Dialogue-based Recommendation Method and System for Incomplete Knowledge Graphs

Abstract

This paper proposes a conversational recommendation method and system for incomplete knowledge graphs. The method encodes the relationships between product entities in the original incomplete knowledge graph. Based on historical conversational context, the encoded original incomplete knowledge graph is filtered for head and tail entities to identify a conversational context-specific subgraph. The context-specific subgraph is reconstructed using approximate posterior and prior relationships. Recommended responses are generated based on the tail entities in the reconstructed network that are related to the head entity. This method leverages the large-scale conversational corpus naturally associated with conversational recommendation tasks to enhance incomplete knowledge graphs, thereby fully promoting semantic fusion in conversational systems.

Description

Technical Field

[0001] The present invention belongs to the technical field related to personalized recommendation, and in particular to a conversational recommendation method and system for incomplete knowledge graphs. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Recommender systems have become an essential tool for information seekers. Traditional recommendation systems often predict user preferences for products by evaluating historical offline user behavior. Although widely used in industry, they still face two significant challenges: 1. User information learned from historical data is noisy, making traditional recommendation models unable to accurately capture user preferences; 2. Numerous factors influence user consumption behavior in real life, making it difficult for traditional recommendation models to understand the underlying reasons behind user behavior. In general, the static nature of traditional recommendation systems fundamentally limits how users can express their intent, leading to information asymmetry between users and the system. To address these issues, conversational recommender systems (CRSs) have emerged. They infer user interests and recommend products through multiple rounds of online interaction with users. Compared to traditional static recommender systems, conversational recommender systems can more easily capture users' current interests and understand the motivations behind their consumption behavior. Conversational recommender systems can be categorized as attribute-based and topic-guided. Attribute-based recommender systems focus on recommendation strategies, including "ask or recommend," "ask which attributes," or "recommend which products." Currently, most attribute-based conversational recommendation systems are based on deep reinforcement learning frameworks. Topic-guided conversational recommendation systems focus more on leveraging natural language to interact with users, emphasizing fluent responses and precise recommendations. Unlike attribute-based conversational recommendation systems, topic-guided conversational recommendation systems gain greater control over how to influence the direction of the conversation.

[0004] To help systems understand complex conversational semantics, external knowledge, such as knowledge graphs (KGs), has been widely used in conversational recommendation systems, significantly improving their performance. Knowledge graphs typically represent a piece of factual information as a triplet of the form (head entity, relation, tail entity).

[0005] However, the sparsity, incompleteness, and noise of knowledge graphs limit their full utilization by conversational recommendation systems. First, there are a large number of missing relationships in the knowledge graph, resulting in multiple disconnected subgraphs within it. Figure 1As shown, given a knowledge graph, the solid line represents the original relationship in the knowledge graph. It can be seen that the lower half of the knowledge graph is an independent subgraph, which lacks the relationship with the "Dark Knight" entity. This subgraph with missing existence relationship hinders the recommendation task in human-computer dialogue 2, because the "Dark Knight" cannot be inferred based on the original knowledge graph. Traditional path-based and embedding-based knowledge graph completion methods cannot handle multiple disconnected subgraphs: path-based methods cannot effectively learn information between entities due to the lack of key paths, and embedding-based graph completion methods cannot learn graph representations well because they cannot obtain structural information. Secondly, existing conversational recommendation research simply uses the entire knowledge graph to recommend products to users, and cannot adaptively select appropriate knowledge for personalized recommendations. Similarly, Figure 1 As shown in Figure 2, although both human-computer dialogues mention "Wonder Woman", the bold italic text in the dialogues represents different dialogue contexts, so recommendations and replies need to be generated based on context-specific subgraphs. Summary of the Invention

[0006] The present invention mainly solves the problem of a large number of missing relationships and useless noise knowledge in the knowledge graph introduced by the conversational recommendation system. The present invention provides a conversational recommendation method and system for incomplete knowledge graphs, which uses the large-scale conversational corpus naturally carried by the conversational recommendation task to enhance the incomplete knowledge graph, so as to fully promote semantic fusion in the conversational system.

