CN115269984A

CN115269984A - Professional information recommendation method and system

Info

Publication number: CN115269984A
Application number: CN202210900591.2A
Authority: CN
Inventors: 肖喜; 李毅; 苏良才; 夏树涛; 江勇; 郑海涛
Original assignee: Shenzhen International Graduate School of Tsinghua University
Current assignee: Shenzhen International Graduate School of Tsinghua University
Priority date: 2022-07-28
Filing date: 2022-07-28
Publication date: 2022-11-01

Abstract

The invention discloses a professional information recommendation method, which comprises the following steps: acquiring an emotional statement and an intelligence mechanism representation; based on a feature cross network of self attention, performing high-order modeling on the feature to obtain a matching score; arranging according to the matching scores from high to low so as to generate a final information recommendation list, wherein the information representation is formed by fusing information label representation and information content representation; the intelligence label representation is obtained by a label enhancement method based on a graph convolution neural network; the information content representation is obtained through multi-mode adaptive model pre-training based on candidate diversified information data; the representation of the intelligence institution is obtained through the pre-training of the multi-mode alignment model based on the intelligence data issued by the intelligence institution. The invention can efficiently and accurately recommend corresponding information to professional information institutions, is beneficial to enhancing the information representation capability, and effectively improves the information recommendation effect.

Description

Professional information recommendation method and system

Technical Field

The invention relates to the field of professional information recommendation, in particular to a professional information recommendation method and system.

Background

With the development of the internet, professional intelligence is no longer limited to official reports and scientific papers, but also comprises a large number of self-media articles, microblogs and the like. Traditional scientific and technical literature provides comprehensive and professional knowledge reference, and emerging network information can provide timely and multivariate information. However, the amount of network information is large and the coverage is wide, and it is difficult to quickly and accurately retrieve information required by the information organization by means of simple keyword retrieval, matching and the like.

Unlike the scenarios of merchandise, news recommendations, etc., professional intelligence recommendations have unique challenges. First, professional information is expressed in various forms, and it is difficult to uniformly model in a simple manner. Especially, the network information may include videos, pictures, characters, etc., or may be just comment information with few characters. On the other hand, interactive information between the intelligence organization and the history intelligence is lacking because of information security and the like. However, existing personalized recommendation algorithms mainly focus on mining the characteristics of users and items from their historical interaction records (clicks, purchases, likes, etc.), and require sufficient historical interaction information. The existing recommendation algorithm is difficult to be suitable for professional intelligence recommendation. For example:

in order to improve accuracy of scientific and technological information recommendation, a recommendation type, a target direction and a recommendation period of scientific and technological information are determined from a retrieval history and a retrieval frequency of a target account. Specifically, a target type of the user is determined based on the account attribute of the user, and then a main recommendation type is determined by combining the proportion of each retrieval type in the user retrieval history; after the recommendation type is determined, clustering the keywords in the retrieval history, and determining the recommendation direction; and finally, determining a recommendation period corresponding to the content to be recommended according to the retrieval frequency. The method mainly focuses on information mining of the user side, for example, account information, retrieval history and retrieval frequency are all related information of the user side, and no relation is provided for how massive candidate information contents are modeled. Meanwhile, the method is a coarse-grained recommendation method by determining the recommendation form of the target type and the target direction, and accurate recommendation is not achieved.

The second existing scheme discloses a knowledge graph construction method and a recommendation method for human-read threat information. Specifically, based on a human-read threat intelligence set, a main word set is constructed by using an LDA (document theme generation model) main model and a special entity set is generated; obtaining a knowledge graph for recommending human read threat information based on the three data sets and the relation of the three data sets; then adopting a TransE (an algorithm for representing embedded representation of nodes and relations in a graph structure) series knowledge representation method to obtain low-dimensional representation of entities and entity relations; and finally training a long-short term memory neural network for recommendation based on the intelligence vector and the user vector. The method obtains entity representation in the information in a mode of constructing the knowledge graph so as to further obtain the representation of the information, and richer semantic information can be obtained, but the method cannot fully learn effective information.

Disclosure of Invention

In order to make up for the defects of the prior art, the invention provides a professional information recommendation method and system to solve the problem of inaccurate information recommendation.

