Patent text search and measurement method, system and device based on language model enhancement

Abstract

This invention belongs to the field of natural language processing and provides a method, system, and device for patent text search measurement based on language model enhancement. It can obtain the semantic non-linear decentering curve of the search object and the semantic convergence centering curve of each text data in the search database by curve fitting the semantic non-linear decentering array of the search object and the semantic convergence centering curve of each text data. Based on curve similarity measurement, the text data with the most similar semantic non-linear decentering curves is selected as output. While retaining the semantic expressive power of the pre-trained language model, a structured measurement of semantic centering relationships within the text is introduced, achieving fine-grained matching at the semantic structure level of patent text information, significantly improving the accuracy and stability of patent retrieval and similarity analysis.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing, and specifically relates to a patented text search measurement method, system and device based on language model enhancement. Background Technology

[0002] With the intensification of global technological competition and industrial upgrading, the number of patent applications in various countries continues to grow rapidly. Enterprises rely heavily on the automatic retrieval and similarity analysis of massive amounts of patent texts in scenarios such as project initiation and R&D, patent layout, infringement risk screening, and high-value patent evaluation. Existing patent retrieval systems have roughly evolved from traditional Boolean logic retrieval to semantic retrieval based on concept and vector space models, and then to semantic vector retrieval based on pre-trained language models and vector databases that have emerged in recent years.

[0003] In the traditional stage, a Boolean search algorithm combined with manual parameter tuning is commonly used. Complex search queries are constructed using keywords and classification numbers, and results are then ranked using inverse document frequency (IVF) and vector space models. This approach is highly dependent on the expertise of the search personnel and struggles to balance recall and precision, particularly when dealing with patent texts that are diverse in expression and cross-domain. To improve semantic representation capabilities, one approach introduces structured knowledge such as knowledge ontology and domain dictionaries to abstract concepts and extend the semantics of patent texts. For example, patent document CN107247780A proposes a patent document similarity measurement method based on knowledge ontology. This method extracts the so-called core technical solutions using the structural position features of the patent, constructs a classification number-subject term relationship model, extracts keywords using TF-IDF and TextRank, trains word vectors using Word2Vec or FastText, and finally calculates the semantic distance between patent documents based on keyword sets, word weights, and word vectors using algorithms such as EMD, thereby improving the similarity measurement effect.

[0004] In recent years, with the development of deep learning and large language models, more and more solutions have attempted to encode patent texts using pre-trained language models and perform retrieval and recommendation based on vector similarity. For example, patent document CN118193726B discloses a visual patent retrieval method based on a pre-trained language model. It uses RoBERTa and BiLSTM to extract keywords and corresponding word vectors from patent texts, adds multiple keyword vectors to obtain a high-dimensional vector representation of the patent, calculates the cosine similarity between the query keyword vector and the patent vector to obtain candidate results, and further uses Barnes-Hutt-SNE for dimensionality reduction. The similarity and clustering between patents are then visualized in a two-dimensional scatter plot to improve the intuitiveness and interactivity of the retrieval results. Meanwhile, some literature has proposed a more general retrieval framework such as LLM plus a vector database. For example, patent document CN117435696A proposes that the text to be retrieved and the query text are uniformly input into a large language model and encoded into vectors. After being stored in a vector database, the target text is quickly returned through similarity calculation, addressing the efficiency and accuracy issues of large-scale unstructured text retrieval.

[0005] However, taking the aforementioned technical solution CN107247780A as an example, although it considers the structural position, classification number, and subject term relationships of patent documents, and constructs a semantic similarity measure by combining domain ontology and word vectors, it ultimately represents the patent as a weighted set of keywords plus weights plus word vectors, and calculates the overall distance between the keyword distributions of the two patents using EMD. Essentially, it still belongs to the first-order statistical level of representation. This type of method is difficult to distinguish the interrelationships between multiple semantic topics within the same patent, and it is also difficult to explicitly reflect the relative position and organization of different semantic centers in the text. It still has limitations in distinguishing between patent documents with similar technical solutions and those with only similar keywords. While the technical solutions disclosed in CN118193726B and CN117435696A have introduced pre-trained language models to encode patents or general texts and use vector databases for similarity retrieval, most of them only compress the entire text or keyword set into a single high-dimensional vector and sort it using cosine similarity or distance measures. Such an overall vector representation often fails to distinguish the structural differences between different semantic segments or functional modules within a text. For patents containing multiple technical points and complex claim structures, it is easy to see situations where different structures have similar overall vectors, or similar structures have large differences in overall vectors. This makes it difficult to provide stable and reliable structural guidance for applications such as high-value patent evaluation and patent structure comparison.

[0006] Existing technologies typically only measure the similarity between query vectors and document vectors, or keyword vectors and keyword vectors, at most supplemented by simple clustering or two-dimensional visualization. They rarely explicitly mine the divergence relationships, similarity gradients, and ranking patterns among multiple semantic centers within the same text, nor do they attempt to serialize these internal relationships into a measurable one-dimensional structural feature array, and then perform more refined structural-level comparisons of the text based on the similarity between the array or its fitted curve. Therefore, in scenarios requiring differentiation of higher-order structural features such as the number of semantic topics, alignment / reverse relationships between topics, and the distribution of semantically dense and sparse segments, existing methods struggle to provide a robust measurement foundation. As for knowledge ontology-based methods like CN107247780A, they require manual extraction of numerous patent subject terms and classification number relationships to construct and maintain domain ontologies, directed graph models, and domain dictionaries. When their application domains expand, classification systems adjust, or new technologies emerge, continuous manual updates to the knowledge base become necessary, resulting in high maintenance costs and hindering rapid migration to emerging technology fields or cross-domain scenarios. Existing patent text semantic retrieval and similarity measurement technologies either primarily rely on keywords and knowledge ontology or only use first-order vector similarity output by pre-trained language models. There is a lack of a patent text measurement optimization method that can explicitly model the divergence and similarity relationships between multiple semantic centers within a unified pre-trained language model embedding space, abstract these internal relationships into a comparable one-dimensional structural feature array, and then achieve fine-grained semantic structure-level matching between texts based on the similarity between these structural feature arrays. Existing technologies struggle to simultaneously meet the comprehensive requirements of high-dimensional semantic expression capabilities, internal structural interpretability, and retrieval measurement stability. Summary of the Invention

[0007] The purpose of this invention is to propose a patent text search measurement method, system and device based on language model enhancement, so as to solve one or more technical problems existing in the prior art, and at least provide a beneficial option or create conditions.

[0008] To achieve the above objectives, according to one aspect of the present invention, a patent text search measurement method based on language model enhancement is provided, the method comprising the following steps:

[0009] The search object and the search database are obtained, and the search database stores multiple text data. For each text data in the search database, vectorization embedding is performed through a unified pre-trained language model to obtain multiple text embedding vectors. The multiple text embedding vectors are clustered to obtain multiple semantic center clusters.

[0010] For each target semantic center cluster among the plurality of semantic center clusters, the divergence metric between the target semantic center cluster and the other semantic center clusters is calculated. The semantic center cluster with the smallest divergence metric is determined as the semantic base cluster of the target semantic center cluster, and the semantic center cluster with the largest divergence metric is determined as the semantic vertex cluster of the target semantic center cluster. Based on the vector relationship between the target semantic center cluster and the semantic base cluster and the semantic vertex cluster, a first centrality index and a second centrality index are obtained.

[0011] The semantic and axial centrality of the target semantic center cluster is determined based on the first centrality index and the second centrality index, and the semantic and axial centrality of the target semantic center cluster is generated based on the semantic and axial centrality.

