Intelligent recommendation method and system for ipc classification number based on multi-source information fusion

Abstract

The application discloses an IPC classification number intelligent recommendation method and system based on multi-source information fusion, which firstly constructs a corpus through a sorting algorithm based on patent text, and generates a vector representation of each word, and obtains a text classification score of the patent through a convolutional neural network.Secondly, a heterogeneous graph is constructed, and a random walk sequence set is obtained through random walk driven by a multi-element structure, and a graph structure classification score is obtained through a classification network.Finally, a multi-modal deep fusion based on a gating mechanism is adopted to fuse the text classification score and the graph structure classification score of the patent, obtain a final classification prediction score of the patent, output a recommendation list of the corresponding IPC classification number, and perform reverse training.The application realizes adaptive fusion of different modal characteristics, accurately and efficiently generates an IPC classification number recommendation list, and provides multi-dimensional decision reference for patent applicants and related institutions.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and particularly relates to an IPC classification number intelligent recommendation method and system combining text information and graph structure information. BACKGROUND

[0002] With the continuous growth of global patent application volume, not only the patent applicants themselves need to accurately identify the technical subject, but also patent agencies, enterprise intellectual property departments, and local intellectual property protection centers need to preliminarily classify the patent technology at an early stage in the business process. The agent agency often makes an IPC classification prediction according to the received technical disclosure materials, and then assigns the corresponding agent according to the classification results; the internal intellectual property department of the enterprise will preliminarily classify the disclosure materials submitted by each business unit so as to be subsequently processed by the patent engineers of different technical lines; and the local protection center needs to preliminarily judge the technical field of the patent when performing the patent pre-examination service, and the IPC classification is one of the most direct and effective organization bases. Therefore, at the application stage or the preliminary judgment stage, various institutions responsible for the front-end processing urgently need to obtain the IPC classification of the patent as a basis more accurately and efficiently.

[0003] However, the IPC system structure presents a multi-level and fine-grained feature, and it is very difficult for non-classification experts to understand and correctly select. At present, the widely used method still mainly depends on manual retrieval of the database or relies on experience-based classification judgment, and this method has obvious limitations. First, the manual retrieval process is tedious and inefficient, and the effect is highly dependent on the retrieval keywords and personal experience. Second, the classification relies on the experience of professional personnel, and the cost of manual intervention is high. Third, subjective judgment is easy to lead to inconsistent classification results, affecting the accuracy of subsequent technology distribution and examination work. Although existing research has also explored automatic patent classification technology based on deep learning, these technologies basically face the examination or verification stage, and rely heavily on metadata such as assignee information, examiner review information, etc., and these information does not exist at the initial pre-judgment stage or the disclosure material stage, so that the existing technology cannot be truly implemented in the front-end business side. In addition, the existing scheme does not fully utilize the valuable prior information available at the application stage, such as the inventor's historical technical field, the past patent portfolio structure of the applicant, etc. SUMMARY

[0004] Based on the above real problems, the present application proposes an IPC intelligent recommendation method for the application stage and the front-end business link. The present application makes full use of the multi-source information that can be obtained at this stage, including patent text, applicant and inventor's past patent portfolio and other information, designs a double-channel feature extraction network: one channel uses convolutional neural network (CNN) for text semantic modeling, and the other channel uses multi-layer graph neural network to learn the patent relationship graph structure, and sorts and displays the candidate IPC classification numbers in the concept of the recommendation system. The present application not only retains the final selection right of artificial, but also significantly improves the front-end processing efficiency and prediction accuracy through intelligent recommendation.

[0005] In order to effectively realize the IPC classification number recommendation, in one aspect, the present application proposes an IPC classification number intelligent recommendation method based on multi-source information fusion.

[0006] (1) Text information extraction and preprocessing: extracting the text content of the title, abstract and description part from the patent document, using the TextRank algorithm to automatically extract the preset number of core sentences with the most important semantics from the text; selecting the preset number of words with the highest information entropy from the core sentences to form the text sequence of the patent , and removing stop words to build a corpus.

