A semantic-based method for identifying highly inventive patent networks

Abstract

The present invention relates to a highly creative patent network identification method based on semantics, comprising the following steps: acquiring a target functional patent set; extracting keywords in patent texts, taking nouns and verbs in the keywords of each patent in the patent set as concept vocabulary of the patent, and also as nodes of a semantic network in a patent semantic memory model; calculating semantic similarity of concept nodes; judging the creativity level of semantic memory: determining a semantic distance according to a path in a fully connected concept semantic network and the comprehensive semantic similarity on the path, judging the creativity level of the fully connected concept semantic network based on the semantic distance, and pruning the network based on the creativity level, thereby minimizing the scale of the concept semantic network on the basis of ensuring high creativity, and obtaining a highly creative semantic network for concept stimulation.

Description

Technical Field

[0001] The present invention belongs to the information field based on patent knowledge recommendation and semantic memory network, and relates to a patent network identification method. Background Art

[0002] Concept design is in the early stage of product design, and its purpose is to provide product solutions. Concept design not only determines product quality, cost, performance, reliability, safety and environmental protection, but also produces design defects that cannot be compensated by subsequent design processes. Concept design is a process of solving design problems through the interaction of innovative design thinking and design knowledge. Designers need to give full play to their creativity and comprehensively use their own design knowledge and experience to find possible solutions. They also need to break through the limitations of their own design knowledge and experience and expand the knowledge space for searching solutions.

[0003] Patent texts are regarded as a kind of scientific and technological documents. Their format is clear and rigorous. They usually consist of a title page, claims, a specification, and an appendix to the specification. Figure 4 Patents are divided into three main parts. These parts will elaborate on solutions to specific technical problems, and these solutions usually involve the structural composition of a certain product, but do not describe the structural parameters and processing technology information of the product in detail. Patents contain rich and creative technical information, which is of great reference and inspiration value for product innovation design. Among them, design knowledge is an important reference source for designers when designing products, which can effectively expand the designer's knowledge space [1]. However, designers need to spend a lot of time and energy to search and read patent documents. In order to alleviate this problem, patent analysis-related technologies have emerged. These technologies use information technologies such as artificial intelligence, big data, and machine learning to efficiently search and extract knowledge from patent documents. This can not only significantly reduce the burden of designers' manual knowledge retrieval, but also improve the efficiency of patent information collection and analysis.

[0004] Semantic memory is a long-term representation of human knowledge of the world. It encodes and stores knowledge of various concepts used to represent the world and the semantic relationships between concepts, and is available for human retrieval at any time to complete various cognitive activities. Classical cognitive theory proposes that semantic memory can be represented in the form of a semantic network, where nodes represent various concepts and edges represent semantic relationships between different concepts. The semantic relationships between different concepts are usually connected by semantic similarities of different strengths, and the different strengths reflect the semantic distance between concepts. The farther the semantic distance between two concepts, the more difficult it is to associate one with the other. The semantic memory structure of highly creative individuals can reduce the difficulty of such long semantic distance associations [2]. Individual creativity is reflected in the network representation of semantic memory: the more semantic connections between concepts in the semantic network, the shorter the semantic distance between them, and the higher the individual creativity.

[0005] References

[0006] [1]SARICA S,LUO J,WOOD K L.TechNet:Technology semantic network based patent data[J].Expert Systems with Applications,2020,142:112995

[0007] [2]Beaty RE,Silvia PJ,Nusbaum EC,et al.The roles of associative and executive processes in creative cognition[J].Memory&Cognition,2014,42:1186–1197. Summary of the invention

[0008] In order to assist designers in expanding their knowledge space and efficiently extract valuable design knowledge from existing patent knowledge, the purpose of the present invention is to provide a highly creative patent semantic network generation method based on semantic similarity calculation. The present invention aims to construct a patent semantic memory network by extracting patent key concept texts and combining similarity calculations between concept nodes. Pruning is performed on the basis of the fully connected concept semantic memory network to provide a smaller, more efficient and highly creative patent semantic memory network for subsequent concept innovation and reorganization. The main technical solutions are as follows:

[0009] A semantic-based method for identifying highly inventive patent networks comprises the following steps:

[0010] Step 1: Acquisition of target functional patent set

[0011] Step 2: Extract keywords from patent texts, and use the nouns and verbs in the keywords of each patent in the patent collection as the conceptual vocabulary of the patent, and also as the nodes of the semantic network in the patent semantic memory model;

[0012] Step 3: Calculate the semantic similarity of concept nodes. The method is as follows:

[0013] After extracting all the conceptual words in the patent collection and determining the nodes in the semantic network, the nodes need to be connected with edges to form a complete semantic network. The semantic similarity is calculated by combining the semantic dictionary-based and corpus-based methods.

[0014] The two semantic similarities are weighted summed to obtain the comprehensive semantic similarity of the two concepts; after calculating the comprehensive semantic similarity between all concepts, all concepts are connected pairwise to construct a fully connected concept semantic network;

[0015] Step 4: Determine the creativity level of semantic memory: Determine a semantic distance based on the paths in the fully connected concept semantic network and the comprehensive semantic similarity on the paths, and judge the creativity level of the fully connected concept semantic network based on this semantic distance. Prune the network based on the creativity level, thereby minimizing the size of the concept semantic network while ensuring high creativity, and obtain a highly creative semantic network for concept stimulation. The specific steps are as follows:

[0016] (1) The comprehensive semantic similarity threshold is determined by using the network's maximum shortest semantic distance max_dist, average shortest semantic distance avg_dist, and invalid edge number invalid_num to ensure that the fully connected network still maintains high creativity after pruning with this threshold;

[0017] (2) Based on the comprehensive semantic similarity threshold, prune all edges below the threshold to obtain the candidate concept semantic network;

[0018] (3) Remove all invalid edges in the candidate concept semantic network to obtain a highly creative semantic network that meets the high creativity conditions and has the smallest scale.

[0019] Furthermore, the method of the second step is as follows: extract keywords from the patent text: use the title and abstract of the patent as the original text, and use the TF-IDF algorithm to extract keywords; select conceptual vocabulary as concepts from the extracted keywords, and use the nouns and verbs in the keywords of each patent in the patent collection as the conceptual vocabulary of the patent, and also as the nodes of the semantic network in the patent semantic memory model.

[0020] Furthermore, the method of using the TF-IDF algorithm for keyword extraction is as follows: in the preprocessing stage before calculating TF-IDF, perform word segmentation and stop word removal preprocessing operations, and also perform part-of-speech tagging and word form restoration operations; amplify the TF value of the vocabulary in the patent title; and expand the patent text set for calculating IDF to all patents.

[0021] Furthermore, in the third step, the semantic similarity based on the semantic dictionary adopts the semantic similarity based on the wordnet path distance; the semantic similarity based on the corpus adopts the semantic similarity based on word2vec. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 The overall framework of the network identification method for highly inventive patents

[0023] Figure 2 Comprehensive semantic similarity calculation process based on word2vec and wordnet

[0024] Figure 3 Semantic distance of analogy resistor series model

[0025] Figure 4 Concept semantic distance example

[0026] Figure 5 Fully connected concept semantic network pruning process

[0027] Figure 6 Curve for determining the comprehensive semantic similarity threshold

[0028] Figure 7 Patent statistics chart of main function expression template coverage

[0029] Figure 8 Calculation results of semantic similarity of concept words

[0030] Fig. 9 k-max_dist, k-avg_dist, k-invalid_num graph

[0031] Fig.10 Highly Inventive Patent Knowledge Semantic Network DETAILED DESCRIPTION

[0032] To make the technical solution of the present invention clearer, the present invention is further described below with reference to the accompanying drawings. The overall framework of the semantic-based highly creative patent network identification method of the present invention is as follows: Figure 1 As shown, the present invention is implemented in the following steps: 1. Based on the target functional patent set, a fully connected concept semantic network is constructed.

[0033] (1) Acquisition of target function patent set

[0034] The main functional verbs of each patent are extracted, and a patent set with high semantic similarity between the main functional verbs and the target functional verbs is found as the target functional patent set.

