Technology trend prediction method and device, computer system and medium

Abstract

The invention provides a technology trend prediction method and device, a computer system and a medium, and aims to excavate key technology evolution trends in various technical fields and identify future hot topics. The method comprises the following steps: performing document matching on generalized topics of various technical fields based on a highly cited document database, and extracting documents related to the generalized topics; performing text preprocessing on the literature under the generalized theme, wherein the text preprocessing comprises field proper noun standardization, word segmentation, stop word processing and word form merging and text vectorization; adopting an LDA topic clustering model to perform detailed topic clustering on the pre-processed literature, generating a keyword set and keyword contribution degrees, and condensing the keyword set and the keyword contribution degrees into detailed topics; and training and predicting the popularity data of the detailed theme over the years based on a Chronos time sequence model, and identifying the detailed theme with high popularity in the future.

Description

TECHNICAL FIELD

[0001] The present application relates to a technical trend prediction method, device, computer system and medium. BACKGROUND

[0002] Traditional technology prediction relies on expert experience (such as Delphi method), which has subjective bias, and the technology prediction result lacks real-time updating ability, and the coverage and granularity are limited by the expert field knowledge, which is difficult to adapt to the multi-dimensional technology evolution demand of power system. Although the existing text mining methods (such as LDA topic model and SAO structure analysis) can extract patent and paper keywords, the technical theme condensation still depends on artificial subjective judgment, which cannot realize dynamic hotspot identification, and lacks semantic correlation analysis ability of cross-source data (patent / paper). In the existing time series technology prediction, the regression model has strong interpretability but low prediction accuracy, while the machine learning method has improved accuracy but has "black box" characteristics, which is difficult to balance prediction accuracy and decision interpretability. In addition, the existing technology prediction mostly uses a single data source (only patent or paper), lacks a heterogeneous data fusion mechanism for cross-domain technical knowledge base, resulting in insufficient completeness and forward-looking of technology evolution path analysis. SUMMARY

[0003] The present application proposes a technical trend prediction method, device, computer system and medium, aiming to mine the key technology evolution trend of each technical field and identify future hot topics.

[0004] In the first aspect, a technical trend prediction method is proposed, comprising: based on a highly cited literature database, matching the generalized topics of each technical field with the literature, and extracting the literature related to the generalized topics; performing text preprocessing on the literature under the generalized topics, including field-specific name standardization, word segmentation and stop word processing, and word form merging and text vectorization; using an LDA topic clustering model to cluster the preprocessed literature into detailed topics, generating a keyword set and keyword contribution, and condensing into detailed topics; based on the Chronos time series model, training and predicting the annual heat data of the detailed topics, and identifying future high-heat detailed topics.

[0005] Secondly, a technology trend prediction device is proposed, including: a document matching module, configured to perform document matching on broad topics in various technical fields based on a highly cited document database, and extract documents related to the broad topics; a text preprocessing module, configured to perform text preprocessing on the documents under the broad topics, including standardization of domain-specific nouns, word segmentation and stop word processing, and word form grouping and text vectorization; an LDA topic clustering model, configured to perform detailed topic clustering on the preprocessed documents, generate keyword sets and keyword contributions, and condense them into detailed topics; a Chronos time series model, configured to train and predict the historical popularity data of the detailed topics, and identify detailed topics with high popularity in the future.

[0006] In some examples, the method for extracting literature related to the broad topics includes: obtaining manually determined broad topics in various technical fields; using the topic tree provided by Open Alex as a basis, identifying and extracting literature related to each broad topic in the Open Alex database based on the topic tree coding; combining the keyword space distance matching method to include literature from non-Open Alex database sources in the screening range, and screening out literature related to each broad topic; ensuring that all included documents meet the highly cited literature scoring index P both , which can be retrieved from the highly cited literature database.

[0007] In some examples, another Chronos time series model is used to train and predict the historical popularity data of each broad topic to screen out high-popularity broad topics, and the LDA topic clustering model performs detailed topic clustering on the documents corresponding to the high-popularity broad topics.

[0008] In a third aspect, a computer system is proposed, comprising a processor and a memory, wherein the memory stores one or more computer program modules, and the computer program modules are configured to be executed by the processor to implement the technology trend prediction method described in the item.

[0009] In a fourth aspect, a non-volatile computer-readable storage medium is proposed, which stores a computer program, and when the computer program is executed by a computer, it can implement the technology trend prediction method.