[0007] (2) Word vector representation generation: using the Word2Vec method of the Skip-Gram model to train the text sequence , generating the vector representation of each word, and concatenating all word vectors into a two-dimensional matrix , where is the vector dimension, is the number of words.

[0008] (3) Text feature extraction and classification: inputting the text matrix into the convolutional neural network, using four parallel convolution kernels of different sizes for convolution operation, performing maximum pooling operation on each group of convolution outputs to extract significant features, concatenating the pooled significant features through the full connection layer, and outputting the text classification score of the patent , where is the number of categories.

[0009] (4) Patent relationship graph construction: constructing a heterogeneous graph, which contains three types of nodes, i.e. patent nodes , applicant nodes and inventor nodes , and three types of edges, i.e. patent-applicant edges , patent-inventor edges and patent-patent citation edges .

[0010] (5) Multi-structure driven random walk: starting from all patent nodes in the heterogeneous graph, a multi-structure driven random walk strategy is performed to obtain a random walk sequence set with the same number of elements as the number of patents.

[0011] (6) Graph structure learning and relationship modeling: using the random walk sequence set to generate the initial representation of each node; using the aggregation method based on the node weight and degree perception mechanism to aggregate the neighbor node features, learning the multi-order neighbor features of the nodes through the multi-layer network structure, and using the hierarchical incremental feature fusion to output the graph structure classification score of the patent . .

[0012] (7) Adaptive fusion of multi-modal classification scores: a multi-modal deep fusion method based on a gating mechanism is adopted to adaptively fuse the text classification score of the patent and the graph structure classification score , to obtain the final classification prediction score of the patent . .

[0013] (8) Model training and optimization: using binary cross-entropy as the loss function, using the Adam optimizer to update the model parameters, using multi-modal adversarial learning to enhance the discriminability of the model, and training the entire network through the multi-label classification task.

[0014] (9) Recommended result output: according to the final classification score , the probability prediction result of the patent belonging to each technical category is obtained and the recommended list of IPC classification numbers is output.

[0015] In another aspect of the present application, an IPC classification number intelligent recommendation system based on multi-source information fusion is provided, which includes the following modules:

[0016] Corpus construction module: based on patent text, a corpus is constructed through a sorting algorithm.

[0017] Text classification module: based on the text sequence in the corpus, the vector representation of each word is generated, and the text classification score of the patent is obtained through a convolutional neural network.

[0018] Graph structure classification module: used for constructing a heterogeneous graph, obtaining a random walk sequence set through a multi-structure driven random walk, and obtaining a graph structure classification score through a classification network.

[0019] ​The IPC classification number recommendation module adopts a multimodal deep fusion based on a gating mechanism to adaptively fuse the text classification score and graph structure classification score of the patent to obtain the final classification prediction score of the patent. It obtains the probability prediction result of the patent belonging to each technology category and outputs the corresponding IPC classification number recommendation list, and performs reverse training.

[0020] Compared with traditional manual retrieval or single-source auxiliary methods, the beneficial effects of this invention are: (1) Enhanced decision credibility through multi-dimensional recommendation basis: By integrating multi-source information such as the semantics of patent texts and the applicant's and inventor's historical technical field preferences, an IPC classification number recommendation list is generated accurately and efficiently, providing multi-dimensional decision references for patent applicants and related institutions; (2) Adaptive feature weight allocation: Through the final classification score fusion mechanism based on gating mechanism, the system can dynamically adjust the contribution weight of different information sources in the final recommendation, realizing the adaptive fusion of different modal features; (3) Difference constraint fusion mechanism: Through multi-modal adversarial learning, the learning objective of the fusion channel is compared with the output of the single channel, forcing the model to learn complementary information from the two modalities, greatly improving the model's discrimination ability. Attached Figure Description