[0035] The Derwent Patent Database has rewritten patent data from different organizations and countries written in different languages ​​into English, and has also standardized some patent texts that differ greatly from the language style of scientific and technological literature, which has greatly facilitated the analysis of patent data. Therefore, the Derwent Patent Database is used as the source of patent data. Through data collection, patents from different organizations and countries are obtained from the Derwent Patent Database, and then invention patents are extracted from them through data screening and data classification, and divided into invention names, abstracts, claims and other data according to the composition of the patents. Finally, the various parts of the invention patent data are saved in the form of text in an Excel file to provide a data basis for subsequent operations.

[0036] Based on the main function expression template, the main functions expressed in verbs, present participles and derived nouns in the patent text data are extracted and restored to the form of verbs. After determining the main function verbs of each patent, the target function patent set for building the semantic memory model is obtained according to the similarity between the main function verbs and the target function verbs of the patent. The similarity between the main function verbs and the target function verbs adopts the semantic similarity based on wordnet path distance.

[0037] (2) Patent concept extraction based on TF-IDF and part of speech

[0038] The concepts in patents are words used to express the structure, function, operation, scenario and other information contained in patents. They are keywords that can express the subject content of patents, so the keywords in patent texts are extracted first. Keyword extraction also uses the title and abstract of patents as the original texts and uses the TF-IDF (Term Frequency & Inverse Documentation Frequency) algorithm for extraction.

[0039] The TF-IDF algorithm is TF×TDF, where TF stands for term frequency, which refers to the frequency of a word appearing in a text. The more important the word, the more times it appears in the text. IDF stands for inverse document frequency, which refers to the document frequency of a specific word appearing in all documents. Words that appear in only a few documents are more important than words that appear in multiple documents. The calculation method of the TF-IDF algorithm is as follows:

[0040]

[0041]

[0042] TFIDF i,j =TF i,j ×IDF i #(3)

[0043] Where n i,j is the number of times word i appears in document j, |D| is the total number of documents, |{j:w i ∈d j} is the number of documents in which word i appears.

[0044] The TF-IDF algorithm can not only highlight the words with high frequency in the document, but also consider the influence of the words that appear frequently in each document, and can better extract keywords from long texts. However, the TF-IDF algorithm still has shortcomings. When obtaining internal information of the text through TF, it simply considers the frequency of the word in the text, and ignores other features such as the position, range, and part of speech; when obtaining external information of the text through IDF, it ignores the category of the word in all documents. If a word often appears in a type of document, then the word can be used as a keyword for this type of document, but the number of this type of documents in the document set far exceeds that of other types of documents, then IDF will weaken the importance of the word.

[0045] In order to avoid the shortcomings of the TF-IDF algorithm as much as possible, the calculation of TF and IDF is modified as follows: in the preprocessing stage before calculating TF-IDF, in addition to the conventional preprocessing operations such as word segmentation and stop word removal, part-of-speech tagging is also performed to remove prepositions that have no practical meaning, and the plural nouns, non-finite verbs, and comparative adjectives and adverbs are restored to avoid the situation where one word has multiple forms as much as possible; the patent title is a condensed summary of the entire patent content, so the words that appear in the title are more important than the words in other positions. To reflect their importance, the TF value of the words in the patent title is amplified by 1.5 times; since the patents in the patent collection are all patents with similar functions to the target, the field categories they cover will be relatively concentrated, so IDF will weaken the importance of the keywords in this category. This situation can be avoided by expanding the patent text set for calculating IDF to all patents.

[0046] After extracting the keywords, it is necessary to select conceptual words as concepts. The inspiration for product design comes from specific types of words, such as verbs, nouns and other content words. In patent texts, nouns mainly express certain objects, structures, and characteristics, while verbs mainly express certain functions, behaviors, and operations. Adjectives and adverbs are mainly used to semantically supplement the content expressed by nouns and verbs, and it is difficult to express concepts independently. Therefore, the nouns and verbs in the keywords of each patent in the patent collection are used as the conceptual vocabulary of the patent, and also as the nodes of the semantic network in the patent semantic memory model.

