A professional intelligence recommendation method and system

Abstract

This invention discloses a professional intelligence recommendation method, comprising the following steps: acquiring intelligence representations and intelligence agency representations; performing explicit high-order modeling of features based on a self-attention feature cross-network to obtain matching scores; and arranging the features from high to low according to the matching scores to generate a final intelligence recommendation list. The intelligence representation is derived from the fusion of intelligence label representations and intelligence content representations. The intelligence label representation is obtained using a label enhancement method based on graph convolutional neural networks; the intelligence content representation is obtained through pre-training a multimodal adaptive model based on diverse candidate intelligence data; and the intelligence agency representation is obtained through pre-training a multimodal alignment model based on intelligence data released by the intelligence agency itself. This invention can efficiently and accurately recommend relevant intelligence information to professional intelligence agencies, enhancing intelligence representation capabilities and effectively improving intelligence recommendation results.

Description

Technical Field

[0001] This invention relates to the field of professional intelligence recommendation, and in particular to a method and system for professional intelligence recommendation. Background Technology

[0002] With the development of the internet, professional intelligence is no longer limited to official documents and scientific papers, but also includes a large number of self-media articles and microblogs. Traditional scientific and technological literature provides comprehensive and professional knowledge references, while emerging online intelligence can provide timely and diverse information. However, due to the sheer volume and wide coverage of online intelligence, simple keyword searches and matching methods are insufficient to quickly and accurately retrieve the information needed by intelligence agencies.

[0003] Unlike product and news recommendations, professional intelligence recommendations present unique challenges. First, professional intelligence manifests in diverse forms, making it difficult to model uniformly using simple methods. This is especially true for online intelligence, which may include videos, images, text, or even just brief comments. Second, due to information security concerns, there is a lack of interaction information between intelligence agencies and historical intelligence. However, existing personalized recommendation algorithms primarily focus on mining user and item features from historical interaction records (clicks, purchases, likes, etc.), requiring ample historical interaction information. Existing recommendation algorithms are ill-suited for professional intelligence recommendations. For example:

[0004] Existing Solution 1 discloses a big data-based method for recommending scientific and technological intelligence. To improve the accuracy of recommendations, this method determines the recommendation type, target direction, and recommendation cycle based on the target account's search history and frequency. Specifically, it determines the target type based on the user's account attributes, then determines the primary recommendation type by combining the proportion of each search type in the user's search history. After determining the recommendation type, it clusters keywords from the search history to determine the recommendation direction. Finally, it determines the recommendation cycle corresponding to the content to be recommended based on the search frequency. This method primarily focuses on user-side information mining, such as account information, search history, and search frequency, but it doesn't address how to model massive amounts of candidate intelligence content. Furthermore, this method, which determines the recommendation format based on target type and target direction, is a coarse-grained recommendation method and does not achieve precise recommendations.

[0005] Existing Solution 2 discloses a method for constructing a knowledge graph and a recommendation method for human-readable threat intelligence. This method constructs a knowledge graph for recommending human-readable threat intelligence, abstracting entities and entity relationships in the knowledge graph into low-dimensional vector representations, and then trains a Long Short-Term Memory (LSTM) neural network for recommendation. Specifically, based on a set of human-readable threat intelligence, an LDA (a document topic generation model) subject model is used to construct a subject word set and generate a special entity set; based on these three datasets and their relationships, a knowledge graph for recommending human-readable threat intelligence is obtained; then, the TransE (an algorithm for representing the embedding representation of nodes and relationships in a graph structure) series of knowledge representation methods are used to obtain low-dimensional representations of entities and entity relationships; finally, an LSM neural network is trained based on intelligence vectors and user vectors for recommendation. This method obtains entity representations in the intelligence by constructing a knowledge graph, thereby further obtaining the representation of the intelligence and acquiring richer semantic information; however, this method cannot fully learn effective information. Summary of the Invention

[0006] To overcome the shortcomings of the existing technology, this invention proposes a professional intelligence recommendation method and system to solve the problem of inaccurate intelligence recommendations.

[0007] The technical problem of this invention is solved by the following technical solution:

[0008] A professional intelligence recommendation method includes the following steps:

[0009] S1. Acquisition of intelligence representation and intelligence agency representation;

[0010] S2. Based on a self-attention feature cross-network, perform explicit high-order modeling of the intelligence representation and intelligence agency representation;

[0011] S3. Generate an intelligence recommendation list based on the matching score output by the model;

[0012] In step S1,

[0013] The intelligence representation is derived from the fusion of intelligence tag representation and intelligence content representation;

[0014] The intelligence tag representation is obtained based on a tag enhancement method using graph convolutional neural networks;

[0015] The intelligence content representation is obtained through pre-training of a multimodal adaptive model based on diverse candidate intelligence data.

