WO2021213448A1 - Determination of map for information recommendation - Google Patents

Abstract

A method, system and apparatus for determining a map for information recommendation. The method comprises: acquiring a plurality of nodes used for constructing a target map, wherein the nodes at least comprise a word node and a knowledge point node (202), wherein if the node is a word node, a vector representation of a word corresponding to the node is taken as a vector representation of the node, and if the node is a knowledge point node, on the basis of a vector representation of a word related to the knowledge point node, a vector representation corresponding to the knowledge point node is determined; for any two nodes, determining an edge weight between the two nodes on the basis of the types of the two nodes, and taking the edge weight as an association relationship between the two nodes (204); and on the basis of the vector representation of the node and the association relationship between the two nodes, performing at least one round of map aggregation iteration, and updating the vector representation of the node in the map (206).

Description

Determine the atlas used for information recommendation Technical field

This specification relates to the field of data processing, and in particular to a method, system and device for determining an atlas for information recommendation.

Background technique

With the development of science and technology, the emergence of artificial intelligence provides new solutions for industries that used to consume a lot of labor costs, such as manual customer service. Intelligent customer service robots can answer simple text questions from users, but they are not good at handling complex and vague questions. Because users will send complex or ambiguous questions, the intelligent customer service robot cannot recommend accurate information to the user, which increases the processing difficulty of the intelligent customer service robot and reduces the user's experience.

Summary of the invention

One of the embodiments of this specification provides a method for determining the atlas for information recommendation. The method includes: obtaining a plurality of nodes for constructing a graph; the nodes include at least a word node and a knowledge point node; if the node is a word node, the vector representation of the word corresponding to the node is used as the vector representation of the node; If the node is a knowledge point node, determine the vector representation corresponding to the knowledge point node based on the vector representation of the words related to the knowledge point node; for any two nodes: based on the types of the two nodes, Determine the edge weight between the two nodes, and use the edge weight as the association relationship between the two nodes; based on the vector representation of the node and the association relationship between the nodes, perform at least one round Graph aggregation iterations to update the vector representation of nodes in the graph.

One of the embodiments of this specification provides an information recommendation method using a determined map. The method includes: acquiring input information; using the graph to determine the node corresponding to the input information in the graph; the graph is determined by a method for determining a graph for information recommendation; based on the vector representation of the node , And the vector representation of the adjacent nodes of the node, determine the recommended node; output the information related to the recommended node.

One of the embodiments of this specification provides a system for determining an atlas for information recommendation. The system includes a first acquisition module, a first determination module, and an update module; the first acquisition module is used to acquire a plurality of nodes for constructing a graph; the nodes include at least a word node and a knowledge point node; The node is a word node, and the vector representation of the word corresponding to the node is taken as the vector representation of the node; if the node is a knowledge point node, based on the vector representation of the word related to the knowledge point node, it is determined to correspond to the The vector representation of the knowledge point node; for any two nodes: the first determining module is configured to determine the edge weight between the two nodes based on the types of the two nodes, and use the edge weight as The association relationship between the two nodes; the update module is configured to perform at least one round of graph aggregation iterations based on the vector representation of the node and the association relationship between the nodes to update the nodes in the graph Vector representation.

One of the embodiments of this specification provides an information recommendation system using atlas. The system includes a second acquisition module, a second determination module, a third determination module, and an output module; the second acquisition module is used to acquire input information; the second determination module is used to use the atlas to determine The node corresponding to the input information in the graph; the graph is determined by a method such as determining the graph for information recommendation; the third determining module is configured to be based on the vector representation of the node, and the node The vector representation of the adjacent nodes of, determines the recommended node; the output module is used to output the information related to the recommended node.

One of the embodiments of this specification provides a device for determining a map for information recommendation. The device includes a processor, and the processor is configured to execute the above-mentioned method for determining a graph for information recommendation.

One of the embodiments of this specification provides a device for information recommendation using a determined map. The device includes a processor, and the processor is configured to execute the above-mentioned method for information recommendation using the determined atlas.

Description of the drawings

This specification will be further described in the form of exemplary embodiments, and these exemplary embodiments will be described in detail with the accompanying drawings. These embodiments are not restrictive. In these embodiments, the same number represents the same structure, in which:

FIG. 1 is a schematic diagram of an application scenario 100 of an information recommendation system according to some embodiments of this specification;

Fig. 2 is an exemplary flow chart of determining the atlas for information recommendation according to some embodiments of the present specification;

Fig. 3 is an exemplary flowchart of updating the initial atlas according to some embodiments of the present specification;

Fig. 4 is an exemplary flow chart of information recommendation using a target graph according to some embodiments of the present specification;

Fig. 5 is a block diagram of a system for determining an atlas for information recommendation according to some embodiments of this specification;

Fig. 6 is a block diagram of a system for recommending information by using a target graph according to some embodiments of this specification; and

Figure 7 is a schematic diagram of a map according to some embodiments of the present specification.

Detailed ways

In order to more clearly describe the technical solutions of the embodiments of the present specification, the following will briefly introduce the accompanying drawings used in the description of the embodiments. Obviously, the drawings in the following description are only some examples or embodiments of this specification. For those of ordinary skill in the art, without creative work, this specification can also be applied to these drawings. Other similar scenarios. Unless it is obvious from the language environment or otherwise stated, the same reference numerals in the figures represent the same structure or operation.

It should be understood that the “system”, “device”, “unit” and/or “module” used herein is a method for distinguishing different components, elements, parts, parts, or assemblies of different levels. However, if other words can achieve the same purpose, the words can be replaced by other expressions.

As shown in this specification and claims, unless the context clearly suggests exceptions, the words "a", "an", "one" and/or "the" do not specifically refer to the singular, but may also include the plural. Generally speaking, the terms "include" and "include" only suggest that the clearly identified steps and elements are included, and these steps and elements do not constitute an exclusive list, and the method or device may also include other steps or elements.

In this specification, a flowchart is used to illustrate the operations performed by the system according to the embodiment of this specification. It should be understood that the preceding or following operations are not necessarily performed exactly in order. Instead, the steps can be processed in reverse order or at the same time. At the same time, other operations can be added to these processes, or a certain step or several operations can be removed from these processes.

In some application scenarios, intelligent customer service robots can provide bubble recommendation functions, and users can click on bubbles to obtain knowledge or services. The bubble can be understood as a text box, which has a certain shape, such as a circle, a rectangle, etc., which corresponds to a text with a specific meaning. In some embodiments, fixed bubbles are provided for the user, and each bubble corresponds to a fixed function. This requires specialized configuration and development. In still other embodiments, when the user clicks on a bubble, refined knowledge or services associated with the bubble will be recommended for the user. However, this solution relies on manual labeling for bubble generation, and no connection can be established between words without co-occurrence. The methods for determining the atlas used for information recommendation and performing information recommendation based on the above atlases involved in other embodiments disclosed in this specification rely on unsupervised data and do not require manual annotation. In addition, this method uses a graph structure to establish connections between words that do not co-occur, and can dig deep-level representation information.

Fig. 1 is a schematic diagram of an application scenario 100 of an information recommendation system according to some embodiments of this specification.