[0007] In order to achieve the above object, the present invention mainly includes the following aspects:

[0008] The first aspect is to provide a conversational recommendation method for incomplete knowledge graphs, including:

[0009] Encode the relationships between product entities in the original incomplete knowledge graph;

[0010] Based on the historical conversation context, the original incomplete knowledge graph after encoding is filtered to select the head entity and tail entity, and determine the entities in the specific subgraph of the conversation context;

[0011] Relationship reconstruction is performed on specific subgraphs in the dialogue context based on the approximate posterior relationship reconstruction network and the prior relationship reconstruction network;

[0012] Generate recommended responses based on the tail entities related to the head entity in the specific subgraph of the reconstructed conversation context.

[0013] The second aspect is to provide a conversational recommendation system for incomplete knowledge graphs, including:

[0014] Encoding module: encodes the relationships between entities in the original incomplete knowledge graph;

[0015] Subgraph determination module: This module filters the head and tail entities of the original incomplete knowledge graph encoded based on the historical conversation context and determines the entities in the specific subgraph of the conversation context.

[0016] Network reconstruction module: Reconstructs the relationship of the specific subgraph in the dialogue context based on the approximate posterior relationship reconstruction network and the prior relationship reconstruction network;

[0017] Recommended reply module: Generates recommended replies by reconstructing the tail entities related to the head entity in the specific subgraph of the conversation context.

[0018] In a third aspect, an embodiment of the present invention provides a computer device comprising: a processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor and the memory communicate through the bus, and when the machine-readable instructions are executed by the processor, the steps of the conversational recommendation method for incomplete knowledge graphs as described in the first aspect above are performed.

[0019] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, executes the steps of the conversational recommendation method for incomplete knowledge graphs as described in the first aspect above.

[0020] One or more of the above technical solutions have the following beneficial effects:

[0021] This invention leverages the large-scale conversational corpus inherent in conversational recommendation tasks to dynamically infer knowledge from incomplete knowledge graphs, completing missing but task-relevant knowledge relationships while adaptively retaining task-relevant knowledge based on the conversational context, thereby enabling personalized recommendations. Unlike existing conversational recommendation methods that treat conversational information as a "query," this invention leverages the large-scale conversational corpus inherent in conversational recommendation tasks to enhance incomplete knowledge graphs, fully promoting semantic fusion within conversational systems.

[0022] The present invention proposes a variational inference method, which dynamically enhances and refines task-advantageous knowledge based on the conversation context, and infers the conversation context-specific subgraph of each interaction in an unsupervised manner to achieve personalized recommendations.

[0023] Advantages of additional aspects of the present invention will be given in part in the following description and in part will be obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0025] Figure 1 The conversational examples listed for the present invention;

[0026] Figure 2 This is a framework diagram of the conversational recommendation method for incomplete knowledge graphs in Example 1 of the present invention. DETAILED DESCRIPTION

[0027] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0028] It should be noted that the terms used herein are for describing particular embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0029] In the absence of conflict, the embodiments of the present invention and the features thereof may be combined with each other.

[0030] Example 1

[0031] like Figure 2 As shown, this embodiment discloses a conversational recommendation method for incomplete knowledge graphs, including:

[0032] Step 1: Encode the relationships between product entities in the original incomplete knowledge graph;

[0033] Step 2: Based on the historical conversation context, the original incomplete knowledge graph after encoding is filtered to select the head entity and tail entity, and determine the entities in the specific subgraph of the conversation context;

[0034] Step 3: Reconstruct the relationship of the specific subgraph in the dialogue context based on the approximate posterior relationship reconstruction network and the prior relationship reconstruction network;

[0035] Step 4: Generate recommended responses based on the tail entities related to the head entity in the reconstructed conversation context-specific subgraph.