[0012] Based on the semantic and convergent centrality of each semantic center cluster corresponding to the text data, the semantic center clusters are sorted to obtain a semantic and convergent centrality sequence. A reference semantic center cluster is selected in the semantic and convergent centrality sequence, and the vector similarity between the semantic and convergent centrality of the reference semantic center cluster and the semantic and convergent centrality of each semantic center cluster in the sequence is calculated. The vector similarities are arranged in order to form the semantic and convergent centrality array of the text data.

[0013] The search object is vectorized and clustered using the pre-trained language model to obtain multiple search semantic center clusters. For each of the multiple search semantic center clusters, a statistic of the divergence measure between the search semantic center cluster and the other search semantic center clusters is calculated, and the statistic is used as the semantic non-centrality of the search semantic center cluster.

[0014] Based on the semantic non-centrality, the multiple retrieval semantic center clusters are sorted to obtain a semantic non-centrality cluster center sequence. In the semantic non-centrality cluster center sequence, the semantic non-centrality corresponding to each position is determined according to the vector similarity between adjacent retrieval semantic center clusters. The semantic non-centrality of each position is arranged in order to form the semantic non-centrality array of the retrieval object.

[0015] Curve fitting is performed on the semantic non- and decentralized array of the search object and the semantic convergent center array of each text data in the search database to obtain the semantic non- and decentralized curve of the search object and the semantic convergent center curve of each text data. Based on the curve similarity measure, the target semantic convergent center curve that is most similar to the semantic non- and decentralized curve is determined, and the text data corresponding to the target semantic convergent center curve is output as the optimal text data.

[0016] As can be seen, the method described in this invention does not directly compare the similarity between the "query vector" and the "text vector" as in previous existing technologies. Instead, it compares whether the curve shapes of the internal semantic structure of the query (non-centralized mode) and the internal semantic structure of the text (centralized mode) are similar. This is the second-order semantic structure similarity retrieval proposed in this invention.

[0017] This invention does not directly calculate query-doc similarity, but rather compares "semantic center relationship curves." Most conventional language model retrieval methods vectorize the query and the document or chunk, directly using cosine similarity and dot product to rank similarity. This invention adds several layers of processing: First, it performs internal clustering on each text to extract multiple semantic centers; then, it constructs an internal structure array curve between these semantic centers using cross-entropy and similarity; finally, the query side also constructs its own internal structure array (a non-centralized array); and finally, it compares the similarity using curve-level Frechet Distance.

[0018] This is equivalent to performing a semantic relationship topology matching layer, rather than a simple vector dot product. Therefore, it can better capture the interrelationships between multiple topics and technical elements in complex patent texts while maintaining retrieval efficiency, thereby improving the accuracy and robustness of semantic retrieval.

[0019] Furthermore, the text data in the search database includes patent specification text and / or patent claim text, which are string data of published patent documents obtained through web crawling and / or exporting from patent databases.

[0020] Furthermore, the language model includes a language model pre-trained and / or fine-tuned on the patent corpus, and the sub-text unit corresponding to the text embedding vector is at least one of sentence, paragraph and / or character segment of preset length. The clustering of multiple text embedding vectors adopts a clustering algorithm based on vector distance or vector similarity, and the clustering algorithm includes at least one of K-means clustering, hierarchical clustering and density clustering.

[0021] Furthermore, the divergence metric is the cross-entropy metric. When calculating the cross-entropy metric, the values ​​of the semantic center clusters involved in the calculation are normalized in each dimension to convert them into a probability distribution. Specifically, after performing an exponential operation on the values ​​of each dimension, the sum of the exponents of each dimension is normalized to make the value of each dimension non-negative and the sum of the values ​​of all dimensions equal to 1.

[0022] Further, the semantic convergence is obtained by comparing the vector differences between the target semantic center cluster size and the target semantic center cluster size with the largest and smallest divergence metric values, respectively. The semantic convergence is obtained by subtracting the semantic convergence from the target semantic center cluster size, specifically including:

[0023] For each target semantic center cluster among the plurality of semantic center clusters, the divergence metric between the target semantic center cluster and the other semantic center clusters is calculated. The semantic center cluster with the smallest divergence metric is determined as the semantic base cluster of the target semantic center cluster, and the semantic center cluster with the largest divergence metric is determined as the semantic vertex cluster of the target semantic center cluster. Based on the vector relationship between the target semantic center cluster and the semantic base cluster and the semantic vertex cluster, a first centrality index and a second centrality index are obtained.

[0024] The semantic and axial centrality of the target semantic center cluster is determined based on the first centrality index and the second centrality index, and the semantic and axial centrality of the target semantic center cluster is generated based on the semantic and axial centrality.

[0025] The present invention describes a design for both lateral and non-lateral centrality, along with its basis and vertex structures. This design aims to find the most similar basis and the least similar vertex for each semantic center using cross-entropy. Then, a scalar lateral centrality is constructed by combining vector differences with similarity, and this scalar is used to shift the original center vector. This operation utilizes the two closest and furthest reference points to measure the balance or directional bias of the semantic center within the entire semantic space. The basis clustering reflects the commonalities between the semantic cluster and other highly similar semantic clusters in the patent text, while the vertex clustering reflects the limits of difference between the semantic cluster and other semantic clusters. Constructing lateral centrality using the relative centrality of these two factors essentially compresses common noise in the overall text, highlighting central semantics that are neither extreme nor mediocre within the overall semantic framework. These semantics often correspond to distinguishing data features in the core technical points of the text. In patent text retrieval, this automatically highlights the core semantic features of the technical solutions in the patent text, weakening non-critical background semantics and redundant descriptions, making subsequent retrieval and matching more focused on the technical substance.

[0026] Furthermore, the first centrality index and the second centrality index are centrality indices based on vector similarity, specifically:

[0027] The first centrality index is determined based on the vector similarity between the difference vector between the target semantic center cluster and its semantic basis cluster and the vector similarity between the target semantic center cluster.

[0028] The second centrality index is determined based on the vector similarity between the difference vector between the target semantic center cluster and its semantic vertex cluster and the vector similarity between the target semantic center cluster;

[0029] The preferred vector similarity is cosine similarity.

[0030] Furthermore, the semantic centrality is a scalar, constructed based on the relationship between the first centrality index and the second centrality index. The smaller of the two values ​​is used as the numerator and the larger value is used as the denominator to form a proportionality coefficient, which is used as the semantic centrality. The semantic centrality is a vector, obtained by subtracting the semantic centrality from the value of the target semantic centrality in each dimension.

[0031] Further, the semantic disjoint centrality is the arithmetic mean of the divergence measures between the retrieved semantic center cluster and the remaining retrieved semantic center clusters; when determining the semantic disjoint decentrality of each position in the semantic disjoint cluster center sequence:

[0032] For a retrieval semantic center cluster located at the first or last position, its semantic non-centrality is determined by the vector similarity between it and the unique adjacent retrieval semantic center cluster.

[0033] For a retrieval semantic center cluster located in the middle, its semantic non-centrality is determined by the arithmetic mean of its vector similarity with the previous rank retrieval semantic center cluster and its vector similarity with the next rank retrieval semantic center cluster. The vector similarity is preferably cosine similarity.

[0034] Furthermore, when constructing the semantic convergence center array, the reference semantic center convergence is the semantic center convergence at a preset extreme position of semantic convergence, which can preferably be the semantic center convergence with the smallest semantic convergence. Each element of the semantic convergence center array is the cosine similarity between the semantic convergence of the reference semantic center convergence and the semantic convergence of each semantic center convergence, arranged in order of increasing semantic convergence.

[0035] Furthermore, when using curve similarity measurement: the Fraser distance is used to calculate the distance between the semantic non-centralized curve and each semantic convergent curve; the smaller the distance value, the closer the curve shapes are.