As shown in FIG. 1, the application scenario 100 may include a processing device 110, a network 120, a user terminal 130, and a storage device 140. The application scenario 100 may include at least a cloud customer service scenario. The user sends the consultation data to the processing device 110 by using the user terminal 130, and the processing device 110 can determine the recommendation information most relevant to the received consultation data, and return the recommendation information to the user terminal 130.

The processing device 110 may perform one or more functions described in this specification. For example, the processing device 110 may be used to construct a target map, and use the target map to recommend information to the user. The user of the processing device 110 may be a service provider, and the service provider may construct a target map based on the service content provided by itself or the consulting data of multiple users in history, and recommend information to new and old users based on the target map. The recommended information may be knowledge related to the service provided by the service provider, or a link to request a service, etc. In some embodiments, the processing device 110 may be an independent server or a server group. The server group may be centralized or distributed (for example, the processing device 110 may be a distributed system). In some embodiments, the processing device 110 may be regional or remote. For example, the processing device 110 may access the information and/or data stored in the user terminal 130 and the storage device 140 via the network. In some embodiments, the processing device 110 may be directly connected to the user terminal 130 and the storage device 140 to access the information and/or data stored therein. In some embodiments, the processing device 110 may be executed on a cloud platform. For example, the cloud platform may include one or any combination of private cloud, public cloud, hybrid cloud, community cloud, decentralized cloud, internal cloud, etc.

In some embodiments, the processing device 110 may include one or more processing devices (for example, a single-core processing device or a multi-core and multi-core processing device). Merely as an example, the processing device may include a central processing unit (CPU), an application specific integrated circuit (ASIC), an application specific instruction processor (ASIP), a graphics processing unit (GPU), a physical processor (PPU), and a digital signal processor. (DSP), Field Programmable Gate Array (FPGA), Editable Logic Circuit (PLD), Controller, Microcontroller Unit, Reduced Instruction Set Computer (RISC), Microprocessor, etc. or any combination of the above.

The network 120 can facilitate the exchange of data and/or information among various components in the application scenario 100. For example, the processing device 110 may send the recommended information to the user terminal 130 through the network 120. In some embodiments, one or more components (user terminal 130, storage device 140) in the application scenario 100 may send data and/or information to other components in the application scenario 100 via the network 120. In some embodiments, the network 120 may be any type of wired or wireless network. For example, the network 120 may include a wired network, an optical fiber network, a telecommunication network, an internal network, the Internet, a local area network (LAN), a wide area network (WAN), a wireless local area network (WLAN), a metropolitan area network (MAN), a wide area network (WAN), public Switched Telephone Network (PSTN), Bluetooth Network, Zigbee Network, Near Field Communication (NFC) Network, Global System for Mobile Communications (GSM) Network, Code Division Multiple Access (CDMA) Network, Time Division Multiple Access (TDMA) Network, General Packet Wireless service (GPRS) network, enhanced data rate GSM evolution (EDGE) network, broadband code division multiple access (WCDMA) network, high-speed downlink packet access (HSDPA) network, long-term evolution (LTE) network, user datagram protocol (UDP) network, transmission control protocol/Internet protocol (TCP/IP) network, short message service (SMS) network, wireless application protocol (WAP) network, ultra-wideband (UWB) network, mobile communication (1G, 2G, 3G, 4G, 5G) network, Wi-Fi, Li-Fi, Narrowband Internet of Things (NB-IoT), infrared communication, etc., one or more combinations. In some embodiments, the network 120 may include one or more network access points. For example, the network 120 may include wired or wireless network access points. Through these access points, one or more components in the application scenario 100 can be connected to the network 120 to exchange data and/or information.

The user terminal 130 may be a device with information sending and/or receiving functions. For example, the user terminal 130 may send the consultation data entered by the user to the processing device 110, and receive a reply regarding the consultation data returned by the recommendation system 110. In some embodiments, the user terminal may include one or any combination of a smart phone 130-1, a tablet computer 130-2, a notebook computer 130-3, and the like. The above examples are only used to illustrate the breadth of the scope of the user terminal 130 and not to limit its scope. In some embodiments, a variety of application programs may be installed on the user terminal 130, for example, a computer program, a mobile application program (mobile phone APP), and so on. The application program may be produced and published by a service provider, and the user may download and install it in the user terminal 130. And users can consult service providers through the application.

The storage device 140 may store data and/or instructions. The data may include the data required to construct the graph, the completed graph, knowledge points, and user-oriented recommendation data, such as descriptions of the services provided by the service provider. The instruction may be an instruction required by the processing device 110 to implement the functions as disclosed in this specification. In some embodiments, the storage device 140 may also obtain data from the user terminal 130, for example, consultation/query data input by the user in history. In some embodiments, the storage device 140 may include mass storage, removable storage, volatile read-write storage, read-only storage (ROM), etc., or any combination thereof. Exemplary mass storage devices may include magnetic disks, optical disks, solid state disks, and the like. Exemplary removable storage may include flash drives, floppy disks, optical disks, memory cards, compact disks, magnetic tapes, and the like. An exemplary volatile read-write memory may include random access memory (RAM). Exemplary RAM may include dynamic random access memory (DRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), static random access memory (SRAM), thyristor random access memory (T-RAM), and zero capacitance Random access memory (Z-RAM), etc. Exemplary read-only memory may include mask-type read-only memory (MROM), programmable read-only memory (PROM), erasable programmable read-only memory (PEROM), electrically erasable programmable read-only memory (EEPROM), CD-ROM and digital versatile disk read-only memory, etc. In some embodiments, the storage device 140 may be implemented in a single central server, or multiple servers or multiple personal devices connected through a communication link. The storage device 140 may also be implemented by multiple personal devices and cloud servers. The storage device 140 may also be implemented on a cloud platform. For example, the cloud platform may include private cloud, public cloud, hybrid cloud, community cloud, decentralized cloud, internal cloud, etc. or any combination of the above.

In some embodiments, the storage device 140 may be connected to the network 120 to communicate with one or more components in the application scenario 100 (for example, the processing device 110, the user terminal 130, etc.). One or more components in the application scenario 100 can access data or instructions stored in the storage device 140 via the network 120. In some embodiments, the storage device 140 may directly connect or communicate with one or more components in the application scenario 100 (for example, the processing device 110, the user terminal 130, etc.). In some embodiments, the storage device 140 may be part of the processing device 110.

It should be noted that the description of each component in the above application scenario 100 is only for example and explanation, and does not limit the scope of application of this specification. For those skilled in the art, under the guidance of this specification, components in the application scenario 100 can be added or reduced. However, these changes are still within the scope of this specification.

Fig. 2 is an exemplary flow chart of determining a map (or referred to as a target map) for information recommendation according to some embodiments of the present specification. In some embodiments, the process 200 may be implemented by the information recommendation system 500 or the processing device 110 shown in FIG. 1. For example, the process 200 may be stored in a storage device (such as the storage device 140) in the form of a program or instruction, and the program or instruction may implement the process 200 when the program or instruction is executed. As shown in FIG. 2, the process 200 may include the following steps.

Step 202: Obtain multiple nodes for constructing the target graph; the nodes include at least word nodes and knowledge point nodes.

This step may be performed by the first acquisition module 510.