[0010] The present application broadens the data sources of related literatures in various technical fields (such as power electronic power systems), ensures sufficient data and fast and efficient retrieval process, and thus improves the accuracy of topic identification and the reliability of hot technology trend prediction. The present application uses a coarse and fine granularity topic clustering model to perform topic clustering of literatures from two dimensions of coarse and fine, to solve the problem of rough topic clustering results of traditional topic clustering methods, and improves the fineness of the clustering results, which is more helpful to grasp the future hot topics in various technical fields. The present application uses a topic future heat prediction model based on a time series algorithm, trains a time series model using a Chronos model, enhances the compatibility of the model, makes the model show good zero sample learning ability, and improves the prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0011] Figure 1 is a logic block diagram of the technical trend prediction method and device in an embodiment of the present application.

[0012] Figure 2 is a comparison diagram of part of the topic heat prediction results and actual heat in an embodiment of the present application.

[0013] Figure 3 is a part of the general topic heat prediction result diagram in an embodiment of the present application.

[0014] Figure 4 is a part of the fine topic heat prediction result diagram in an embodiment of the present application. DETAILED DESCRIPTION

[0015] As shown in Figure 1 , the principle of the present application is that the traditional topic future heat technology prediction result lacks real-time updating ability, and the coverage range and granularity are limited by expert field knowledge, which is difficult to adapt to the multi-dimensional technology evolution needs of various technical fields. The present application aims to provide a technology trend prediction method and system based on literature metrology analysis. First, the related literatures in various technical fields are obtained, and a highly cited literature database is constructed; then a coarse and fine granularity topic clustering model is proposed to obtain more accurate fine topics; secondly, the topics are matched with the corresponding paper information, the topic heat evaluation standard is constructed, and the annual topic heat of each fine topic is calculated; finally, a time series prediction model is trained to predict the future topic heat of the fine topic, so as to identify the hot topics. Through the present application, the objective hot trend of the technical field can be obtained, the prediction accuracy is improved, and reference and suggestions are provided for resource allocation.

[0016] Step 1, construct a highly cited literature database as the data basis for technology prediction.

[0017] Specifically, papers, patents and conference reports are obtained from open-source databases, authoritative patent databases and important academic conferences at home and abroad to ensure sufficient and efficient data while taking into account copyright compliance. On this basis, high-cited literature screening standards are developed, taking into account the average citation frequency in recent years (ACFY) and the average citation frequency growth rate in recent years (ACFGY).

[0018] The calculation formula of the screening standard is as follows:

[0019]

[0020] In the formula, is the average citation frequency in recent years, is the average citation frequency growth rate in recent years, C n10 is the cumulative citation frequency of the paper in the past 10 years, C all is the total citation frequency of the paper, G n10 is the annual citation frequency growth rate of the paper in the past 10 years, G all is the total growth rate of the citation frequency of the paper, , y n is the cumulative publication duration of the paper.

[0021] It should be noted that 10 years is only an example time range, and the time span can be adjusted according to the characteristics of the technical field, data availability or predicted needs in actual application (such as 5 years, 15 years, etc.).

[0022] Based on artificial intelligence and expert opinions, weights are assigned to ACFY and ACFGY k 1 and k 2, to finally form a high-cited literature score index:

[0023] In the formula, is the maximum average citation frequency in recent years, is the maximum average citation frequency growth rate in recent years.

[0024] The specific determination method of weights k 1 and k 2 belongs to the conventional technology in the art, and the present application will not be described in detail.

[0025] Identify high-cited papers using mean division method. According to the high-cited literature scoring index, use the mean division method to identify high-cited papers. Specifically, first calculate the mean of the number of citations of all papers, and divide the papers into "above mean group" and "below mean group". This division can help the project to preliminarily identify papers with strong academic influence. Then, in the "above mean group", use the same mean division method to divide again, and further divide these papers into high, medium and low cited paper groups. Through this iterative process, the high-cited paper subset can be accurately screened out, and the high-cited literature database is built through Clickhouse.

[0026] Step 2, screen the general topics of each technical field.

[0027] First, select the general topics of each technical field (manually screened, for example, determined by experts in each technical field), such as: energy technology (renewable energy, energy storage technology, smart grid), advanced manufacturing (additive manufacturing, industrial internet, flexible electronics), information and communication technology (5G / 6G, quantum computing, edge computing), biomedicine (gene editing, nanomedicine, AI-assisted diagnosis), environment and sustainable technology (carbon capture, green chemical industry, ecological restoration).