[0036] Figure 1As shown in Figure 2, based on the dialogue data, the missing relationship between "The Dark Knight" and "superhero" can be explicitly inferred, which is the dotted line. This inferred relationship is then utilized in human-computer dialogue 2 to help the system successfully recommend the movie "The Dark Knight." The utilized knowledge subgraph is shown in the shaded area in human-computer dialogue 2. At the same time, as shown in human-computer dialogue 1, relevant knowledge is dynamically selected based on the specific dialogue context, as shown in the shaded area, thereby reducing the noise in the knowledge graph caused by useless knowledge, such as the gray-shaded area representing "Fast and Furious" and "Gal Gadot."

[0037] The main purpose of this embodiment is to combine the large-scale dialogue corpus naturally carried in conversational recommendation tasks, perform dynamic knowledge reasoning on incomplete knowledge graphs, complete missing but task-useful knowledge relationships, and adaptively retain task-related knowledge based on the dialogue context, thereby performing personalized recommendations.

[0038] This example uses a knowledge graph subgraph specific to the conversation context as a discrete latent variable, the recommended items as observed variables, the conversation history and the original knowledge graph as conditions, and performs reasoning using variational Bayes. The stochastic gradient variational Bayes estimator is used to optimize the derived evidence lower bound (ELBO).

[0039] Latent variable: latent variable, or latent variable, latent variable, in statistics represents unobservable random variable, as opposed to observed variable.

[0040] In step 1 of this embodiment, first, based on a given specific conversation turn and the corresponding conversation history word sequence and target recommendation product word sequence, the conversation history encoding stage first uses the BERT encoder to encode the conversation history into a representation vector d t Next, the conversation history and the target recommended product word sequence are concatenated, and the two sequences are separated by special words, namely special tokens, and then input into the same BERT encoder to obtain the concatenated vector

[0041] In this embodiment, RGCN is used as the basic framework to encode the entities and the relationships between entities in the original incomplete knowledge graph.

[0042] Specifically, for any entity e in the graph, the process of propagating information from neighboring entity e′ to e is as follows:

[0043]

[0044] Among them, σ represents the si gmo id activation function, is the vector representation of entities e, e', represents the set of neighbor entities of e specific to relation r, represents a set of relations, W r and W is the learnable parameter matrix, Z e,r is the normalization factor.

[0045] It should be noted that entity e refers to any entity in the graph. For example, if the recommended product is a movie, the entity can be all movie-related nouns such as movie style, genre, director, actor, movie, etc.

[0046] In step 2 of this embodiment, in order to improve computational efficiency and identify more important entities, this embodiment assumes that only entities that appear in the conversation history can become the head entity e in the specific subgraph of the conversation context. h , for the tail entity e t Filtered from two strategies:

[0047] (1) Strategy 1 uses mutual information to measure the probability of each entity e being selected as the tail entity. The calculation formula is as follows:

[0048]

[0049] Among them, P o (e) is the probability of occurrence of entity e, P(e|e h ) is the entity P(e|e h ) with the given header entity e h The co-occurrence probability of ,both of which are statistically derived from the training dataset.,This embodiment selects the top k entities with high probability and adds them to the,entity set at the end of the subgraph.

[0050] (2) In strategy 2, the entities that have relationships with the head entity in the original incomplete knowledge graph are retained in the tail entity set.

[0051] It should be noted that knowledge graphs generally represent a piece of factual information as a triple of the form (head entity, relationship, tail entity). Therefore, the head and tail entities are ordinary entities selected by the strategy, and the head and tail specifically refer to their position in the knowledge.

[0052] In step 3 of this embodiment, a relation reconstruction network is trained to infer relations within a specific subgraph in a conversational context. This allows for dynamic knowledge inference based on the context. Specifically, given a set of entities and their corresponding entity representations within a specific subgraph in a conversational context, the relation reconstruction task is defined as a classification task. The network is reconstructed using prior relations and approximate posterior relations.