In some embodiments, the target graph may refer to a graph used for information recommendation for users, which includes multiple nodes and associated information between nodes, and each node may correspond to a piece of information. When in use, the target graph can determine the most relevant node according to the user's input, and recommend the information corresponding to the node to the user. The multiple nodes constituting the target graph may include at least word nodes and knowledge point nodes. The information corresponding to the word node may be a word, and the word corresponding to the word node may be directly recommended to the user during information recommendation. The information corresponding to the knowledge point node may be a knowledge point. The knowledge points can be composed of a title and a text. The title can be a question, and the text can be the answer to the question. When recommending information, you can determine whether it is most relevant to the user's input according to the title, and if so, recommend the text to the user. In the target graph, any two nodes have a certain association relationship. When performing information recommendation, the association relationship between nodes can be used to determine the node most relevant to the user's input.

Referring to FIG. 7, FIG. 7 is a schematic diagram of a target map according to some embodiments of the present specification. As shown in Figure 7, the box is used to represent a word node, and the content in the box is the word corresponding to the word node. As shown in Figure 7, "Album", "Size", "Photo", "Free Shipping", "Coupon" and so on. The words corresponding to the word nodes may be keywords used in information recommendation, which are closely related to the information to be recommended. It can also be a high-frequency word input by the user during information consultation, which can be linked to one or more pieces of information to be recommended. The round box in Figure 7 is used to represent the knowledge point node, and the content in the round box is the title of the knowledge point node, as shown in Figure 7, "What are the sizes of the photo album", "What are the sizes of the photos", and "Can it be packaged? Mail", "how to use coupons", etc. The question and answer content (ie title and text) corresponding to the knowledge point node can be the information support content that the user wants to obtain when the information is recommended, and it can be the service range or The service content is determined. For example, assuming that the user of the information recommendation 100 is a photography studio, the question and answer content corresponding to the knowledge point node can be photography-related such as the opening time, the type of photography provided, such as the ID photo, the size of the finished photo frame, whether It can be mailed, whether it is free shipping, etc.

The connection between the nodes in Figure 7 can represent the association relationship between the two nodes. For example, the connection between the word node "album" and "size" may indicate the frequency of two words co-occurring, such as the frequency of co-occurring in a sentence or a paragraph. The higher the frequency, the closer the relationship between the two. For another example, the connection between the word node "album" and the knowledge node "what are the sizes of the album" may indicate whether the album accounts for the key to the answer or explanation of the "what sizes of the album". The key to the proportion indicates that the answer or explanation is closely related to the album. When performing information recommendation, the most relevant content of the target map and the user's input can be recommended to the user according to the user's input. The construction of the target map is described in detail in the subsequent part of this flowchart, and you can refer to Figure 3. For the description of information recommendation, please refer to Figure 4 of this manual.

In some embodiments, the multiple nodes may be pre-stored in a storage device, for example, the storage device that comes with the processing device 110, or the storage device 140. It can be determined and stored in advance according to the user's historical consultation or his own service scope. The first acquiring module 510 may read the multiple nodes after communicating with the storage device.

In some embodiments, the vector representation corresponding to each node may be determined based on the type of each node. It can be understood that the content (for example, words or knowledge points) corresponding to each node can be expressed in the form of vectors. For example, mapping words, phrases, sentences, or paragraphs into numbers through word embedding, and expressing them in vector space through mathematics, is beneficial to data processing. The relationship between nodes and nodes may also be embodied in this specification using a piece of quantitative data to indicate the closeness of the relationship between the two nodes.

In some embodiments, the first obtaining module 510 may determine the corresponding vector representation for each node according to the difference of node types (word nodes and knowledge node nodes). If the node is a word node, the first obtaining module 510 can use the vector representation of the word corresponding to the node as the vector representation of the node, and can use the word vector representation model to determine the vector representation corresponding to the word. The word vector representation model includes a machine learning model, for example, an artificial neural network. An exemplary word vector representation model may be a word embedding model, including but not limited to word2vec, glove, ELMo, BERT, etc. The input can be a word, and the output can be the word vector corresponding to the word. The first acquisition module 510 may determine the vector corresponding to each word through the word embedding model, and then each corresponding vector representation is a vector representation of the word node corresponding to the word. For example, assume that the two words corresponding to two word nodes are "album" and "size". The first obtaining module 510 can input the above two words into the word embedding model, obtain the word vectors V ₁ and V ₂ corresponding to "album" and "size" respectively, and separate the two vectors _{V 1} and V ₂ As a vector representation of the above two word nodes.

In some embodiments, if the node is a knowledge point node, the first obtaining module 510 may determine the vector representation of the knowledge point node based on the vector representation of the words related to the knowledge point node. The words related to the knowledge point node may be words included in the knowledge point, or may be words corresponding to the word node having an association relationship with the knowledge point node. For example, the words included in the knowledge point node "What are the sizes of the album" may include "album" and "size", and the words related to the knowledge point node are "album" and "size". For another example, the words corresponding to the word node associated with the knowledge point node "what size of the album" are "album" and "size", the above two words can be used as the words related to the knowledge point node.

In some embodiments, the first obtaining module 510 may first obtain one or more words from the knowledge point corresponding to the knowledge point node, and determine the vector representation of the one or more words. Subsequently, the first obtaining module 510 may perform operations on one or more of the vector representations, and use the operation results as the vector representations corresponding to the knowledge point nodes. The operation may be a sum operation or an average operation represented by one or more vectors, and the average operation may include a weighted average or an arithmetic average. As an example, it is assumed that the words from the knowledge point node "what sizes of albums" include "album" and "size", and the word vectors corresponding to the two words are V ₁ and V _{2 respectively} , which can be determined based on the word vector representation model. _{The first obtaining module 510 may calculate the vector V 3} by averaging the two word vectors, for example, arithmetic average calculation. Then V ₃ will be represented as a vector of the knowledge point node "what size is the album".

Step 204: For any two nodes: determine an edge weight between the two nodes based on the types of the two nodes, and use the edge weight as an association relationship between the nodes. This step may be executed by the first determining module 520.

In some embodiments, when determining the association relationship between the node and the node, the first determining module 520 may perform different processing based on the types of the two nodes. The first determining module 520 may first determine whether the two nodes are the same type of node, and based on the determination result, determine the edge weight between the two nodes, and then use the edge weight as the association relationship between the two nodes.

In some embodiments, if the two nodes are both word nodes, the first determining module 520 may determine the edge weight between the two nodes based on the co-occurrence frequency between words corresponding to the two word nodes. The co-occurrence frequency may refer to the probability of two words appearing at the same time in the text. The greater the probability, the closer the relationship between the two words, and the higher the degree of association. The first determining module 520 may determine the co-occurrence frequency through a point-wise mutual information algorithm (PMI, point-wise mutual information). If one of the two nodes is a word node, and the other node is a knowledge point node, the first determining module 520 may be based on the importance of the word of the word node relative to the knowledge point (including title and text) corresponding to the knowledge point node. Determine the edge weight between two nodes. This degree of importance can be understood as the degree to which the word is explained in the content of the knowledge node. For example, assuming that the content of a knowledge point node is an explanation of a word (for example, a word is a service provided by a service provider and the knowledge point node explains it), then the word can be considered relative to the word The importance of knowledge nodes is high. Conversely, if a word is only a constituent element of a knowledge node, it can be considered that the word has a low degree of importance relative to the knowledge node. The first determining module 520 may use term frequency-inverse document frequency (TF-IDF, term frequency-inversed document frequency) to measure the importance of the term based on the term node relative to the knowledge point corresponding to the knowledge point node. If the two nodes are both knowledge point nodes, the first determining module 520 can directly determine the edge weight between the two nodes as zero. Referring to Figure 7, in the target graph, the connection between two nodes can indicate that there is an association relationship between the two nodes, which can use the PMI value (the connection between two boxes (word nodes)), or TF-IDF value (the line between the box and the circle (word node and knowledge node)). There may also be no connection between two nodes. For example, if the relationship between two knowledge point nodes is 0, there is no connection between the two knowledge point nodes.

