An emerging technology evolution analysis method, device and storage medium

Abstract

The application discloses an emerging technology evolution analysis method and device and a storage medium. First, the patent data of the target technology is obtained, and is divided according to time periods to observe the development state of the technology at different time nodes. Then, the patent data is preprocessed and text vector conversion is performed, main technical topics in each time period are extracted through dimension reduction processing and clustering analysis. Secondly, the abstract texts of each topic are spliced to obtain target abstract texts, and keywords are extracted from the target abstract texts by calculating the cosine similarity between the word vectors and the document vectors, so that the topic name of each topic is determined. Finally, the association relationship between the topics is determined according to the cosine similarity between the topics, and the evolution path of the target technology is determined. Thus, the evolution path of the target technology can be efficiently and accurately captured, the accuracy of the technology evolution analysis is improved, and the technical problem of low accuracy of the technology evolution analysis in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of big data, and in particular, relates to a new technology evolution analysis method and device and a storage medium. BACKGROUND

[0002] New technology evolution analysis is a method of studying the process and rules of technology from its emergence to maturity, diffusion, and integration with other technologies. New technology evolution analysis can help individuals, enterprises, governments, and other different subjects better understand the trends, rules, and impacts of technology development, so as to make more intelligent decisions. Enterprises can predict which new technologies may become mainstream technologies in the future through technology evolution analysis, so as to allocate research and development resources in advance, formulate technology strategies, and avoid falling behind or being eliminated by the market. Governments can use technology evolution analysis to formulate science and technology policies, guide resources to new technology fields with strategic significance, and promote industrial upgrading and economic restructuring.

[0003] Traditionally, technology evolution analysis relies on expert experience and qualitative description. This method is not only time-consuming and labor-intensive, but also susceptible to subjective factors, making it difficult to comprehensively and objectively reflect the true evolution process of technology. In recent years, with the continuous progress of big data and artificial intelligence technologies, methods of analyzing technology evolution using patent data have gradually emerged. Patent data, as an important carrier of technological innovation, contains rich technical information and development trends, providing a new perspective and means for technology evolution analysis.

[0004] However, existing technology evolution analysis methods based on patent data still have many shortcomings. For example, some methods simply count the number of patents, ignoring the technical details and theme changes in patent content, resulting in low accuracy of analysis results. Other methods attempt to analyze the content of patent text, but often have low accuracy in theme extraction and keyword recognition due to the high dimensionality and noise interference of text vectors.

[0005] Therefore, how to efficiently process and analyze patent data and accurately capture the evolution path of technology has become a problem to be solved in the field of technology evolution analysis. SUMMARY

[0006] Embodiments of the present disclosure provide a new technology evolution analysis method to at least solve the problem of low accuracy of technology evolution analysis in the prior art.

[0007] According to an aspect of embodiments of the present disclosure, there is provided a method for emerging technology evolution analysis, comprising: obtaining patent data of a target technology for evolution analysis, dividing the patent data into S time periods according to time periods; preprocessing the patent data of each time period, converting the abstract text of each patent into a first text vector of a first preset dimension; performing dimension reduction processing on the first text vector to obtain a second text vector of a second preset dimension, and clustering the second text vector to obtain M topics of each time period; for each time period, splicing the abstract text of all patent data in each topic to obtain a target abstract text of each topic; converting each word in the target abstract text of each topic of each time period into a word vector of the first preset dimension, and converting the entire document of each patent of each topic of each time period into a document vector of the first preset dimension; calculating the cosine similarity between the word vector and the document vector, and extracting N keywords from the target abstract text as key words according to the N highest results of the cosine similarity, to obtain N key words of each topic of each time period; determining the topic name of each topic of each time period according to the N key words, calculating the cosine similarity between all topics of the previous and next two time periods according to the topic name, and determining the association relationship between each topic according to the cosine similarity between the topics; determining the evolution path of the target technology according to the association relationship; wherein S is an integer greater than or equal to 3, M is an integer greater than or equal to 2, and N is an integer greater than or equal to 2; the first preset dimension and the second preset dimension are pre-set, and the first preset dimension is greater than the second preset dimension.

[0008] According to another aspect of embodiments of the present disclosure, there is also provided a storage medium, the storage medium comprising a stored program, wherein the program is executed by a processor when the program is running.

[0009] According to another aspect of the embodiments of the present disclosure, there is also provided an emerging technology evolution analysis device, comprising: a patent data acquisition module configured to acquire patent data of a target technology for evolution analysis, and divide the patent data into S time periods according to time periods; an abstract vectorization module configured to pre-process the patent data of each time period, and convert the abstract text of each patent into a first text vector of a first preset dimension; a dimension reduction module configured to perform dimension reduction processing on the first text vector to obtain a second text vector of a second preset dimension, and perform clustering on the second text vector to obtain M topics of each time period; a splicing module configured to splice the abstract text of all patent data in each topic to obtain a target abstract text of each topic for each time period; a document vectorization module configured to convert each word in the target abstract text of each topic of each time period into a word vector of the first preset dimension, and convert the entire document of each patent of each topic of each time period into a document vector of the first preset dimension; a keyword extraction module configured to calculate the cosine similarity between the word vector and the document vector, and extract N keywords from the target abstract text as keywords according to the N highest results of the cosine similarity, to obtain N keywords of each topic of each time period; an association determination module configured to determine the topic name of each topic of each time period according to the N keywords, calculate the cosine similarity between all topics of two adjacent time periods according to the topic name, and determine the association relationship between the topics according to the cosine similarity between the topics; an evolution module configured to determine the evolution path of the target technology according to the association relationship; wherein S is an integer greater than or equal to 3, M is an integer greater than or equal to 2, and N is an integer greater than or equal to 2; the first preset dimension and the second preset dimension are preset, and the first preset dimension is greater than the second preset dimension.

[0010] According to another aspect of the embodiments of the present disclosure, there is also provided an emerging technology evolution analysis device, comprising: a processor; and a memory connected with the processor, configured to provide the processor with instructions to process the following processing steps: obtaining patent data of a target technology for evolution analysis, dividing the patent data into S time periods according to time periods; preprocessing the patent data of each time period, converting the abstract text of each patent into a first text vector of a first preset dimension; performing dimension reduction processing on the first text vector to obtain a second text vector of a second preset dimension, and clustering the second text vector to obtain M topics of each time period; for each time period, splicing the abstract text of all patent data in each topic to obtain a target abstract text of each topic; converting each word in the target abstract text of each topic of each time period into a word vector of the first preset dimension, and converting the entire document of each patent of each topic of each time period into a document vector of the first preset dimension; calculating the cosine similarity between the word vector and the document vector, and extracting N keywords from the target abstract text as keywords according to the N highest results of the cosine similarity, to obtain N keywords of each topic of each time period; determining the topic name of each topic of each time period according to the N keywords, calculating the cosine similarity between all topics of the previous and subsequent two time periods according to the topic name, and determining the association relationship between the topics according to the cosine similarity between the topics; determining the evolution path of the target technology according to the association relationship; wherein the S is an integer greater than or equal to 3, the M is an integer greater than or equal to 2, and the N is an integer greater than or equal to 2; the first preset dimension and the second preset dimension are preset, and the first preset dimension is greater than the second preset dimension.

