of a segment in the semantic space, used to measure the similarity and relevance between segments; The weight of term t; Let be the word vector of term t; Cosine similarity; For domain similarity; represents the vector similarity weights, with values ​​ranging from [0,1]. This is the domain similarity weight; in this embodiment, , , , , .

[0022] In this embodiment, the weight W0(t) of term t is used to quantify the importance of the term to the information expression within the current segment. ; is the word frequency, representing the frequency of occurrence of word t in a specific segment; is the inverse document frequency, used to represent the rarity of term t in the entire patent document set; The weights represent the technical field features to reflect the technical field relevance of term t.

[0023] Word vector of term t WordVector(t) is a high-dimensional vector representing the semantic features of a term. The word vector of each term t covers both the basic meaning of the term and the relationship between it and other terms in the technical field. Specific vocabulary in the patent field (such as common patent terms and technical terms) is used as input elements and fed into a pre-trained word vector model. The pre-trained word vector model is pre-trained to reflect the semantic features of the technical field and can use Word2Vec, GloVe, or domain-defined word vectors. Finally, the word vector WordVector(t) of term t is obtained.

[0024] In step S3, the technical association probability between fragment s1 and fragment s2 is... ; For domain relevance, For relevance to technology categories, For the relevance of the classification number, ; To emphasize the importance of technology when constructing the GMM model, this approach weights the term t. Specifically, for each term t in segments s1 and s2, the word vector of term t is weighted... When training the GMM, weights are assigned to each term t. This results in technical keywords dominating the probability distribution. The resulting P1 and P2 are probability distributions carrying technical weight information, where P1 is the probability distribution of fragment s1 in the GMM model, and P2 is the probability distribution of fragment s2 in the GMM model.

[0025] Semantic similarity probability between fragment s1 and fragment s2 ; P1 represents the KL divergence of the GMM distributions of fragments s1 and s2; P2 represents the probability distribution of fragment s1 in the GMM model; P2 represents the probability distribution of fragment s2 in the GMM model. Technical feature similarity probability between fragment s1 and fragment s2 ; in, The calculation process is as follows: Extract the set of technical features from fragment s1 In this context, m0 represents the total number of technical features of segment s1, and the set of technical features is extracted from segment s2. , where m1 is the total number of technical features of segment s2; Then for each feature in the technical feature set F1 Find the feature that best matches the given set of technical features F2. Specifically, a complete calculation should involve iterating through each feature in the technical feature set F1. For features Then iterate through all features in the technical feature set F2. And calculate the matching degree between them. Find the feature in the technical feature set F2. The feature with the highest matching degree is used as the matching degree relative to the feature. Best Match Score ; Features The corresponding weight is directly adopted from the weight of the term t as defined above. ,in Corresponding to the term t defined above, the technical field feature weights are thus assigned. Weight of term t The influence of this is introduced into the feature matching process.

[0026] N01 is the normalization parameter. Specifically, the scores of the best matching pairs corresponding to all features in the technical feature set F1 are weighted and accumulated, and then normalized by the normalization parameter N01 so that the technical feature similarity probability is between 0 and 1. The normalization parameter N01 can be half of the sum of the weights of all features in the technical feature set F1 and the technical feature set F2, or the maximum value between the sum of the weights of all features in the technical feature set F1 and the sum of the weights of all features in the technical feature set F2.

[0027] Domain relevance probability between fragment s1 and fragment s2 ; Similarity to IPC classification numbers; ; For the degree of overlap in the technical fields; ; This indicates that t is a term in segment s1. Let t represent a term in segment s2; DS(t), TC(t), and CS(t) are the three components of the domain feature weights DW(t). Let S1 be the length of the segment. For the length of segment S2, more specifically, It is the number of terms in segment S1. It represents the number of terms in segment S2; cos is the cosine similarity, where cos is the number of terms in the segment S2. and Both are two-dimensional vectors. The overlap of technical fields is to determine whether the relative distribution patterns of two segments are consistent across different technical fields and technical categories.

[0028] In this way, the three components of the domain feature weight DW(t), DS(t), TC(t), and CS(t), are used to calculate different dimensions of domain relevance. The domain relevance score reflects the relevance of term t in a specific domain, and its value range is [0,1]. For technology category relevance, it reflects the technology category to which term t belongs, and the value range is [0,1]. The patent classification number relevance reflects the importance of term t in the international patent classification, with a value range of [0,1]. In this embodiment, the technology relevance TechRelevance(s1,s2) is composed of vector similarity and domain similarity. Vector similarity, as the weight W0(t) of the comprehensive term t, is reflected in the technology feature similarity probability Ptech(s1,s2). The domain similarity component, as the three components DS(t), TC(t), and CS(t) of the technology feature weight DW(t), is reflected in the domain relevance probability Pdom(s1,s2), thus fully integrating the information of TechRelevance(s1,s2) into the calculation of the overall technology relevance probability P(s1,s2).