Step 206, based on the vector representation of the node and the association relationship between the node and the node, perform at least one round of graph aggregation iteration to update the vector representation of the node in the graph.

This step may be performed by the update module 530.

In some embodiments, the vector representation and edge weights of the nodes determined in step 202 and step 204 can be regarded as the initial expression of the graph, and the graph with the initial expression can be understood as a graph that does not yet have the information recommendation function. The vector representation of its nodes needs to be updated to get a more complete representation of the graph.

In some embodiments, the initial expression of the atlas may be represented by a matrix. As an example, a graph matrix X composed of vector representations of multiple nodes and a relationship matrix R composed of association relationships between multiple nodes can be used to represent the initial expression of the graph. Assuming that there are a total of N nodes constituting the graph, and the vector of each node is a 300-dimensional vector, the graph matrix X can be a matrix of N*300 or a matrix of 300*N. For the relation matrix R, it can be an N*N matrix, and each row or column can be an association relationship between a node and other nodes (for example, edge weights). When the node is compared with itself, the edge weight can be 1.

In some embodiments, the update module 530 may perform at least one round of graph aggregation iterations on the expression of the graph to update the expression of the graph. In some embodiments, graph aggregation can be understood as a process of performing operations based on the vector representation of at least one node and/or edge weight in the graph, and updating at least one other node and/or edge weight vector representation in the graph with the result of the operation. For example, for each node, in a round of iteration, the update module 530 may use the vector representation of the adjacent node of the node to update the vector representation of the node. As an example, the update module 530 may perform operations on the vector representation of the neighboring nodes of the node in the current iteration round, for example, a weighted average operation (the edge weight between the node and the neighboring nodes is used as the weight), and use the result of the operation to update the vector representation. The vector representation of the node.

In some embodiments, the update module 530 may use the relationship matrix R to update the vector representation of the nodes in the graph, so as to achieve the purpose of updating the expression of the graph. In a round of iteration, the update module 530 may use the vector representation of multiple nodes in the current iteration round to obtain a vector representation matrix, for example, the graph matrix X in the foregoing example. At the same time, the update module 530 may determine the adjacency matrix corresponding to the multiple nodes based on the association relationship between the nodes, for example, the relationship matrix R in the foregoing example. Subsequently, the update module 530 may perform operations on the vector representation matrix and the adjacency matrix, and use the results of the operations to update the vector representation of each node in the graph. For example, the relationship matrix R is used to perform weighted aggregation on the graph matrix X to update the graph matrix X.

In some embodiments, the update module 530 uses a neural network-based aggregation model to update the vector representation of the nodes in the graph. The update module 530 may use a neural network-based aggregation model to process the vector representation matrix obtained by using the vector representation of the multiple nodes, and the adjacency corresponding to the multiple nodes determined based on the association relationship between the nodes Matrix to obtain an updated vector representation matrix, and update vector representations of nodes in the graph based on the updated vector representation matrix. The neural network-based aggregation model may include GCN (Graph Convolutional Network, Graph Convolutional Network), GAT (Graph Attention Networks, Graph Neural Network), and the like. Suppose that the vector representation matrix is represented by X (for example, the graph matrix X), and the adjacency matrix is represented by R (for example, the relation matrix R). Taking GCN as an example, the update module 530 can input X and R into the GCN. Inside the GCN, After the vector representation matrix X, the adjacency matrix R and the model parameter W of the GCN are calculated, the GCN can convert the vector representation of the map node from X to X'. X'can refer to the updated vector representation matrix. It can be understood that whether the updated vector representation matrix X'can accurately represent the information of the map depends on the accuracy of the GCN model parameter W to a certain extent.

In some embodiments, GCN needs to be trained to optimize its model parameter W. In practical applications, the prediction task of the GCN can be determined according to specific application scenarios, and the GCN can be trained based on the prediction task. Taking the prediction task of predicting the correlation of two nodes as an example, GCN can be used as a part of the prediction model. The input of the prediction model is two nodes. The prediction model can be based on the vector representation of the two nodes by GCN (such as vector representation matrix). X') Calculate the correlation between these two nodes and output. In the GCN training stage, the GCN model parameter W is a random initial value, and X'is also inaccurate at this time. The input layer of the prediction model receives the input nodes A and B of the training sample, based on the two input nodes corresponding to X' The vector represents determining the similarity y of two nodes, constructing a loss function based on the difference between the true value of the correlation between y and the training sample, and adjusting the model parameter W of GCN to minimize the loss function. Among them, the true value can be expressed as "0" or "1". For example, a recommendation system outputs A to the user, and then clicks on B, indicating that node A is related to node B, and its true value is 1, otherwise it is 0. As the training progresses, the model parameter W is trained well. At the same time, the vector representation matrix X'of the map node can also better reflect the information of the map. It should be noted that the loss function can be determined based on a specific training task, and this specification does not make any restrictions on this.

For other descriptions of the update of the vector representation of the nodes in the graph, please refer to Figure 3 of this specification.

It should be noted that the foregoing description of the process 200 is only for example and description, and does not limit the scope of application of this specification. For those skilled in the art, various modifications and changes can be made to the process 200 under the guidance of this specification. However, these corrections and changes are still within the scope of this specification.

Fig. 3 is an exemplary flow chart of updating the initial expression of the map according to some embodiments of the present specification. In some embodiments, the process 300 may be implemented by the information recommendation system 500 or the processing device 110 shown in FIG. 1. For example, the process 200 may be stored in a storage device (such as the storage device 140) in the form of a program or instruction, and the program or instruction may implement the process 200 when the program or instruction is executed. In some embodiments, the process 300 may describe a specific process of a round of iteration. In some embodiments, the process 300 may be executed by the update module 530. As shown in FIG. 3, the process 300 may include the following steps.

Step 302: Use the vector representation of the multiple nodes in the current iteration round to obtain a vector representation matrix.

In some embodiments, the update module 530 may arrange the vector representations of the multiple nodes in the current iteration round to obtain the vector representation matrix. As an example, assuming that there are a total of N nodes constituting the graph, and the vector of each node is a 300-dimensional vector, the update module 530 can arrange the vector of the nodes in rows to form an N*300 vector representation matrix, or press Arrange the vector of nodes in columns to form a 300*N vector to represent the matrix.

Step 304: Determine an adjacency matrix corresponding to the multiple nodes based on the association relationship between the nodes.