[0036] Furthermore, the construction of semantic clustering center arrays corresponding to text data in the retrieval database is completed in the offline batch processing stage, and the semantic clustering center arrays corresponding to each text data are pre-stored in the index library. When a retrieval object is received, a curve similarity comparison is performed only based on the semantic non-centralized array of the retrieval object and the pre-stored semantic clustering center arrays in the index library. This invention adopts a system design of offline pre-computation and online lightweight matching. Specifically, the semantic clustering center arrays on the database side can be pre-processed and cached offline; during online queries, only one clustering and array construction is performed on the query, and then curve similarity calculation is performed with the cached array. This design ensures improved retrieval accuracy while having a smaller impact on online computing resource consumption and response latency, making it suitable for real-time retrieval scenarios in large-scale patent databases.

[0037] This invention also provides a language model-enhanced patent text search measurement system. The language model-enhanced patent text search measurement system includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the language model-enhanced patent text search measurement method. The language model-enhanced patent text search measurement system can run on computing devices such as desktop computers, laptops, handheld computers, and cloud data centers. The runnable system may include, but is not limited to, processors, memory, and server clusters. The processor executes the computer program within the following system units:

[0038] The data acquisition unit is used to acquire the retrieval object and the retrieval database, which stores multiple text data; each text data in the retrieval database generates multiple text embedding vectors through a language model, and the multiple text embedding vectors are clustered to obtain multiple semantic center clusters;

[0039] The option processing unit is used to calculate the divergence metric between each target semantic center cluster and the other semantic center clusters for each target semantic center cluster. It then compares the target semantic center cluster with the target semantic center clusters with the largest and smallest divergence metrics to obtain the semantic convergence degree. The target semantic center cluster is subtracted from the semantic convergence degree to obtain its semantic convergence quantity. Based on the semantic convergence degree of each semantic center cluster in the text data, the semantic center clusters are sorted, and a reference semantic center cluster is selected from the sort. The vector similarity between the semantic convergence quantity of the reference semantic center cluster and the semantic convergence quantities of each semantic center cluster in the sort is calculated. The values ​​of each vector similarity form the semantic convergence quantity center array of the text data.

[0040] The retrieval processing unit is used to perform vectorized embedding and clustering on the retrieval object to obtain multiple retrieval semantic center clusters. For each of the multiple retrieval semantic center clusters, a statistical measure of the divergence between the retrieval semantic center cluster and the other retrieval semantic center clusters is calculated as the semantic non-centrality of the retrieval semantic center cluster. Based on the semantic non-centrality, the multiple retrieval semantic center clusters are sorted, and the semantic non-centrality corresponding to each position is obtained according to the vector similarity between the retrieval semantic center clusters in the sorted sequence, forming a semantic non-central array of the retrieval object.

[0041] The curve matching unit is used to perform curve fitting on the semantic non-linear decentering array of the search object and the semantic convergent central array of each text data in the search database, respectively, to obtain the semantic non-linear decentering curve of the search object and the semantic convergent central curve of each text data. Based on the curve similarity metric, the text data with the most similar semantic non-linear decentering curves is selected as the output.

[0042] Correspondingly, the present invention also provides an electronic device, a readable storage medium, and a computer program product:

[0043] An electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the language model-enhanced patent text search measurement method and the method for each step therein.

[0044] A non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to perform the language model-enhanced patented text search measurement method and the methods for each step therein.

[0045] A computer program product includes a computer program that, when executed by a processor, implements the language model-enhanced patent text search measurement method and the methods for each step thereof.

[0046] The beneficial effects of this invention are as follows: This invention provides a patent text search measurement method, system, and device based on language model enhancement. It can perform curve fitting on the semantic non-linearly decentralized array of the search object and the semantic convergent central array of each text data in the search database, respectively, to obtain the semantic non-linearly decentralized curve of the search object and the semantic convergent central curve of each text data. Based on curve similarity measurement, the text data with the most similar semantic non-linearly decentralized curves is selected as the output. While maintaining search efficiency, it can better capture the interrelationships between multiple topics and technical elements in complex patent texts, thereby improving the accuracy and robustness of semantic retrieval. It can automatically highlight the core semantic features of the technical solutions in the patent text, weaken non-critical background semantics and redundant descriptions, making subsequent search matching more focused on the technical essence in the semantic information. While ensuring improved search accuracy, it has a smaller impact on the consumption of online computing resources and response latency, making it suitable for deployment in real-time search scenarios with large-scale patent databases. Attached Figure Description

[0047] The above and other features of the present invention will become more apparent from the detailed description of the embodiments shown in conjunction with the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same or similar elements. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without any creative effort. In the drawings:

[0048] Figure 1 The flowchart shown is a patent text search measurement method based on language model enhancement;

[0049] Figure 2 The diagram shows the system architecture of a patent text search measurement system based on language model enhancement. Detailed Implementation

[0050] The following will provide a clear and complete description of the concept, specific structure, and technical effects of the present invention in conjunction with the embodiments and accompanying drawings, so as to fully understand the purpose, solution, and effects of the present invention. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0051] In the description of this invention, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0052] like Figure 1 The diagram shown is a flowchart of the patent text search measurement method based on language model enhancement according to the present invention. The following is in conjunction with... Figure 1 This paper describes a patent text search measurement method, system, and device based on language model enhancement according to embodiments of the present invention.

[0053] This invention proposes a patent text search measurement method based on language model enhancement, the method specifically including the following steps:

[0054] The search object and the search database are obtained. The search database stores multiple text data. Each text data in the search database generates multiple text embedding vectors through a language model. The multiple text embedding vectors are clustered to obtain multiple semantic center clusters.

[0055] For each target semantic center cluster, calculate the divergence metric between the target semantic center cluster and the other semantic center clusters. Compare the vector differences between the target semantic center cluster and the target semantic center clusters with the largest and smallest divergence metrics to obtain the semantic convergence centrality. Subtract the semantic convergence centrality from the target semantic center cluster to obtain its semantic convergence mass. Based on the semantic convergence centrality of each semantic center cluster in the text data, sort the semantic center clusters and select a reference semantic center cluster from the sort. Calculate the vector similarity between the semantic convergence mass of the reference semantic center cluster and the semantic convergence centrality of each semantic center cluster in the sort. Use the values ​​of each vector similarity to form the semantic convergence centrality array of the text data.

[0056] The search objects are vectorized, embedded, and clustered to obtain multiple search semantic center clusters. For each of the multiple search semantic center clusters, a statistical measure of the divergence between the search semantic center cluster and the other search semantic center clusters is calculated as the semantic non-centrality of the search semantic center cluster. The multiple search semantic center clusters are sorted based on the semantic non-centrality, and the semantic non-centrality corresponding to each position is obtained according to the vector similarity between search semantic center clusters with adjacent positions in the sorted sequence, forming a semantic non-central array of the search objects.

[0057] Curve fitting is performed on the semantic non-linear decentering array of the search object and the semantic convergence centering array of each text data in the search database to obtain the semantic non-linear decentering curve of the search object and the semantic convergence centering curve of each text data. Based on the curve similarity measure, the text data with the most similar semantic non-linear decentering curves is selected as the output.

[0058] Furthermore, the text data in the search database includes patent specification text and / or patent claim text, which are string data of published patent documents obtained through web crawling and / or exporting from patent databases.

[0059] Furthermore, the language model includes a language model pre-trained and / or fine-tuned on the patent corpus, and the sub-text unit corresponding to the text embedding vector is at least one of sentence, paragraph and / or character segment of preset length. The clustering of multiple text embedding vectors adopts a clustering algorithm based on vector distance or vector similarity, and the clustering algorithm includes at least one of K-means clustering, hierarchical clustering and density clustering.