[0053] Both the prior and approximate posterior reconstruction networks are based on multi-layer perceptrons (MLPs). In order to fully integrate the semantics of the head and tail entities and the conversation history, the representation vector e of the given entity pair and the conversation history is h ,e t ,d t , the intermediate vector m of the prior relationship reconstruction network is calculated as follows:

[0054] f o1 =[e h ;e t ;d t ],

[0055] f o2 =[e h ⊙e t ;e t ⊙d t ;e h ⊙d t ],

[0056] f o3 =e h ⊙e t ⊙d t ,

[0057] m=[f o1 ;f o2 ;f o3 ]W o , (3)

[0058] Among them, [·;·] represents vector connection, ⊙ represents dot product, W o is a learnable parameter matrix. Then, the vector m is mapped to the dimension of the number of relationship types through a multi-layer perceptron, which means the prior probability distribution of the relationship type between the entity pair, i.e., the head entity and the tail entity, is obtained. The calculation method is:

[0059]

[0060] in, Represent the conversation history and the original incomplete knowledge graph respectively.

[0061] In order to adapt to the downstream recommendation and generation subtasks, this embodiment simplifies the relationship between entities into two types, namely, whether there is a connection between the head and tail entities. If there is a connection, it is represented as Otherwise = 0.

[0062] The structure of the approximate posterior relation reconstruction network is similar to that of the prior network. The approximate posterior relation reconstruction network can additionally perceive the target recommended products. Therefore, the d in the calculation formula (3) is t Replace with Get vector m′. During the training process, in the approximate posterior distribution In the training process, the relationships between the head and tail entities are sampled, but during the test, only the prior distribution is calculated and the relationships between entities are sampled based on the prior distribution.

[0063] Because it is assumed that the conversation context specific subgraph is inferred based on the original knowledge graph, in order to better utilize the information of the original knowledge graph, Kullback-Leibler divergence is used as the regularization loss to pre-train the above RGCN and relation reconstruction network. The loss calculation is as follows:

[0064]

[0065] in, is the prior probability distribution of the head entity and tail entity types calculated by formula (4), is the approximate posterior distribution between the head and tail entities obtained by the above calculation.

[0066] If the head and tail entities have a relationship in the original knowledge graph, then Otherwise = 0. This regularization loss forces the relation reconstruction network to not lose the relations in the original knowledge graph. During the fine-tuning phase, the regularization loss is still retained as a regularization constraint for the knowledge graph reconstructor to prevent it from overfitting.

[0067] In step 4 of this embodiment, in the recommendation phase, given the inferred dialog context specific subgraph, the goal of this embodiment is to select accurate products to recommend to the user. u Specifically, the probability of the "existence of connection" type of relationship in the approximate posterior relationship reconstruction network is used as the weight of the tail entity vector representation, and the weighted sum is used to obtain the user representation. The calculation process is as follows:

[0068]

[0069] Since items connected to the head entity in the specific subgraph of the dialogue context are more likely to be recommended, we calculate all candidate items I i The recommendation probability is as follows:

[0070]

[0071] in, represents a specific subgraph in the dialogue context, Z rec is the normalization factor, α is the weight hyperparameter, e i It is commodity I i The vector representation of . If the product I i If it is not in the tail entity set,

[0072] During testing, the probability of reconstructing the network using the prior relationship Replace the posterior probability distributions in (6) and (7) in the above two equations.

[0073] The goal of this example is to maximize the recommendation probability of the target product while minimizing the KL divergence between the prior and approximate posterior networks, while also maintaining the regularization constraint. Therefore, the final target loss function of the recommendation task is:

[0074]

[0075] Among them, β, γ and λ are hyperparameters. The probability of the target recommended product being recommended.

[0076] In the reply generation stage, the parameters within BERT and the knowledge graph reconstructor in the dialogue history encoding stage are frozen, and the tail entities that are inferred by the prior relationship reconstruction network to be related to the head entity are filtered out for generating reply sentences, believing that they help generate context-consistent and diverse replies.

[0077] The decoder in the reply generation phase follows the transformer framework, which is the current mainstream generative model and contains multiple identical decoding layers. However, this embodiment improves the basic transformer framework to enhance the role of knowledge in the decoding process. Specifically, the representations of the head entity and the filtered tail entity that appear in the conversation history extracted in step 2 are concatenated into a matrix and After the self-attention calculation, the entity representation matrix is fused through two knowledge-based attention layers. The calculation process is as follows:

[0078]

[0079]

[0080]

[0081]

[0082]

[0083] Where MHA[,,] represents the multi-head attention mechanism. FFN(·) represents the fully connected feedforward network, X is the conversation history word vector representation matrix obtained by the transfer encoder, and C (l-1)represents the output matrix of the lth decoding layer and serves as the input matrix of the l+1th decoding layer. In addition, a replication mechanism is used during the decoding process to generate a reply containing more information.