In some embodiments, the association relationship between the multiple nodes may be expressed in the form of a matrix, such as the relationship matrix R mentioned in step 206. In this specification, the relationship matrix R can also be referred to as the adjacency matrix A, which is used to represent the relationship between a certain node and all other nodes. Assuming that there are a total of N nodes, the adjacency matrix A is an N*N matrix. The number in the i-th row and j-th column of the matrix represents the relationship between node i and node j, such as edge weights. For illustrative purposes, a simplified adjacency matrix A is shown below: For the relational matrix R, it can be an N*N matrix, and each row or column can be the association relationship between a node and other nodes ( For example, Bian Quan). And the node is compared with itself, the edge weight can be 1

Among them, A _ij represents the relationship between the i-th node and the j-th node. When the i-th node and the j-th node are both word nodes, A _ij =PMI(i,j); when the i-th node is a word node and the j-th node is a knowledge node, A _ij =TF- IDF(i,j); when i=j, that is, the i-th node relative to itself, A _ij =1; when both the i-th node and the j-th node are knowledge point nodes, A _ij =0 , Indicating that there is no correlation between the two knowledge point nodes.

Step 306: Perform operations on the vector representation matrix and the adjacency matrix, and use the results of the operations to update the vector representation of each node in the graph.

In some embodiments, the update module 530 may use the adjacency matrix A to perform a weighted average calculation on the vector representation matrix (here assigned the label X). For example, according to the formula aggregate(X)=A*X of the weighted average algorithm, the vector representation matrix X is calculated through the adjacency matrix A, and the vectors contained in the calculation result X'are used as the vectors of the nodes after the current iteration. Express.

In some embodiments, in one round of iteration, the update module 530 may also be updated separately for each node. For any node, the update module 530 may determine the adjacent node of the node based on the association relationship between the node and the node. The adjacent node may be a node directly connected to the node, which can be understood as an association relationship between two nodes (for example, there is an edge weight between the two nodes, such as PMI or TD-IDF). Referring to FIG. 7, the adjacent nodes of the word node "photo" shown in FIG. 7 may be the word node "size", the word node "Free shipping", and the knowledge node "what size is the photo". The word node "photo" and the above nodes are directly connected by a line. After determining the adjacent node of the node, the update module 530 may perform a weighted average operation on the vector representation of the adjacent node based on the edge weight between the node and the adjacent node, and use the calculation result as the updated vector representation of the node. For example, the adjacent nodes of the word node "photo" are the word node "size", the word node "Free shipping", and the knowledge node "what size is the photo". When the vector representation of the word node "photo" is updated, these three The vector representation of the adjacent nodes is weighted and averaged, and the calculation result is used as the updated vector representation of the word node "photo". Among them, the weight represented by the vector of each adjacent node in the weighted average algorithm can be determined based on the association relationship between the node and each adjacent node. For example, the value of the element in the adjacency matrix A can be used as the weight.

The above describes a round of iterative process. The update module 530 may update the initial expression in the graph for one or more iterations (for example, update the vector representation of the nodes one or more times) according to the above description to obtain the final expression of the graph. It can be understood that the vector representation of each node in the graph can be updated in the manner of step 304. When each node has been updated a set number of times, it can be considered that the update is complete. Or, continue to update until the change in the vector representation of each node is less than the set threshold. As an example, after one update iteration, the graph matrix X will be updated as X'=aggregate(X)=A*X. In the next iteration, the graph matrix X will be updated to X”=aggregate(X')=A*X'. In the third iteration, the graph matrix X will be updated to X”'=aggregate(X”) =A*X". And so on. The number of rounds of the iteration can be preset, for example, 3 times, 5 times, 7 times, etc., which are not limited in this specification. After the iteration is completed, the graph matrix X and the joint relation matrix R (ie, the adjacency matrix A) after several updates can be used as the target graph.

It should be noted that the foregoing description of the process 300 is only for example and description, and does not limit the scope of application of this specification. For those skilled in the art, various modifications and changes can be made to the process 300 under the guidance of this specification. However, these corrections and changes are still within the scope of this specification.

Fig. 4 is an exemplary flow chart of information recommendation using a target graph according to some embodiments of the present specification. In some embodiments, the process 300 may be implemented by the information recommendation system 600 or the processing device 110 shown in FIG. 1. For example, the process 400 may be stored in a storage device (such as the storage device 140) in the form of a program or instruction, and when the program or instruction is executed, the process 400 may be implemented. As shown in FIG. 4, the process 300 may include the following steps.

Step 402: Obtain input information.

This step may be performed by the second information acquisition module 610.

In some embodiments, the input information may be one or more words selected by the user from candidate words provided to the user in advance. For example, when performing information recommendation, the processing device 110 (or the information recommendation system 600) may send the candidate words provided to the user to the user terminal 130 and display them. The display format can be multiple bubble recommendations, and each bubble corresponds to a candidate word. The user can click one or more of the candidate words to provide click feedback to the processing device 110 (or the information recommendation system 600). The feedback content is the input information. For example, the candidate words provided to the user in advance include "photo", "top", "shoes", "size", etc., and the user selects the word "photo" among them, and the input information is the word "photo". When the user selects the two words "photo" and "size", the input information is the words "photo" and "size". In some embodiments, the candidate words provided to the user in advance may be high-frequency words that appeared during the user's consultation in history, or may be a user of the processing device 110 (or the information recommendation system 600) (for example, a service provider).者) Terms related to the service provided. Assuming that the service provided by the service provider is an online clothing vendor, the candidate words provided to the user in advance may include "size", "discount", "free shipping", and so on.

Step 404: Using the graph, determine the node corresponding to the input information in the graph.

This step may be executed by the second determining module 620.

In some embodiments, the map may be the target map. For a specific description of the target map, you can refer to the relevant content in Figures 2 and 3 of this specification.

In some embodiments, the second determining module 620 may compare the words in the input information with the words corresponding to the word nodes in the target graph to determine the node corresponding to the input information. For example, assuming that the input information includes the word "photo", the second determining module 620 may determine the word node "photo" corresponding to the word "photo" in the target atlas as the node corresponding to the input information. Assuming that the input information includes the words "photo" and "size", the second determining module 620 may determine the word node "photo" and the word node "size" corresponding to the words "photo" and "size" in the target atlas as the words "photo" and "size". Describe the node corresponding to the input information.

Step 406: Determine a recommended node based on the vector representation of the node and the vector representation of the adjacent nodes of the node.

This step may be performed by the third determining module 630.

For the related content of the vector representation of the node and the vector representation of the adjacent nodes of the node, please refer to the related descriptions in FIG. 2 and FIG. 3 in this specification.

In some embodiments, the third determining module 630 may respectively determine the distance between the vector representation of the node and the vector representation of each adjacent node of the node. The distance may be Minkowski distance, Euclidean distance, Manhattan distance, Chebyshev distance, angle cosine, Hamming distance, Jackard similarity coefficient, and the like. The third determining module 630 may determine the node corresponding to the closest distance (for example, the smallest distance value) as the recommended node. 7, assuming that the input information is the word "photo", the third determining module 630 can determine the vector representation of the word node "photo", the word node "size" and the word node "Free shipping" adjacent to the word node "photo" , Knowledge point node "what size of the photo", the respective vector represents the distance between them, and determine the one or more nodes with the closest corresponding distance as the recommended node.

Step 408: Output the information related to the recommended node.

This step can be performed by the output module 640.