[0016] The intelligence agency representation is obtained through pre-training of a multimodal alignment model based on intelligence data released by the intelligence agency itself.

[0017] In some embodiments, the following technical features are also included:

[0018] Step S1 further includes: modeling the query term-intelligence-label as a tripartite graph, transforming the task of labeling intelligence into a link prediction problem between intelligence nodes and label nodes on the graph; and propagating all information in the tripartite graph based on the graph convolutional neural network, so that intelligence nodes and label nodes can fuse neighbor information to obtain better representation.

[0019] In the aforementioned tag enhancement method, similarity is calculated based on the representation of intelligence and tags, and one or more tags with the highest similarity are used as new tags for that intelligence.

[0020] The multimodal includes image modality and text modality.

[0021] The multimodal alignment model uses a Co-Attentional Transformer module to fuse information from both image and text modalities; both the text and image sides obtain representations that incorporate other modalities; and the representations output from these two sides are spliced ​​together to obtain the final intelligence agency representation.

[0022] The multimodal alignment model includes two parallel Transformer models operating on images and text respectively; feature fusion between the image and text modalities is accomplished using a Co-Attention module; the text content in the intelligence is fed into the Transformer encoder via an embedding layer to extract contextual information, resulting in a text representation TeE; for the image portion of the intelligence, the image is first divided into blocks and then flattened, with each image block flattened into a one-dimensional vector; then, a linear transformation is performed on each vector, and the resulting vector is input into the Transformer encoder to obtain the image representation PiE; the Co-Attentional Transformer module is used to fuse the information from the two modalities; the image-side Co-Attentional Transformer module uses the text representation TeE to obtain the Key and Value matrices, while the Query matrix is ​​generated from the image representation PiE; the text-side Co-Attentional Transformer module's Key and Value matrices originate from the image representation PiE, and the Query matrix originates from the text representation TeE.

[0023] In step S1, the multimodal adaptive model learns the intelligence content representation, and at the same time, through pre-training technology, the multimodal adaptive model can extract effective information based on large-scale data, enabling the downstream model to converge quickly.

[0024] The multimodal adaptive model includes a policy network.

[0025] This invention also provides the following technical solutions:

[0026] A professional intelligence recommendation system includes a processor and a memory, wherein the memory stores a computer program that can be executed by the processor to implement the aforementioned method.

[0027] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0028] The beneficial effects of this invention compared to the prior art include:

[0029] The professional intelligence recommendation method provided by this invention, based on intelligence data released by the intelligence agency itself, can fully learn effective information through pre-training of a multimodal alignment model, and thus can accurately recommend relevant intelligence information to professional intelligence agencies.

[0030] In addition, some embodiments also have the following beneficial effects:

[0031] This invention provides a label enhancement method based on graph convolutional neural networks, which is beneficial for enhancing intelligence representation capabilities;

[0032] The embodiments of the present invention design a multimodal alignment model and a multimodal adaptive model to obtain the representation of intelligence agencies and the representation of intelligence content, which effectively improves the intelligence recommendation effect.

[0033] Other beneficial effects of the embodiments of the present invention will be further described below. Attached Figure Description

[0034] Figure 1 This is a flowchart of a professional intelligence recommendation method in an embodiment of the present invention. Detailed Implementation

[0035] Before introducing the specific implementation methods, the concept of the present invention is described as follows:

[0036] Replacing manual search with "recommendation" is a more practical and convenient intelligence acquisition solution. Therefore, this method focuses on how to efficiently and accurately recommend relevant intelligence information to professional intelligence agencies. Addressing the unique challenges of the intelligence field and the shortcomings of existing recommendation algorithms, this invention proposes a professional intelligence recommendation method based on multimodal pre-training and label enhancement. First, a label enhancement technique based on GCN (Graph Convolutional Neural Network) is used to match labels to intelligence, enabling intelligence with partially or completely missing labels to obtain corresponding label sets, which is beneficial for enhancing intelligence representation capabilities. Furthermore, given the extremely sparse interactive data in the intelligence field, most existing recommendation methods are not suitable for this scenario. Therefore, this solution employs multimodal pre-training to learn the representations of intelligence agencies and intelligence content, allowing it to fully learn effective information. There are significant differences in the formats of reports issued by intelligence agencies and the diverse sources of online intelligence candidates; reports issued by intelligence agencies have standardized writing structures and language, while online intelligence may lack image information or rich text information. Therefore, a multimodal alignment model and a multimodal adaptive model are designed to obtain the representations of intelligence agencies and intelligence content, respectively. The pre-trained intelligence agency representation and intelligence content representation are input into a feature cross-network based on a self-attention mechanism to explicitly model higher-order feature cross-interactions, ultimately obtaining a more accurate matching score. This involves the following specific considerations:

[0037] The idea of ​​domain tag enhancement for intelligence recommendation and the method of GCN tag enhancement

[0038] Tags can effectively characterize the key content of intelligence. However, in the field of intelligence recommendation, few people pay attention to the problem of incomplete or missing intelligence tags. This invention is the first to naturally model the user's query terms, intelligence, and tags into a ternary graph, transforming the tag prediction problem into a link prediction problem of nodes on the graph. It then utilizes GCN technology to propagate information on the graph, allowing various nodes to aggregate information from neighboring nodes and acquire rich semantic information. This completes the task of matching tags to intelligence.

[0039] Multimodal pre-training ideas in the field of intelligence recommendation

[0040] Intelligence content often contains information across multiple modalities, but existing intelligence recommendation methods only use single-modal information to represent intelligence, resulting in the underutilization of much valuable information within the intelligence. To address this, this invention designs a multimodal method for modeling intelligence content. Furthermore, due to the sparsity of interactive data in the intelligence domain, directly training the model based on the collected training set would lead to insufficient learning of the representations of intelligence content and intelligence agencies. Therefore, this invention employs a pre-training approach to first learn the representations of intelligence agencies and intelligence content.

[0041] The idea of ​​multimodal alignment in the intelligence field

[0042] In addition to using some basic information to characterize intelligence agencies, this invention also uses intelligence data released by the intelligence agencies themselves to characterize them. Because the intelligence data released by intelligence agencies has a standardized and uniform format, a multimodal alignment method is designed to extract information from it for characterizing the intelligence agencies.

[0043] The concept of multimodal adaptation in the intelligence field

[0044] This invention addresses the issue that candidate intelligence datasets come from diverse sources and lack a unified format. This is significantly different from the point mentioned in point three above, and using a single model for processing would inevitably compromise model performance. Therefore, this invention designs a multimodal adaptive method for representation learning of user candidate intelligence datasets.

[0045] In light of the above considerations, this invention proposes a specialized intelligence recommendation method to address the challenges in intelligence recommendation, including multimodal pre-training methods, label enhancement methods, and self-attention-based feature cross-networks. To alleviate the problem of sparse interactive data in the intelligence field, embodiments of this invention utilize intelligence data released by intelligence agencies to construct intelligence agency representations, while diverse intelligence data from external sources utilizes its multimodal information to construct intelligence content representations. Simultaneously, intelligence data released by intelligence agencies has a standardized format, while candidate intelligence sources are diverse and lack a unified format. Therefore, embodiments of this invention design multimodal alignment pre-training models and multimodal adaptation pre-training models for learning intelligence agency and intelligence content representations, respectively. Finally, a self-attention-based feature cross-network is used to explicitly perform high-order modeling of features to obtain accurate matching scores. These scores are then ranked from high to low to generate the final intelligence recommendation list.

[0046] The present invention will be further described below with reference to the accompanying drawings and preferred embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other.

[0047] In the description of the embodiments of the present invention, "multiple" means two or more, unless otherwise explicitly specified.

[0048] This invention provides a specialized information recommendation method, such as... Figure 1 As shown, it includes the following steps:

[0049] S1. A label enhancement method based on graph convolutional neural networks is used to obtain intelligence label representations; based on diverse candidate intelligence data, intelligence content representations are obtained through pre-training of a multimodal adaptive model; based on intelligence data released by the intelligence agency itself, intelligence agency representations are obtained through pre-training of a multimodal alignment model; intelligence label representations and intelligence content representations are fused to obtain intelligence representations.

[0050] S2. Based on a self-attention feature cross-network, perform explicit high-order modeling of the intelligence representation and intelligence agency representation;

[0051] S3. Generate an intelligence recommendation list based on the matching score output by the model.

