Technical opportunity identification method of double-layer GTM map based on papers and patents

Abstract

The invention relates to the field of technical opportunity recognition, in particular to a technical opportunity recognition method of a double-layer GTM map based on papers and patents. According to the recognition method, a BERTopic topic model, a double-layer generation type topological mapping map and link prediction are mainly combined to carry out recognition on technical opportunities, and calculation and analysis are carried out based on the BERTopic topic model to obtain the trend of a technical topic; processing and calculating the intersection of the important keywords of the thesis and the patent based on GTM to obtain a thesis map and a patent map, and analyzing vacuum points and non-vacuum points in the two maps to identify technical vacuum and technical opportunities; and a prediction similarity index is constructed by combining cosine similarity calculation of keyword vectors and a link prediction index, and selection and interpretation of a target technical opportunity are realized. Compared with an existing identification method, the identification method has the advantages that intervention of field experts is removed, information interpretation of the patent vacuum point is more objective and reasonable, and meanwhile a more comprehensive result can be obtained.

Description

Technical Field

[0001] The present invention relates to the field of technology opportunity identification, and in particular to a technology opportunity identification method based on a double-layer GTM map of papers and patents. Background Art

[0002] At present, the methods of technology opportunity identification mainly include morphological analysis, gap analysis based on science and technology, anomaly detection, link prediction and patent map. Morphological analysis is a typical qualitative technology prediction analysis method based on parameter decomposition and problem construction. It is a non-quantitative modeling method that guides constructive invention. However, morphological analysis requires the participation of experts in related fields to establish morphological matrices and modeling suggestions. The gap method based on science and technology analyzes and compares scientific papers and patents to identify gaps and determine technology opportunities. Researchers combine expert opinions and text mining to compare papers and patents from a thematic perspective to predict the future development direction of technology in the short term. Since the topic is coarse-grained, most of the literature focuses on predicting technology trends rather than providing specific information for technology opportunity identification. Anomaly detection aims to identify individuals that are significantly different from the overall population, also known as outliers. In technology opportunity identification, common unsupervised anomaly detection methods include distance-based k-means, clustering-based DBSCAN and density-based local outlier factor (LOF). Qualitative evaluation of abnormal patents usually relies on expert judgment. Link prediction predicts whether there is a link between two potentially related nodes based on known network structure and other related node information. The advantage of link prediction is that it can reveal links that do not yet exist in the network, thereby facilitating the tracking of dynamic changes in technology and the incorporation of new knowledge. However, link prediction can only predict potential links between existing nodes, and cannot predict the emergence of new nodes or links between old and new nodes. Patent maps are generally defined as a form of visualizing patent relationships, such as charts, tables, or graphs. The main methods for identifying blank technologies by constructing patent maps include principal component analysis (PCA), self-organizing maps (SOM), and generative topology maps (GTM). Patent maps based on PCA and SOM require researchers to subjectively interpret patent gaps to find technological opportunities, while patent maps based on GTM can automatically identify blank grids on the map and have the ability to reversely interpret the map. GTM has gradually become popular in identifying technological opportunities, so we also propose a method based on GTM.

[0003] Most of the existing GTM-based studies have combined semantic relations or clustering methods for technology opportunity identification (TOI). Yoon and Magee in TechnologicalForecasting and Social Change 132:105–17 proposed to visualize patent information through GTM and compared the prediction performance of support vector machine (SVM) and distance-based methods. Technology opportunity identification should consider completely new technology fields, but link prediction tools that use existing nodes in a single network are not appropriate. Liu et al. in TechnologicalForecasting and Social Change 192:122565 proposed a clear path to obtain technology opportunities through cross-domain patent analysis and link prediction. However, since the two technologies belong to different fields, problems of technology mismatch or patent infringement may arise during technology transfer. Recently, researchers have tried to combine topic models and GTM to identify technology opportunities by using patent and paper data in the same technology field. For example, Feng et al. used Latent Dirichlet Allocation (LDA) in PLOS ONE 18(10):e0293309 to analyze the topic trends of patents and papers and mapped them onto a single GTM graph, and then regarded the vacuum points surrounded by patents or papers as technology opportunities. However, the results of the identification rely heavily on the opinions of domain experts. In addition, the LDA model requires auxiliary indicators to determine the number of topics. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art that it is difficult to accurately predict technology trends and requires expert intervention, and to provide a technology opportunity identification method based on a double-layer GTM map of papers and patents.