In some embodiments, when the recommended nodes only include knowledge point nodes, the output module 640 may output the knowledge point text corresponding to the knowledge point node as related information. For example, suppose that the user selects two words "photo" and "size" as input information, according to step 404 and step 406, the knowledge point node "what size of photo" is determined as the recommended node. The recommended node contains only the knowledge point node, and the output module 640 can read the text of the knowledge point corresponding to the knowledge point "what size of the photo", such as "1 inch 2.5*3.5 (cm), 2 inch 3.6*4.7 (cm), 3 inch" 5.8*8.4(cm)” as output and recommended to users.

In some embodiments, when the recommendation node includes a word node, the processing device 110 (or the information recommendation system 600) may recommend the word corresponding to the recommendation node to the user again, allowing the user to select a word from it, and based on the user's selection Determine the recommended node again. For example, when the recommendation node is determined to be the word node "size" and the word node "photo", the processing device 110 (or the information recommendation system 600) may recommend the words "size" and "photo" to the user again for selection. If the user selects the word "photo" again, the processing device 110 (or the information recommendation system 600) may repeat steps 402 to 406 to re-determine the recommendation node. If the re-determined recommended nodes include the knowledge point node "What size of the photo", the output module 640 can convert the knowledge point "What size of the photo" to the text, such as "1 inch 2.5*3.5 (cm), 2 inch 3.6*4.7 ( cm), 3 inches 5.8*8.4(cm)” as output and recommended to users. If the re-determined recommended node still does not include the word node, the above process will be repeated again until the recommended node includes at least one knowledge point node.

It should be noted that the above description of the process to determine the map method used for information recommendation is only for example and explanation, and does not limit the scope of application of this specification. For those skilled in the art, under the guidance of this specification, various corrections and changes can be made to the map method for determining the process for information recommendation. However, these corrections and changes are still within the scope of this specification. For example, other steps are added to the map method used for information recommendation in the process of determining, for example, storage steps, verification steps, and so on.

FIG. 5 is a block diagram of a system 500 for determining a graph for information recommendation according to some embodiments of the present specification.

As shown in FIG. 5, the system 500 for determining a graph for information recommendation may include a first obtaining module 510, a first determining module 520, and an updating module 530.

The first obtaining module 510 may be used to obtain multiple nodes for constructing the target graph. The target graph may refer to a graph used when making information recommendations for users, which includes multiple nodes, and each node may correspond to a piece of information. The nodes include at least word nodes and knowledge point nodes. The information corresponding to the word node may be a word. The information corresponding to the knowledge point node may be a knowledge point. The knowledge points can be composed of a title and a text. The title can be a question, and the text can be the answer to the question. In some embodiments, the multiple nodes may be pre-stored in a storage device, for example, the storage device that comes with the processing device 110, or the storage device 140. It can be determined and stored in advance according to the user's historical consultation or his own service scope. The first acquiring module 510 may read the multiple nodes after communicating with the storage device.

In some embodiments, the first obtaining module 510 may determine the corresponding vector representation for each node according to the difference of node types (word nodes and knowledge node nodes). If the node is a word node, the first obtaining module 510 may use the vector representation of the word corresponding to the node as the vector representation of the node. If the node is a knowledge point node, the first obtaining module 510 may determine the vector representation of the knowledge point node based on the vector representation of the words related to the knowledge point node.

The first determining module 520 may determine the edge weight between the two nodes based on the types of the two nodes, and use the edge weight as the association relationship between the nodes. The first determining module 520 can perform the above operations on any two nodes. In some embodiments, when determining the association relationship between the node and the node, the first determining module 520 may perform different processing based on the types of the two nodes. The first determining module 520 may first determine whether the two nodes are the same type of node, and based on the determination result, determine the edge weight between the two nodes, and then use the edge weight as the association relationship between the two nodes. If the two nodes are both word nodes, the first determining module 520 may determine the edge weight between the two nodes based on the co-occurrence frequency between the words corresponding to the two word nodes. If one of the two nodes is a word node, and the other node is a knowledge point node, the first determining module 520 may be based on the importance of the word of the word node relative to the knowledge point (including title and text) corresponding to the knowledge point node. Determine the edge weight between two nodes. If the two nodes are both knowledge point nodes, the first determining module 520 can directly determine the edge weight between the two nodes as zero.

The update module 530 may perform at least one round of graph aggregation iteration based on the vector representation of the node and the association relationship between the nodes to update the vector representation of the nodes in the graph. In some embodiments, for each node, the update module 530 can update the vector representation of the node by using the vector representation of the neighboring nodes of the node. As an example, the update module 530 may perform an operation on the vector representation of the adjacent node, for example, a weighted average operation, and update the vector representation of the node with the result of the operation. The update module 530 may also update the vector representation of the nodes in the graph by using the association relationship between the nodes to determine the target graph. The update module 530 may also use a neural network-based aggregation model to update the vector representation of the nodes in the initial map.

For more description of the modules of the system 500, please refer to the flowchart part of this specification, for example, FIG. 2 to FIG. 3.

FIG. 6 is a block diagram of a system 600 for information recommendation using target graphs according to some embodiments of this specification.

As shown in FIG. 6, the information recommendation system 600 using the determined atlas may include a second acquisition module 610, a second determination module 620, a third determination module 630, and an output module 640.

The second obtaining module 610 may be used to obtain input information. In some embodiments, the input information may be one or more words selected by the user from candidate words provided to the user in advance. For example, when performing information recommendation, the processing device 110 (or the information recommendation system 600) may send the candidate words provided to the user to the user terminal 130 and display them. The display format can be multiple bubble recommendations, and each bubble corresponds to a candidate word. The user can click one or more of the candidate words to provide click feedback to the processing device 110 (or the information recommendation system 600). The feedback content is the input information.

The second determining module 620 may be configured to use the graph to determine the node corresponding to the input information in the graph. In some embodiments, the map may be the target map. The second determining module 620 may compare the words in the input information with the words corresponding to the word nodes in the target graph to determine the node corresponding to the input information.

The third determining module 630 may be configured to determine the recommended node based on the vector representation of the node and the vector representation of the adjacent nodes of the node. In some embodiments, the third determining module 630 may respectively determine the distance between the vector representation of the node and the vector representation of each adjacent node of the node, and correspond to the closest distance (for example, the smallest distance value). The node is determined as the recommended node.

The output module 640 may be used to output information related to the recommended node. In some embodiments, when the recommended nodes only include knowledge point nodes, the output module 640 may output the knowledge point text corresponding to the knowledge point node as related information. When the recommended node includes a word node, the system 600 can obtain the user's input information again, and determine the recommended node again until the recommended node includes at least one knowledge point node. At this time, the output module 640 may output the at least one knowledge point node to the user.

For more description of the modules of the system 600, please refer to the flowchart part of this specification, for example, FIG. 4.

It should be understood that the system and its modules shown in FIG. 5 and FIG. 6 can be implemented in various ways. For example, in some embodiments, the system and its modules may be implemented by hardware, software, or a combination of software and hardware. Among them, the hardware part can be implemented using dedicated logic; the software part can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art can understand that the above-mentioned methods and systems can be implemented using computer-executable instructions and/or included in processor control codes, for example on a carrier medium such as a disk, CD or DVD-ROM, such as a read-only memory (firmware Such codes are provided on a programmable memory or a data carrier such as an optical or electronic signal carrier. The system and its modules in this specification can not only be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, etc., or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc. It may also be implemented by software executed by various types of processors, or may be implemented by a combination of the foregoing hardware circuit and software (for example, firmware).