[0029] In step S3, the probability storage structure for the technical association probability between segment s1 and segment s2 is a sparse matrix storage, using key-value pairs to store non-zero probabilities. The key format is (segment_id1, segment_id2), and the value includes the probability value, timestamp, and number of updates.

[0030] S4: Construct retrieval numbers for semantic blocks; build a fast retrieval tree based on the constructed retrieval numbers to achieve... Contextual location with time complexity forms a hierarchical numbered index system. In this embodiment, each block is assigned a unique number in the recursive block partitioning algorithm and recorded in the metadata, thereby forming the semantic block location, which is classified as IPC / CPC classification number.

[0031] In step S4, a composite numbering structure is used to construct the retrieval number. FileID is the ID of patent document d (that is, the patent publication number of the patent document), which is the unique identifier of patent document d; ChunkID is the unique number of semantic block, which is the identifier of the position of semantic block of patent document d; TechClass is the technology classification code, which is the official IPC / CPC classification number of patent document d itself. Location mapping relationship of Search ID CompositeID ; offset is the physical storage location offset, length is the content length of the semantic chunk, metadata is the metadata information of the semantic chunk, the metadata information is the structured information describing the attributes of the semantic chunk, including the unique number of the semantic chunk, the ID of the patent document d, the technology classification code and the technology topic membership vector; CompositeID is the retrieval number built based on the position of the chunk, and the corresponding Position mapping relationship is also the location information for the specific semantic chunk.

[0032] In step S4, the multi-level hash index in the fast retrieval tree includes a first-level index (document level), a second-level index (block level), and a third-level index (category level). The first-level index , Where N1 is the number of cached documents and X is the total number of documents; Represents a probability function. This represents the hit condition in conditional probability; This is a very standard and common way of expressing information in computer system design and analysis. It describes the system's performance in a macroscopic, long-term, average sense, rather than a microscopic, instantaneous, or specific state. Therefore... In this scheme, it is a variable value, not a constant value; The secondary index , Where N2 is the number of cache blocks and M is the total number of blocks; The third-level index , ,in Classification by Technology The set of IDs of all patent documents d under; Classification of computing technologies The proportion of the number of documents included to the total number of documents in the database, X; where X is the total number of documents. Function: Prioritize the elimination of high-risk individuals Technical Classification Prioritize retaining low Technical Classification This helps the system decide what is worth caching, i.e., for high-value items... Technical Classification It takes up too much space and is difficult to cache, therefore it needs to be phased out; when processing user search requests... Used to help the query optimizer estimate query costs and select the optimal execution path; The retrieval probability model of the obtained fast search tree ;in From the classification of technologies Access statistics ; ;in For access time complexity, A boolean value (true / false) indicating whether the index was hit. The time complexity upon a hit (constant time). Let X be the time complexity (logarithmic time) when there is a miss, and X be the total number of documents. and All are conditional operators; In step S4, the contextual retrieval mechanism of the fast retrieval tree is as follows: Window definition This is used to implement contextual retrieval in the fast search tree, that is, to obtain the surrounding area of ​​the current block; w is the window size, ChunkID is the position of the current block, and i is used to obtain the ChunkID of the current block and the contents of the w blocks before and after it to form a context window; S5: Select semantic blocks in order of importance, then determine the context window size of each semantic block according to the dynamic adjustment algorithm, and use a fast search tree to search and expand the context of the corresponding semantic block according to the context window size of each semantic block. Finally, output the expanded result (i.e., use the "small to large document search" method to search small fragments first and then expand the context), realize the intelligent routing layer, and use the search-enhanced KV cache to reduce redundant calculations.

[0033] In step S2, when assigning a unique number to each block using the recursive block partitioning algorithm, the technical features, technical fields, and innovation points of the analysis target are first obtained from Query2doc. Domain features are extracted by combining the patent classification number (IPC / CPC) to obtain the query volume q. The blocks are then sorted based on their similarity to the query volume q, and then numbered according to the sorting. In this embodiment, the method of sorting blocks based on their similarity to the query volume q is similar to the method of sorting patent document d based on its similarity to the query volume q in step S1, except that patent document d is replaced with a block in step S1. In step S5, document importance is obtained based on the block sorting. The higher the similarity between a block and the query volume q, the higher the document importance. Higher document importance also means higher priority for dynamically adjusting the context window size.