[0047] (2) Comprehensive semantic similarity calculation based on word2vec and wordnet

[0048] After extracting all the concept words in the patent collection and determining the nodes in the semantic network, the nodes need to be connected with edges to form a complete semantic network. The edge in the semantic network is the semantic similarity between the two concept words corresponding to the node.

[0049] There are many methods for calculating semantic similarity, including dictionary-based, encyclopedia-based, neural network-based, ontology-based, etc., but generally these methods can be divided into two types: semantic dictionary-based and corpus-based. The former is based on artificially constructed semantic dictionaries such as wordnet, Wikipedia, mindnet, etc., and calculates semantic similarity by calculating the distance between words in tree-structured or graph-structured dictionaries; the latter is based on the contextual information of words in large-scale corpora, and calculates semantic similarity by representing words as vectors containing contextual information and calculating the cosine similarity between vectors.

[0050] Semantic similarity based on semantic dictionary is a linguistic method. Compared with semantic similarity based on corpus, it does not rely on corpus, but on semantic dictionary compiled by linguistic experts. It does not have noise interference, is not limited to a certain field, and has a wider scope of application. However, for a specific corpus in a specific field, the semantic dictionary-based method is difficult to find the similarity of information between lexical contexts. The corpus-based method can make up for this shortcoming. Therefore, a combination of semantic dictionary-based and corpus-based methods is used to calculate semantic similarity. The semantic similarity based on semantic dictionary adopts semantic similarity based on wordnet path distance. The calculation method is shown in the following formula:

[0051]

[0052] The corpus-based semantic similarity uses the semantic similarity based on word2vec. Word2vec is a distributed word representation method based on deep learning for word embedding proposed by Google in 2013. It can learn high-quality word vectors containing contextual information from massive corpora. Word2vec mainly has two models: Continuous Bag-of-Words Model (CBOW) and Continuous Skip-gram Model (Skip-gram). CBOW generates the word vector of the target word by calculating the context of the target word, while Skip-gram generates the word vector of its context based on the target word. Although the two models have different processes for generating word vectors, they are essentially calculating and generating multi-dimensional word vectors through the contextual relationship between words.

[0053] As a method of representing words in text, word2vec is widely used in the field of natural language processing for tasks such as text classification, entity recognition, and machine translation. The word vectors generated by word2vec can also calculate semantic similarity through vector cosine distance, and by comparing with wordnet, it is found that word2vec can express richer semantic information than other word representation methods.

[0054] The present invention uses the word2vec model of the Skip-gram model to generate a 300-dimensional word vector, and the similarity of the vocabulary is calculated using the cosine similarity shown in the following formula.

[0055]

[0056] In the formula, w1 and w2 are two words, and their word vectors are A(a1,a2,a3,…,a 300 ) and B(b1,b2,b3,…,b 300 ).

[0057] The comprehensive semantic similarity calculation process based on word2vec and wordnet is as follows Figure 2 As shown in the figure, firstly, a word sequence corpus is constructed from the target patent text set through sentence segmentation, word segmentation, word form restoration, and stop word removal operations to train the word2vec Skip-gram model. Then, two concepts are selected from the concept set and the semantic similarity Sim based on the wordnet path distance is calculated according to formula (4). wn At the same time, the Skip-gram model is used to obtain the 300-dimensional word vectors of the two concepts, and the semantic similarity Sim based on word2vec is calculated according to formula (5) wv ; Finally, give the semantic similarity Sim wn and Sim wv The weight of 0.5 is assigned to calculate the weighted sum of the two concepts, and the result is the comprehensive semantic similarity of the two concepts. According to the above process, the comprehensive semantic similarity between the concepts in the concept set is calculated in sequence. The obtained comprehensive semantic similarity not only reflects the similarity of the concepts from the linguistic point of view, but also reflects the similarity of the concepts from the contextual semantic information of the target functional patent set.

[0058] After calculating the comprehensive semantic similarity between all concepts, all concepts can be connected to construct a fully connected concept semantic network. A fully connected concept semantic network can meet the high creativity condition to the greatest extent, and any concept only needs to pass through one edge to another concept.