It should be noted that the above description of the candidate item display, determination system and its modules is only for convenience of description, and does not limit this specification to the scope of the examples mentioned. It can be understood that for those skilled in the art, after understanding the principle of the system, it is possible to arbitrarily combine various modules, or form a subsystem to connect with other modules without departing from this principle. For example, the first determination module 520 and the second determination module 530 disclosed in FIG. 5, or the third determination module 620 and the fourth determination module 630 disclosed in FIG. One module implements the functions of the two or more modules mentioned above. For another example, each module may share a storage module, and each module may also have its own storage module. Such deformations are all within the protection scope of this specification.

The possible beneficial effects of the embodiments of this specification include but are not limited to: (1) This specification recommends more accurate and distinguishable words to the user for the user to choose, and then responds to the user with more accurate information and improves the response information The accuracy of the cloud customer service robot is reduced, and the user experience is improved. (2) This specification optimizes the vector representation of each node by using the adjacent nodes of each node to obtain a more accurate degree of association between two nodes, so that the words recommended for the user and the reply information are more accurate. (3) This manual trains the model through the adjacent information of the graph and relies on unsupervised data, avoiding the dependence on manual marking data. It should be noted that different embodiments may have different beneficial effects. In different embodiments, the possible beneficial effects may be any one or a combination of the above, or any other beneficial effects that may be obtained.

The basic concepts have been described above. Obviously, for those skilled in the art, the above detailed disclosure is only an example, and does not constitute a limitation to this specification. Although it is not explicitly stated here, those skilled in the art may make various modifications, improvements and amendments to this specification. Such modifications, improvements, and corrections are suggested in this specification, so such modifications, improvements, and corrections still belong to the spirit and scope of the exemplary embodiments of this specification.

Meanwhile, this specification uses specific words to describe the embodiments of this specification. For example, "one embodiment", "an embodiment", and/or "some embodiments" mean a certain feature, structure, or characteristic related to at least one embodiment of this specification. Therefore, it should be emphasized and noted that “one embodiment” or “one embodiment” or “an alternative embodiment” mentioned twice or more in different positions in this specification does not necessarily refer to the same embodiment. . In addition, some features, structures, or characteristics in one or more embodiments of this specification can be appropriately combined.

In addition, those skilled in the art can understand that various aspects of this specification can be explained and described through a number of patentable categories or situations, including any new and useful process, machine, product, or combination of substances, or a combination of them. Any new and useful improvements. Correspondingly, various aspects of this specification can be completely executed by hardware, can be completely executed by software (including firmware, resident software, microcode, etc.), or can be executed by a combination of hardware and software. The above hardware or software can all be referred to as "data block", "module", "engine", "unit", "component" or "system". In addition, various aspects of this specification may be embodied as a computer product located in one or more computer-readable media, and the product includes computer-readable program codes.

The computer storage medium may contain a propagated data signal containing a computer program code, for example on a baseband or as part of a carrier wave. The propagated signal may have multiple manifestations, including electromagnetic forms, optical forms, etc., or a suitable combination. The computer storage medium may be any computer readable medium other than the computer readable storage medium, and the medium may be connected to an instruction execution system, device, or device to realize communication, propagation, or transmission of the program for use. The program code located on the computer storage medium can be transmitted through any suitable medium, including radio, cable, fiber optic cable, RF, or similar medium, or any combination of the above medium.

The computer program codes required for the operation of each part of this manual can be written in any one or more programming languages, including object-oriented programming languages such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, Python Etc., conventional programming languages such as C language, Visual Basic, Fortran 2003, Perl, COBOL 2002, PHP, ABAP, dynamic programming languages such as Python, Ruby and Groovy, or other programming languages. The program code can be run entirely on the user's computer, or run as an independent software package on the user's computer, or partly run on the user's computer and partly run on a remote computer, or run entirely on the remote computer or server. In the latter case, the remote computer can be connected to the user's computer through any network form, such as a local area network (LAN) or a wide area network (WAN), or connected to an external computer (for example, via the Internet), or in a cloud computing environment, or as a service Use software as a service (SaaS).

In addition, unless explicitly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or the use of other names in this specification are not used to limit the order of the processes and methods in this specification. Although the foregoing disclosure uses various examples to discuss some embodiments of the invention that are currently considered useful, it should be understood that such details are only for illustrative purposes, and the appended claims are not limited to the disclosed embodiments. On the contrary, the rights are The requirements are intended to cover all modifications and equivalent combinations that conform to the essence and scope of the embodiments of this specification. For example, although the system components described above can be implemented by hardware devices, they can also be implemented only by software solutions, such as installing the described system on an existing server or mobile device.

For the same reason, it should be noted that, in order to simplify the expressions disclosed in this specification and help the understanding of one or more embodiments of the invention, in the foregoing description of the embodiments of this specification, multiple features are sometimes combined into one embodiment. In the drawings or its description. However, this method of disclosure does not mean that the subject of the specification requires more features than those mentioned in the claims. In fact, the features of the embodiment are less than all the features of the single embodiment disclosed above.

In some embodiments, numbers describing the number of ingredients and attributes are used. It should be understood that such numbers used in the description of the embodiments use the modifier "about", "approximately" or "substantially" in some examples. Retouch. Unless otherwise stated, "approximately", "approximately" or "substantially" indicates that the number is allowed to vary by ±20%. Correspondingly, in some embodiments, the numerical parameters used in the specification and claims are approximate values, and the approximate values can be changed according to the required characteristics of individual embodiments. In some embodiments, the numerical parameter should consider the prescribed effective digits and adopt the method of general digit retention. Although the numerical ranges and parameters used to confirm the breadth of the ranges in some embodiments of this specification are approximate values, in specific embodiments, the setting of such numerical values is as accurate as possible within the feasible range.

For each patent, patent application, patent application publication and other materials cited in this specification, such as articles, books, specifications, publications, documents, etc., the entire contents are hereby incorporated into this specification as a reference. The application history documents that are inconsistent or conflict with the content of this specification are excluded, and the documents that restrict the broadest scope of the claims of this specification (currently or later appended to this specification) are also excluded. It should be noted that if there is any inconsistency or conflict between the description, definition, and/or use of terms in the auxiliary materials of this manual and the content of this manual, the description, definition and/or use of terms in this manual shall prevail. .

Finally, it should be understood that the embodiments described in this specification are only used to illustrate the principles of the embodiments of this specification. Other variations may also fall within the scope of this specification. Therefore, as an example and not a limitation, the alternative configuration of the embodiment of the present specification can be regarded as consistent with the teaching of the present specification. Accordingly, the embodiments of this specification are not limited to the embodiments explicitly introduced and described in this specification.