[0034] In step S5, the dynamic adjustment algorithm includes: Initial window determined: ; ; where Sc is the selected semantic block, and Neighbors(Sc) is the neighboring semantic blocks of the selected semantic block; Window boundary Window_Size constraint: ; ; Calculation of the extended step size Step_Size(t): ; The step size Step_Size(t) guides the physical operations, i.e., how many semantic blocks to select each time expansion is performed. Window update rules: Q_threshold represents the quality threshold. Assess contextual coherence: Where P(Si|Si-1) is the conditional probability of adjacent semantic blocks; Assess technology coverage: ;where Tech_Features() is the set of extracted technical features; Dynamically adjust probability threshold: ;in The dynamically adjusted probability threshold is a dynamic admission threshold set to filter which semantic blocks are eligible to be considered for window expansion. Dynamically adjust quality threshold: ; Stop_Condition condition for termination: ; Best_Window selection rules: ; ; Where base_size is the base window size, which is the baseline value for initial window calculation; P(Sc,Si) is the association probability between the selected semantic block Sc and its neighboring semantic blocks Si, used to calculate Core_Score; Neighbors(Sc) is the set of neighboring semantic blocks of the selected semantic block Sc and the surrounding semantic blocks directly related to the selected semantic block; MIN_WINDOW is the minimum window size; MAX_WINDOW is the maximum window size. This indicates that the window size (Window_Size) must be a power of 2. In this context, W(t) represents the window at time t, and Si and Sj are any two semantic blocks within the window W(t); Tech_Features_Total is the total set of all technical features, used as the benchmark for calculating technical coverage; New_Segments(t) is the set of newly added blocks at time t; base_threshold is the base probability threshold, which is the benchmark value for dynamic threshold calculation; and min_quality is the minimum quality threshold, which is the lower limit for quality assessment. Used for dynamic adjustment of quality thresholds, It is the time decay factor. It is a time step; The difference between the maximum and minimum quality values ​​within the range of the quality threshold; Quality(W(t)) is the quality score of the window at time t; In this context, t represents the time step. This represents the maximum number of time steps, used to define the time range for evaluation; Size(W(t)) is the size of the window at time t, used for window size constraint judgment. When using it, input Core_Segment and P_Matrix, calculate the initial Core_Score to determine the initial window Initial_Window, and then initialize the quality threshold Q_threshold(t) to obtain the window Initial_Window after adjusting the context window size; where Core_Segment is the selected semantic block, that is, the semantic block that needs to be processed, and serves as the center point of the context expansion; P_Matrix is ​​the probability matrix between semantic blocks, storing the association probability between all semantic block pairs, where any two different semantic blocks in the same patent document d form a semantic block pair.

[0035] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and these variations still fall within the protection scope of the present invention.

Claims

1. A method for patent data analysis based on dynamic context window, characterized in that, The method comprises the following steps: S1: obtaining technical features, technical fields and innovation points of an analysis target, extracting field features in combination with a patent classification number to obtain a query amount q; and sorting patent documents d based on similarity of the patent documents d and the query amount q; S2: In the order of the ranking, apply the recursive chunking algorithm to each patent document d to decompose it into a plurality of semantic chunks with hierarchical relationship, denoted as , until the preset minimum chunk is reached, and give each semantic chunk a unique number ChunkID, which is recorded in the metadata of the semantic chunk, and the number is used to establish the context reference relationship between the chunks; then apply the hierarchical reduction algorithm to output the chunk set after algorithm processing, so as to maintain the document structure of each patent document d; S3: selecting two different algorithm-processed block sets, then selecting a semantic block from each of the two algorithm-processed block sets to form a second semantic block pair, taking the two different semantic blocks forming the second semantic block pair as a segment s1 and a segment s2 respectively, and calculating and storing a technical correlation probability between the two segments; Then, soft clustering is performed on all semantic blocks according to the technical correlation probability to obtain a technical theme membership vector containing a probability of each semantic block belonging to each preset technical theme; S4: constructing a retrieval number based on the technical theme membership vector for the semantic blocks; and establishing a fast retrieval tree based on the constructed retrieval number; S5: selecting semantic blocks in turn according to semantic block importance, then determining a context window size of each semantic block according to a dynamic adjustment algorithm, and retrieving and expanding a context of the corresponding semantic block according to the context window size of each semantic block by using the fast retrieval tree, and finally outputting an expanded result.