[0084] It should be noted that to facilitate sentence generation, X here is obtained by tokenizing the conversation history sentences using a different dictionary (different from the one used by the BERT encoder above) and re-encoding them using the transfer encoder. This is a vector matrix, meaning each word corresponds to a vector. Unlike the conversation history vectors generated by the BERT encoder above, X here corresponds to a vector for each sentence.

[0085] Assuming the length of the target response is n, the cross entropy loss function is used to optimize the parameters in the dialogue generation process. The loss calculation is as follows:

[0086]

[0087] in, represents a specific subgraph in the dialogue context, D t is the conversation history, w j is the reply word predicted at time j, w 0:j-1 is the historical reply word predicted before time j.

[0088] Example 2

[0089] This embodiment provides a conversational recommendation system for incomplete knowledge graphs, including:

[0090] Encoding module: encodes the relationships between entities in the original incomplete knowledge graph;

[0091] Subgraph determination module: This module filters the head and tail entities of the original incomplete knowledge graph encoded based on the historical conversation context and determines the entity set in the specific subgraph of the conversation context;

[0092] Network reconstruction module: Reconstructs the relationship of the specific subgraph in the dialogue context based on the approximate posterior relationship reconstruction network and the prior relationship reconstruction network;

[0093] Recommended reply module: Generates recommended replies by reconstructing the tail entities related to the head entity in the specific subgraph of the conversation context.

[0094] Example 3

[0095] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the steps of the above method are implemented when the processor executes the program.

[0096] Example 4

[0097] The purpose of this embodiment is to provide a computer-readable storage medium.

[0098] A computer-readable storage medium stores a computer program, which, when executed by a processor, performs the steps of the above method.

[0099] The steps involved in the apparatuses of Examples 2, 3, and 4 above correspond to those of Method Example 1. For detailed implementations, please refer to the relevant description of Example 1. The term "computer-readable storage medium" should be understood to mean a single medium or multiple media containing one or more instruction sets; it should also be understood to include any medium capable of storing, encoding, or carrying an instruction set for execution by a processor and causing the processor to perform any of the methods of the present invention.

[0100] Those skilled in the art will appreciate that the modules or steps of the present invention described above can be implemented using a general-purpose computer device. Alternatively, they can be implemented using program code executable by a computing device, which can then be stored in a storage device and executed by the computing device. Alternatively, they can be fabricated into separate integrated circuit modules, or multiple modules or steps can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0101] Although the above describes the specific embodiments of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without any creative work are still within the scope of protection of the present invention.

Claims

1. A conversational recommendation method for incomplete knowledge graphs, characterized by: include: Encode the relationships between product entities in the original incomplete knowledge graph; Based on the historical conversation context, the original incomplete knowledge graph after encoding is filtered to select the head entity and tail entity, and determine the entities in the specific subgraph of the conversation context; Relationship reconstruction is performed on specific subgraphs in the dialogue context based on the approximate posterior relationship reconstruction network and the prior relationship reconstruction network; Among them, the relationship reconstruction of the specific subgraph of the network dialogue context based on the approximate posterior relationship reconstruction is specifically: based on the splicing vector e of entity pairs and dialogue history h ,e t , Calculate the intermediate vector m′ of the approximate posterior relationship reconstruction, map the intermediate vector m′ to the dimension of the number of relationship types through a multi-layer perceptron, and obtain the posterior probability of the reconstructed relationship between the head and tail entities; use the posterior probability as the weight of the tail entity representation, and obtain the user representation by weighted summation of the tail entity representation weights; the concatenated vector of the conversation history is obtained by concatenating the representation vector of the conversation history with the target recommended word sequence; Relationship reconstruction is performed on the specific subgraph of the dialogue context based on the prior relationship reconstruction network, specifically: the representation vector e based on the entity pair and dialogue history h ,e t ,d t Calculate the intermediate vector m in the specific subgraph of the dialogue context for prior relationship reconstruction, map the intermediate vector m to the dimension of the number of relationship types through a multi-layer perceptron, and obtain the prior probability distribution of the relationship between entities; Generate recommended responses based on the tail entities related to the head entity in the specific subgraph of the reconstructed conversation context.