[0021] Figure 1 This is a flowchart of the intelligent recommendation method for IPC classification numbers based on multi-source information fusion according to the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1 The present invention provides a detailed description of an intelligent recommendation method for IPC classification numbers based on multi-source information fusion, comprising the following steps:

[0023] (1) Text information extraction and preprocessing:

[0024] 1-1. Core Sentence Generation: From Patents The text content of the title, summary and description sections of the document is extracted, and the TextRank algorithm is used to automatically extract the 10 most semantically important core sentences from the text;

[0025] 1-2. Corpus Construction: Core sentences are segmented into words, and the meaning of each word is calculated. probability of occurrence ; Calculate each word based on information entropy theory The information entropy, as shown in the formula below:

[0026]

[0027] in for Information entropy; select the 100 words with the highest information entropy to form the text sequence of the patent, and remove stop words to construct a corpus;

[0028] (2) Word vector representation generation: The text sequence is trained using the Skip-Gram model's Word2Vec method to generate a vector representation for each word, and the patent... All word vectors are concatenated into a two-dimensional matrix. ,in For the number of words, For vector dimensions;

[0029] (3) Text feature extraction and classification:

[0030] 3-1. Embedded layer processing: Patented. text matrix Then, it is first fed into an embedding layer, and this matrix is ​​set to be updatable during the training phase to allow the model to adjust the word vector representations according to task requirements. The output of the embedding layer is represented as... ;

[0031] 3-2. Convolution Processing: Using four parallel sets of convolution kernels of different sizes. right Perform a convolution operation to obtain four sets of feature vectors. The formula is as follows:

[0032] ,

[0033]

[0034]

[0035] in, for The number of convolution kernels, represent The first in One convolutional kernel, For the first The width of the group convolution kernel, Representative use The implemented convolution operation, convolution kernel The output feature vector set, represent The first in 1 eigenvector , is the dimension of the row vectors in the matrix;

[0036] 3-3. Max pooling processing: For max-pooling operation is performed on the feature vector in, and the maximum value is selected from the output of each convolution kernel, as follows:

[0037]

[0038] wherein, represents the significant features obtained by performing a max-pooling operation on ; subsequently, the four groups of significant features after max-pooling are spliced into a feature vector , as follows:

[0039]

[0040] 3-4, text classification score generation: inputting into a dropout layer and a fully connected layer in sequence, to obtain the classification score of the patent based on text information , , wherein, is the number of categories, as follows:

[0041]

[0042] wherein, represents the fully connected layer;

[0043] (4) patent relationship graph construction: a patent relationship heterogeneous graph is constructed, which contains three types of nodes, i.e., patent nodes , applicant nodes and inventor nodes , and three types of edges, i.e., patent-applicant edges , patent-inventor edges and patent-patent citation edges ;

[0044] (5) multi-structure-driven random walk: taking all patent nodes in the heterogeneous graph as the starting point, a multi-structure-driven random walk strategy is performed to obtain a random walk sequence set with the same number of elements as the number of patents;

[0045] 5-1, structure feature vector construction: taking the corresponding node of the patent as the starting point for random walk; for the current walk position node , the node degree , PageRank value , clustering coefficient and neighbor average degree of each neighbor node of the node are calculated, as follows:

[0046]

[0047]

[0048]

[0049]

[0050] wherein, is the first-order neighbor (i.e., direct neighbor) node set of the node , denotes the pre-computed PageRank value, denotes the number of triangles with the node as a vertex; the structural feature vector of each neighbor node is constructed according to the above information ;

[0051] 5-2, structural feature embedding mapping: input the above feature vector into a linear mapping layer to obtain a 32-dimensional structural embedding representation, as follows:

[0052]

[0053] wherein is a trainable parameter matrix, is a bias term;

[0054] 5-3, next-hop node selection: perform softmax normalization on the embedding representation of all neighbor nodes of the current node to obtain the transition probability of the neighbor node being selected as the next hop, as follows:

[0055]

[0056] wherein is a learnable direction vector; according to the probability distribution, a random sampling is performed to determine the next-hop node;