2. The method of claim 1, wherein the method is based on a dynamic context window. In the step S1, a dual-view retrieval paradigm is adopted in combination with two levels of coarse granularity and fine granularity to obtain text similarity and technical correlation, and a similarity of the patent documents d and the query amount q is obtained through a fine-tuning data construction pipeline.

3. The method of claim 2, wherein the method further comprises: The step S1 comprises: S11: calculating a text similarity of the query amount q and the patent documents d: Text similarity ; and All are fusion weight parameters, among which , ; The preferred value range is 0.3-0.

4. When the document structure is complete and the overall integrity is strong, the value can be appropriately increased. When processing short documents, the value of can be slightly reduced. The possible values ​​of ; The preferred value range is 0.6-0.

7. The more detailed the technical feature description, the better. The higher the value, the better it can be when precise matching is required. The possible values ​​of ; Coarse-grained text similarity ; Fine-grained text similarity ; is the text coarse-grained for the query q, is the text coarse-grained for the patent document d, is the text fine-grained for the i-th fine-grained unit of the query q, is the text fine-grained for the i-th fine-grained unit of the patent document d; N is the number of fine-grained units, qi is the i-th fine-grained unit of the query q, di is the i-th fine-grained unit of the patent document d; when i = 1, q1 is the 1st fine-grained unit of the query q, representing a specific technical parameter, d1 is the 1st fine-grained unit of the patent document d, representing a specific technical feature of the independent claim; when i = 2, q2 is the 2nd fine-grained unit of the query q, representing a specific technical implementation, d2 is the 2nd fine-grained unit of the patent document d, representing an additional feature of the dependent claim; when i = 3, q3 is the 3rd fine-grained unit of the query q, representing a key technical component, d3 is the 3rd fine-grained unit of the patent document d, representing a specific step of the technical solution; sim() is the similarity function, cos() represents the cosine similarity, cos() is an index for measuring the similarity of two vectors, wherein represents the vector dot product, and represents the Euclidean norm of the vector, is the angle between the two vectors. S12: calculating a technical similarity of the query amount q and the patent documents d: Technical similarity ; and are fusion weight parameters, wherein , ; The preferred value range is 0.4-0.5, and the influencing factor is the overall classification of the technical field and the IPC / CPC classification level; The preferred value range is 0.5-0.6, and the influencing factor is the description depth of the technical features and the specific implementation details of the innovation points; Coarse-grained technical similarity ; Fine-grained technical similarity ; is the technical coarse granularity for query q, is the technical coarse granularity for patent document d, is the technical fine granularity for the i-th fine granularity unit of query q, is the technical fine granularity for the i-th fine granularity unit of patent document d; N is the number of fine granularity units, qi is the i-th fine granularity unit of query q, di is the i-th fine granularity unit of patent document d, sim() is the similarity function; cos() represents the cosine similarity. S13: Calculate the overall relevance of the query volume q and the patent document d ; and are fusion weight parameters, wherein , ; S14: calculating a similarity of each patent document d and the query amount q after constructing a fine-tuning data construction pipeline according to the overall correlation: First, the evaluation data of the fine-tuning data construction pipeline is constructed: ; q is the query amount q: the description of the query target, usually input by the user, containing technical features, technical fields, innovation points, and other descriptive information, which is the core content for calculating the relevance of patent documents; d is the patent document d: a specific document in the patent database, usually the patent text to be evaluated, including patent claims, abstracts, specifications, and classification numbers; y is the true relevance: the true relevance label of the query amount q and the patent document d, , y is the overall relevance degree obtained in step S13 ; Then, a predicted correlation score is calculated as the similarity of each patent document d and the query amount q: ; ; ; ； ； is the i-th fine-grained unit in the patent document d that is most similar to the query q, then is the text fine-grain of the fine-grained unit in the patent document d, is the technology fine-grain of the fine-grained unit in the patent document d; sim() is the similarity function; and cos() represents the cosine similarity.