The technical problem of the invention is solved by the following technical scheme:

a professional information recommendation method comprises the following steps:

s1, acquiring an information representation and an information mechanism representation;

s2, performing high-order modeling on the intelligence representation and the intelligence mechanism representation based on a feature cross network of self attention;

s3, generating an intelligence recommendation list according to the matching scores output by the model;

in the step S1, the first step is performed,

the intelligence representation is formed by fusing intelligence label representation and intelligence content representation;

the intelligence label representation is obtained by a label enhancement method based on a graph convolution neural network;

the information content representation is obtained through multi-mode adaptive model pre-training based on candidate diversified information data;

the intelligence agency representation is obtained through multi-mode alignment model pre-training based on intelligence data issued by the intelligence agency.

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

the step S1 further includes: modeling the query word-information-label into a three-part graph, and converting an information labeling task into a link prediction problem between information nodes and label nodes on the graph; all information in the three-part graph is propagated based on the graph convolution neural network, so that the intelligence node and the label node can be fused with neighbor information to obtain better representation.

In the label enhancement method, similarity calculation is carried out based on the information and the representation of the label, and one or more labels with the highest similarity are used as the newly added label of the information.

The multiple modalities include an image modality and a text modality.

The multi-mode alignment model fuses information of two modes of an image and a text by utilizing a Co-extensive Transformer module; the text and the image side respectively obtain the representation fused with other modes; and splicing the representations output from the two sides to obtain the final representation of the intelligence mechanism.

The multi-mode alignment model comprises two parallel transform models which respectively act on an image and a text; the feature fusion of the image and the text is completed by utilizing a Co-Attention module; text content in the information is sent to a Transformer encoder through an embedded layer to extract context information, and a text representation TeE is obtained; for an image part in the information, firstly partitioning the image into blocks, flattening the image, and flattening each image block into a one-dimensional vector; then, carrying out linear transformation on each vector, inputting the obtained vector into a Transformer encoder, and obtaining the representation PiE of the image; fusing the information of the two modes by utilizing a Co-extensive Transformer module; a Co-extensive Transformer module on the image side obtains Key and Value matrixes by using text representation TeE, and a Query matrix is generated by image representation PiE; the Key and Value matrixes of the Co-extensive Transformer module on the text side are derived from image representation PiE, and the Query matrixes are derived from text representation TeE.

In the step S1, the multi-modal adaptive model learns information content representation, and meanwhile, effective information can be extracted from the multi-modal adaptive model based on large-scale data through a pre-training technology, so that a downstream model can be rapidly converged.

The multi-modal adaptive model includes a policy network.

The invention also provides the following technical scheme:

a professional intelligence recommendation system comprising a processor and a memory, the memory having stored therein a computer program executable by the processor to perform the method as described above.

A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the aforementioned method.

Compared with the prior art, the invention has the advantages that:

according to the professional information recommendation method provided by the invention, because effective information can be fully learned through multi-mode alignment model pre-training based on information data issued by an information institution, corresponding information can be accurately recommended to the professional information institution.

In addition, in some embodiments, the following beneficial effects are also achieved:

the embodiment of the invention provides a label enhancement method based on a graph convolution neural network, which is beneficial to enhancing intelligence representation capability;

the embodiment of the invention respectively designs the multi-mode alignment model and the multi-mode adaptive model to obtain the information mechanism representation and the information content representation, thereby effectively improving the information recommendation effect.

Other advantages of embodiments of the present invention will be further described below.

Drawings

FIG. 1 is a flowchart of a method for recommending professional intelligence according to an embodiment of the present invention.

Detailed Description

Before describing the embodiments, the idea of the present invention is described as follows:

replacing human search with "recommendations" is a more practical and convenient intelligence acquisition solution. Therefore, the method aims at efficiently and accurately recommending corresponding information for professional information institutions. Aiming at unique challenges in the information field and the defects of the existing recommendation algorithm, the invention provides a professional information recommendation method based on multi-mode pre-training and label enhancement. Firstly, the label enhancement technology based on GCN (graph convolution neural network) is used for matching labels for intelligence, so that the intelligence with partially or completely missing labels obtains a corresponding label set, which is beneficial to the enhancement of intelligence characterization capability. In addition, most of the existing recommendation methods are not suitable for the situation in which the interactive data in the intelligence field is extremely sparse. Therefore, the scheme adopts a multi-mode pre-training mode to learn the representation of the intelligence mechanism and the intelligence content, so that the intelligence mechanism and the intelligence content can fully learn the effective information. The report issued by the intelligence organization has obvious difference with the network intelligence format with various sources, namely the report line structure and the language issued by the intelligence organization have standard format, and the network intelligence has the condition of missing picture information or rich text information. Therefore, a multimodal alignment model and a multimodal adaptive model are respectively designed to obtain the intelligence agency representation and the intelligence content representation. And inputting the information mechanism representation and the information content representation obtained by pre-training into a feature crossing network based on a self-attention mechanism, modeling high-order feature crossing in a display manner, and finally obtaining a more accurate matching score. Among these, some specific considerations are involved:

idea for label enhancement in information recommendation field and GCN label enhancement method