[0060] Furthermore, the divergence metric is the cross-entropy metric. When calculating the cross-entropy metric, the values ​​of the semantic center clusters involved in the calculation are normalized in each dimension to convert them into a probability distribution. Specifically, after performing an exponential operation on the values ​​of each dimension, the sum of the exponents of each dimension is normalized to make the value of each dimension non-negative and the sum of the values ​​of all dimensions equal to 1.

[0061] Further, the semantic convergence is obtained by comparing the vector differences between the target semantic center cluster size and the target semantic center cluster size with the largest and smallest divergence metric values, respectively. The semantic convergence is obtained by subtracting the semantic convergence from the target semantic center cluster size, specifically including:

[0062] For each target semantic center cluster among the plurality of semantic center clusters, the divergence metric between the target semantic center cluster and the other semantic center clusters is calculated. The semantic center cluster with the smallest divergence metric is determined as the semantic base cluster of the target semantic center cluster, and the semantic center cluster with the largest divergence metric is determined as the semantic vertex cluster of the target semantic center cluster. Based on the vector relationship between the target semantic center cluster and the semantic base cluster and the semantic vertex cluster, a first centrality index and a second centrality index are obtained.

[0063] The semantic and axial centrality of the target semantic center cluster is determined based on the first centrality index and the second centrality index, and the semantic and axial centrality of the target semantic center cluster is generated based on the semantic and axial centrality.

[0064] Furthermore, the first centrality index and the second centrality index are centrality indices based on vector similarity, specifically:

[0065] The first centrality index is determined based on the vector similarity between the difference vector between the target semantic center cluster and its semantic basis cluster and the vector similarity between the target semantic center cluster.

[0066] The second centrality index is determined based on the vector similarity between the difference vector between the target semantic center cluster and its semantic vertex cluster and the vector similarity between the target semantic center cluster;

[0067] The preferred vector similarity is cosine similarity.

[0068] Furthermore, the semantic centrality is a scalar, constructed based on the relationship between the first centrality index and the second centrality index. The smaller of the two values ​​is used as the numerator and the larger value is used as the denominator to form a proportionality coefficient, which is used as the semantic centrality. The semantic centrality is a vector, obtained by subtracting the semantic centrality from the value of the target semantic centrality in each dimension.

[0069] Further, the semantic disjoint centrality is the arithmetic mean of the divergence measures between the retrieved semantic center cluster and the remaining retrieved semantic center clusters; when determining the semantic disjoint decentrality of each position in the semantic disjoint cluster center sequence:

[0070] For a retrieval semantic center cluster located at the first or last position, its semantic non-centrality is determined by the vector similarity between it and the unique adjacent retrieval semantic center cluster.

[0071] For a retrieval semantic center cluster located in the middle, its semantic non-centrality is determined by the arithmetic mean of its vector similarity with the previous rank retrieval semantic center cluster and its vector similarity with the next rank retrieval semantic center cluster. The vector similarity is preferably cosine similarity.

[0072] Furthermore, when constructing the semantic convergence center array, the reference semantic center convergence is the semantic center convergence at a preset extreme position of semantic convergence, which can preferably be the semantic center convergence with the smallest semantic convergence. Each element of the semantic convergence center array is the cosine similarity between the semantic convergence of the reference semantic center convergence and the semantic convergence of each semantic center convergence, arranged in order of increasing semantic convergence.

[0073] Furthermore, when using curve similarity measurement: the Fraser distance is used to calculate the distance between the semantic non-centralized curve and each semantic convergent curve; the smaller the distance value, the closer the curve shapes are.

[0074] Furthermore, the construction of the semantic and convergence center array corresponding to the text data in the retrieval database is completed in the offline batch processing stage, and the semantic and convergence center array corresponding to each text data is pre-stored in the index library; when the retrieval object is received, the curve similarity comparison is performed only based on the semantic non-convergent center array of the retrieval object and the semantic and convergence center array pre-stored in the index library.

[0075] In some embodiments, the method of the present invention includes the following steps: First, acquiring the search object in real time, wherein the search object is the entire content of the search term or the search statement entered by the user, or part of the content of the search statement, wherein the search term or search statement may be structured data and / or the corresponding natural language description in an industrial scenario.

[0076] A retrieval database is obtained, which stores multiple different text data, preferably string data such as specification text and claim text obtained from publicly available patent documents through a web crawler. To improve vectorization and clustering quality, the text data preferably undergoes natural language preprocessing before being entered into the database to remove redundant information. The natural language preprocessing may include, but is not limited to, segmentation, sentence segmentation, word segmentation, stop word removal, regular expression cleaning, and removal of specific format symbols.

[0077] Then, based on each piece of text data stored in the retrieval database, vectorized embedding processing is performed through a unified pre-trained language model to split the text data into several sub-text units (e.g., sentences, paragraphs, or segments with a fixed number of words), and vectorized embedding is performed on each sub-text unit to obtain multiple different text embedding vectors.

[0078] For multiple text embedding vectors corresponding to the same text data, a clustering algorithm is used to cluster them to obtain several cluster center vectors, which serve as several semantic center clusters corresponding to the text data. Preferably, the number of semantic center clusters corresponding to each text data is more than three, so as to facilitate subsequent statistical selection of the basis, vertices and medians.

[0079] In specific clustering implementations, distance- or similarity-based clustering algorithms can be used, such as K-means clustering, hierarchical clustering, or density clustering. The number of clusters can be determined according to preset rules or optimized based on historical retrieval results. This invention does not limit the specific clustering algorithm.

[0080] In some embodiments, for any text data, the set of its semantic center clusters is denoted as several semantic center vectors. For a selected semantic center cluster, it is denoted as A. For ease of description, the remaining semantic center clusters that are different from A can be denoted as candidate semantic center clusters.

[0081] To measure the degree of difference between semantic center clusters, this embodiment uses cross-entropy as the metric. Specifically, before calculating the cross-entropy between A and any candidate semantic center cluster, the values ​​of each dimension of both are normalized and converted into a probability distribution. For example, the values ​​of each dimension can be exponentially calculated first, and then normalized by the sum of the exponents of each dimension, so that the value of each dimension is non-negative and the sum of all dimensions equals 1.

[0082] Based on this, using the normalized vector of A as the true probability distribution and the normalized vector of the candidate semantic center cluster as the predicted probability distribution, the logarithmic products of the true and predicted probabilities for each dimension are multiplied and summed, and the negative value is taken, according to the calculation method of the cross-entropy loss function commonly used in machine learning. This yields the cross-entropy value between A and the candidate semantic center cluster. This calculation is performed sequentially for A and all candidate semantic center clusters to obtain several cross-entropy values.

[0083] In some embodiments, the cross-entropy between A and each candidate semantic center cluster is sorted according to the numerical value based on the cross-entropy values ​​described above.

[0084] The candidate semantic center cluster with the smallest cross-entropy value is defined as the semantic basis cluster of A, that is, the semantic center cluster that is closest to A in the semantic space.

[0085] The candidate semantic center cluster with the largest cross-entropy value is defined as the semantic vertex cluster of A, that is, the semantic center cluster that is most different from A in the semantic space and has the most dispersed semantics.

[0086] Preferably, candidate semantic center clusters whose cross-entropy values ​​are at the median position can also be selected as the semantic median cluster of A, to supplement the characterization of the intermediate distribution of A in the overall semantic structure. In this embodiment of the invention, at least the basis clusters and vertex clusters are used in subsequent calculations.

[0087] In some embodiments, the semantic center aggregation of A is subtracted from the semantic base aggregation of A in each dimension to obtain the semantic base spacing of A; similarly, the semantic center aggregation of A is subtracted from the semantic vertex aggregation of A in each dimension to obtain the semantic vertex spacing of A.

[0088] Subsequently, the semantic similarity between the semantic basis distance of A and the semantic center cluster of A is calculated, and this similarity is defined as the semantic basis centrality of A; the semantic similarity between the semantic vertex distance of A and the semantic center cluster of A is calculated, and this similarity is defined as the semantic vertex centrality of A.