[0057] 5-4, single node sequence generation: repeat steps 5-1-5-3 until the preset walk length 30 is reached, to generate a node sequence with the patent corresponding node as the starting point;

[0058] 5-5, overall sequence generation: repeat steps 5-1-5-4 to perform random walk with all patent nodes in the heterogeneous graph as the starting point, to obtain a random walk sequence set with the same number of elements as the number of patents; the multi-structure-driven random walk strategy proposed in the present application realizes adaptive walk modeling of the structural relationship between nodes in the graph by fusing multiple structural statistical features and introducing a trainable probability generation network, thereby improving the expression ability of the graph embedding and the performance of the downstream task.

[0059] (6) Graph structure learning and relation modeling:

[0060] 6-1, Random walk sequence vectorization: input the random walk node sequence set obtained in the previous step into the Skip-Gram model to train the initial vector representation of each node, and the initial vector representation of the node in the 0th layer heterogeneous graph is ;

[0061] 6-1, neighbor node weight calculation: for each node , the importance weight of the first-order neighbor node of the node is calculated using the node's own feature and degree information, and the formula is as follows:

[0062]

[0063]

[0064] wherein is the importance weight of the node , is a learnable global parameter vector, is the vector representation of the node in the th layer, is a hyperparameter for balancing the contribution of features and degree information; by introducing the node degree information, the application can dynamically balance the semantic contribution and structural centrality of the node; at the same time, combined with the softmax normalization mechanism, the application can efficiently adaptively model the local graph structure;

[0065] 6-2, feature weighted aggregation: for each node , its own feature and the weighted features of all neighbor nodes are aggregated to obtain the feature of the node in the next layer heterogeneous graph, and the calculation formula is as follows:

[0066]

[0067] wherein is a nonlinear activation function, and are learnable parameter matrices, is the importance weight of the node ;

[0068] 6-3, hierarchical incremental feature fusion: the output features of the patent in all layers are weighted and summed to obtain , which is used as the classification score based on the graph structure, wherein For the number of categories, the formula is as follows:

[0069]

[0070] wherein represents the layer number where the final layer is located, is a weight coefficient corresponding to each layer vector, wherein and The hierarchical incremental feature fusion mechanism provided by the application effectively alleviates the over-smoothing problem while retaining the high-layer semantic abstraction ability and the low-layer local structure information, thereby improving the perception ability of the model to the multi-scale graph structure.

[0071] (7) Multi-modal classification score adaptive fusion: a multi-modal deep fusion method based on a gating mechanism is adopted to adaptively fuse the classification information of the text mode and the graph structure mode.

[0072] 7-1, Gating signal generation: first, the text score vector and the graph score vector are spliced, and a gating vector is generated through a full connection layer and a Sigmoid activation function, which is used to dynamically adjust the contribution weight of each mode on different categories, and the formula is as follows:

[0073]

[0074] wherein, represents a vector splicing operation, and are trainable parameter matrices and bias terms, is a Sigmoid activation function;

[0075] 7-2, Cross-modal feature interaction: first, an element-wise multiplication feature interaction term is used to capture the synergistic effect between the text and the graph structure features; then, the weighted modal features and the interaction features are integrated to obtain a fusion feature vector , and the formula is as follows:

[0076]

[0077] wherein, represents element-wise multiplication, is a trainable parameter matrix used for linear transformation of the interaction features;

[0078] 7-3, Final classification prediction: the fusion feature vector is input into a full connection layer and a Sigmoid activation function to obtain a patent final prediction probability score vector for all categories , the formula is as follows:

[0079]

[0080] wherein, and are trainable parameter matrices and bias terms;

[0081] (8) Model training and optimization: use multi-modal adversarial learning to calculate the loss function;

[0082] 8-1, text channel loss calculation: use text classification score to calculate binary cross-entropy loss , the formula is as follows:

[0083]

[0084] wherein and represent the number of training samples and the number of labels respectively, is the actual label value (0 or 1), is the probability value predicted using text information;