[0059] 2. Highly creative semantic network construction based on semantic computing

[0060] Although the constructed fully connected concept semantic network can meet the high creativity conditions to the greatest extent, and any concept only needs to go through one edge to reach another concept, the scale of the fully connected concept semantic network is huge, and even fewer nodes will generate a large number of triples. Moreover, the semantic distance between concepts cannot be simply calculated by the length of the network path between the two concept nodes. For paths of the same length, the comprehensive semantic similarities of the edges on the path are different, and their semantic distances cannot be considered the same.

[0061] Therefore, it is necessary to determine a semantic distance based on the paths in the concept semantic network and the comprehensive semantic similarity on the paths, and judge the creativity level of the fully connected concept semantic network based on this semantic distance. The network is pruned based on the creativity level, so as to minimize the size of the concept semantic network while ensuring high creativity, and obtain a highly creative semantic network for concept stimulation.

[0062] In an unweighted network, the network path length can better measure the distance between two nodes, but in a weighted network, measuring the distance between two nodes by path length will ignore the weight information contained in the path. In a conceptual semantic network, what is ignored is the semantic information contained in the comprehensive semantic similarity. Figure 3 As shown in the figure, the same number of resistors are connected in series, but the resistance values ​​of the resistors are different, and the potential difference between the two ends is also different. The semantic distance between concepts is similar. For the same path length, the smaller the comprehensive semantic similarity on the path, the larger the semantic distance. When the comprehensive semantic similarity of all edges on the path is 1, the semantic distance is 0. When the comprehensive semantic distance approaches 0, the semantic distance approaches infinity. Therefore, formula (6) is used to convert the comprehensive semantic similarity on the path into semantic distance.

[0063]

[0064] In the formula, Sim i is the comprehensive semantic similarity of the i-th edge on the path.

[0065] According to the semantic distance calculated by the above method, the path with a length of 1 directly connecting concepts is not necessarily the path with the shortest semantic distance, such as Figure 4As shown, from the concept node "grind" to "grass", path 1 is a directly connected path, and path 2 is the path with the shortest semantic distance. The directly connected path is farther than the path passing through the concept node "mash", which can also be explained semantically. It is difficult to directly associate "grind" with "grass", while it is easier to associate "grind" with "mash" and then with "grass". In this case, when associating other concepts with the concept node "grind", path 1 will never be selected, but path 2 with a shorter semantic distance will be selected. Path 1 is like a short-circuited resistor in a circuit and becomes an invalid edge. There are a large number of such edges in the fully connected concept semantic network. By removing these invalid edges to prune the fully connected concept semantic network, it can not only ensure that the shortest distance between concepts in the network is not affected, but also effectively reduce the size of the network. The specific pruning process is as follows. Figure 5 As shown in Figure 1, it includes two pruning operations: pruning I and pruning II. Pruning I is to prune the edges below the comprehensive semantic similarity threshold, and pruning II is to further prune the invalid edges based on the previous pruning.

[0066] Pruning I prunes all edges with low comprehensive semantic similarity, so an appropriate comprehensive semantic similarity threshold should be selected. The threshold should be able to ensure that the fully connected network still has high creativity after pruning the edges below this threshold, that is, any two concepts in the network have a short semantic distance. The selection of the threshold requires judging the creativity level of the fully connected concept semantic network after pruning under different thresholds. In order to obtain a semantic network with the smallest possible scale without losing high creativity, the maximum comprehensive semantic similarity threshold k that satisfies the high creativity of the concept semantic network is taken as the pruning threshold of pruning I.

[0067] The creativity level of the concept semantic network is determined by the three indicators of the maximum and shortest semantic distance max_dist, the average shortest semantic distance avg_dist and the invalid edge number invalid_num. The maximum and shortest semantic distance max_dist reflects that the shortest semantic distance between any two concepts in the concept semantic network does not exceed this value. The smaller this value is, the smaller the scale of the concept semantic network is and the higher the creativity is. The average shortest semantic distance avg_dist reflects the average of the shortest semantic distances between concepts in the concept semantic network. When the maximum and shortest semantic distance max_dist is the same, the smaller the average shortest semantic distance avg_dist is, the closer the relationship between the concepts is, and the higher the creativity of the concept semantic network is. The invalid edge number invalid_num reflects the total number of invalid edges in the concept semantic network. The fewer invalid edges are, the smaller the scale of the concept semantic network is, and the less information is lost in the network. Therefore, the final comprehensive semantic similarity threshold k should make the concept semantic network have as small a maximum and shortest semantic distance max_dist, a small average shortest semantic distance avg_dist and a small invalid edge number invalid_num as possible.