[0011] In the embodiments of the present disclosure, first, the patent data of the target technology is obtained and divided according to time periods, so as to observe the development state of the technology at different time nodes. Then, the patent data is preprocessed and text vector conversion is performed, the main technical topics in each time period are extracted through dimension reduction processing and clustering analysis. Further, the abstract text of each topic is spliced to obtain a target abstract text, and keywords are extracted from the target abstract text by calculating the cosine similarity between the word vector and the document vector, so as to determine the topic name of each topic. Finally, the association relationship is determined according to the cosine similarity between the topics, and the evolution path of the target technology is determined. The present application can efficiently and accurately capture the evolution path of the target technology, thereby improving the accuracy of the technology evolution analysis and solving the technical problem of low accuracy of the technology evolution analysis in the prior art. BRIEF DESCRIPTION OF DRAWINGS

[0012] The accompanying drawings, which are included to provide a further understanding of the present disclosure and are incorporated in and constitute a part of this application, illustrate embodiments of the present disclosure and serve to explain the principles of the present disclosure, and do not limit the present disclosure. In the drawings:

[0013] Figure 1 is a hardware structure block diagram of a computing device for implementing the method according to Embodiment 1 of the present disclosure;

[0014] Figure 2 is a flowchart of the emerging technology evolution analysis method according to Embodiment 1 of the present disclosure;

[0015] Figure 3 is a schematic diagram of the dimensionality reduction result according to the first aspect of Embodiment 1 of the present disclosure;

[0016] Figure 4 is a schematic diagram of the clustering result according to the first aspect of Embodiment 1 of the present disclosure;

[0017] Figure 5 is a schematic diagram of the fine-tuned BERT model according to the first aspect of Embodiment 1 of the present disclosure;

[0018] Figure 6 is a schematic diagram of the PBAT field technology evolution according to the first aspect of Embodiment 1 of the present disclosure;

[0019] Figure 7 is a schematic diagram of the emerging technology evolution analysis device according to Embodiment 2 of the present disclosure;

[0020] Figure 8 is a schematic diagram of the emerging technology evolution analysis device according to Embodiment 3 of the present disclosure. DETAILED DESCRIPTION

[0021] In order to enable persons skilled in the art to better understand the technical solutions of the present disclosure, the technical solutions in the embodiments of the present disclosure will be described clearly and completely below with reference to the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, not all the embodiments. Based on the embodiments in the present disclosure, all other embodiments obtained by persons skilled in the art without creative labor should be within the scope of protection of the present disclosure.

[0022] It should be noted that the terms "first", "second", and the like in the description and claims of the present disclosure and above-mentioned drawings are intended to distinguish similar objects, and are not necessarily used to describe a particular order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present disclosure described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0023] First, some of the nouns or terms that appear in the description of the embodiments of the present disclosure are applicable to the following explanations:

[0024] Distilbert-base-nli-mean-tokens model: a natural language processing model developed based on the Sentence-Transformers library, mainly used to map sentences and paragraphs to a 768-dimensional dense vector space, suitable for tasks such as semantic similarity calculation, text clustering, information retrieval, etc.

[0025] UMAP: Uniform Manifold Approximation and Projection, is a nonlinear dimensionality reduction technique used to map high-dimensional data to low-dimensional space while preserving the local and global structure of the data.

[0026] Kmeans algorithm: a clustering algorithm used to divide a data set into K clusters, so that the data points within the cluster are as similar as possible, while the data points between different clusters are as dissimilar as possible.

[0027] nltk tokenization: Natural Language Toolkit, is a widely used Python library for natural language processing (NLP).

[0028] NLI dataset refers to a dataset used for natural language inference (Natural Language Inference, NLI) tasks. Natural language inference is a natural language processing (NLP) task that aims to determine the logical relationship between two sentences (premise and hypothesis), which is usually divided into three types: entailment, contradiction and neutral.

[0029] BERT, Bidirectional Encoder Representations from Transformers, is a pre-trained language model based on the Transformer architecture, designed to improve the performance of natural language processing (NLP) tasks through bidirectional contextual information.

[0030] AdamW Optimizer: The AdamW optimizer is an improved version of the Adam optimizer, incorporating weight decay regularization to enhance model training stability and generalization ability. The Adam (Adaptive Moment Estimation) optimizer is an adaptive learning rate optimization algorithm that combines the advantages of momentum and RMSProp optimization methods. Its core idea is to adjust the learning rate individually for each parameter, while utilizing momentum to accelerate convergence and reduce oscillations.

[0031] Example 1

[0032] According to this embodiment, a method embodiment for analyzing the evolution of emerging technologies is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0033] The method embodiments provided in this example can be executed on mobile terminals, computer terminals, servers, or similar computing devices. Figure 1 A hardware block diagram of a computing device for implementing an evolution analysis method for emerging technologies is shown. Figure 1 As shown, a computing device may include one or more processors (processors may include, but are not limited to, microprocessors such as MCUs or programmable logic devices such as FPGAs), memory for storing data, transmission devices for communication functions, and input / output interfaces. The memory, transmission devices, and input / output interfaces are connected to the processor via a bus. In addition, it may also include a display, keyboard, and cursor control device connected to the input / output interfaces. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, a computing device may also include... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0034] It should be noted that the one or more processors and / or other data processing circuitry described above can be referred to herein generically as "data processing circuitry." The data processing circuitry can be embodied as, include or otherwise be associated with software, hardware, firmware, or any combination thereof. Further, the data processing circuitry can be a single independent processing module or any combination of plural processing modules, wholly or partially incorporated into other elements of the computing device. As referred to in the embodiments of the present disclosure, the data processing circuitry functions as a processor to control, for example, selection of the variable resistance terminal path connected to the interface.

[0035] The memory can be used to store software programs of application software and modules, such as program instructions / data storage means corresponding to the emerging technology evolution analysis method in the embodiments of the present disclosure. The processor performs various functional applications and data processing by running the software programs and modules stored in the memory, i.e., implements the emerging technology evolution analysis method of the application program described above. The memory can include a high-speed random access memory, and can also include a non-volatile memory, such as one or more magnetic storage devices, flash memories, or other non-volatile solid-state memories. In some examples, the memory can further include a memory remotely arranged with respect to the processor, which can be connected to the computing device through a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and a combination thereof.

[0036] The transmission device is used to receive or send data via a network. Specific examples of the above-mentioned network can include a wireless network provided by a communication provider of the computing device. In one example, the transmission device includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to be able to communicate with the Internet. In one example, the transmission device can be a radio frequency (Radio Frequency, RF) module, which is used to communicate with the Internet in a wireless manner.

[0037] The display can be, for example, a touch screen type liquid crystal display (LCD), which can enable a user to interact with the user interface of the computing device.

[0038] It should be noted that in some optional embodiments, the above-mentioned Figure 1 The computing device shown can include hardware elements (including circuitry), software elements (including computer code stored on a computer readable medium), or combinations of both hardware and software elements. It should be noted that in some embodiments, the functions of the above-mentioned Figure 1 is merely one example of a particular implementation and is intended to illustrate the types of components that can be present in the computing device described above.

[0039] Under the above operating environment, according to a first aspect of the embodiment, there is provided a new technology evolution analysis method, which is implemented by a computing device as shown in Figure 1 . Figure 2 A flowchart of the method is shown, referring to Figure 2 , the method comprises:

[0040] S202, obtaining patent data of a target technology for evolution analysis, and dividing the patent data into S time periods according to time periods;

[0041] S204, preprocessing the patent data of each time period, and converting the abstract text of each patent into a first text vector of a first preset dimension;

[0042] S206, performing dimension reduction processing on the first text vector to obtain a second text vector of a second preset dimension, and clustering the second text vector to obtain M topics of each time period;

[0043] S208, for each time period, splicing the abstract text of all patent data in each topic to obtain a target abstract text of each topic;

[0044] S210, converting each word in the target abstract text of each topic of each time period into a word vector of the first preset dimension, and converting the entire document of each patent of each topic of each time period into a document vector of the first preset dimension;

[0045] S212, calculating the cosine similarity between the word vector and the document vector, and extracting N keywords from the target abstract text as keywords according to the N highest results of the cosine similarity, to obtain N keywords of each topic of each time period;

[0046] S214, determining the topic name of each topic of each time period according to the N keywords, calculating the cosine similarity between all topics of the previous and subsequent two time periods according to the topic name, and determining the association relationship between the topics according to the cosine similarity between the topics;

[0047] S216, determining the evolution path of the target technology according to the association relationship;

[0048] Wherein, the S is an integer greater than or equal to 3, the M is an integer greater than or equal to 2, and the N is an integer greater than or equal to 2;

[0049] The first preset dimension and the second preset dimension are preset, and the first preset dimension is greater than the second preset dimension.