[0085] 8-2, graph channel loss calculation: use graph structure classification score to calculate binary cross-entropy loss , the formula is as follows:

[0086]

[0087] wherein and represent the number of training samples and the number of labels respectively, is the actual label value (0 or 1), is the probability value predicted using graph structure information;

[0088] 8-3, adaptive fusion loss calculation: use the final prediction probability score obtained after adaptive fusion to calculate binary cross-entropy loss , the formula is as follows:

[0089]

[0090] wherein and represent the number of training samples and the number of labels respectively, is the actual label value (0 or 1), is the probability value predicted by adaptive fusion of text information and graph structure information;

[0091] 8-4, multi-modal adversarial learning: combine , and The final loss function is calculated The fusion process is explicitly constrained by controlling the difference between the final fusion loss and the single-modal loss, as follows:

[0092]

[0093] wherein, is a weighting coefficient, is a tiny constant to prevent the denominator from being zero; the multi-modal adversarial learning proposed by the present application can prompt the fusion module to simultaneously focus on the information contribution of the text and graph channels during the optimization process, and is no longer dominated by a single mode, but maintains a dynamic balance in the loss space, thereby obtaining more discriminative fusion features and improving the discriminability and robustness of the model in complex contract scenarios;

[0094] 8-5, model optimization: using the Adam optimizer to update the model parameters, and training the entire network through a multi-label classification task;

[0095] (9) recommendation result output: according to the final classification score , the probability prediction result of the patent belongs to each technical category, and the top 10 IPC classification numbers with the highest prediction probability are output as the recommendation list.

[0096] The present application will further demonstrate the detailed implementation process and technical effects of the IPC classification number intelligent recommendation method based on multi-source information fusion shown in steps (1)~(9) above on specific data sets through a specific embodiment, so as to facilitate the understanding of the technical effects of the present application.

[0097] In another aspect of the present application, an IPC classification number intelligent recommendation system based on multi-source information fusion is provided, comprising the following modules:

[0098] A corpus construction module constructs a corpus based on patent text through a sorting algorithm;

[0099] A text classification module generates a vector representation of each word based on the text sequence in the corpus, and obtains a text classification score of the patent through a convolutional neural network;

[0100] A graph structure classification module: used to construct a heterogeneous graph, a random walk sequence set is obtained by a random walk driven by a multi-element structure, and a graph structure classification score is obtained through a classification network;

[0101] An IPC classification number recommendation module adopts multi-modal deep fusion based on a gating mechanism to adaptively fuse the text classification score and the graph structure classification score of the patent, obtains the final classification prediction score of the patent, obtains the probability prediction result of the patent belonging to each technical category, and outputs the recommendation list of the corresponding IPC classification number, and performs reverse training.

[0102] Embodiment

[0103] The steps of this embodiment are the same as the IPC classification number intelligent recommendation method based on multi-source information fusion shown in the foregoing steps (1)-(9), and will not be repeated here. This embodiment mainly demonstrates the specific data set, part of the specific parameter setting and implementation results. For ease of description, the method shown in steps (1)-(9) is referred to as the method of the present application, and the IPC classification number recommender used is referred to as MinPac.

[0104] Experimental data set: The data set used by the method of the present application contains 100,000 patent data obtained from the USPTO public database in the past five years. Each data covers key fields such as title, abstract, applicant, inventor and citation information, and uses the IPC code as the classification label for supervised learning, which is used to train and evaluate the IPC classification number intelligent recommendation model.

[0105] Evaluation index: precision@k (top-k precision), recall@k (top-k recall), F1@k (top-k F1 score) and NDCG@k (top-k normalized discounted cumulative gain). Precision@k focuses on the "accuracy" of the top-k items in the result list, recall@k measures how many total relevant items are captured in the top-k items, which focuses on the "recall rate", F1@k is the harmonic mean of precision@k and recall@k, which provides a single comprehensive index, and NDCG@k considers the ranking order quality in the result list. The higher the value of these indicators, the better the performance of the model. In the present application, k is 1, 3 and 5.