[0068] By drawing a curve graph of the relationship between the above three indicators and the threshold k, the curve graph is as follows Figure 6 As shown in the figure, in the concept semantic network corresponding to the area before the point where the slope of the max_dist-k curve and the avg_dist-k curve increases sharply, the network scale is small, and there is a small semantic distance between any two concepts. This area is selected as the threshold target area; then the value with the least number of invalid edges in this area is selected as the comprehensive semantic similarity threshold, and based on this, pruning is performed to obtain the candidate concept semantic network. Then, all invalid edges in the candidate concept semantic network are removed through pruning II, so as to obtain a highly creative semantic network that meets the high creativity conditions and has the smallest scale.

[0069] 3. Example Application

[0070] To construct a patent semantic memory model, it is necessary to extract the target function patent text set used to construct the patent semantic memory model from the patent database. Seven main function expression templates are used to extract the main functions of all patents in the patent database. The main function expression templates proposed in this invention cover the patents in the patent database as follows: Figure 7 As shown in the figure, the sum of the coverage of each template exceeds 100% because the same patent may involve multiple templates. As can be seen from the figure, only 8.79% of the patents are not covered by the template, and the coverage rate exceeds 90%.

[0071] According to the above example description, the target main function is "crush". The semantic similarity calculation method based on wordnet is used to calculate the semantic similarity between the main functions of each patent in the patent database and the target main function "crush". According to the meaning of functional words under different semantic similarities, 0.6 is used as the similarity threshold to take all patents greater than the threshold to form the target function patent text set. The target function patent text set has a total of 3175 patent data involving 11 main functions, including "crush", "pulverize", "grind", "extrude", "smash", "squeeze", "break", "shred", "demolish", "mash", and "fragment". These data will be used to construct the patent semantic memory model in the future.

[0072] First, a semantic network was constructed. TF-IDF was used to extract keywords from the titles and abstracts of 3,175 patents in the target functional patent text set, and nouns and verbs were taken out as concept words, which were also used as nodes in the semantic network. A total of 1,106 concept words were extracted in this step. Then, the title abstracts of all patents in the target functional patent text set were segmented, word segmented, stop words removed, and morphological restoration was performed to construct the corpus for training the word2vec Skip-gram model. After the training, the 1,106 concept words were input into the word2vec model to obtain the corresponding 1,106 300-dimensional vectors, and the cosine similarity was used to calculate the semantic similarity Sim between the concept words based on word2vec. wv At the same time, wordnet is used to calculate the semantic similarity Sim between concept words based on wordnet path distance wn , and calculate the weighted sum to get the comprehensive semantic similarity Sim between concept words, Figure 8 The comprehensive semantic similarity between some concept words is listed.

[0073] After calculating the comprehensive semantic similarity between all 1106 conceptual words, a fully connected conceptual semantic network can be constructed. The fully connected conceptual semantic network also needs to be pruned to construct a semantic network that meets high creativity conditions and minimizes its size.

[0074] The first pruning is to determine the appropriate threshold k by calculating the semantic distance between concepts and drawing a curve chart of the relationship between the network's maximum and shortest semantic distance max_dist, average shortest semantic distance avg_dist, and invalid edge number invalid_num and the comprehensive semantic similarity threshold k, and remove all edges less than k to achieve pruning. Take the threshold k interval as [0.0, 0.25] and draw the following Fig. 9The three curves shown. From the k-max_dist graph, we can see that when k is in the interval [0.0, 0.13], the network has the smallest maximum shortest semantic distance, so we can first narrow the value range of k to the interval [0.0, 0.13]; from the k-avg_dist graph, we can see that the average shortest semantic distance of the network rises slowly in the interval [0.0, 0.1], and the point k = 0.1 is an obvious inflection point. After this point, the average shortest semantic distance rises rapidly as k increases, so the value range of k can be further narrowed to the interval [0.0, 0.1]; from the k-invalid_num graph, we can see that in the interval [0.0, 0.1], the number of invalid edges decreases as the k value increases, and k = 0.1 is the point with the smallest number of invalid edges. Therefore, based on the above analysis, the threshold of the comprehensive semantic similarity is 0.1, and all edges less than 0.1 in the network will be deleted. After the first pruning, the concept semantic network has been pruned from a fully connected concept semantic network with 611,065 edges to a candidate concept semantic network with 317,654 edges. The network size has been reduced by 48%, but the semantic distance between concepts in the candidate concept semantic network is still small, ensuring that the network has high creativity.