[0005] In the first aspect, the present invention provides a technical identification method combining a BERTopic topic model, a two-layer generative topology map and link prediction, and constructs a more suitable similarity index by combining link prediction and cosine similarity. Compared with the existing identification method, the method removes the intervention of domain experts and can make a more reasonable and objective interpretation of patent vacuum point information, while also obtaining more comprehensive results. The identification method specifically includes the following steps:

[0006] S1. Collect and preprocess data;

[0007] S2, extract topics from papers and patents and analyze topic trends based on the BERTopic neural network topic model;

[0008] S3. Identify technology vacuum points and technology opportunities through patent maps and paper maps generated by GTM;

[0009] S4. Based on the patent map and the paper map generated in S3, semantic link relationships and citation link relationships are constructed to select and interpret technical opportunities.

[0010] Preferably, step S1 specifically includes the following steps:

[0011] S11. Search papers and patent data through the lens database;

[0012] S12: Use title, abstract, publication date and journal source as specific search strategies to collect important structured data, and extract keywords from paper and patent abstracts through segmentation, tokenization, lemmatization and stop word removal.

[0013] Preferably, step S2 generates document embedding representation based on the BERTopic neural network topic model, reduces the dimension of the embedding, clusters the embedding, and then constructs the topic representation.

[0014] Further preferably, the step S2 specifically comprises the following steps:

[0015] S21, document embedding using the all-MiniLM-L6-v2 sentence transformer model;

[0016] S22, using uniform manifold approximation and projection to reduce the dimension of the document embedding and obtain cluster dimension-reduced document embedding;

[0017] S23, applying noisy density-based hierarchical spatial clustering to the cluster-reduced document embedding to capture structures at different densities;

[0018] S24. Use the literature topic document generated by the BERTopic model to determine the topic of each paper and patent, and conduct comparative analysis on the paper and patent topics on a yearly basis to observe and analyze the evolution trend of topics in specific fields.

[0019] Preferably, step S3 is to discover technology vacuums through GTM visualization and obtain target technology opportunities.

[0020] Further preferably, the step S3 specifically includes the following steps:

[0021] S31, using word frequency inverse document frequency to extract keywords from papers and patents respectively and obtain two groups of keywords, obtaining the intersection elements of the two groups of keywords by analyzing the abstract data of papers and patents, using the intersection elements to respectively construct a paper document keyword matrix d1 and a patent document keyword matrix d2, and mapping the paper document keyword matrix d1 and the patent document keyword matrix d2 respectively through GTM to obtain a paper map and a patent map;

[0022] S32, based on the GTM map generated in step S31, the vacuum point and the non-vacuum point are identified, and the original vectors of the vacuum point and the non-vacuum point are converted into binary keyword vectors by using a threshold after GTM inverse mapping, and the terms contained in the vacuum point and the non-vacuum point are determined by GTM inverse mapping and setting a threshold, and the meanings contained in the vacuum point and the non-vacuum point are explained;

[0023] S33, constructing a link through a semantic relationship and a reference relationship; the semantic relationship is obtained by calculating the cosine similarity of the keyword vector obtained in S32 to obtain the semantic similarity between the paper non-vacuum point and the patent non-vacuum point, and connecting the two non-vacuum points whose cosine similarity calculation results are less than 0.3 to construct the semantic link relationship; the reference relationship is to use cosine similarity to measure the similarity between the patent vacuum point and the surrounding points to obtain the surrounding points with the highest similarity to the patent vacuum point, and regard the reference link relationship between the surrounding points and the paper non-vacuum point as the reference link relationship between the patent vacuum point and the paper non-vacuum point;

[0024] S34, randomly dividing the semantic link relationship obtained in step S33 into a training set and a test set in a ratio of 8:2 for link prediction, and obtaining a link prediction index; selecting the link prediction index with the highest AUC value, and predicting the connection between the patent non-vacuum points around the patent vacuum point and the paper non-vacuum point according to the selected link prediction index;

[0025] S35. The patent vacuum points around the patent non-vacuum points connected to the paper non-vacuum points are taken as target technology opportunities.

[0026] Preferably, the step S4 is to perform cosine similarity calculation on the keyword vector obtained in S32, and at the same time, combine the link prediction index obtained in S34 to construct a prediction similarity index for selecting technology opportunities.