The label can effectively depict the key content of the intelligence. However, in the field of information recommendation, people pay attention to the problem that the information label is incomplete or missing. The invention firstly models the query words, the intelligence and the labels of the user into a three-part graph naturally, and converts the label prediction problem into the link prediction problem of the nodes on the graph. And the GCN technology is utilized to carry out information transmission on the graph, so that various nodes gather information of adjacent nodes and have rich semantic information. Thereby completing the task of matching tags for intelligence.

Multi-modal pre-training philosophy in the field of intelligence recommendation

The content of the intelligence often contains information of a plurality of modes, but the existing intelligence recommendation methods only use information of a single mode to represent the intelligence, so that much valuable information in the intelligence is not well utilized. Therefore, the invention designs a multi-modal method for modeling the intelligence content. In addition, because of the problem of interactive data and sparseness in the information field, training the model directly based on the collected training set can make the information content and the characterization learning of the information organization insufficient. Therefore, the invention adopts a pre-training mode to firstly learn the representation of the intelligence organization and the intelligence content.

Idea of multimodal alignment in the field of intelligence

The invention not only uses some basic information to represent the intelligence organization, but also uses the intelligence data issued by the intelligence organization to represent the intelligence organization. As the intelligence data issued by the intelligence agency has the characteristic of uniform format specification, a multi-mode alignment method is designed for extracting the information in the information for representing the intelligence agency.

Idea of multimodal adaptation in the intelligence domain

The invention focuses on the fact that the sources of candidate information data sets are various and the candidate information data sets do not have uniform formats. This is clearly a great difference from the third point mentioned above, which inevitably compromises model performance if the same model is used for processing. Therefore, the invention designs a multi-modal adaptive method for the characterization learning of the user candidate intelligence data set.

In view of the above considerations, the present invention provides a special intelligence recommendation method to solve the challenges in intelligence recommendation, including a multi-modal pre-training method, a tag enhancement method, and a feature crossing network based on self-attention. In order to solve the problem of sparse interactive data in the information field, the embodiment of the invention utilizes the information data issued by the information mechanism to construct the information mechanism representation, and the information data with diversified external sources utilizes the multimodal information to construct the information content representation. Meanwhile, the information data issued by the information organization has a standard format, and the candidate information has various sources and does not have a uniform format. Therefore, the embodiment of the invention respectively designs a multi-mode aligned pre-training model and a multi-mode adaptive pre-training model for learning of intelligence institutions and intelligence content representation. Finally, performing high-order modeling on the features through a feature cross network based on self attention to obtain an accurate matching score. Ranking from high to low matching scores to produce a final intelligence recommendation list.

The invention will be further described with reference to the drawings and preferred embodiments. It should be noted that the embodiments and features of the embodiments in the present application may be combined with each other without conflict.

In the description of the embodiments of the present invention, "a plurality" means two or more unless specifically limited otherwise.

The embodiment of the invention provides a special information recommendation method, as shown in fig. 1, comprising the following steps:

s1, obtaining an intelligence tag representation by a tag enhancement method based on a graph convolution neural network; acquiring information content representation through multi-mode adaptive model pre-training based on candidate diversified information data; obtaining the representation of the information mechanism through the multi-mode alignment model pre-training based on the information data issued by the information mechanism; the information tag representation and the information content representation are fused to obtain an information representation;

and S3, generating an intelligence recommendation list according to the matching scores output by the model.