[0075] The second pruning minimizes the size of the concept semantic network by calculating the semantic distance between concepts and removing the invalid edges that still exist in the candidate concept semantic network. The semantic network obtained after the second pruning has 201,873 edges, and the network size is reduced by 36.4% compared with the candidate concept semantic network and 67% compared with the initial fully connected concept semantic network, but the shortest semantic distance between concepts is still consistent with the candidate concept semantic network. The second pruning minimizes the size of the candidate concept semantic network without changing the high creativity of the candidate concept semantic network. The resulting semantic network is as follows: Fig.10 shown.

Claims

1. A semantic-based method for identifying highly inventive patent networks, comprising the following steps: Step 1: Acquisition of target functional patent set Step 2: Extract keywords from patent texts, and use the nouns and verbs in the keywords of each patent in the patent collection as the conceptual vocabulary of the patent, and also as the nodes of the semantic network in the patent semantic memory model; Step 3: Calculate the semantic similarity of concept nodes. The method is as follows: After extracting all the conceptual words in the patent collection and determining the nodes in the semantic network, the nodes need to be connected with edges to form a complete semantic network. The semantic similarity is calculated by combining the semantic dictionary-based and corpus-based methods. The two semantic similarities are weighted summed to obtain the comprehensive semantic similarity of the two concepts; after calculating the comprehensive semantic similarity between all concepts, all concepts are connected pairwise to construct a fully connected concept semantic network; Step 4: Determine the creativity level of semantic memory: Determine a semantic distance based on the paths in the fully connected concept semantic network and the comprehensive semantic similarity on the paths, and judge the creativity level of the fully connected concept semantic network based on this semantic distance. Prune the network based on the creativity level, thereby minimizing the size of the concept semantic network while ensuring high creativity, and obtain a highly creative semantic network for concept stimulation. The specific steps are as follows: (1) The comprehensive semantic similarity threshold is determined by using the network's maximum shortest semantic distance max_dist, average shortest semantic distance avg_dist, and invalid edge number invalid_num to ensure that the fully connected network still maintains high creativity after pruning with this threshold; (2) Based on the comprehensive semantic similarity threshold, prune all edges below the threshold to obtain the candidate concept semantic network; (3) Remove all invalid edges in the candidate concept semantic network to obtain a highly creative semantic network that meets the high creativity conditions and has the smallest scale.

2. The method for identifying highly inventive patent networks according to claim 1, characterized in that: The method of the second step is as follows: extract keywords from the patent text: use the title and abstract of the patent as the original text, and use the TF-IDF algorithm to extract keywords; select conceptual vocabulary as the concept from the extracted keywords, and use the nouns and verbs in the keywords of each patent in the patent collection as the conceptual vocabulary of the patent, and also as the nodes of the semantic network in the patent semantic memory model.

3. The method for identifying highly inventive patent networks according to claim 2, characterized in that: The method of keyword extraction using the TF-IDF algorithm is as follows: in the preprocessing stage before calculating TF-IDF, perform word segmentation and stop word removal preprocessing operations, as well as part-of-speech tagging and word form restoration operations; amplify the TF value of the vocabulary in the patent title; and expand the patent text set for calculating IDF to all patents.

4. The method for identifying highly inventive patent networks according to claim 1, characterized in that: In the third step, the semantic similarity based on the semantic dictionary adopts the semantic similarity based on wordnet path distance; the semantic similarity based on the corpus adopts the semantic similarity based on word2vec.

Images