[0027] Preferably, the step S4 is to calculate the link prediction index and the similarity prediction index through the semantic relationship constructed in the step S33, and select and interpret the technical opportunity in combination with the reference relationship constructed in the step S33.

[0028] Further preferably, the step S4 specifically comprises the following steps:

[0029] S41, based on the semantic link relationship, the link prediction index and the similarity prediction index are calculated, wherein the link prediction index is obtained in the step S34; the similarity prediction index is to link the patent vacuum point in claim 6 with the paper non-vacuum point, and the higher the similarity prediction index, the higher the similarity between the patent vacuum point and the paper non-vacuum point, and the calculation formula is as follows:

[0030]

[0031] Where: It represents the cosine similarity between the vacuum point and the surrounding points in the patent map, similarityscore (V mt ,V n ) represents the similarity value between the non-vacuum point and the surrounding points in the paper map obtained by the link prediction index, prediction similarity (V m ,V n ) represents the similarity between the vacuum point in the patent map and the corresponding non-vacuum point in the paper map;

[0032] S42: Select and interpret the target opportunity based on the link prediction index and the similarity prediction index obtained in step S41 and the reference link relationship calculated in step S33.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] The present invention provides a technology opportunity identification method based on a double-layer GTM map of papers and patents, predicts the trend of technology topics based on the BERTopic neural network topic model, discovers technology vacuums through GTM visualization, and obtains target technology opportunities; the present invention constructs a prediction similarity index and selects and interprets the target technology opportunities, which can remove the intervention of field experts and make the information interpretation of patent vacuum points more objective and reasonable, and can also obtain more comprehensive results. Description of the drawings:

[0035] Figure 1 A diagram showing the steps of a system method for identifying technology opportunities in an embodiment of the present invention;

[0036] Figure 2 An example diagram of selecting corresponding thesis points for target technology opportunities based on semantic links in an embodiment of the present invention;

[0037] Figure 3a It is a GTM map of the paper based on drone technology in the embodiment of the present invention;

[0038] Figure 3b Patented GTM map based on drone technology in the embodiment of the present invention

[0039] Figure 4 It is a schematic diagram of the semantic links and reference links of the patent vacuum V165 point in the present invention. DETAILED DESCRIPTION

[0040] The present invention is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0041] Example 1

[0042] This paper takes the research on the application of IoT sensors and computer vision technology in flood monitoring and mapping as a case where technology opportunity identification is needed.

[0043] A two-layer GTM map technology opportunity identification method based on papers and patents, the flowchart is as follows: Figure 1 As shown, the specific steps include:

[0044] S1. Collect and preprocess data;

[0045] S2, extract topics from papers and patents and analyze topic trends based on the BERTopic neural network topic model;

[0046] S3. Identify technology vacuum points and technology opportunities through patent maps and paper maps generated by GTM;

[0047] S4. Based on the patent map and the paper map generated in S3, semantic link relationships and citation link relationships are constructed to select and interpret technical opportunities.

[0048] The step S1 of collecting and preprocessing data in the identification method specifically includes the following steps:

[0049] S11. Search papers and patent data through the lens database;

[0050] S12. Use the Natural Language Toolkit (NLTK) package in the Python environment to collect important structured data using titles, abstracts, publication dates, and journal sources as specific search strategies, and extract keywords from paper and patent abstracts through segmentation, word segmentation, lemmatization, and stop word removal.

[0051] Step S3 is to generate patent maps and paper maps about the application of IoT sensors and computer vision technology in flood monitoring and mapping through GTM visualization, and to establish semantic link relationships and reference link relationships by conducting research on vacuum points and non-vacuum points on the two maps, thereby discovering technology vacuums and obtaining the target technology opportunities.

[0052] Step S4 is to calculate the link prediction index and the similarity prediction index through the semantic relationship constructed in step S3, and select and interpret the technical opportunities for the application of IoT sensors and computer vision technology in flood monitoring and mapping in combination with the reference relationship constructed in step S3, which specifically includes the following steps:

[0053] S41: The semantic link between the target technology opportunity and the non-vacuum point of the paper is constructed through a similarity index, namely a similarity prediction index, as follows: Figure 2 As shown, the similarity prediction index is to link the patent vacuum point with the paper non-vacuum point. The higher the similarity prediction index is, the higher the similarity between the patent vacuum point and the paper non-vacuum point is. The calculation formula is as follows:

[0054]

[0055] Where: It represents the cosine similarity between the vacuum point and the surrounding points in the patent map, similarity score (V mt ,V n ) represents the similarity value between the non-vacuum point and the surrounding points in the paper map obtained by the link prediction index, prediction similarity (V m ,V n ) represents the similarity between the vacuum point in the patent map and the corresponding non-vacuum point in the paper map;

[0056] S42. Select and interpret the target opportunity based on the link prediction index and the similarity prediction index obtained in step S41, and the reference link relationship calculated in step S3. The embodiment of the present invention interprets patent vacuum points (i.e., target technology opportunities) through relevant non-vacuum points in the paper map. Since each non-vacuum point in the paper map contains multiple papers, it is necessary to evaluate the papers to narrow the scope of explaining the target technology opportunities. The embodiment of the present invention evaluates and screens papers based on journal quality and timeliness. JCR (Journal Citation Reports) partitioning is based on data in the Web of Science database. According to different impact factors, JCR divides journals into four levels: Q1, Q2, Q3 and Q4. In general, the higher the impact factor, the greater the academic status and influence of the journal. Among them, Q1 is the journal with the highest impact factor, and Q4 is the journal with the lowest impact factor. The present invention divides journals into three levels: Q1 and Q2, Q3 and Q4 and others, and assigns values ​​of 3, 2 and 1 as journal quality scores respectively. The publication date is divided into three periods: 2001-2010, 2011-2020 and 2021-2023, and assigns values ​​of 1, 2 and 3 as journal timeliness scores respectively. The average of the journal quality score and the journal timeliness score is taken as the final score of the paper. For the target technology opportunity, the paper with the highest score will be selected to explain it.

[0057] Figure 3a The keyword matrix of the paper document is mapped onto the graph through the GTM model, and the coordinate position is determined by the parameters set by GTM; Figure 3b The patent document keyword matrix is ​​mapped onto the graph through the GTM model. The red circle indicates that there is no relevant patent data at this position, which is the patent vacuum point. The label is the label of the patent vacuum point, and the labels are sorted from left to right and from top to bottom.

[0058] like Figure 3b As shown in the figure, vacuum point No. 10 on the patent map is V165. Taking patent vacuum point V165 as an example, Figure 4 As shown, the paper map related to V165 is obtained by using the recognition method proposed in the present invention (such as Figure 3aThe two non-vacuum points contain a total of 6 paper information. The journal quality score and timeliness score are used as the measurement criteria for the papers, and the papers with the highest scores are selected to explain the patent vacuum. For V165, the paper 10.3390 / s19225012 titled computer vision and IoT-based sensors in flood monitoring and mapping: a systematic review and the paper 10.1002 / rob.21526 titled autonomous exploration of urban environments using unmanned aerial vehicles are used. The paper 10.3390 / s19225012 is a literature review that systematically reviews the research on the application of IoT sensors and computer vision technology in flood monitoring and mapping. Therefore, this paper focuses on the analysis of paper 10.1002 / rob.21526, which introduces a self-developed UAV that combines advanced hardware such as an inertial navigation system equipped with a two-dimensional laser rangefinder and an innovative software algorithm that can generate the best viewing angle for the next step, aiming to improve the autonomous exploration capability in an urban environment, which is of great significance to the fields of urban planning, emergency management and environmental protection.

[0059] Vacuum Point V165 indicates that drones equipped with IoT sensor data and computer vision technology have promoted new advances in drone technology for complex environment monitoring, highlighting their importance in solving practical constraints, flood monitoring, early warning systems, and urban environment monitoring. In the 2024 patent application, we found a total of three related patents. Compared with the existing technical solutions, the present invention removes the intervention of field experts and interprets the information of the patent vacuum point more objectively and reasonably, while also obtaining more comprehensive results.

[0060] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A technology opportunity identification method based on a two-layer GTM map of papers and patents, characterized by: The following steps are involved: S1. Collect and preprocess data; S2, extract topics from papers and patents and analyze topic trends based on the BERTopic neural network topic model; S3. Identify technology vacuum points and technology opportunities through patent maps and paper maps generated by GTM; S4. Based on the patent map and the paper map generated in S3, semantic link relationships and citation link relationships are constructed to select and interpret technical opportunities.