[0089] The semantic similarity is preferably calculated using cosine similarity, which involves summing the products of the two vectors in each dimension, calculating the magnitude of each vector, and dividing the sum of the products by the product of the magnitudes of the two vectors to obtain a similarity value between negative one and positive one. Of course, in other embodiments, semantic similarity can also be achieved using Euclidean distance, dot product similarity, etc., and this invention does not limit this.

[0090] In some embodiments, a proportionality coefficient between zero and one is formed by using the smaller of the semantic basis centrality and the semantic vertex centrality of A as the numerator and the larger of the two as the denominator. This proportionality coefficient is defined as the semantic centrality of A.

[0091] Subsequently, the semantic and centrality of A are subtracted from the values ​​of each dimension in the semantic center cluster of A to obtain a new vector, which is defined as the semantic and centrality corresponding to the semantic center cluster of A. By uniformly subtracting the semantic and centrality, the position of A in the vector space can be consistently translated without changing the overall structure of the semantic space, thereby weakening the common semantic bias between different texts and highlighting the representative semantic structure within the text.

[0092] For all semantic center clusters corresponding to the same text data, repeat the above steps to obtain the semantic centrality and semantic centrality of each semantic center cluster.

[0093] In some embodiments, for any text data, based on the semantic and centrality of each semantic center cluster corresponding to the text data, multiple semantic center clusters are sorted in ascending order of their semantic and centrality values ​​to obtain a semantic and centrality cluster sequence; at the same time, the semantic and centrality corresponding to each semantic center cluster in the sequence are arranged sequentially for subsequent similarity calculation.

[0094] In the aforementioned sequence of semantic convergence centers, the semantic center with the smallest semantic convergence degree is selected as the reference semantic center, and its corresponding semantic convergence value is used as the reference semantic convergence value. Subsequently, the semantic similarity between the reference semantic convergence value and the semantic convergence values ​​corresponding to each semantic center in the sequence is calculated sequentially, and these similarity values ​​are arranged into a one-dimensional array according to the sequence order. This one-dimensional array is defined as the semantic convergence center array corresponding to the text data.

[0095] Preferably, the semantic similarity mentioned above is also calculated using cosine similarity. Through the semantic convergence center array, the overall convergent structural relationship between multiple semantic center clusters in the text can be characterized in the form of a one-dimensional numerical sequence, facilitating subsequent comparison of the semantic structures within different texts.

[0096] In some embodiments, the search objects undergo the same vectorization and clustering processing as the text data. Specifically, the search objects are segmented, preprocessed, and vectorized using a unified pre-trained language model to obtain multiple different text embedding vectors. These multiple text embedding vectors corresponding to the search objects are then clustered using a clustering algorithm to obtain several cluster center vectors, which serve as several semantic center clusters corresponding to the search objects. Preferably, the number of semantic center clusters corresponding to the search objects is consistent with the number of semantic center clusters corresponding to the text data in the search database.

[0097] For each semantic center cluster corresponding to the search object, the cross-entropy value is calculated with the other semantic center clusters corresponding to the search object according to the cross-entropy calculation rules described above. The arithmetic mean of the cross-entropy values ​​of the semantic center cluster and the other semantic center clusters is used as the semantic non-centrality of the semantic center cluster.

[0098] The semantic center clusters corresponding to the search object are sorted in ascending order of their semantic non-linear centrality values ​​to obtain the semantic non-linear cluster center sequence corresponding to the search object.

[0099] In the semantic disjoint clustering center sequence, for each positional semantic center cluster, the semantic similarity between it and the semantic center clusters of its adjacent positions is calculated as the semantic disjoint decentralization degree of that position. Specifically, for the first and last semantic center clusters, their adjacent positions can be taken as their single next or previous semantic center clusters, respectively; for the semantic center clusters located in the middle, their semantic disjoint decentralization degree can be taken as the arithmetic mean of the semantic similarity between it and the semantic center cluster of its previous position and the semantic similarity between it and the semantic center cluster of its next position.

[0100] Following the sequence of semantic disjoint cluster centers, the semantic disjoint decentralization degrees corresponding to each order are arranged sequentially to form a one-dimensional numerical array. This array is defined as the semantic disjoint decentralization array corresponding to the retrieval object. This array reflects the disjoint structural relationships between the semantic cluster centers within the retrieval object in the form of decentralization degrees.

[0101] In some embodiments, the semantic non-centralized array corresponding to the search object is subjected to curve fitting to obtain the semantic non-centralized curve corresponding to the search object; the semantic convergence center array corresponding to each text data in the search database is subjected to curve fitting to obtain the semantic convergence center curve corresponding to each text data.

[0102] Subsequently, using curve similarity, the semantic non-centralized curve corresponding to the search object is compared one by one with the semantic convergent centralized curve corresponding to each text data in the search database. Based on the magnitude of the curve similarity, the target curve that is most similar to the curve corresponding to the search object is selected, and the text data corresponding to the target curve is output to the user as the optimal text data in the search database.

[0103] Preferably, the curve similarity can be calculated using the Fraser distance method, which gives a distance quantification value based on the shape difference between the two curves over the entire domain. The smaller the distance value, the closer the shapes of the two curves are. Of course, in other embodiments, dynamic time warping, Euclidean distance, or cosine similarity based on sampling points can also be used to measure curve similarity. This invention does not limit this method.

[0104] Through the above steps, in the context of patent text quantitative optimization, this invention, while preserving the semantic expressive power of the language model, introduces a structured array and a non-structured array based on semantic center clustering. This abstracts the semantic structural relationship between the retrieval object and the internal text of the database into a curve shape for comparison, thereby achieving a fine characterization of the relationship between multiple themes and multiple technical elements in the patent text. This significantly improves the accuracy and stability of the retrieval results under the condition of a large-scale patent database.

[0105] The present invention provides an embodiment one:

[0106] 1) Obtain the search object and the search database, wherein the search database stores multiple text data;

[0107] 2) For each piece of text data in the retrieval database:

[0108] a) The text data is vectorized and embedded using a unified pre-trained language model, dividing the text data into multiple sub-text units and generating corresponding text embedding vectors for each sub-text unit.

[0109] b) Perform clustering operations on the multiple text embedding vectors corresponding to the text data to obtain multiple cluster center vectors, and use each cluster center vector as the multiple semantic center clustering quantity corresponding to the text data;

[0110] c) For each semantic center cluster A of this text data:

[0111] c1) Normalize the values ​​of each dimension of A and each candidate semantic center cluster that is different from A, and convert them into a probability distribution form;

[0112] c2) Using the probability distribution corresponding to A as the true distribution and the probability distribution corresponding to each candidate semantic center cluster as the predicted distribution, calculate the cross-entropy value between A and each candidate semantic center cluster.

[0113] c3) In the sorting results based on the cross-entropy values, the candidate semantic center cluster with the smallest cross-entropy value is selected as the semantic basis cluster of A, and the candidate semantic center cluster with the largest cross-entropy value is selected as the semantic vertex cluster of A.

[0114] c4) Subtract A from the semantic basis aggregate and semantic vertex aggregate in each dimension to obtain the semantic basis distance and semantic vertex distance of A. Calculate the first semantic similarity between the semantic basis distance and A and the second semantic similarity between the semantic vertex distance and A, where the semantic similarity is calculated based on vector similarity.

[0115] c5) Using the smaller value of the first semantic similarity and the second semantic similarity as the numerator and the larger value as the denominator, calculate the proportionality coefficient between zero and one, and define the proportionality coefficient as the semantic centrality of A.