[0050] In the emerging technology evolution analysis method of the embodiment, first, patent data of a target technology (such as biotechnology, information technology, etc.) to be subjected to evolution analysis is acquired, and the patent data is divided into S time periods (for example, S=3, that is, divided into three periods: budding stage, high-speed development stage, and stable development stage) according to time periods. Then, a pre-trained SentenceBERT model is used to vectorize the summary text of the patent data of the target technology, and after dimension reduction and clustering, a plurality of topics are obtained. The summary text of each topic is spliced to obtain a target summary text, and the target summary text is converted into a word vector. The entire document of each patent of each topic in each time period is converted into a document vector. The cosine similarity of the word vector and the document vector is calculated, the keywords of each topic are determined according to the cosine similarity, and the topic name of each topic is determined according to the keywords of each topic. The cosine similarity between all topics in the two time periods before and after is calculated, the association relationship between the topics is determined according to the cosine similarity between the topics, and thus the evolution path of the target technology is determined according to the association relationship, so that the evolution path of the target technology is efficiently and accurately captured, and the accuracy of the technology evolution analysis is greatly improved, thereby solving the technical problem of low accuracy of the technology evolution analysis in the prior art.

[0051] In the embodiment of the application, the target technology is a technology that needs to be subjected to evolution analysis, which can be a subdivided technical field or an industry technology. For example, it can be PBAT subdivided technology in the biotechnology field (PBAT, Polybutylene Adipate Terephthalate, is a biodegradable thermoplastic polymer, and its Chinese name is polybutylene adipate terephthalate. It is a copolymer of polybutylene adipate (PBA) and polybutylene terephthalate (PBT), which has the characteristics of both), printed circuit board technology in the electronic information technology, etc., and can also be big data technology, artificial intelligence technology, biomedicine technology, etc.

[0052] In the embodiment of the application, the patent data of the target technology to be subjected to evolution analysis is obtained by searching, screening, etc.

[0053] In the embodiment of the present application, the patent data is divided into S time periods according to the time period, which can be divided into S time periods from the earliest date of application date in the patent data to the date of obtaining the patent data according to the patent application date. It can also be divided into S time periods according to the history of technological development from the appearance of the target technology to a specified date, for example, the target technology appeared in 1885, and a specified date such as 2020, then 1885 to 2020 is divided into S time periods. As an optional example, S time periods can be divided into different stages according to the law and current situation of technological development, for example, it can be divided into three time periods of budding stage, high-speed development stage and stable development stage, and it can also be divided into four time periods of budding stage, high-speed development stage, stable development stage and decline stage. As an optional example, for PBAT technology, it can be divided into three time periods, which are: budding stage (1900-2011), high-speed development stage (2012-2019), and stable development stage (2020-2023).

[0054] Optionally, in step S204, the preprocessing of the patent data in each time period comprises:

[0055] The summary text of the patent data is processed by using the nltk word segmentation tool;

[0056] Filter out stop words, punctuation, conjunctions, prepositions, personal pronouns and exclamatory words;

[0057] Among them, the filtering is performed in a combination of manual filtering and automatic filtering.

[0058] In the embodiment of the present application, the patent data is preprocessed, also known as data cleaning. One of the most important links of emerging technology topic identification and evolution analysis is data cleaning, which is closely related to the accuracy and efficiency of subsequent topic clustering results. Therefore, data preprocessing such as information extraction, word segmentation, stop word removal, and word standardization of patent data is essential. The present application uses the nltk word segmentation tool to process the summary text of the patent, and filters out stop words, punctuation, and other words without actual meaning such as conjunctions, prepositions, personal pronouns, and exclamatory words. In the process of filtering words, both automatic and manual word filtering methods can be used to achieve better word filtering effect.

[0059] Optionally, in step S204, converting the abstract text of each patent into a first text vector of a first preset dimension comprises:

[0060] The Distilbert-base-nli-mean-tokens model is fine-tuned using the patent data;

[0061] transform the summary text into a vectorized representation using the fine-tuned Distilbert-base-nli-mean-tokens model.

[0062] The Distilbert-base-nli-mean-tokens model is a SentenceBERT model that is fine-tuned on an NLI dataset using a pre-trained Distilbert-base model.

[0063] The first preset dimension is 768.

[0064] For example, for the PBAT technology, the Distilbert-base-nli-mean-tokens model is used and fine-tuned for PBAT domain data. The Distilbert-base-nli-mean-tokens model uses a mean pooling strategy to calculate sentence representation. As a pre-trained model, the quality of the model in embedding sentences and paragraphs and embedding search queries has been widely evaluated.

[0065] SentenceBERT (SBERT for short) as a sentence vector calculation model, by generating the embedding vector of the sentence, realizes the similarity calculation of the patent literature, effectively solves the problem of sparse semantic features of patent abstract literature. SentenceBERT is an improvement of the BERT language model, mainly to solve the problem of huge time overhead of BERT model for text semantic similarity calculation. Sentence-BERT extends the pre-trained BERT model, and the model uses Sentence Transformer, which can create document embeddings from a set of documents by loading a pre-trained model.

[0066] Optionally, fine-tuning the Distilbert-base-nli-mean-tokens model using the patent data includes:

[0067] Randomly select two pieces of data from the ownership claims, titles and abstract text content of the patent data as positive and negative samples respectively, and then input the Distilbert-base-nli-mean-tokens model to obtain the corresponding vector representation;

[0068] Calculate the cosine similarity between the vectors of the positive sample and the negative sample, and set the label according to the cosine similarity;

[0069] Input the patent data with the set label as training data into the Distilbert-base-nli-mean-tokens model.

[0070] As an optional example, the Distilbert-base-nli-mean-tokens model in the embodiment of the application is fine-tuned as a pre-training model, two data are randomly extracted from all the claims and title and abstract content of the original data as a positive sample and a negative sample respectively, and then input into the model to obtain the corresponding vector representation. Next, the cosine similarity of the positive sample and the negative sample vector is calculated, and the label is set according to the cosine similarity. Finally, the patent data set with the label is input into the Distilbert-base-nli-mean-tokens model for iteration. For example, set the AdamW optimizer, and the BCEWithLogitsLoss as the loss function, and the iteration number is 30 times, to complete the fine-tuning of the pre-training model.

[0071] Wherein, BCEWithLogitsLoss is a loss function for binary classification problems, which combines the Sigmoid activation function and the binary cross entropy loss (Binary Cross Entropy Loss) together, providing a more efficient and numerically stable way to calculate the loss.

[0072] Assuming that the input is the output Logits x of the model and the true label y, where y∈{0,1}, the calculation formula of BCEWithLogitsLoss is:

[0073]

[0074] Wherein, is the Sigmoid function. By combining the Sigmoid function and the loss calculation, the numerical stability is improved.

[0075] Wherein, Logits is the output of the last layer in the neural network, which is a numerical vector. P is the number of data, i is the data number, and the value is 1 to P. i The i-th output, y i is the i-th true label, x is the set of x i , and y is the set of y i .

[0076] For example, when the target technology is PBAT technology, the result of converting the abstract text of each patent into a first text vector of a first preset dimension is as shown in Table 1:

[0077] Table 1: PBAT technology abstract text vectorization representation example

[0078]

[0079] Optionally, in step S206, the dimension reduction processing on the first text vector comprises:

[0080] The UMAP method is adopted to perform dimension reduction operation on the first text vector, and a second text vector of a second preset dimension is obtained after the dimension reduction operation.

[0081] The second preset dimension can be 3.