[0052] The following technical methods were used:

[0053] GCN-based tag enhancement technology

[0054] Tags are typically defined as keywords describing key information about an item, such as category, style, target audience, and related entities. In many industrial applications, item tags are a key element for better organizing items for easier recommendation. In the intelligence field, tags are used to concisely describe key intelligence information, such as industry and issuing organization. Item tag generation often employs time-consuming and inefficient methods such as manual labeling. To replace or supplement manual labeling methods, much research has been conducted on methods for automatically generating item recommendation tags. These methods for generating item tags can be broadly categorized into two types: keyword extraction and multi-tag classification. Keyword extraction methods, such as TF-IDF (a commonly used weighted technique for information retrieval and data mining), TextRank (a graph-based ranking algorithm for text), and PositionRank (an unsupervised key phrase extraction model for academic documents), are widely used in text documents or websites. These methods identify keywords that best describe the document's topic from the original content, primarily following two stages: candidate tag extraction and tag ranking. Since tags may not appear in the item description, they are suitable for items with long textual descriptions but not for items without detailed textual descriptions. Due to the limitations of keyword extraction methods, item label recommendation is increasingly modeled as a multi-label classification problem. Multi-label classification models have been extensively studied in the literature, and many have been successfully applied to text classification tasks. However, directly applying traditional multi-label classification models to item label recommendation is not optimal, especially in intelligence recommendation tasks. Specifically, existing methods suffer from three main problems: First, most traditional multi-label classification models cannot fully utilize the correlation between labels. Second, intelligence descriptions are often brief and noisy, making it difficult to extract high-quality semantic information for classification. Third, in reality, some new intelligence lacks any existing labels and requires complete label prediction; some older intelligence has partially incomplete labels and only requires label completion. Existing models cannot uniformly handle these two situations.

[0055] Because intelligence tags play a crucial role in intelligence recommendation tasks, and existing methods for generating corresponding tags for intelligence have the aforementioned shortcomings, this invention proposes a tag enhancement technique based on GCN (Graph Convolutional Neural Network) for obtaining intelligence tag representations in step S1. The query term-intelligence-tag structure can be naturally modeled as a ternary graph, thus transforming the task of tagging intelligence into a link prediction problem between intelligence nodes and tag nodes in the graph. Specifically, GCN is used to propagate all information in the graph, enabling intelligence nodes and tag nodes to fuse neighbor information to obtain better representations. Finally, similarity is calculated based on the representations of intelligence and tags, and one or more tags with the highest similarity are used as new tags for that intelligence.

[0056] Pre-training methods for multimodal alignment

[0057] In step S1 of this embodiment, the intelligence agency's representation is obtained based on intelligence data released by the intelligence agency itself, using a multimodal alignment pre-training method. Since intelligence content released by the intelligence agency itself generally has a standardized format, the multimodal alignment method allows information from multiple modalities to complement and enhance each other. The multimodal alignment model in this scheme mainly consists of two parallel Transformer models (a neural network architecture based on a self-attention mechanism) acting on the image and text respectively. Feature fusion of the two modalities is accomplished using a Co-Attention module. Specifically, the text content in the intelligence is fed into the Transformer encoder after passing through the Embedding layer to extract contextual information, resulting in the text representation TeE. For the image portion of the intelligence, the image is first divided into blocks and then flattened, with each image block flattened into a one-dimensional vector; then, a linear transformation is performed on each vector, and the resulting vector is input into the Transformer encoder to obtain the image representation PiE. Subsequently, the Co-Attentional Transformer module is used to fuse the information from the two modalities. Specifically, the image-side Co-Attentional Transformer module uses the text representation TeE to obtain the Key and Value matrices, while the Query matrix is ​​generated from the image representation PiE. Correspondingly, the Key and Value matrices of the text-side Co-Attentional Transformer module originate from the image representation PiE, and the Query matrix originates from the text representation TeE. After passing through the Co-Attentional Transformer module, both the text and image sides obtain representations that incorporate other modalities. The representations output from both sides are then concatenated to obtain the final intelligence agency representation.

[0058] Multimodal adaptive pre-training method

[0059] Compared to news recommendation and item recommendation, the intelligence domain faces a more pronounced problem of sparse interactive data. This is one of the reasons why existing recommendation methods fail to achieve good results in the intelligence domain. To alleviate this problem, step S1 of this embodiment proposes a multimodal adaptive pre-training method to obtain intelligence content representations, thereby improving recommendation performance. When learning intelligence content representations, multimodal adaptive techniques are employed due to the diverse sources and content of intelligence. Simultaneously, the pre-training technique enables the model to extract effective information based on large-scale data, allowing downstream models to converge quickly. The multimodal adaptive model differs from the multimodal alignment model in that it adds a lightweight policy network at the very beginning of the process. This policy network determines which modal information is selected for subsequent operations. Otherwise, the processes of the two models are identical.