Wherein, the following technical means are used:

GCN-based label enhancement technology

A tag is generally defined as a key that describes key information for an item. Such as categories, styles, target audiences, related entities, etc. In many industrial applications, item tagging is a key element in better organizing items for recommendation. In the intelligence field, tags are used to briefly describe intelligence critical information, such as: industry, distribution agencies, etc. Label generation of articles often employs time consuming and inefficient methods such as manual labeling. In order to replace or supplement the manual labeling method, many researches on a method for automatically generating an item recommendation label have been carried out. These methods of generating labels for articles can be broadly divided into two categories: keyword extraction and multi-label classification. Keyword extraction methods, such as TF-IDF (a common weighting technique for information retrieval and data mining), textRank (a graph-based ranking algorithm for text), posionrank (an unsupervised key phrase extraction model for academic documents), have been widely used for text documents or websites, and identify keywords that best describe the topic of the document from the original content, mainly following two stages, i.e., candidate tag extraction and tag ranking. Since labels may not be present in the article description, they are suitable for articles with longer text, but not for articles without detailed text. Due to the shortcomings of keyword extraction methods, the tag recommendation task for an article is more modeled as a multi-tag classification problem. Multi-label classification models have been extensively studied in the literature, many of which have been successfully applied to text classification tasks. However, applying the conventional multi-label classification model directly to label recommendation of an item is not an optimal solution, especially in intelligence recommendation tasks. Specifically, the existing method has the following three problems. First, most conventional multi-label classification models fail to fully exploit the correlation between labels. Second, the related description of intelligence is usually short and noisy, and it is difficult to extract high-quality semantic information from it for classification. Thirdly, in practical situations, some new information does not have any existing label, and needs to be subjected to complete label prediction; some old information has partially incomplete labels and only needs to be completed. Existing models cannot uniformly handle both cases.

Because the intelligence tag plays an important role in the intelligence recommendation task, and the existing method for generating the corresponding tag for intelligence has the above-mentioned disadvantages, the embodiment of the present invention provides a tag enhancement technology based on GCN (graph convolutional neural network) for obtaining the intelligence tag representation in step S1. The query term-intelligence-tag can be naturally modeled as a three-part graph, thereby converting the intelligence-tagging task into a link prediction problem between intelligence nodes and tag nodes on the graph. In particular, all information in the graph is propagated based on the GCN, so that both informative nodes and tag nodes can fuse neighbor information for better characterization. And finally, carrying out similarity calculation based on the intelligence and the representation of the label, wherein one or more labels with the highest similarity are used as new labels of the intelligence.

Multi-modal alignment pre-training method

In step S1 of the embodiment of the present invention, a multi-mode alignment pre-training method is used to obtain a representation of an intelligence organization based on intelligence data issued by the intelligence organization itself. Most of the intelligence contents issued by the intelligence organization have a standard format, so that the multi-mode alignment method can mutually supplement and enhance the information before a plurality of modes. The multi-modal alignment model body in the scheme is composed of two parallel transform (a neural network architecture based on a self-attention mechanism) models which respectively act on images and texts. The feature fusion of the two modes is completed by using a Co-Attention module. Specifically, text content in the information passes through an Embedding layer (Embedding layer) and then is sent to a Transformer encoder to extract context information, and text representation TeE is obtained. For the image part in the intelligence, firstly partitioning the image, and then flattening, wherein each picture block is flattened into a one-dimensional vector; and then, performing linear transformation on each vector, inputting the obtained vector into a Transformer encoder, and obtaining the representation PiE of the image. Then, the information of the two modalities is fused by using a Co-extensive Transformer module. Specifically, the image-side Co-extensive Transformer module obtains Key and Value matrixes by using the text representation TeE, and the Query matrix is generated by the image representation PiE. Correspondingly, the Key and Value matrixes of the Co-extensive Transformer module on the text side are derived from the image representation PiE, and the Query matrix is derived from the text representation TeE. Through the Co-extensive Transformer module, the text side and the image side respectively obtain the representation fused with other modes. And splicing the representations output from the two sides to obtain the final representation of the intelligence mechanism.

Multi-modal adaptive pre-training method

Compared with the fields such as news recommendation and article recommendation, the intelligence field has more prominent problem of sparse interactive data. This is one of the reasons why the existing recommendation method cannot achieve a good effect in the intelligence field. In order to alleviate the problem, the embodiment of the present invention provides a multi-modal adaptive pre-training method in step S1 to obtain information content representation, so as to improve the recommendation effect. When the information content representation is learned, the multimodal self-adaptive technology is adopted for learning due to the fact that information sources are various and contents are diversified. Meanwhile, the pre-training technology enables the model to extract effective information based on large-scale data, so that the downstream model can be converged quickly. The multi-modal adaptive model differs from the multi-modal alignment model in that it adds a lightweight policy network at the beginning of the procedure. The policy network is used for judging which selected modal information is used for subsequent operation. In addition, the flows of the two models are consistent.