[0082] Since the dimension of the first text vector is too high, the UMAP algorithm is adopted to perform data dimension reduction on the first text vectorization matrix. The UMAP algorithm is a very effective scalable dimension reduction algorithm. The algorithm maps high-dimensional probability distribution to low-dimensional space while preserving more global structure information of the summary text, so as to extract keywords and perform topic clustering.

[0083] The UMAP algorithm mainly uses local manifold approximation and local fuzzy simplex set representation to construct a topological representation of high-dimensional data, which can reduce the computational complexity and memory usage, and also can preserve the characteristics of the original data to the greatest extent. The scatter plot can well represent the relationship and structure between high-dimensional embedding vectors, as shown in Figure 3 Each point in the figure represents an embedding vector (i.e. a summary text), and the distance between the points reflects their similarity in the original high-dimensional space. For example, if two points are very close in three-dimensional space, their embedding vector representations in high-dimensional space can be very similar. As an optional example, the UMAP algorithm is adopted to reduce the 768-dimensional vector processed by SentenceBERT to a three-dimensional vector, i.e. the second text vector is a three-dimensional vector.

[0084] Optionally, in step S206, clustering the second text vector to obtain M topics in each time period comprises:

[0085] The Kmeans algorithm is adopted to cluster the second text vector to obtain M topics in each time period;

[0086] The Kmeans algorithm for clustering the second text vector comprises:

[0087] K data points are randomly selected from the second text vector as K cluster centers;

[0088] The following three steps are iterated until the target function converges to a square error value less than a preset error value:

[0089] The distance between the remaining data points and the K cluster centers is calculated, and the remaining data points are divided into the cluster closest to the K cluster centers.

[0090] Calculate the mean of all data points in the K clusters and take it as the new cluster center;

[0091] Calculate the convergence square error value of the objective function;

[0092] Wherein, the objective function is defined as:

[0093]

[0094] Where i is the number of clusters, J is the objective function, K is the number of clusters, S i is the set of data points in the ith cluster, x is the data point, μ i is the ith cluster center, and μ i satisfies:

[0095]

[0096] Where K is the number of clusters.

[0097] That is, in the above steps, the second text vector is the dimensionality-reduced representation of the patent data, which is used as input for the K-means algorithm. The detailed steps of the clustering process are as follows:

[0098] 1) Initialize cluster centers: randomly select K data points from the second text vector as the initial cluster centers; here K corresponds to the number of topics M desired for each time period;

[0099] 2) The algorithm enters the iteration process until the convergence condition is met (i.e., the objective function convergence square error value is less than the preset error value):

[0100] Calculate the distance and divide the cluster: for each second text vector (i.e., each data point), calculate its distance from the K cluster centers. Then, divide the data point into the cluster corresponding to the nearest cluster center;

[0101] Update the cluster center: for each cluster, calculate the mean of all data points in the cluster and take this mean as the new cluster center. This step ensures that the cluster center can represent the central position of the data points in the cluster;

[0102] Calculate the objective function: the objective function is usually defined as the sum of the squares of the distances from all data points to their corresponding cluster centers. After each iteration, the objective function value is recalculated to evaluate the quality of the current cluster division;

[0103] 3) Convergence condition: the iteration process continues until the change in the objective function value is less than the preset error value, indicating that the cluster division has been relatively stable and the algorithm converges;

[0104] 4) Obtain M topics for each time period: when the Kmeans algorithm converges, the data points of each time period are divided into K (i.e. M) clusters, each cluster represents a topic.

[0105] Thus, by clustering the second text vector using the Kmeans algorithm, M topics for each time period can be obtained, which reflect the technical content and trends of patent data in different time periods.

[0106] wherein the objective function is defined as:

[0107]

[0108] where i is the number of clusters, J is the objective function, K is the number of clusters, S i is the set of data points in the i-th cluster, x is the data point, u i is the i-th cluster center, and μ i satisfies:

[0109]

[0110] wherein K is the number of clusters.

[0111] The Kmeans algorithm requires the number of clusters to be generated to be determined before execution. For example, according to the actual situation of the PBAT field, each stage is clustered into 6 clusters. The results of part of the data after Kmeans clustering are shown in Figure 4 such as patents EP2631060A1, EP2712889A1, EP2803753A1, EP2804908A1, etc. in the X-shaped cluster on the left side.

[0112] Optionally, in step S208, for each time period, the abstract text of all patent data in each topic is spliced to obtain the target abstract text of each topic, including:

[0113] Splice the abstract text of all patents in each topic to obtain the target abstract text of each topic.

[0114] Optionally, in step S210, each word in the target abstract text of each topic of each time period is converted into a word vector of the first preset dimension, and the entire document of each patent of each topic of each time period is converted into a document vector of the first preset dimension, including:

[0115] The first step is data preprocessing, which involves removing stop words and synonyms using predicate logic from the input target summary text. The second step is vectorization, where a finely tuned BERT model is used to convert each word in the target summary text into a word vector of the first preset dimension. A pre-trained Distilbert-base-nli-mean-tokens model is then used to convert the entire document into a document vector of the first preset dimension, which is 768 dimensions.

[0116] Optionally, in step S212, calculating the cosine similarity between the word vector and the document vector includes:

[0117] The cosine similarity is calculated using the following formula:

[0118]

[0119] Where x 1k Let x represent the k-th dimension of the first computational quantity. 2k Let k represent the k-th dimension of the second computational quantity, where k ranges from [1, 768], n is the number of dimensions and n = 768, and cos(θ) ranges from [-1, 1].

[0120] cos(θ) represents the cosine similarity between the first and second computational quantities.

[0121] For example, x 1k Represents word vectors, x 2k Let represent the document vector. The cosine similarity between the word vector and the document vector can be calculated using the formula above.

[0122] In this embodiment, combined with Figure 5 As shown, the BERT model is fine-tuned and trained in the following way:

[0123] (1) Obtain the historical dataset, in which each historical data item includes all topics of an emerging technology in various time periods, the topic name of each topic, the patent documents of all patents under each topic, and the topic summary text corresponding to each topic;

[0124] (2) Perform word segmentation on the topic summary text corresponding to each topic, extract all candidate words A1~Am (such as noun phrases and verb phrases), and record the position of each candidate word in the sentence (i.e., the start and end token indexes);

[0125] (3) For each candidate word, the sentence in which it is located is intercepted as a context sentence, and an importance score (such as a 0-1 value) is assigned to each candidate word according to artificial annotation or pseudo-label; wherein the importance score is used to indicate the contribution value of the candidate word to the topic name;

[0126] (4) The topic summary text data and the patent documents under the topic are taken as a set of training samples. The topic summary text data is the feature information of all candidate words, which includes the context sentence, the candidate word position, and the corresponding label (i.e., the importance score);

[0127] (5) The context sentence and the candidate word position in the topic summary text data are input into the BERT model to be trained, and the patent documents under the topic are input into the Distilbert-base-nli-mean-tokens model that has been trained. The BERT model outputs the word vectors X1~Xm (768 dimensions) of all candidate words, and the Distilbert-base-nli-mean-tokens model outputs the document vector (768 dimensions) of the patent documents;

[0128] (6) The cosine similarity between the word vector and the document vector is calculated by a cosine similarity calculation module;

[0129] (7) The parameters of the Distilbert-base-nli-mean-tokens model are frozen, and the model parameters of the BERT model are back-propagated according to the following loss function:

[0130]

[0131] wherein m is the number of candidate words, Si is the cosine similarity between the i-th candidate word and the document vector, is the importance score of the i-th candidate word.

[0132] (8) The above steps are iteratively trained until the loss function converges or the preset number of iterations is reached.

[0133] In this way, the word vector output by the BERT model obtained through fine-tuning can reflect the importance of the subsequent topic name determination, so as to ensure the accuracy of the keywords selected from the target summary text.

[0134] Optionally, in step S212, N keywords are extracted from the target summary text according to the N results with the highest cosine similarity, to obtain N keywords of each topic in each time period, wherein N is an integer greater than or equal to 2, and optionally N is equal to 15, that is, the 15 words with the highest cosine similarity between the word vector and the document vector are taken as the keywords.