[0116] c6) Subtract the semantic and centrality from the values ​​of A in each dimension to obtain a new vector, and define the new vector as the semantic and clustering quantity corresponding to A;

[0117] d) Based on the semantic and axial centrality of all semantic center clusters corresponding to the text data, sort each semantic center cluster in ascending order of semantic and axial centrality to obtain the semantic and axial cluster center sequence, and take the semantic and axial cluster corresponding to the first semantic center cluster in the sequence as the reference semantic and axial cluster.

[0118] e) Calculate the second vector similarity between the reference semantics and the semantics and the corresponding semantics and the clustering quantity of each semantic center according to the order of the semantics and the clustering quantity sequence. Arrange the second vector similarities in order to form a one-dimensional numerical array, which serves as the semantics and clustering quantity center array corresponding to the text data.

[0119] 3) Perform the following processing on the retrieved object:

[0120] a) The pre-trained language model is used to perform vectorized embedding processing on the retrieval object to obtain multiple text embedding vectors corresponding to the retrieval object. The multiple text embedding vectors corresponding to the retrieval object are then clustered to obtain multiple semantic center clusters corresponding to the retrieval object, which can be called retrieval semantic center clusters.

[0121] b) For each semantic center cluster corresponding to the search object, calculate the cross-entropy value with the other semantic center clusters corresponding to the search object according to the rules of steps 2)c1) and 2)c2), and take the arithmetic mean of the cross-entropy values ​​of the semantic center cluster relative to the other semantic center clusters as the semantic non-centrality of the semantic center cluster.

[0122] c) Based on the semantic non-centrality, sort the multiple semantic center clusters corresponding to the search object from small to large to obtain the semantic non-centrality cluster center sequence;

[0123] d) In the semantic disjoint clustering center sequence, for each position's semantic center cluster, calculate the third vector similarity between it and the semantic center clusters of its adjacent positions, and use the third vector similarity as the semantic disjoint decentralization degree of that position; wherein, for the first and last semantic center clusters, only the third vector similarity between it and its single adjacent position's semantic center cluster is used as the corresponding semantic disjoint decentralization degree, and for the semantic center clusters located in the middle, the arithmetic mean of the third vector similarity between it and the semantic center cluster of the previous position and the third vector similarity between it and the semantic center cluster of the next position is used as the corresponding semantic disjoint decentralization degree;

[0124] e) Arrange the semantic non-centrality of each bit order of the search object in the order of the semantic non-centrality clustering center sequence to obtain the semantic non-centrality array corresponding to the search object;

[0125] 4) Perform curve fitting on the semantic non-linear decentralization array corresponding to the search object and the semantic convergent centralization array corresponding to each text data in the search database to obtain the semantic non-linear decentralization curve corresponding to the search object and the semantic convergent centralization curve corresponding to each text data. Based on the curve similarity, compare the semantic non-linear decentralization curve with each semantic convergent centralization curve, select the target curve with the best curve similarity, and output the text data corresponding to the target curve as the optimal text data.

[0126] Preferably, the text data in the search database includes patent specification text and / or patent claim text, and is preferably string data of published patent documents obtained through web crawling and / or exporting from patent databases.

[0127] Preferably, in steps 2)a) and 3)a), the pre-trained language model is a language model pre-trained on patent domain corpus, and the sub-text unit is at least one of sentence, paragraph, and / or character fragment of preset length.

[0128] Preferably, in steps 2)b) and 3)a), the clustering operation on the text embedding vector is performed using a distance- or similarity-based clustering algorithm, wherein the clustering algorithm includes at least one of K-means clustering, hierarchical clustering and density clustering, and the number of semantic center clusters corresponding to each text data is greater than three.

[0129] Preferably, in step 2)c1), in order to convert the semantic center aggregate into a probability distribution form, the numerical values ​​of each semantic center aggregate in each dimension are subjected to exponential operation, and normalized by the sum of the exponents of each dimension, so that the numerical value of each dimension is non-negative and the sum of the numerical values ​​of all dimensions is equal to 1.

[0130] Preferably, in step 2)c4), the first semantic similarity and the second semantic similarity are cosine similarity, that is, the product of each dimension of the two vectors involved in the calculation is summed, and the magnitude of the two vectors is calculated respectively. The sum of the products is divided by the product of the magnitudes of the two vectors to obtain a similarity value between negative one and positive one.

[0131] Preferably, the second vector similarity and the third vector similarity in steps 2)e) and 3)d) are calculated using cosine similarity; in other embodiments, the second vector similarity and / or the third vector similarity can also be calculated using dot product similarity and / or reciprocal similarity based on Euclidean distance.

[0132] Preferably, in step 4), the curve similarity is calculated based on the Fraser distance, that is, the distance quantization value is obtained based on the shape difference of the two curves in the global domain, and the smaller the distance value, the closer the curve shapes are; in other embodiments, the curve similarity can also be calculated using dynamic time warping distance, Euclidean distance and / or cosine similarity based on sampling points.

[0133] Preferably, in step 2), the operation of constructing semantic and convergent central arrays for each text data in the search database is completed in the offline batch processing stage, and the semantic and convergent central arrays corresponding to each text data are pre-stored in the index library; steps 3) and 4) are executed online after receiving the search object, and the curve similarity comparison is performed only based on the semantic non-convergent central array corresponding to the search object and the semantic and convergent central arrays pre-stored in the index library.

[0134] The present invention also provides an embodiment two:

[0135] This embodiment, based on the aforementioned Embodiment 1, provides a specific application of the method of the present invention in a high-value patent evaluation scenario, illustrating how the technical solution in Embodiment 1 can be applied to the relevant patent screening and reordering stage in the patent value evaluation process, specifically including:

[0136] Obtain the target patent document to be evaluated. The target patent should include at least the invention title, abstract, claims, and specification.

[0137] Based on the fixed text structure of patent writing, the target patent is structured and parsed to extract text data content containing strings such as "invention title text", "abstract text", "claims text", and "specification text" from the text document. Source markers are recorded within these text data to distinguish texts from different sources later.

[0138] In this embodiment, the concatenated content of the abstract text and claims text of the target patent is used as the "search object" text data content in the method of the present invention, while the specification text of the target patent is retained for subsequent use as contextual supplementation during large language model evaluation.

[0139] Keyword extraction is performed on the text of each field of the target patent. The algorithm used can be TF-IDF, TextRank and / or keyword extraction algorithm based on a large language model. The source dimension is introduced in the calculation process so that the keyword sets for "Abstract", "Claims" and "Specification" are obtained respectively.

[0140] The above keywords are expanded using synonyms. The expansion methods can include: dictionary / knowledge base-based expansion, semantic co-occurrence network-based expansion, vector representation-based expansion, and end-to-end expansion based on large language models, etc., to obtain keyword clusters that cover different expressions.

[0141] The retrieval query is generated based on the combination of keywords and their synonyms. The generation method can be a combination of preset rules / templates, or a sequence-to-sequence generation model based on deep learning or large language models. This embodiment does not limit this method.

[0142] Using the aforementioned Query, the local patent search engine and the external general search engine are invoked successively to retrieve several candidate patent documents related to the target patent from the invention patent data published in recent years, forming a candidate patent set; each candidate patent includes at least its title, abstract, claims and specification text.

[0143] The same structured analysis is performed on each candidate patent, and the concatenated text of its abstract and claims is used as the text data in the method of this invention; the specification text is retained as supplementary material for subsequent large language model evaluation.

[0144] The core improvement of this invention begins with the deep semantic structure sorting within the candidate patent set. The semantics of the candidate patents are constructed into a cluster center array, which can be applied to the database-side process of Implementation Example 1.

[0145] In this embodiment, each piece of text data in the obtained candidate patent set is regarded as text data in the retrieval database in Embodiment 1. The complete database-side processing flow in Embodiment 1 is performed on each piece of text data, including but not limited to:

[0146] The text data is vectorized and embedded using a unified pre-trained language model, which splits the text into multiple sub-text units and generates multiple text embedding vectors.