[0106] The specific experimental results of this embodiment are as follows:

[0107] (1) Algorithm comparison results

[0108] Table 1 Comparison of the present application with the current patent classification algorithm

[0109]

[0110] Table 1 is a comparison of the patent classification results of the present application and other comparative methods on the collected data set. From the summary results in Table 1, it can be seen that the present application has obvious advantages over other comparative methods.

[0111] (2) Ablation experiment results

[0112] Table 2 Effect of different information

[0113]

[0114] The w / o graph information in Table 2 represents removing the graph structure information, and the w / o text information represents removing the text information.

[0115] As can be seen from the results in the table, removing any kind of structure information will cause the model performance to decrease, among which removing the graph structure information causes a larger decrease in model performance, and removing the text information causes a smaller decrease in model performance.

[0116] Table 3 Effects of different metadata

[0117]

[0118] In Table 3, w / o inventor represents removing the inventor information, and w / o applicant represents removing the applicant information.

[0119] As can be seen from the results in the table, removing any kind of metadata information will cause the model performance to decrease, among which removing the inventor information causes a larger decrease in model performance, and removing the applicant information causes a smaller decrease in model performance.

Claims

1. A method for intelligent recommendation of IPC classification numbers based on multi-source information fusion, characterized in that, Includes the following steps: Step 1: Based on the patent text, construct a corpus using a ranking algorithm. Specifically, this involves: starting from the patent... The text content of the title, abstract, and description sections of the document is extracted. The TextRank algorithm, based on graph ranking, is used to automatically extract a predetermined number of semantically most important core sentences from the text. From these core sentences, a predetermined number of words with the highest information entropy are selected to constitute the patent. The text sequence was analyzed, and stop words were removed to construct a corpus; Step 2: Based on the text sequences in the corpus, generate a vector representation of each word, and obtain the text classification score of the patent through a convolutional neural network; Step 3: Construct a heterogeneous graph. Obtain a set of random walk sequences through a multi-structure-driven random walk. Obtain the graph structure classification score through a classification network. The specific implementation process is as follows: Construct a heterogeneous graph containing three types of nodes: patent nodes. Applicant Node And inventor node And three types of edges, namely the patent-applicant edge. Patent - Inventor Bian Patent-Patent Citation Edge ; Multi-structure driven random walk: Starting from all patent nodes in the heterogeneous graph, a multi-structure driven random walk strategy is performed to obtain a set of random walk sequences with the same number of elements as the number of patents. Initial representations for each node are generated using a set of random walk sequences; neighbor node features are aggregated using an aggregation method based on node weights and degree awareness; multi-level neighbor features of nodes are learned through a multi-layer network structure; and the patent is output using hierarchical incremental feature fusion. Graph structure classification score ; Step 4: Employ multimodal deep fusion based on a gating mechanism to adaptively fuse the text classification score and graph structure classification score of the patent to obtain the final classification prediction score of the patent. Obtain the probability prediction result of the patent belonging to each technology category and output the corresponding IPC classification number recommendation list, and perform reverse training.

2. The intelligent recommendation method for IPC classification numbers based on multi-source information fusion according to claim 1, characterized in that, The specific implementation process of step 2 is as follows: Word vector representation generation: The Word2Vec word vector generation method is used to generate word vector representations. The text sequence is used for training to generate a vector representation of each word, and all word vectors are concatenated into a two-dimensional matrix. ,in For vector dimensions, For the number of words; Text feature extraction and classification: extracting text matrices... The input is a convolutional neural network, which performs convolution operations using four sets of convolutional kernels of different sizes in parallel. Max pooling is then performed on the output of each convolutional kernel to extract salient features. These pooled salient features are concatenated and passed through a fully connected layer to output the patent. Text classification score ,in This represents the number of categories.