[0135] For example, taking PBAT technology as an example, the number of keywords is 10 keywords per topic, the multiple stages are respectively the budding stage (1900-2011), the high-speed development stage (2012-2019), and the stable development stage (2020-2023), and each topic is labeled, and each stage contains 6 topics. The keywords of each stage are shown in Table 2. Among them, the keywords of each topic are shown in the Keywords column, and the Weight column is the weight value of the keyword. The higher the weight value, the more important the keyword in the abstract.

[0136] Table 2: Example of keywords of each topic in each stage of PBAT technology field

[0137]

[0138]

[0139] Optionally, in step S214, determining the topic name of each topic in each time period according to the N keywords comprises:

[0140] According to the N keywords in each topic, input the N keywords and the related weight into a large language model (LLM) to obtain the topic name of each topic, and the final topic name is obtained through expert review according to the keywords.

[0141] For example, taking PBAT technology as an example, the determined topic name is shown in Table 3.

[0142] Table 3: Topic name of each topic in each stage of PBAT technology field

[0143]

[0144]

[0145] Optionally, in step S214, according to the topic name, calculating the cosine similarity between all topics in the two adjacent time periods, and determining the association relationship between the topics according to the cosine similarity between the topics can comprise:

[0146] Starting from a time period, the cosine similarity between all topics in adjacent two time periods is calculated in turn. Then check all the cosine similarities, if the cosine similarity is greater than a preset similarity threshold, it is determined that there is an association relationship between the two topics.

[0147] As a preferred example, the similarity threshold is 0.4.

[0148] For example, if a total of 4 time periods are divided, respectively, as the budding stage S0, the high-speed development stage S1, the stable stage S2 and the decline stage S3. The budding stage S0 has N0 topics, the high-speed development stage S1 has N1 topics, the stable stage S2 has N2 topics, and the decline stage S3 has N3 topics. Then the following calculations are made in turn:

[0149] 1. Calculate the cosine similarity between each of the N0 topics and each of the N1 topics, a total of N0*N1 cosine similarity values, if a certain cosine similarity is greater than the preset similarity threshold, it is determined that there is an association relationship between the two topics calculated by the cosine similarity.

[0150] 2. Calculate the cosine similarity between each of the N1 topics and each of the N2 topics, a total of N1*N2 cosine similarity values, if a certain cosine similarity is greater than the preset similarity threshold, it is determined that there is an association relationship between the two topics calculated by the cosine similarity.

[0151] 3. Calculate the cosine similarity between each of the N2 topics and each of the N3 topics, a total of N2*N3 cosine similarity values, if a certain cosine similarity is greater than the preset similarity threshold, it is determined that there is an association relationship between the two topics calculated by the cosine similarity.

[0152] After the above steps, the association relationship between the adjacent two time periods and the topics can be obtained, and the evolution relationship of the emerging technology can be visualized and displayed in a graphical way. The existence of an association relationship between two topics means that there is an evolution relationship between the two topics.

[0153] Taking PBAT technology as an example, the evolution trend of PBAT technology is drawn using a Sankey diagram as shown in Figure 6 . (Only technology topics with evolution relationship are shown.) Among them, the first stage (1900-2011) is mainly the manufacture of polymer technology and material production, the main content of the second stage (2012-2019) is the improvement and innovation of polymers, which has developed significantly compared with the first stage, in order to obtain high-performance polymer products, and some are applied to biodegradable film products. The third stage (2020-2023) is the further development of polymer materials, mainly the improvement and application of polymers. Compared with the changes in the first and second stages, the changes between the second and third stages are relatively small, and the development speed is slow, but PBAT technology is still a popular direction.

[0154] From Figure 6 The following information can also be interpreted:

[0155] Most of the technical subjects with evolutionary relationship are related to biodegradable and polymer, which belong to the field of polymer and its manufacturing process, including 14 subjects such as S0, M5, L2, etc. 5 subjects such as S3, M1, L0 are mostly related to the field of film and product. It can be explained that the two fields are the most important two technical fields in PBAT, and the polymer and its manufacturing process have been the focus of PBAT since 2000, and until recently in 2023, it is still a popular research direction, and the film and its products have appeared in the three stages, which is a perennial research direction in this field.

[0156] The evolution direction between these subjects located in adjacent time slices is mainly inheritance and fusion, and the evolution intensity (i.e. the similarity between technical subjects) is also relatively high, for example: S1→M0, M2→L2. It can be known that the biodegradable polymer related technology PBAT field continues to receive widespread attention from the industry, has strong vitality and high scientific research value and application value. Among them, M2 is Biodegradable Polymer-Based Polyester Composition Invention, because PBAT is a kind of Polyester, so this subject has an impact on many subsequent subjects (L0, L2, L3, L4, L5).

[0157] In addition, according to a second aspect of the embodiment, a storage medium is provided. The storage medium includes a stored program, wherein the program is executed by a processor when the program is running.

[0158] According to the method of the embodiment, first, the patent data of the target technology is obtained, and is divided according to time periods, so as to observe the development state of the technology at different time nodes. Then, the patent data is preprocessed and text vector converted, the main technical subjects in each time period are extracted through dimension reduction processing and clustering analysis. Further, the abstract texts of each subject are spliced to obtain target abstract texts, and the cosine similarity between the word vectors and the document vectors is calculated to extract keywords from the target abstract texts, so as to determine the subject name of each subject. Finally, the association relationship between the subjects is determined according to the cosine similarity between the subjects, and the evolution path of the target technology is determined. The application can efficiently and accurately capture the evolution path of the target technology, thereby improving the accuracy of the technical evolution analysis and solving the technical problem of low accuracy of the technical evolution analysis in the prior art.

[0159] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited to the action sequence described, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily necessary for the present application.

[0160] Through the description of the above embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be realized by means of software and the necessary general hardware platform, of course, it can also be realized by hardware, but in many cases the former is a better embodiment. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes a plurality of instructions for making a terminal device (which can be a mobile phone, computer, server, or network device, etc.) execute the method described in each embodiment of the present application.

[0161] Embodiment 2

[0162] Figure 7 A new technology evolution analysis device 700 according to the present embodiment is shown, which corresponds to the method according to the first aspect of embodiment 1. As shown in the figure, the device 700 includes: Figure 7

[0163] A patent data acquisition module 701 configured to acquire patent data of a target technology for evolution analysis, and divide the patent data into S time periods according to time periods;

[0164] An abstract vectorization module 702 configured to preprocess the patent data of each time period, and convert the abstract text of each patent into a first text vector of a first preset dimension;

[0165] A dimension reduction module 703 configured to perform dimension reduction processing on the first text vector to obtain a second text vector of a second preset dimension, and perform clustering on the second text vector to obtain M topics of each time period;

[0166] A splicing module 704 configured to splice the abstract text of all patent data in each topic to obtain a target abstract text of each topic for each time period;

[0167] ​The document vectorization module 705 is configured to convert each word in the target abstract text of each topic of each time period into a word vector of the first preset dimension, and convert the entire document of each patent of each topic of each time period into a document vector of the first preset dimension;

[0168] The keyword extraction module 706 is configured to calculate the cosine similarity between the word vector and the document vector, and extract N keywords from the target abstract text as keywords according to the N results with the highest cosine similarity, to obtain N keywords of each topic of each time period;

[0169] The association determination module 707 is configured to determine the topic name of each topic of each time period according to the N keywords, calculate the cosine similarity between all topics of the previous and subsequent time periods according to the topic name, and determine the association relationship between the topics according to the cosine similarity between the topics.

[0170] The evolution module 708 is configured to determine the evolution path of the target technology according to the association relationship.

[0171] Wherein, the S is an integer greater than or equal to 3, the M is an integer greater than or equal to 2, and the N is an integer greater than or equal to 2.