[0147] Clustering is performed on multiple text embedding vectors to obtain several semantic center clusters;

[0148] For each semantic center cluster, cross-entropy is calculated with other semantic center clusters. The cluster with the smallest cross-entropy is the semantic basis cluster, and the cluster with the largest cross-entropy is the semantic vertex cluster. The semantic basis distance and semantic vertex distance are calculated.

[0149] Based on the semantic similarity between the semantic basis distance / semantic vertex distance and the semantic center clustering, the semantic basis centrality and semantic vertex centrality are obtained, and the semantic centrality between 0 and 1 is obtained accordingly.

[0150] Subtract the semantic centrality from each dimension of the semantic centrality to obtain the semantic centrality;

[0151] Sort all semantic center clusters of the same text data in ascending order of semantic centrality. Select the semantic central cluster with the smallest semantic centrality as the reference semantic central cluster. Calculate the semantic similarity between the reference semantic central cluster and all other semantic central clusters in sequence, and arrange them in order to form the semantic central cluster array of the text data.

[0152] The above steps are completely consistent with the steps described in Example 1 for searching the database. This example will not repeat the formula derivation, but only specifies the application scenario: the text data here is the combined text file of the candidate patent abstract and the claims.

[0153] In this embodiment, the constructed target patent abstract and claims combined text are used as the search object, and the complete search object-side processing flow is performed on the search object, including but not limited to:

[0154] Using a unified pre-trained language model, the retrieved objects are vectorized, embedded, and clustered to obtain several semantic center clusters.

[0155] For each semantic center cluster of the retrieved object, cross-entropy is calculated with the other semantic center clusters, and the arithmetic mean of the cross-entropy is used as the semantic non-centrality of that semantic center cluster.

[0156] Sort all semantic center clusters in ascending order of semantic non-centrality to obtain the semantic non-centrality cluster center sequence;

[0157] In the semantic non-coaxial clustering center sequence, for each position's semantic center clustering, the semantic similarity between it and the semantic center clustering of adjacent positions is calculated, and the single adjacent similarity of the first / last position or the average of the upper and lower similarities of the middle positions is taken as the semantic non-coaxial decentralization degree of that position.

[0158] Arrange the semantic non-centrality of each position into a one-dimensional array according to the sequence order to obtain the semantic non-centralized array corresponding to the target patent.

[0159] Curve fitting is performed on the semantic non-centralized array corresponding to the target patent to obtain the semantic non-centralized curve of the target patent;

[0160] Curve fitting is performed on the semantics and clustering center array corresponding to each text data in the candidate patent set to obtain the semantics and clustering center curve of each candidate patent;

[0161] Curve similarity measures such as Fraser distance and / or dynamic time warping distance are used to calculate the curve similarity between the semantic non-centralized curve of the target patent and the semantic convergent curve of each candidate patent.

[0162] Candidate patents are reordered based on curve similarity, and the best-performing candidate patents are selected as the key comparison patent set in the subsequent high-value evaluation process.

[0163] Compared with traditional sorting schemes that rely solely on keyword or first-order vector similarity, this embodiment achieves deep matching of target and candidate patents at the semantic structure level through processing of semantic center clustering, oriented / non-oriented centrality, and curve similarity. This significantly improves the accuracy of comparative patent screening for high-value search recommendations and provides a more reliable semantic evidence basis for subsequent value assessment by large language models.

[0164] In one optional implementation, the original content of the target patent and the selected set of key comparative patents are organized into prompts for input into a large language model using a search-enhanced generation (RAG) method. The prompts include multi-dimensional evaluation guidance such as legal, technical, and market dimensions. The large language model outputs the value level of the target patent in each dimension and the corresponding reasons. The comparative evidence texts provided by the large language model are preferably from the key comparative patents screened by the method of this invention, thereby making the high-value evaluation results more interpretable and stable.

[0165] The language model-enhanced patent text search and measurement system runs on any computing device, such as a desktop computer, laptop computer, handheld computer, or cloud data center. The computing device includes a processor, a memory, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps in the language model-enhanced patent text search and measurement method. The runnable system may include, but is not limited to, a processor, a memory, and a server cluster.

[0166] The embodiments of the present invention provide a patent text search measurement system based on language model enhancement, such as... Figure 2 As shown, the language model-enhanced patent text search measurement system of this embodiment includes: a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the above-described language model-enhanced patent text search measurement method embodiment. The processor executes the computer program in the following system unit:

[0167] The data acquisition unit is used to acquire the retrieval object and the retrieval database, which stores multiple text data; each text data in the retrieval database generates multiple text embedding vectors through a language model, and the multiple text embedding vectors are clustered to obtain multiple semantic center clusters;

[0168] The option processing unit is used to calculate the divergence metric between each target semantic center cluster and the other semantic center clusters for each target semantic center cluster. It then compares the target semantic center cluster with the target semantic center clusters with the largest and smallest divergence metrics to obtain the semantic convergence degree. The target semantic center cluster is subtracted from the semantic convergence degree to obtain its semantic convergence quantity. Based on the semantic convergence degree of each semantic center cluster in the text data, the semantic center clusters are sorted, and a reference semantic center cluster is selected from the sort. The vector similarity between the semantic convergence quantity of the reference semantic center cluster and the semantic convergence quantities of each semantic center cluster in the sort is calculated. The values ​​of each vector similarity form the semantic convergence quantity center array of the text data.

[0169] The retrieval processing unit is used to perform vectorized embedding and clustering on the retrieval object to obtain multiple retrieval semantic center clusters. For each of the multiple retrieval semantic center clusters, a statistical measure of the divergence between the retrieval semantic center cluster and the other retrieval semantic center clusters is calculated as the semantic non-centrality of the retrieval semantic center cluster. Based on the semantic non-centrality, the multiple retrieval semantic center clusters are sorted, and the semantic non-centrality corresponding to each position is obtained according to the vector similarity between the retrieval semantic center clusters in the sorted sequence, forming a semantic non-central array of the retrieval object.

[0170] The curve matching unit is used to perform curve fitting on the semantic non-linear decentering array of the search object and the semantic convergent central array of each text data in the search database, respectively, to obtain the semantic non-linear decentering curve of the search object and the semantic convergent central curve of each text data. Based on the curve similarity metric, the text data with the most similar semantic non-linear decentering curves is selected as the output.

[0171] In order to better unify the linear relationship and probabilistic connection between physical quantities with different units of measurement, dimensionless processing can be performed on different physical quantities.

[0172] Preferably, all undefined variables in this invention, if not explicitly defined, can be manually set thresholds.

[0173] The language model-enhanced patent text search and measurement system can run on computing devices such as desktop computers, laptops, handheld computers, and cloud data centers. The language model-enhanced patent text search and measurement system includes, but is not limited to, processors and memory. Those skilled in the art will understand that the examples described are merely illustrations of the language model-enhanced patent text search and measurement method, system, and device, and do not constitute a limitation on the language model-enhanced patent text search and measurement method, system, and device. It may include more or fewer components, or combine certain components, or different components. For example, the language model-enhanced patent text search and measurement system may also include input / output devices, network access devices, buses, etc.

[0174] The present invention also provides an electronic device, a readable storage medium, and a computer program product:

[0175] An electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the language model-enhanced patent text search measurement method and the method for each step therein.

[0176] A non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to perform the language model-enhanced patented text search measurement method and the methods for each step therein.

[0177] A computer program product includes a computer program that, when executed by a processor, implements the language model-enhanced patent text search measurement method and the methods for each step thereof.

[0178] The term "electronic device" is intended to refer to various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices can also refer to various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0179] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0180] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0181] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0182] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0183] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with embodiments of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0184] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other.