In other embodiments, the difference from the previous embodiments is that,

multimodal is not limited to both text and images modalities;

the label enhancement technology is not limited to the query word-intelligence-label three-part graph, and can be a two-part graph or other reasonable three-part graphs;

the characteristic cross network is not limited to a specific number of layers and the number of neurons, and is adjusted according to a specific data set and a scene;

the image encoding module is not limited to a specific number of segmentation blocks, and may be different according to a specific data set.

As will be appreciated by one skilled in the art, embodiments of the present invention may be provided as a method, system, or computer program product. Accordingly, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present invention may 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, and the like) having computer-usable program code embodied therein.

The present invention has been 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 flow and/or block of the flow diagrams and/or block diagrams, and combinations of flows and/or blocks in the flow diagrams and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may 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, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

The foregoing is a more detailed description of the invention in connection with specific/preferred embodiments and is not intended to limit the practice of the invention to those descriptions. It will be apparent to those skilled in the art that various substitutions and modifications can be made to the described embodiments without departing from the spirit of the invention, and these substitutions and modifications should be considered to fall within the scope of the invention. In the description of the specification, references to the description of "one embodiment," "some embodiments," "preferred embodiments," "an example," "a specific example" or "some examples" or the like are intended to mean that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the invention. In this specification, the schematic representations of the terms used above are not necessarily intended to refer to the same embodiment or example. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples. Various embodiments or examples and features of various embodiments or examples described in this specification can be combined and combined by one skilled in the art without contradiction. Although embodiments of the present invention and its advantages have been described in detail, it should be understood that various changes, substitutions and alterations can be made herein without departing from the scope of the application.

Claims

1. A professional information recommendation method is characterized by comprising the following steps:

s3, generating an intelligence recommendation list according to the matching score output by the model;

in the step S1, the process is carried out,

2. The professional intelligence recommendation method of claim 1, wherein step S1 further comprises: modeling a query word-information-label into a three-part graph, and converting an information labeling task into a link prediction problem between an information node and a label node on the graph; all information in the three-part graph is propagated based on the graph convolution neural network, so that the intelligence node and the label node can be fused with neighbor information to obtain better representation.

3. The method of claim 1, wherein the label enhancement method comprises similarity calculation based on the characteristics of the intelligence and the labels, and one or more labels with the highest similarity are used as new labels for the intelligence.

4. The professional intelligence recommendation method of claim 1, wherein the multiple modalities comprise an image modality and a text modality.

5. The professional intelligence recommendation method of claim 4, wherein the multi-modal alignment model utilizes a Co-extensive Transformer module to fuse information of two modalities, an image and a text; the text side and the image side respectively obtain representations fused with other modalities; and splicing the representations output from the two sides to obtain the final representation of the intelligence mechanism.

6. The professional intelligence recommendation method of claim 5, wherein the multi-modal alignment model comprises two parallel transform models acting on images and text, respectively; the feature fusion of the image and the text is completed by utilizing a Co-Attention module; text content in the information is sent to a Transformer encoder through an embedded layer to extract context information, and a text representation TeE is obtained; for the image part in the intelligence, firstly partitioning the image, and then flattening, wherein each picture block is flattened into a one-dimensional vector; then, carrying out linear transformation on each vector, inputting the obtained vector into a Transformer encoder, and obtaining the representation PiE of the image; fusing the information of the two modes by utilizing a Co-extensive Transformer module; a Co-extensive Transformer module on the image side obtains Key and Value matrixes by using text representation TeE, and a Query matrix is generated by image representation PiE; the Key and Value matrixes of the Co-extensive Transformer module on the text side are derived from image representation PiE, and the Query matrixes are derived from text representation TeE.

7. The method for recommending specialized information according to claim 1, wherein in step S1, the multi-modal adaptive model learns information content characterization, and enables the multi-modal adaptive model to extract effective information based on large-scale data through a pre-training technique, so that a downstream model can be converged quickly.

8. The professional intelligence recommendation method of claim 7, wherein the multi-modal adaptive model comprises a policy network.

9. A professional intelligence recommendation system comprising a processor and a memory, said memory having stored therein a computer program, characterized in that said computer program is executable by the processor to implement the method of any of the preceding claims 1 to 8.

10. A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the method according to any one of claims 1 to 8.