[0172] The first preset dimension and the second preset dimension are preset, and the first preset dimension is greater than the second preset dimension.

[0173] Optionally, the abstract vectorization module 702 is further configured to:

[0174] Perform word segmentation processing on the abstract text of the patent data using an nltk word segmentation tool;

[0175] Filter out stop words, punctuation, conjunctions, prepositions, personal pronouns, and exclamatory words;

[0176] Wherein, the filtering is performed in a combination of manual filtering and automatic filtering.

[0177] Optionally, the abstract vectorization module 702 is further configured to:

[0178] Fine-tune the Distilbert-base-nli-mean-tokens model using the patent data;

[0179] Convert the abstract text into a vectorized representation using the fine-tuned Distilbert-base-nli-mean-tokens model;

[0180] The Distilbert-base-nli-mean-tokens model is a SentenceBERT model in which a pre-trained Distilbert-base model is fine-tuned on an NLI data set.

[0181] Optionally, the fine-tuning of the Distilbert-base-nli-mean-tokens model using the patent data comprises:

[0182] In the ownership claims, titles, and abstract text content of the patent data, two pieces of data are randomly extracted as positive and negative samples, respectively, and then input into the Distilbert-base-nli-mean-tokens model to obtain corresponding vector representations.

[0183] The cosine similarity between the vectors of the positive sample and the negative sample is calculated, and a label is set according to the cosine similarity.

[0184] The patent data with the set label is input into the Distilbert-base-nli-mean-tokens model as training data.

[0185] Optionally, the dimension reduction module 703 is further configured to:

[0186] The UMAP method is used to perform dimension reduction operation on the first text vector.

[0187] Optionally, the dimension reduction module 703 is further configured to:

[0188] The Kmeans algorithm is used to cluster the second text vector, and M topics in each time period are clustered.

[0189] The Kmeans algorithm is used to cluster the second text vector, and M topics in each time period are clustered.

[0190] K data points are randomly selected from the second text vector as K cluster centers.

[0191] The following three steps are iteratively repeated until the target function converges with a square error value less than a preset error value:

[0192] The distances between the remaining data points and the K cluster centers are calculated, and the remaining data points are divided into the cluster closest to the data point and the K cluster centers.

[0193] The mean values of all data points in the K clusters are calculated and taken as new cluster centers.

[0194] The square error value of the target function convergence is calculated.

[0195] The target function is defined as:

[0196]

[0197] Where i is the number of clusters, J is the target function, K is the number of clusters, S i is the set of data points in the ith cluster, x is a data point, μ i is the ith cluster center, and μ i satisfies:

[0198]

[0199] Where K is the number of clusters.

[0200] Optionally, the keyword extraction module 706 is further configured to calculate the cosine similarity according to the following formula:

[0201]

[0202] Where x 1k represents the kth dimension of the first calculation amount, x 2k represents the kth dimension of the second calculation amount, k ranges from 1 to 768, n is the number of dimensions and n = 768, and the value of cos(θ) ranges from -1 to 1.

[0203] cos(θ) represents the cosine similarity between the first calculation amount and the second calculation amount.

[0204] Optionally, the association determination module 707 is further configured to:

[0205] If the cosine similarity of two topics is greater than a preset similarity threshold, it is determined that there is an association relationship between the two topics.

[0206] It should be noted that the emerging technology evolution analysis device of embodiment 2 and the emerging technology evolution analysis method of embodiment 1 belong to the same inventive concept, solve the same technical problem, and obtain the same technical effect. The same parts will not be described again.

[0207] According to the device embodiment, firstly, patent data of a target technology is acquired and divided according to time periods, so as to observe the development state of the technology at different time nodes. Then, the patent data is preprocessed and converted into a text vector, and through dimension reduction processing and clustering analysis, the main technical topics in each time period are extracted. Further, the summary text of each topic is spliced to obtain a target summary text, and keywords are extracted from the target summary text by calculating the cosine similarity between the word vector and the document vector, so as to determine the topic name of each topic. Finally, the association relationship is determined according to the cosine similarity between the topics, and the evolution path of the target technology is determined. The application can efficiently and accurately capture the evolution path of the target technology, thereby improving the accuracy of the technical evolution analysis and solving the technical problem of low accuracy of the technical evolution analysis in the prior art.

[0208] Embodiment 3

[0209] Figure 8 An emerging technology evolution analysis device according to the embodiment is shown, which corresponds to the method according to the first aspect of embodiment 1. As shown in Figure 8 The device comprises:

[0210] a processor 810; and

[0211] a memory 820 connected with the processor 810, configured to provide the processor 810 with instructions for processing the following processing steps:

[0212] acquire patent data of a target technology for evolution analysis, and divide the patent data into S time periods according to time periods;

[0213] preprocess the patent data of each time period, and convert the summary text of each patent into a first text vector of a first preset dimension;

[0214] perform dimension reduction processing on the first text vector to obtain a second text vector of a second preset dimension, and perform clustering on the second text vector to obtain M topics in each time period;

[0215] for each time period, splice the summary texts of all patent data in each topic to obtain a target summary text of each topic;

[0216] convert each word in the target summary text of each topic in each time period into a word vector of the first preset dimension, and convert the entire document of each patent of each topic in each time period into a document vector of the first preset dimension;

[0217] calculate cosine similarity between the word vector and the document vector, and extract N keywords from the target abstract text as keywords according to the N highest cosine similarity results, to obtain N keywords of each topic in each time period;

[0218] According to the N keywords, determine the topic name of each topic in each time period, and calculate the cosine similarity between all topics in the two time periods before and after according to the topic name, and determine the association relationship between the topics according to the cosine similarity between the topics;

[0219] According to the association relationship, determine the evolution path of the target technology;

[0220] Wherein, the S is an integer greater than or equal to 3, the M is an integer greater than or equal to 2, and the N is an integer greater than or equal to 2;

[0221] The first preset dimension and the second preset dimension are preset, and the first preset dimension is greater than the second preset dimension.

[0222] Optionally, the preprocessing of the patent data in each time period comprises:

[0223] The abstract text of the patent data is processed by using the nltk word segmentation tool;

[0224] Filter out stop words, punctuation, conjunctions, prepositions, personal pronouns and exclamatory words;

[0225] Wherein, the filtering is performed by combining manual filtering and automatic filtering.

[0226] Optionally, the converting of the abstract text of each patent into a first text vector of a first preset dimension comprises:

[0227] The Distilbert-base-nli-mean-tokens model is fine-tuned using the patent data;

[0228] The fine-tuned Distilbert-base-nli-mean-tokens model is used to convert the abstract text into a vector representation;

[0229] Wherein, the Distilbert-base-nli-mean-tokens model is a SentenceBERT model that fine-tunes the pre-trained Distilbert-base model on the NLI dataset.

[0230] Optionally, the fine-tuning of the Distilbert-base-nli-mean-tokens model using the patent data comprises:

[0231] In the ownership claims, titles and abstract text content of the patent data, two data are randomly extracted as positive and negative samples respectively, and then input into the Distilbert-base-nli-mean-tokens model to obtain the corresponding vector representation;

[0232] The cosine similarity between the vectors of the positive sample and the negative sample is calculated, and the label is set according to the cosine similarity;

[0233] The patent data with the set label is input into the Distilbert-base-nli-mean-tokens model as training data.

[0234] Optionally, the dimensionality reduction processing of the first text vector comprises:

[0235] The UMAP method is used to perform dimensionality reduction operation on the first text vector.