Claims (20)

1. A method for determining an atlas for information recommendation, wherein the method includes:

   Obtain multiple nodes for constructing the graph; the node includes at least a word node and a knowledge point node; if the node is a word node, the vector representation of the word corresponding to the node is used as the vector representation of the node; if the node is A knowledge point node, based on the vector representation of the words related to the knowledge point node, determine the vector representation corresponding to the knowledge point node;

   For any two nodes: determine the edge weight between the two nodes based on the types of the two nodes, and use the edge weight as the association relationship between the two nodes;

   Based on the vector representation of the node and the association relationship between the node and the node, at least one round of graph aggregation iteration is performed to update the vector representation of the node in the graph.
2. The method of claim 1, wherein:

   The vector representation of the word is determined in the following way:

   Using a word vector representation model to determine a vector representation corresponding to the word, and the word vector representation model includes a machine learning model;

   The determining the vector representation corresponding to the knowledge point node based on the vector representation of the word related to the knowledge point node includes:

   Obtain one or more words from the knowledge point corresponding to the knowledge point node;

   Determine the vector representation of the one or more words;

   An operation is performed on one or more of the vector representations, and the operation result is used as a vector representation corresponding to the knowledge point node.
3. The method according to claim 1, wherein the determining the edge weight between the two nodes based on the types of the two nodes comprises:

   If the two nodes are both word nodes, determine the edge weight between the two nodes based on the co-occurrence frequency between the words corresponding to the two nodes;

   If one of the two nodes is a word node and the other node is a knowledge point node, the edge weight between the two nodes is determined based on the importance of the word corresponding to the word node relative to the knowledge point corresponding to the knowledge point node;

   If the two nodes are both knowledge point nodes, the edge weight between the two nodes is determined to be zero.
4. The method according to claim 1, wherein one of the at least one round of graph aggregation iterations includes:

   For any node:

   Based on the association relationship between the node and the node, determine the adjacent node of the node;

   The vector representation of the neighboring node in the current iteration round is weighted based on the edge weight between the node and the neighboring node, and the vector representation of the node is updated with the result of the calculation.
5. The method according to claim 1, wherein one of the at least one round of graph aggregation iterations includes:

   Using the vector representations of the multiple nodes in the current iteration round to obtain a vector representation matrix;

   Determine the adjacency matrix corresponding to the multiple nodes based on the association relationship between the nodes;

   The vector representation matrix and the adjacency matrix are operated on, and the vector representation of each node in the map is updated with the result of the operation.
6. The method according to claim 1, wherein the performing at least one round of graph aggregation iteration based on the vector representation of the node and the association relationship between the nodes to update the vector representation of the nodes in the graph comprises:

   Using the vector representations of the multiple nodes to obtain a vector representation matrix;

   Determine the adjacency matrix corresponding to the multiple nodes based on the association relationship between the nodes;

   Processing the vector representation matrix and the adjacency matrix by using a neural network-based aggregation model to obtain an updated vector representation matrix; the neural network-based aggregation model includes at least GCN or GAT;

   The vector representation of the node in the graph is updated based on the updated vector representation matrix.
7. An information recommendation method using graphs, wherein the method includes:

   Get input information;

   Use the graph to determine the node corresponding to the input information in the graph; the graph is determined by the method according to any one of claims 1-6;

   Determine a recommended node based on the vector representation of the node and the vector representation of the adjacent nodes of the node;

   The information related to the recommended node is output.
8. 8. The method according to claim 7, wherein the input information is one or more words selected by the user from candidate words provided to the user in advance.
9. The method according to claim 7, wherein the information related to the recommending node includes knowledge points related to the recommending node.
10. A system for determining an atlas for information recommendation, wherein the system includes a first acquisition module, a first determination module, and an update module;

    The first acquisition module is configured to acquire a plurality of nodes for constructing a graph; the nodes include at least a word node and a knowledge point node; if the node is a word node, the vector representation of the word corresponding to the node is used as the node If the node is a knowledge point node, determine the vector representation corresponding to the knowledge point node based on the vector representation of the words related to the knowledge point node;

    For any two nodes: the first determining module is configured to determine the edge weight between the two nodes based on the types of the two nodes, and use the edge weight as the value between the two nodes ’S relationship;

    The update module is configured to perform at least one round of graph aggregation iteration based on the vector representation of the node and the association relationship between the nodes to update the vector representation of the nodes in the graph.
11. The system according to claim 10, wherein, in order to obtain a vector representation of a word, the first obtaining module is configured to:

    Using a word vector representation model to determine a vector representation corresponding to the word; the word vector representation model includes a machine learning model;

    In order to determine the vector representation corresponding to the knowledge point node based on the vector representation of the words related to the knowledge point node, the first acquisition module is configured to:

    Obtain one or more words from the knowledge point corresponding to the knowledge point node;

    Determine the vector representation of the one or more words;

    An operation is performed on one or more of the vector representations, and the operation result is used as a vector representation corresponding to the knowledge point node.
12. The system according to claim 10, wherein, to determine the edge weight between the two nodes based on the types of the two nodes, the first determining module is configured to:

    If the two nodes are both word nodes, determine the edge weight between the two nodes based on the co-occurrence frequency between the words corresponding to the two nodes;

    If one of the two nodes is a word node and the other node is a knowledge point node, the edge weight between the two nodes is determined based on the importance of the word corresponding to the word node relative to the knowledge point corresponding to the knowledge point node;

    If the two nodes are both knowledge point nodes, the edge weight between the two nodes is determined to be zero.
13. The system according to claim 10, wherein, to perform one of the at least one round of graph aggregation iteration, the update module is configured to:

    For any node:

    Based on the association relationship between the node and the node, determine the adjacent node of the node;

    The vector representation of the neighboring node in the current iteration round is weighted based on the edge weight between the node and the neighboring node, and the vector representation of the node is updated with the result of the calculation.
14. The system according to claim 10, wherein, to perform one of the at least one round of graph aggregation iteration, the update module is configured to:

    Using the vector representations of the multiple nodes in the current iteration round to obtain a vector representation matrix;

    Determine the adjacency matrix corresponding to the multiple nodes based on the association relationship between the nodes;

    The vector representation matrix and the adjacency matrix are operated on, and the vector representation of each node in the graph is updated with the result of the operation.
15. The system according to claim 10, wherein at least one round of graph aggregation iteration is performed to update the vector representation of the nodes in the graph based on the vector representation of the nodes and the association relationship between the nodes and the nodes, and the update Modules are used for:

    Using the vector representations of the multiple nodes to obtain a vector representation matrix;

    Determine the adjacency matrix corresponding to the multiple nodes based on the association relationship between the nodes;

    Processing the vector representation matrix and the adjacency matrix by using a neural network-based aggregation model to obtain an updated vector representation matrix; the neural network-based aggregation model includes at least GCN or GAT;

    The vector representation of the node in the graph is updated based on the updated vector representation matrix.
16. An information recommendation system using graphs, wherein the system includes a second acquisition module, a second determination module, a third determination module, and an output module;

    The second obtaining module is used to obtain input information;

    The second determining module is configured to use the graph to determine the node corresponding to the input information in the graph; the graph is determined by the method according to any one of claims 1-6;

    The third determining module is configured to determine a recommended node based on the vector representation of the node and the vector representation of the adjacent nodes of the node;

    The output module is configured to output information related to the recommended node.
17. The system according to claim 16, wherein the input information is one or more words selected by the user from candidate words provided to the user in advance.
18. The system according to claim 16, wherein the information related to the recommending node includes knowledge points related to the recommending node.
19. An apparatus for determining an atlas for information recommendation, wherein the apparatus includes a processor, and the processor is configured to execute the method according to any one of claims 1-6.
20. An information recommendation device using graphs, wherein the device includes a processor, and the processor is configured to execute the method according to any one of claims 7-9.

Images