4. The method of claim 3, wherein the method further comprises: In the step S2, for each patent document d, an initial block set corresponding to each patent document d is obtained from a recursive block algorithm, and the recursive block algorithm comprises: ； ； wherein ; ； ； ； ； D is a document which needs to be recursively divided, and can be a patent document d or a sub-document after several divisions of the patent document d, is the maximum semantic division size, represents that all of , is the first semantic division, is the sub-document semantic division result; for , indicates a constraint condition that must be met for all generated semantic divisions , means that the size of each semantic division cannot exceed the preset maximum value ; and Split(D) is a function of performing splitting on the document D which needs to be recursively split; and Split() indicates the process of performing splitting on a document D which needs to be recursively split; In, represents the best cut point of the document D that needs to be recursively chunked; represents finding an index ix such that the value of the expression Score() that follows is maximized; represents the part of the document D from the beginning to the position ix, represents the part of the document D from the position ix to the end; where the score function Score is computed by Coherence represents the degree of coherence or cohesion; cos is the cosine similarity function; this is part of the score function Score that measures the degree of semantic association between the two parts and after segmentation; and respectively represent the function that converts into a mathematical vector and the function that converts into a mathematical vector; Balance indicates balance, indicates the length of indicates the length of is another part of the scoring function Score that assesses whether the two parts and are approximately equal in length; the absolute value of the difference between the lengths of the two parts is calculated, which is normalized by the length of the longer part. a and b are weight parameters calculated as a = 1 - b.

5. The method for patent data analysis based on dynamic context window according to claim 4, characterized in that: In the step S2, the hierarchical reduction algorithm is a Hierarchical Reduction algorithm, and the hierarchical reduction algorithm comprises: obtaining an initial set of tiles from a recursive tiling algorithm constructing an initial tree hierarchy for the initial set of tiles, the initial set of tiles being used as leaf nodes of the hierarchy The similarity of the patent documents d and the query amount q obtained through the step S1 is used to identify similarity between adjacent blocks, and similar blocks with high similarity are merged layer by layer to form parent nodes until a preset simplification degree is met; The alignment result of the local alignment algorithm is used to guide the merging of similar blocks with high similarity to retain semantic integrity.

6. The method of claim 5, wherein: In the step S3, the technical field feature weight ; is the field relevance score, reflecting the relevance of the word item t in a specific field, with a value range of [0, 1]; is the technical category relevance, reflecting the technical category to which the word item t belongs, with a value range of [0, 1]; is the patent classification number relevance, reflecting the importance of the word item t in the international patent classification, with a value range of [0, 1]; is the field relevance, is the technical category relevance, is the classification number relevance, ; t is a word item; In the step S3, the technical relevance between the segment s1 and the segment s2 ; wherein , is a vector representation of a segment, is an overall representation of a segment in a semantic space, used to measure the similarity and relevance between segments; is a weight of a term t; is a word vector of a term t; is a cosine similarity; is a domain similarity; is a vector similarity weight, with a value range of [0, 1]; is a domain similarity weight.

7. The method of claim 6, wherein the method further comprises: In said step S3, the technical association probability between the fragment s1 and the fragment s2 ; for the field relevance, for the technical class relevance, for the classification number relevance, ; semantic similarity probability between segment s1 and segment s2 ; is the KL divergence of the two segment GMM distributions for segment s1 and segment s2; Pi is the probability distribution of segment s1 in the GMM model; P2 is the probability distribution of segment s2 in the GMM model; The technical feature similarity probability between the segment s1 and the segment s2 ; wherein, The calculation process is as follows: extracting a set of technical features from the fragment s1 where m0 is the total number of technical features of the fragment s1, extracting a set of technical features from the fragment s2 where m1 is the total number of technical features of the fragment s2; Then traverse each feature in the technical feature set F1 For the feature Traverse all the features in the technical feature set F2 again And calculate the matching degree between them Find the feature in the technical feature set F2 that matches the feature with the highest matching degree, and take the matching degree of the feature as the best matching pair score relative to the feature ; is a feature corresponding weight, is equal to the weight of the term t wherein is the term t; N01 is a normalization parameter; the field relevance probability between segment s1 and segment s2 ; is the IPC classification number similarity; is the technical field overlap.

8. The method of claim 7, wherein the method further comprises: In the step S4, the retrieval number is constructed by using the composite number structure wherein FileID is the ID of the patent document d, ChunkID is the unique number of the semantic chunk, and TechClass is the technical classification code; CompositeID ; offset is the offset of the physical storage location, length is the content length of the semantic chunk, metadata is the metadata information of the semantic chunk, the metadata information is the structured information describing the attributes of the semantic chunk chunk, including the unique number of the semantic chunk, the ID of the patent document d, the technical classification code and the technical subject membership vector.

9. The method of claim 8, wherein: In the step S4, the multi-level hash index in the fast retrieval tree comprises a first-level index, a second-level index and a third-level index; The primary index , where N1 is the number of cached documents and X is the total number of documents; denotes a probability function (Probability), denotes a hit condition in conditional probability; The secondary index , where N2 is the number of cache blocks and M is the total number of blocks. The tertiary index , wherein is a set of IDs of all patent documents d under the technology classification ; The search probability model of the resulting fast search tree .

Images