[0060] In other embodiments, the difference from the foregoing embodiments is that,

[0061] Multimodal is not limited to text and image modalities;

[0062] Tag enhancement techniques are not limited to the query-information-tag tripartite graph; they can be a two-part subgraph or other reasonable tripartite graphs.

[0063] Feature cross-networks are not limited to a specific number of layers and neurons; they should be adjusted according to the specific dataset and scenario.

[0064] The image encoding module is not limited to a specific number of segments and can vary depending on the specific dataset.

[0065] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0066] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0067] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0068] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0069] The above description provides a further detailed explanation of the present invention in conjunction with specific / preferred embodiments, and it should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various substitutions or modifications can be made to these described embodiments without departing from the concept of the present invention, and all such substitutions or modifications should be considered within the scope of protection of the present invention. In the description of this specification, the reference to terms such as "an embodiment," "some embodiments," "preferred embodiment," "example," "specific example," or "some examples," etc., indicates that the specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples. Although the embodiments of the present invention and their advantages have been described in detail, it should be understood that various changes, substitutions, and modifications can be made herein without departing from the scope of protection of the patent application.

Claims

1. A professional information recommendation method characterized by comprising: The method comprises the following steps: S1, obtaining intelligence representation and intelligence agency representation; S2, performing explicit high-order modeling on the intelligence representation and intelligence agency representation based on a self-attention feature cross network; S3, generating an intelligence recommendation list according to a matching score output by the model; In step S1, The intelligence representation is obtained by fusing intelligence label representation and intelligence content representation; The intelligence label representation is obtained based on a label enhancement method of a graph convolutional neural network; The intelligence content representation is obtained based on candidate diversified intelligence data and by pre-training a multi-modal adaptive model; The intelligence agency representation is obtained based on intelligence data published by an intelligence agency itself and by pre-training a multi-modal alignment model; the multi-modal includes an image modal and a text modal; the multi-modal alignment model includes two parallel Transformer models respectively acting on the image and the text; feature fusion of the two modalities of the image and the text is completed by using a Co-Attention module; the text content in the intelligence is sent to an encoder of the Transformer through an embedding layer to extract context information, and a text representation TeE is obtained; for the image part in the intelligence, the image is first divided into blocks and then flattened, and each picture block is flattened into a one-dimensional vector; then, each vector is linearly transformed, and the obtained vector is input into the encoder of the Transformer to obtain an image representation PiE; the information of the two modalities is fused by using a Co-Attentional Transformer module; the Co-Attentional Transformer module on the image side obtains Key and Value matrices by using the text representation TeE, and a Query matrix is generated from the image representation PiE; the Key and Value matrices of the Co-Attentional Transformer module on the text side are derived from the image representation PiE, and the Query matrix is derived from the text representation TeE; the multi-modal adaptive model is initially added with a light strategy network in the multi-modal alignment model, and the strategy network is used to determine which modal information is selected for subsequent operation.

2. The professional information recommendation method according to Claim 1, wherein, Step S1 further comprises: modeling the query word-intelligence-label as a tripartite graph, and converting the intelligence labeling task into a link prediction problem between intelligence nodes and label nodes on the graph; all information in the tripartite graph is propagated based on the graph convolutional neural network, so that the intelligence nodes and the label nodes can fuse neighbor information to obtain better representation.

3. The professional information recommendation method according to Claim 1, wherein In the label enhancement method, similarity calculation is performed based on the representations of intelligence and labels, and one or more labels with the highest similarity are taken as new labels of the intelligence.

4. The professional information recommendation method according to Claim 1, wherein The multi-modal alignment model fuses the information of the two modalities of the image and the text by using a Co-Attentional Transformer module; the text and the image side each obtain a representation fused with the other modality; the representations output from the two sides are spliced to obtain the final intelligence agency representation.

5. The professional information recommendation method according to Claim 1, wherein In step S1, the multi-modal adaptive model learns intelligence content representation, and through pre-training technology, the multi-modal adaptive model is enabled to extract effective information based on large-scale data, so that the downstream model can quickly converge.

6. The professional information recommendation method according to Claim 5, wherein The multi-modal adaptive model comprises a policy network.

7. A professional information recommendation system comprising a processor and a memory having stored therein a computer program, characterized in that, The computer program can be executed by a processor to implement the method of any one of claims 1 to 6.

8. A computer-readable storage medium storing a computer program, the computer-readable storage medium comprising: The computer program is executed by a processor to implement the method of any one of claims 1 to 6.

Images