[0028] Then, based on the topic tree provided by Open Alex, identify the literature related to each general topic in the Open Alex database based on the topic tree coding. For example, by matching the child nodes of the "energy technology" topic (such as "renewable energy" and "energy storage technology"), the corresponding literature is extracted.

[0029] Combine the keyword spatial distance matching method to include non-Open Alex sources (such as Web of Science, IEEE Xplore, SpringerLink) in the screening range, and also screen out literature related to each general topic from them.

[0030] Ensure that all screened literature (whether from Open Alex or non-Open Alex sources) meets the high-cited literature scoring criteria in Step 1, that is, all screened literature can be searched in the high-cited literature database.

[0031] Finally, classify and store the literature closely related to each technical field according to the general topic.

[0032] Step 3, predict the future popularity of each general topic and screen out high-heat topics (for example, screen out high-heat general topics with an annual proportion of the top 5% in the next 3-5 years, the specific time and number can be dynamically adjusted according to the characteristics of the technical field or the prediction demand).

[0033] Specifically, we use a Chronos-based time series model (also known as a time series algorithm-based topic future popularity prediction model) to detect trends in hot topics in various technology fields. This mainly involves the following steps: (1) Constructing time series training data: First, we extract the general topic frequency in the literature to construct a technology heat index, and then construct time series training data based on the heat index.

[0034] (2) Data preprocessing: Before inputting time series data into the Chronos model, appropriate preprocessing is required. Through calibration and quantization, it is converted into discrete symbols so that it can be used as the input of the model.

[0035] (3) Model training and prediction: The model is trained by inputting the preprocessed time series data into the Chronos-based time series model. After the model training is completed, the model is applied to the future general topic popularity prediction. T gi to identify which broad topics will become future research hotspots.

[0036] (4) Analyze topic trends: Based on the prediction results of Chronos' time series model, calculate the changes in the proportion of each broad topic in various technical fields in the next few years, and identify the most popular broad topics year by year. These topics represent the possible main directions of future academic research.

[0037] Step 4: Further cluster the broad topics to obtain detailed topics.

[0038] First, we perform text preprocessing on the literature within these broad topics. This includes core steps such as defining and standardizing domain-specific terms, segmenting and stopping words, and text vectorization. First, we construct a domain-specific vocabulary or terminology library to extract and standardize key technical terms (such as "perovskite photovoltaics" and "flexible electronics"). This provides a precise semantic foundation for topic analysis and ensures the accuracy and domain-specific adaptability of the topic extraction results from the LDA model.

[0039] Secondly, continuous text is broken down into basic terms or phrases (e.g., "5G communication" and "nano-drug") to improve semantic granularity. Subsequently, common stop words (e.g., "the" and "and") and low-information symbols (e.g., punctuation and special characters) are removed. Standardized term forms are extracted through word form merging and stemming (e.g., "running" → "run," "research" → "research") to reduce term sparsity and enhance semantic consistency.

[0040] Finally, after completing the above preprocessing, the text is converted into a vector representation that can be processed by the machine learning model.

[0041] Specifically, the LDA topic clustering model (LDA model for short) is used for detailed topic clustering, that is, for the first i high heat general topic T gi , the title, abstract and other contents contained therein are converted into a vector representation of words after preprocessing. The basic idea of LDA is to assume that a document is randomly generated by several hidden topics, and these topics are composed of corresponding feature words. There are three layers of word, topic and document. Based on the Gibbs sampling method, the topic distribution probability of the document and the word distribution probability of the topic are iteratively adjusted.

[0042] Therefore, these word vectors are clustered into a certain number of topics by the LDA algorithm, and the form of the topic is a set of keywords and keyword contribution. The keywords and contribution can be condensed into the expression of the detailed topic by the text large model. At the same time, in the process of LDA topic clustering, the literature and the detailed topic can be effectively matched, thereby laying the foundation for the heat index calculation of the detailed topic.

[0043] Step 5, predict the heat change trend of each detailed topic.

[0044] According to the heat data of each detailed topic over the years, the Chronos-based time series model (also known as: topic future heat prediction model based on time series algorithm) is used to predict the future heat change trend of the detailed topic. Mainly includes the following steps: (1) Construct time series training data: first extract the detailed topic frequency in the literature, construct the technology heat index, and construct the time series training data based on the heat index.

[0045] (2) Data preprocessing: before inputting the time series data into the Chronos model, appropriate preprocessing is needed. Through calibration and quantization, it is converted into discrete symbols so as to be able to serve as the input of the model.