2. The technology opportunity identification method based on a double-layer GTM map of papers and patents according to claim 1 is characterized in that: The data collection and preprocessing in step S1 specifically includes collecting structured data using a search strategy including title, abstract, publication date and journal source, and extracting keywords from paper and patent abstracts by segmentation, word segmentation, lemma restoration and removal of stop words.

3. The method for identifying technology opportunities based on a double-layer GTM map of papers and patents according to claim 1 is characterized in that: The step S2 is to analyze the topic trend based on the BERTopic neural network topic model, and specifically includes the following steps: S21, document embedding using the all-MiniLM-L6-v2 sentence transformer model; S22. Using uniform manifold approximation and projection to reduce the dimensionality of the document embedding, and applying hierarchical spatial clustering to capture the structure at different densities; S23. Use the BERTopic model to determine the topic of each paper and patent, and conduct comparative analysis on the topics of the papers and patents in units of preset time periods to observe and analyze the evolution trend of topics in specific fields.

4. The method for identifying technology opportunities based on a double-layer GTM map of papers and patents according to claim 1 is characterized in that: The step S3 is to discover technology vacuums through GTM visualization and obtain target technology opportunities, which specifically includes the following steps: S31, using word frequency inverse document frequency to extract keywords from papers and patents respectively and obtain two groups of keywords and intersection elements, using the intersection elements to respectively construct a paper document keyword matrix d1 and a patent document keyword matrix d2, and using GTM to map d1 and d2 to obtain a paper map and a patent map; S32, defining the vacuum points and non-vacuum points on the paper map as paper vacuum points and paper non-vacuum points, respectively, The vacuum points and non-vacuum points on the patent map are defined as patent vacuum points and patent non-vacuum points; the original vectors of the vacuum points and the non-vacuum points are converted into binary keyword vectors by using a threshold after GTM inverse mapping, and the terms contained in the vacuum points and the non-vacuum points are determined by using the GTM inverse mapping and setting the threshold, and the meanings contained in the vacuum points and the non-vacuum points are explained; S33, constructing a link through a semantic relationship and a reference relationship; the semantic relationship is achieved by calculating the cosine similarity of the keyword vector obtained in S32, connecting the two non-vacuum points whose cosine similarity calculation results are less than a preset threshold and constructing the semantic link relationship; the reference relationship is achieved by using cosine similarity to measure the similarity between the patent vacuum point and the surrounding points, obtaining the surrounding points with the highest similarity to the patent vacuum point, and treating the reference link relationship between the surrounding points and the non-vacuum point of the paper as the reference link relationship between the patent vacuum point and the non-vacuum point of the paper; S34, randomly dividing the semantic link relationship obtained in step S33 into a training set and a test set in a preset ratio, calculating a link prediction index, and selecting the link prediction index with the highest AUC value to predict the connection; S35. The patent vacuum points around the patent non-vacuum points connected to the paper non-vacuum points are taken as target technology opportunities.

5. The method for identifying technology opportunities based on a double-layer GTM map of papers and patents according to claim 1 is characterized in that: The preset threshold used in step S33 in claim 6 is 0.3, and the preset ratio used in step S34 is 8:

2.

6. The method for identifying technology opportunities based on a double-layer GTM map of papers and patents according to claim 1 is characterized in that: The step S4 mainly selects and interprets the technical opportunity based on the semantic relationship and reference relationship constructed in the step S33, and specifically includes the following steps: S41, based on the semantic link relationship, the link prediction index and the similarity prediction index are calculated, wherein the link prediction index is obtained in step S34; the higher the similarity prediction index, the higher the similarity between the patent vacuum point and the paper non-vacuum point, and the calculation formula is as follows: Where: It represents the cosine similarity between the vacuum point and the surrounding points in the patent map, similarity score (V mt ,V n ) represents the similarity value between the non-vacuum point and the surrounding points in the paper map obtained by the link prediction index, prediction similarity (V m ,V n ) represents the similarity between the vacuum point in the patent map and the corresponding non-vacuum point in the paper map; S42: Select and interpret the target opportunity based on the link prediction index and the similarity prediction index obtained in step S41 and the reference link relationship calculated in step S33.

7. A technology opportunity identification method based on a two-layer GTM map of papers and patents, characterized in that: It includes at least one processor and a memory that is communicatively connected to the at least one processor; the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute a technology opportunity identification method based on a double-layer GTM map of papers and patents as described in any one of claims 1 to 6.

Images