[0236] Optionally, the clustering of the second text vector to obtain M topics in each time period comprises:

[0237] The Kmeans algorithm is used to cluster the second text vector to cluster M topics in each time period;

[0238] The Kmeans algorithm for clustering the second text vector comprises:

[0239] Randomly select K data points from the second text vector as K cluster centers;

[0240] Iterate the following three steps until the target function converges to a square error value less than a preset error value:

[0241] Calculate the distance between the remaining data points and the K cluster centers, and divide the remaining data points into the cluster closest to the K cluster centers;

[0242] Calculate the mean of all data points in the K clusters as the new cluster center;

[0243] Calculate the target function convergence square error value;

[0244] Wherein, the target function is defined as:

[0245]

[0246] Where i is the number of clusters, J is the target function, K is the number of clusters, S i is the data point set in the ith cluster, x is the data point, μ iis the i-th cluster center, and μ i satisfies:

[0247]

[0248] wherein K is the number of clusters.

[0249] Optionally, the cosine similarity is calculated according to the following formula:

[0250]

[0251] wherein x 1k represents the k-th dimension of the first calculation amount, x 2k represents the k-th dimension of the second calculation amount, k ranges from 1 to 768, n is the number of dimensions and n = 768, and cos(θ) ranges from -1 to 1;

[0252] cos(θ) represents the cosine similarity between the first calculation amount and the second calculation amount.

[0253] Optionally, the determining the association relationship between each theme according to the cosine similarity between the themes comprises:

[0254] If the cosine similarity between two themes is greater than a preset similarity threshold, it is determined that there is an association relationship between the two themes.

[0255] It should be noted that the emerging technology evolution analysis device of embodiment 3 and the emerging technology evolution analysis method of embodiment 1 belong to the same inventive concept, solve the same technical problem, and obtain the same technical effect. The same parts will not be described again.

[0256] According to the device embodiment, first, the patent data of the target technology is obtained, and is divided according to time periods, so as to observe the development state of the technology at different time nodes. Then, the patent data is preprocessed and text vector conversion is performed, and through dimension reduction processing and clustering analysis, the main technical themes in each time period are extracted. Further, the abstract texts of each theme are spliced to obtain target abstract texts, and the cosine similarity between the word vectors and the document vectors is calculated to extract keywords from the target abstract texts, so as to determine the theme name of each theme. Finally, the association relationship is determined according to the cosine similarity between the themes, and the evolution path of the target technology is determined. The application can efficiently and accurately capture the evolution path of the target technology, thereby improving the accuracy of the technical evolution analysis and solving the technical problem of low accuracy of the technical evolution analysis in the prior art.

[0257] The above-mentioned serial numbers of the embodiments of the application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0258] In the above-mentioned embodiments of the present application, the description of each embodiment is focused on, and the part not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0259] In several embodiments provided in the present application, it should be understood that the disclosed technical contents can be implemented by other ways. Among them, the above-mentioned device embodiments are only schematic, for example, the division of the units is only a logical function division, and actual implementation can have another division way, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units or modules shown or discussed can be indirect coupling or communication connection through some interfaces, units or modules, which can be electrical or other forms.

[0260] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place or distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0261] In addition, each functional unit in each embodiment of the present application can be integrated in a processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0262] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application essentially or the part of the prior art or the whole or part of the technical solutions can be embodied in the form of a software product, which is stored in a storage medium and includes a plurality of instructions for making a computer device (which can be a personal computer, a server or a network device, etc.) execute all or part of the steps of the method described in each embodiment of the present application. The foregoing storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and various program code storage media.

[0263] The above-mentioned is only the preferred embodiment of the present application, and it should be pointed out that for ordinary skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements should be regarded as the protection scope of the present application.

Claims

1. A method for analyzing the evolution of emerging technologies, characterized in that, include: Obtain patent data for the target technology to be used for evolutionary analysis, and divide the patent data into S time periods according to time periods; The patent data for each time period is preprocessed, and the abstract text of each patent is converted into a first text vector with a first preset dimension. The first text vector is reduced in dimensionality to obtain a second text vector with a second preset dimension, and the second text vector is clustered to obtain M topics for each time period; For each time period, the abstract texts of all patent data in each topic are concatenated to obtain the target abstract text for each topic; Each word in the target summary text of each topic in each time period is converted into a word vector of the first preset dimension, and the entire document of each patent of each topic in each time period is converted into a document vector of the first preset dimension. Calculate the cosine similarity between the word vector and the document vector, and extract N words as keywords from the target summary text based on the N results with the highest cosine similarity, to obtain N keywords for each topic in each time period; Based on the N keywords, determine the topic name of each topic in each time period. Based on the topic name, calculate the cosine similarity between all topics in two consecutive time periods. Based on the cosine similarity between the topics, determine the association between each topic. Based on the aforementioned correlation, the evolution path of the target technology is determined; Wherein, S is an integer greater than or equal to 3, M is an integer greater than or equal to 2, and N is an integer greater than or equal to 2; The first preset dimension and the second preset dimension are preset, and the first preset dimension is greater than the second preset dimension; The process of converting the abstract text of each patent into a first text vector of a first preset dimension includes: The Distilbert-base-nli-mean-tokens model was fine-tuned using the patented data. The summarized text is converted into a vectorized representation using a finely tuned Distilbert-base-nli-mean-tokens model; The Distilbert-base-nli-mean-tokens model is a SentenceBERT model that is a pre-trained Distilbert-base model fine-tuned on the NLI dataset. The fine-tuning of the Distilbert-base-nli-mean-tokens model using the patented data includes: Step 1: Obtain the historical dataset. Each piece of historical data in the historical dataset includes all topics of the emerging technology in various time periods, the topic name of each topic, the patent documents of all patents under each topic, and the topic summary text corresponding to each topic. Step 2: Perform word segmentation on the topic summary text corresponding to each topic, extract all candidate words, and record the position of each candidate word in the sentence; Step 3: For each candidate word, extract the sentence containing it as the context sentence, and assign an importance score to each candidate word based on manual annotation or pseudo-labels; wherein, the importance score is used to indicate the contribution value of the candidate word to the topic name; Step 4: Use the topic summary text data and the patent documents under the topic as a set of training samples, wherein the topic summary text data is the feature information of all candidate words, and the feature information includes the context sentence, the position of the candidate word and the corresponding label; Step 5: Input the context sentences and candidate word positions in the topic summary text data into the BERT model to be trained, and input the patent documents under the topic into the pre-trained Distilbert-base-nli-mean-tokens model. The BERT model outputs the word vectors of all candidate words, and the Distilbert-base-nli-mean-tokens model outputs the document vector of the patent document. Step 6: Calculate the cosine similarity between the word vectors and the document vectors using cosine similarity; Step 7: Freeze all parameters of the Distilbert-base-nli-mean-tokens model and backpropagate the BERT model parameters according to the following loss function: Where m is the number of candidate words, and Si is the cosine similarity between the i-th candidate word and the document vector. The importance score is given to the i-th candidate word, where i is the candidate word number; Step 8: Perform iterative training according to steps 1 to 7 above until the loss function converges or the preset number of iterations is reached.

2. The method according to claim 1, characterized in that, The preprocessing of patent data for each time period includes: The abstract text of the patent data was segmented using the nltk word segmentation tool; Filter out stop words, punctuation, conjunctions, prepositions, personal pronouns, and interjections; The filtration process employs a combination of manual and automatic filtration.

3. The method according to claim 1, characterized in that, The clustering of the second text vector to obtain M topics for each time period includes: The K-means algorithm is used to cluster the second text vector to identify M topics for each time period. The step of clustering the second text vector using the K-means algorithm includes: K data points are randomly selected from the second text vector as K cluster centers; Iterate through the following three steps until the squared error of the objective function converges is less than the preset error value: Calculate the distance between the remaining data points and the K cluster centers, and assign the remaining data points to the clusters that are closest to the K cluster centers; Calculate the mean of all data points in K clusters and use it as the new cluster center; Calculate the squared error of convergence of the objective function; The objective function is defined as follows: Where j is the cluster number, J is the objective function, K is the number of clusters, and S i Let x be the set of data points in the i-th cluster, and μ be the data point. i It is the i-th cluster center, and μ i satisfy: Where K is the number of clusters.