[0185] The processor referred to can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete component gate circuits, transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the language model-enhanced patent text search and measurement system, connecting various sub-regions of the system via various interfaces and lines.

[0186] The memory can be used to store the computer programs and / or modules. The processor implements various functions of the language model-enhanced patent text search measurement method, system, and device by running or executing the computer programs and / or modules stored in the memory and calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function (such as sound playback function, image playback function, etc.), etc.; the data storage area may store data created based on the use of the mobile phone (such as audio data, phonebook, etc.). In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.

[0187] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0188] This invention provides a method, system, and device for patent text search measurement based on language model enhancement. The method first embeds each patent text in the retrieval database using a unified pre-trained language model and clusters the embedded vectors to obtain multiple semantic center clusters. Based on the divergence metric and vector similarity between the semantic center clusters within the same text, a first structural feature array representing the semantic structural relationships within the text is constructed. The retrieval object is similarly embedded and clustered to obtain multiple retrieval semantic center clusters, and a second structural feature array is constructed based on their internal divergence and similarity relationships to represent the internal semantic structure of the retrieval object. The system further uses sequence similarity or curve similarity metrics to compare the second structural feature array of the retrieval object with the first structural feature arrays of each text, selecting the text with the highest similarity as the optimal text data output. This invention, while retaining the semantic expressive power of the pre-trained language model, introduces a structured measurement of the semantic center relationships within the text, achieving fine-grained matching at the semantic structure level of patent text information, significantly improving the accuracy and stability of patent retrieval and similarity analysis.

[0189] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A patent text search measurement method based on language model enhancement, characterized in that, The method includes: The search object and the search database are obtained. The search database stores multiple text data. Each text data in the search database generates multiple text embedding vectors through a language model. The multiple text embedding vectors are clustered to obtain multiple semantic center clusters. For each target semantic center cluster, calculate the divergence metric between the target semantic center cluster and the other semantic center clusters. Compare the vector differences between the target semantic center cluster and the target semantic center clusters with the largest and smallest divergence metrics to obtain the semantic convergence centrality. Subtract the semantic convergence centrality from the target semantic center cluster to obtain its semantic convergence mass. Based on the semantic convergence centrality of each semantic center cluster in the text data, sort the semantic center clusters and select a reference semantic center cluster from the sort. Calculate the vector similarity between the semantic convergence mass of the reference semantic center cluster and the semantic convergence centrality of each semantic center cluster in the sort. Use the values ​​of each vector similarity to form the semantic convergence centrality array of the text data. The search objects are vectorized, embedded, and clustered to obtain multiple search semantic center clusters. For each of the multiple search semantic center clusters, a statistical measure of the divergence between the search semantic center cluster and the other search semantic center clusters is calculated as the semantic non-centrality of the search semantic center cluster. The multiple search semantic center clusters are sorted based on the semantic non-centrality, and the semantic non-centrality corresponding to each position is obtained according to the vector similarity between search semantic center clusters with adjacent positions in the sorted sequence, forming a semantic non-central array of the search objects. Curve fitting is performed on the semantic non-linear decentering array of the search object and the semantic convergence centering array of each text data in the search database to obtain the semantic non-linear decentering curve of the search object and the semantic convergence centering curve of each text data. Based on the curve similarity measure, the text data with the most similar semantic non-linear decentering curves are selected as the output. Specifically, the semantic centrality is obtained by comparing the vector differences between the target semantic center cluster size and the target semantic center cluster size with the largest and smallest divergence metric values, respectively. The semantic centrality is obtained by subtracting the semantic centrality from the target semantic center cluster size, which specifically includes: For each target semantic center cluster among the plurality of semantic center clusters, the divergence metric between the target semantic center cluster and the other semantic center clusters is calculated. The semantic center cluster with the smallest divergence metric is determined as the semantic base cluster of the target semantic center cluster, and the semantic center cluster with the largest divergence metric is determined as the semantic vertex cluster of the target semantic center cluster. Based on the vector relationship between the target semantic center cluster and the semantic base cluster and the semantic vertex cluster, a first centrality index and a second centrality index are obtained. The semantic and axial centrality of the target semantic center cluster is determined based on the first centrality index and the second centrality index, and the semantic and axial centrality of the target semantic center cluster is generated based on the semantic and axial centrality. The semantics and centrality are scalars, and are constructed based on the relationship between the first centrality index and the second centrality index; The semantic disjoint centrality is the arithmetic mean of the divergence measures between the retrieved semantic center cluster and the remaining retrieved semantic center clusters; when determining the semantic disjoint decentrality of each position in the semantic disjoint cluster center sequence: For a retrieval semantic center cluster located at the first or last position, its semantic non-centrality is determined by the vector similarity between it and the unique adjacent retrieval semantic center cluster. For a retrieval semantic center cluster located in the middle, its semantic non-centrality is determined by the arithmetic mean of its vector similarity with the previous rank retrieval semantic center cluster and its vector similarity with the next rank retrieval semantic center cluster.

2. The patent text search measurement method based on language model enhancement according to claim 1, characterized in that, The text data in the search database includes patent specification text and / or patent claim text, which are string data of published patent documents obtained through web crawling and / or exporting from patent databases.

3. The patent text search measurement method based on language model enhancement according to claim 1 or 2, characterized in that, The language model includes a language model pre-trained and / or fine-tuned on the patent corpus. The sub-text unit corresponding to the text embedding vector is at least one of sentence, paragraph and / or character segment of preset length. The clustering of multiple text embedding vectors adopts a clustering algorithm based on vector distance or vector similarity. The clustering algorithm includes at least one of K-means clustering, hierarchical clustering and density clustering.

4. The patent text search measurement method based on language model enhancement according to claim 1, characterized in that, The divergence metric is the cross-entropy metric. When calculating the cross-entropy metric, the values ​​of the semantic center clusters involved in the calculation are normalized in each dimension to convert them into a probability distribution. Specifically, after performing an exponential operation on the values ​​of each dimension, the sum of the exponents of each dimension is normalized to make the value of each dimension non-negative and the sum of the values ​​of all dimensions equal to 1.

5. The patent text search measurement method based on language model enhancement according to claim 1, characterized in that, in, The first centrality metric and the second centrality metric are centrality metrics based on vector similarity, specifically: The first centrality index is determined based on the vector similarity between the difference vector between the target semantic center cluster and its semantic basis cluster and the vector similarity between the target semantic center cluster. The second centrality index is determined based on the vector similarity between the difference vector between the target semantic center cluster and its semantic vertex cluster and the vector similarity between the target semantic center cluster.

6. The patent text search measurement method based on language model enhancement according to claim 1, characterized in that, in, When constructing the semantic convergence center array, the reference semantic center convergence is the semantic center convergence at a preset extreme value position of semantic convergence centrality.

7. The patent text search measurement method based on language model enhancement according to claim 1, characterized in that, in, The construction of semantic clustering center arrays corresponding to text data in the retrieval database is completed in the offline batch processing stage. The semantic clustering center arrays corresponding to each text data are pre-stored in the index. When a retrieval object is received, the curve similarity comparison is performed only based on the semantic non-centralized array of the retrieval object and the pre-stored semantic clustering center array in the index.

8. A patent text search and measurement system based on language model enhancement, characterized in that, The language model-enhanced patent text search measurement system operates on any computing device, such as a desktop computer, a laptop computer, or a cloud data center. The computing device includes a processor, a memory, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it implements the steps of the language model-enhanced patent text search measurement method as described in any one of claims 1 to 7.

9. An electronic device, comprising: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions executable by the at least one processor, characterized in that the instructions are executed by the at least one processor to enable the at least one processor to perform the method according to any one of claims 1 to 7.

Images