3. The intelligent recommendation method for IPC classification numbers based on multi-source information fusion according to claim 2, characterized in that, The specific implementation process of the multi-structure driven random walk is as follows: 5-1. Using patents A random walk is performed starting from the corresponding node; for the node at the current walk position... Calculate each of its neighbor nodes Corresponding node degree The node importance measurement algorithm outputs the PageRank value. Clustering coefficient Average degree of neighbors Construct the structural feature vector of each neighbor node. ; 5-2. The feature vector Input a linear mapping layer to obtain a structural embedding representation; 5-3. For the current node The embedding representations of all neighboring nodes are softmax normalized to obtain the transition probability of a neighboring node being selected as the next hop. ; 5-4. Repeat steps 5-1–5-3 until the preset walking length is reached, generating a patent. The corresponding node is the node sequence starting from the node; 5-5. Repeat steps 5-1–5-4, and perform random walks starting from all patent nodes in the heterogeneous graph to obtain a set of random walk sequences with the same number of elements as the number of patents.

4. The intelligent recommendation method for IPC classification numbers based on multi-source information fusion according to claim 3, characterized in that, The specific implementation of obtaining graph structure classification scores through a classification network is as follows: 6-1. Input the set of random walk node sequences into the Skip-Gram model to train and obtain the initial vector representation of each node. In the heterogeneous graph of layer 0, the nodes... The initial features are ; 6-1. For each node Calculate the first-order neighbor nodes using the node's own features and degree information. Importance weights; 6-2. For each node The node in the next layer of the heterogeneous graph is obtained by aggregating its own features with the weighted features of all its neighboring nodes. Feature Calculation ; 6-3. Patent The output features of all layers are weighted and summed to obtain As Classification scores based on graph structures, where This represents the number of categories.

5. The intelligent recommendation method for IPC classification numbers based on multi-source information fusion according to claim 4, characterized in that, The specific implementation process of the multimodal deep fusion based on the gating mechanism is as follows: 7-1. Convert the text score vector And graph fraction vector The concatenation is performed, and a gated vector is generated by passing the concatenation through a fully connected layer and a sigmoid activation function. ; 7-2. Using the feature interaction term of element-wise multiplication Capture the synergistic effect between text and graph structure features; then weight the modal features. , With interaction features Weighted integration is performed to obtain the fused feature vector. ; 7-3. Final classification prediction: The feature vectors will be fused. The patent is obtained through a fully connected layer and a Sigmoid activation function. The final predicted probability score vector for all categories .

6. The intelligent recommendation method for IPC classification numbers based on multi-source information fusion according to claim 5, characterized in that, The loss function used in step 3 for training includes: Text channel loss calculation: Calculate binary cross-entropy loss using text classification scores. ; Graph Channel Loss Calculation: Calculate binary cross-entropy loss using graph structure classification scores. ; Adaptive fusion loss calculation: Calculate the binary cross-entropy loss using the final predicted probability scores obtained after adaptive fusion. ; Multimodal adversarial learning: Combining , and Calculate the final loss function. By controlling the final fusion loss , With single-mode loss The difference explicit constraint fusion process. To prevent constants with a denominator of zero.

7. An intelligent recommendation system for IPC classification numbers based on multi-source information fusion, used to implement the intelligent recommendation method for IPC classification numbers as described in any one of claims 1 to 6, characterized in that, Includes the following modules: The corpus construction module builds a corpus based on patent texts using a sorting algorithm. The text classification module generates a vector representation of each word based on text sequences from a corpus, and obtains the patented text classification score through a convolutional neural network. Graph structure classification module: Constructs heterogeneous graphs, obtains a set of random walk sequences through multi-structure-driven random walks, and obtains graph structure classification scores through a classification network; The IPC classification number recommendation module adopts a multimodal deep fusion based on a gating mechanism to adaptively fuse the text classification score and graph structure classification score of the patent to obtain the final classification prediction score of the patent. It obtains the probability prediction result of the patent belonging to each technology category and outputs the corresponding IPC classification number recommendation list, and performs reverse training.

Images