[0046] (3) Model training and prediction: by inputting the preprocessed time series data into the Chronos-based time series model, the model is trained. After the model training is completed, the model is applied to the future detailed topic heat prediction. Based on the heat change trend of each detailed topic T xi , identify which topics will become the research focus in the future.

[0047] (4) Analyze the theme trend: based on the prediction results of the Chronos time series model, the proportion change of each detailed theme in the next few years is counted, and the high heat detailed theme of each year is identified (for example, identify the top 10% heat detailed theme in the next 3-5 years, the specific time and quantity can be dynamically adjusted according to the characteristics of the technical field or the prediction demand), these detailed themes represent the possible main direction of future academic research.

[0048] The application further provides a technical trend prediction device. The device comprises a highly cited literature database, a literature matching module, a text preprocessing module, an LDA theme clustering model and a Chronos time series model.

[0049] The literature matching module is configured to match the literature related to the general theme of each technical field based on the highly cited literature database.

[0050] The text preprocessing module is configured to preprocess the literature under the general theme, including field-specific name standardization, word segmentation and stop word processing, and word shape merging and text vectorization.

[0051] The LDA theme clustering model is configured to cluster the preprocessed literature into detailed themes, generate a keyword set and keyword contribution, and condense into detailed themes.

[0052] The Chronos time series model is configured to train and predict the heat data of the detailed themes over the years, and identify the detailed themes with high heat in the future.

[0053] The technical trend prediction device is a program product coupled with the technical trend prediction method. The detailed algorithm implementation of each core function module in the device is referred to the above method embodiments.

[0054] The present invention also provides a computer system. The system includes a processor and a memory. The memory and the processor can be interconnected via a bus system and / or other forms of connection mechanisms. The processor can include any one or more of a central processing unit (CPU), a graphics processing unit (GPU), a microprocessor (MP), or a digital signal processor (DSP). The memory can include a volatile memory (volatile memory), such as a random access memory (RAM). The memory can also include a non-volatile memory (non-volatile memory), such as a read-only memory (ROM), a flash memory, a mechanical hard disk drive (HDD), or a solid state drive (SSD). The memory stores executable program code, and the processor executes the executable program code to implement the technology trend forecasting method. That is, the memory stores instructions for executing the technology trend forecasting method.

[0055] The present invention also provides a computer-readable storage medium. For example, the computer-readable storage medium is a non-transitory computer-readable storage medium, such as a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), a magnetic tape, a floppy disk, or an optical data storage device. The computer-readable storage medium is used to store non-transitory computer-readable instructions. When the non-transitory computer-readable instructions are executed by a computer, one or more steps of the above-mentioned technology trend forecasting method can be implemented.

[0056] Application examples: Taking the field of power electronic power system as an example, we first obtain the paper / patent information of various open source databases, and construct a broad topic set of the field of power electronic power system based on the OpenAlex topic number, as shown in Table 1, and screen out the topics in the field of power electronic power system. P both (Here k 1, k 2 were both taken as 0.5), highly cited literature was screened and a highly cited literature database was constructed.

[0057] Table 1 Example: Partial general topics and codes in the field of power electronic power systems

[0058] After rough classification of documents based on the general topic set, the technical heat score of each general topic in 1998-2023 is calculated. The data from 1998 to 2015 is used as the training set, and the data from 2016 to 2023 is used as the prediction set. The difference between the prediction results and the actual values is compared, as shown in Figure 2 , which proves the accuracy of the time series model trained based on the bibliometric method of the present application.

[0059] Then, using the data from 1998 to 2023 as the training set, the heat of each general topic from 2024 to 2033 is predicted. The model parameters are shown in Table 2, and some of the prediction results are shown in Figure 3 . According to the predicted values, the future high-heat general topic set is identified, as shown in Table 3.

[0060] Table 2 Chronos time series model parameters

[0061] Table 3 Example: High-heat general topics

[0062] Further preprocessing and LDA clustering of the text information contained in each high-heat general topic are performed. The keyword set of the detailed topic is obtained, and the corresponding topic name can be condensed through a large language model. Some detailed topic keywords and condensed topic names are shown in Table 4.