4. The method according to claim 1, characterized in that, The cosine similarity is calculated using the following formula: Where x 1k Let x represent the k-th dimension of the first computational quantity. 2k Let k represent the k-th dimension of the second computational quantity, where k ranges from [1, 768], n is the number of dimensions and n = 768, and cos(θ) ranges from [-1, 1]. cos(θ) represents the cosine similarity between the first and second computational quantities.

5. The method according to claim 1, characterized in that, The step of determining the association between topics based on the cosine similarity between the topics includes: If the cosine similarity between two topics is greater than a preset similarity threshold, then it is determined that there is a correlation between the two topics.

6. A storage medium, characterized in that, The storage medium includes a stored program, wherein, when the program is executed, the method described in any one of claims 1 to 5 is performed by a processor.

7. A device for analyzing the evolution of emerging technologies, characterized in that, include: The patent data acquisition module is configured to acquire patent data of the target technology for evolutionary analysis, and divide the patent data into S time periods according to time periods; The abstract vectorization module is configured to preprocess the patent data for each time period, converting the abstract text of each patent into a first text vector of a first preset dimension. The dimensionality reduction module is configured to perform dimensionality reduction processing on the first text vector to obtain a second text vector with a second preset dimension, and to cluster the second text vector to obtain M topics for each time period; The splicing module is configured to splice the summary text of all patent data in each topic for each time period to obtain the target summary text for each topic; The document vectorization module is configured to convert each word in the target summary text of each topic in each time period into a word vector of the first preset dimension, and to convert the entire document of each patent of each topic in each time period into a document vector of the first preset dimension. The keyword extraction module is configured to calculate the cosine similarity between the word vector and the document vector, and extract N words as keywords from the target summary text based on the N results with the highest cosine similarity, thereby obtaining N keywords for each topic in each time period; The association determination module is configured to determine the topic name of each topic in each time period based on the N keywords, calculate the cosine similarity between all topics in two consecutive time periods based on the topic name, and determine the association relationship between each topic based on the cosine similarity between the topics. An evolution module is configured to determine the evolution path of the target technology based on the association relationship; Wherein, S is an integer greater than or equal to 3, M is an integer greater than or equal to 2, and N is an integer greater than or equal to 2; The first preset dimension and the second preset dimension are preset, and the first preset dimension is greater than the second preset dimension; The process of converting the abstract text of each patent into a first text vector of a first preset dimension includes: The Distilbert-base-nli-mean-tokens model was fine-tuned using the patented data. The summarized text is converted into a vectorized representation using a finely tuned Distilbert-base-nli-mean-tokens model; The Distilbert-base-nli-mean-tokens model is a SentenceBERT model that is a pre-trained Distilbert-base model fine-tuned on the NLI dataset. The fine-tuning of the Distilbert-base-nli-mean-tokens model using the patented data includes: Step 1: Obtain the historical dataset. Each piece of historical data in the historical dataset includes all topics of the emerging technology in various time periods, the topic name of each topic, the patent documents of all patents under each topic, and the topic summary text corresponding to each topic. Step 2: Perform word segmentation on the topic summary text corresponding to each topic, extract all candidate words, and record the position of each candidate word in the sentence; Step 3: For each candidate word, extract the sentence containing it as the context sentence, and assign an importance score to each candidate word based on manual annotation or pseudo-labels; wherein, the importance score is used to indicate the contribution value of the candidate word to the topic name; Step 4: Use the topic summary text data and the patent documents under the topic as a set of training samples, wherein the topic summary text data is the feature information of all candidate words, and the feature information includes the context sentence, the position of the candidate word and the corresponding label; Step 5: Input the context sentences and candidate word positions in the topic summary text data into the BERT model to be trained, and input the patent documents under the topic into the pre-trained Distilbert-base-nli-mean-tokens model. The BERT model outputs the word vectors of all candidate words, and the Distilbert-base-nli-mean-tokens model outputs the document vector of the patent document. Step 6: Calculate the cosine similarity between the word vectors and the document vectors using cosine similarity; Step 7: Freeze all parameters of the Distilbert-base-nli-mean-tokens model and backpropagate the BERT model parameters according to the following loss function: Where m is the number of candidate words, and Si is the cosine similarity between the i-th candidate word and the document vector. The importance score is given to the i-th candidate word, where i is the candidate word number; Step 8: Perform iterative training according to steps 1 to 7 above until the loss function converges or the preset number of iterations is reached.

8. A device for analyzing the evolution of emerging technologies, characterized in that, include: processor; as well as A memory, connected to the processor, for providing the processor with instructions to perform the following processing steps: Obtain patent data for the target technology to be used for evolutionary analysis, and divide the patent data into S time periods according to time periods; The patent data for each time period is preprocessed, and the abstract text of each patent is converted into a first text vector with a first preset dimension. The first text vector is reduced in dimensionality to obtain a second text vector with a second preset dimension, and the second text vector is clustered to obtain M topics for each time period; For each time period, the abstract texts of all patent data in each topic are concatenated to obtain the target abstract text for each topic; Each word in the target summary text of each topic in each time period is converted into a word vector of the first preset dimension, and the entire document of each patent of each topic in each time period is converted into a document vector of the first preset dimension. Calculate the cosine similarity between the word vector and the document vector, and extract N words as keywords from the target summary text based on the N results with the highest cosine similarity, to obtain N keywords for each topic in each time period; Based on the N keywords, determine the topic name of each topic in each time period. Based on the topic name, calculate the cosine similarity between all topics in two consecutive time periods. Based on the cosine similarity between the topics, determine the association between each topic. Based on the aforementioned correlation, the evolution path of the target technology is determined; Wherein, S is an integer greater than or equal to 3, M is an integer greater than or equal to 2, and N is an integer greater than or equal to 2; The first preset dimension and the second preset dimension are preset, and the first preset dimension is greater than the second preset dimension; The process of converting the abstract text of each patent into a first text vector of a first preset dimension includes: The Distilbert-base-nli-mean-tokens model was fine-tuned using the patented data. The summarized text is converted into a vectorized representation using a finely tuned Distilbert-base-nli-mean-tokens model; The Distilbert-base-nli-mean-tokens model is a SentenceBERT model that is a pre-trained Distilbert-base model fine-tuned on the NLI dataset. The fine-tuning of the Distilbert-base-nli-mean-tokens model using the patented data includes: Step 1: Obtain the historical dataset. Each piece of historical data in the historical dataset includes all topics of the emerging technology in various time periods, the topic name of each topic, the patent documents of all patents under each topic, and the topic summary text corresponding to each topic. Step 2: Perform word segmentation on the topic summary text corresponding to each topic, extract all candidate words, and record the position of each candidate word in the sentence; Step 3: For each candidate word, extract the sentence containing it as the context sentence, and assign an importance score to each candidate word based on manual annotation or pseudo-labels; wherein, the importance score is used to indicate the contribution value of the candidate word to the topic name; Step 4: Use the topic summary text data and the patent documents under the topic as a set of training samples, wherein the topic summary text data is the feature information of all candidate words, and the feature information includes the context sentence, the position of the candidate word and the corresponding label; Step 5: Input the context sentences and candidate word positions in the topic summary text data into the BERT model to be trained, and input the patent documents under the topic into the pre-trained Distilbert-base-nli-mean-tokens model. The BERT model outputs the word vectors of all candidate words, and the Distilbert-base-nli-mean-tokens model outputs the document vector of the patent document. Step 6: Calculate the cosine similarity between the word vectors and the document vectors using cosine similarity; Step 7: Freeze all parameters of the Distilbert-base-nli-mean-tokens model and backpropagate the BERT model parameters according to the following loss function: Where m is the number of candidate words, and Si is the cosine similarity between the i-th candidate word and the document vector. The importance score is given to the i-th candidate word, where i is the candidate word number; Step 8: Perform iterative training according to steps 1 to 7 above until the loss function converges or the preset number of iterations is reached.

Images