[0063] Table 4 Example: Detailed topic keywords and topic condensation results in the field of power electronic power systems

[0064] Finally, the heat index of the detailed topic in 1998-2023 is calculated, and the data from 1998 to 2023 is used as the training set to predict the heat of each detailed topic from 2024 to 2033. The model parameters are shown in Table 5, and some of the prediction results are shown in Figure 4 . According to the predicted values, the future high-heat detailed topic set is identified, which is the future hot topic of power electronic power systems predicted and identified by the system, as shown in Table 6.

[0065] Table 5 Chronos time series model parameters

[0066] Table 6 Example: Predicted future hot technologies in the field of power electronics

Claims

1. A technology trend prediction method, characterized in that: include: Based on the highly cited literature database, we conduct literature matching on broad topics in various technical fields and extract literature related to the broad topics. Perform text preprocessing on the literature under the broad topic, including domain-specific noun standardization, word segmentation and stop word processing, as well as word form grouping and text vectorization; The LDA topic clustering model is used to perform detailed topic clustering on the pre-processed documents, generate keyword sets and keyword contributions, and condense them into detailed topics; Based on the Chronos time series model, the popularity data of the detailed topics over the years are trained and predicted to identify detailed topics with high popularity in the future.

2. The method according to claim 1, characterized in that The screening criteria for highly cited documents in the highly cited document database include: calculating the average annual citation volume in recent years Compared with the average annual citation growth rate in recent years , and assign weights to them respectively k 1 and k 2. Form a scoring index for highly cited literature P both ,According to the highly cited literature scoring index, the mean partition method is used to identify highly cited literature, P both The calculation formula is: Where, This is the highest average annual citation volume in recent years. This is the highest average annual citation growth rate in recent years.

3. The method according to claim 1, characterized in that Methods for extracting literature related to the broad topic include: Obtaining manually identified broad themes across technical fields; Based on the topic tree provided by Open Alex, we identify and extract documents related to each broad topic in the Open Alex database based on the topic tree coding; Combined with the keyword space distance matching method, the literature from non-Open Alex database sources was included in the screening scope to select the literature related to each broad topic; Ensure that all the documents included in the screening meet the high-cited document scoring indicators P both , which can be retrieved from the highly cited literature database.

4. The method according to claim 1, wherein Another Chronos time series model is used to train and predict the historical popularity data of each broad topic, and high-popularity broad topics are screened out. The LDA topic clustering model performs detailed topic clustering on the documents corresponding to the high-popularity broad topics.

5. A technology trend prediction device, characterized in that: include: a document matching module configured to perform document matching on broad topics in various technical fields based on a highly cited document database and extract documents related to the broad topics; a text preprocessing module configured to perform text preprocessing on the documents under the broad topic, including standardization of domain proper nouns, word segmentation and stop word processing, and word form grouping and text vectorization; The LDA topic clustering model is configured to perform detailed topic clustering on the preprocessed documents, generate keyword sets and keyword contributions, and condense them into detailed topics; The Chronos time series model is configured to train and predict the historical popularity data of the detailed topics and identify detailed topics with high popularity in the future.

6. The device according to claim 5, characterized in that The screening criteria for highly cited documents in the highly cited document database include: calculating the average annual citation volume in recent years Compared with the average annual citation growth rate in recent years , and assign weights to them respectively k 1 and k 2. Form a scoring index for highly cited literature P both ,According to the highly cited literature scoring index, the mean partition method is used to identify highly cited literature, P both The calculation formula is: Where, This is the highest average annual citation volume in recent years. This is the highest average annual citation growth rate in recent years.

7. The device according to claim 5, characterized in that Methods for extracting literature related to the broad topic include: Obtaining manually identified broad themes across technical fields; Based on the topic tree provided by Open Alex, we identify and extract documents related to each broad topic in the Open Alex database based on the topic tree coding; Combined with the keyword space distance matching method, the literature from non-Open Alex database sources was included in the screening scope to select the literature related to each broad topic; Ensure that all the documents included in the screening meet the high-cited document scoring indicators P both , which can be retrieved from the highly cited literature database.

8. The device according to claim 5, characterized in that Another Chronos time series model is used to train and predict the historical popularity data of each broad topic, and high-popularity broad topics are screened out. The LDA topic clustering model performs detailed topic clustering on the documents corresponding to the high-popularity broad topics.

9. A computer system, characterized in that: The method comprises a processor and a memory, wherein the memory stores one or more computer program modules, and the computer program modules are configured to be executed by the processor to implement the technology trend prediction method according to any one of claims 1 to 4.

10. A non-volatile computer-readable storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a computer, the technology trend prediction method according to any one of claims 1 to 4 can be implemented.

Images