2. The conversational recommendation method for incomplete knowledge graphs according to claim 1, characterized in that: RGCN is used to encode the relationship between entities in the original incomplete knowledge graph, specifically: Where e is any entity, represents the set of neighbor entities of e specific to relation r, represents a set of relations, W r and W is the learnable parameter matrix, Z e,r is the normalization factor, and σ represents the sigmoid activation function.

3. The conversational recommendation method for incomplete knowledge graphs according to claim 1, wherein: Based on the historical dialogue context, the original incomplete knowledge graph after encoding is filtered to filter the head entity and tail entity, and determine the entities in the specific subgraph of the dialogue context, specifically: Use entities in historical conversations as head entities of a specific subgraph in the conversation context; Filter tail entities in a specific subgraph of a conversation context based on the head entity.

4. The conversational recommendation method for incomplete knowledge graphs according to claim 1, wherein: Filter the tail entity in a specific subgraph of the conversation context based on the head entity, specifically: Mutual information is used to measure the probability of each entity becoming a tail entity, and the entities with the top k probabilities are selected as the tail entities of the specific subgraph in the dialogue context. The probability calculation formula is: Among them, P o (e) is the probability of occurrence of entity e, P(e|e h ) is the entity e with the given head entity e h The co-occurrence probability of .

5. The conversational recommendation method for incomplete knowledge graphs according to claim 1, wherein: The entities that have relationships with the head entity in the original incomplete knowledge graph are used as the tail entities of the specific subgraph in the dialogue context.

6. The conversational recommendation method for incomplete knowledge graphs according to claim 1, wherein: The final objective loss function of the recommendation task is: Among them, β, γ and λ are hyperparameters. represents a specific subgraph of the dialogue context, For commodity I t The recommendation probability, is the prior probability, is the posterior probability.

7. The conversational recommendation method for incomplete knowledge graphs according to claim 1, characterized in that: In the recommended response, the representation of the head entity and the filtered tail entity inferred by the prior relationship reconstruction network as having a relationship with the head entity are concatenated into a matrix and Entity representation matrices are fused through two knowledge-based attention layers and 8. A conversational recommendation system for incomplete knowledge graphs, characterized by: include: Encoding module: encodes the relationships between entities in the original incomplete knowledge graph; Subgraph determination module: This module filters the head and tail entities of the original incomplete knowledge graph encoded based on the historical conversation context and determines the entities in the specific subgraph of the conversation context. Network reconstruction module: Reconstructs the relationship of the specific subgraph in the dialogue context based on the approximate posterior relationship reconstruction network and the prior relationship reconstruction network; Among them, the relationship reconstruction of the specific subgraph of the network dialogue context based on the approximate posterior relationship reconstruction is specifically: based on the splicing vector e of entity pairs and dialogue history h ,e t , Calculate the intermediate vector m′ of the approximate posterior relationship reconstruction, map the intermediate vector m′ to the dimension of the number of relationship types through a multi-layer perceptron, and obtain the posterior probability of the reconstructed relationship between the head and tail entities; use the posterior probability as the weight of the tail entity representation, and obtain the user representation by weighted summation of the tail entity representation weights; the concatenated vector of the conversation history is obtained by concatenating the representation vector of the conversation history with the target recommended word sequence; Relationship reconstruction is performed on the specific subgraph of the dialogue context based on the prior relationship reconstruction network, specifically: the representation vector e based on the entity pair and dialogue history h ,e t ,d t Calculate the intermediate vector m in the specific subgraph of the dialogue context for prior relationship reconstruction, map the intermediate vector m to the dimension of the number of relationship types through a multi-layer perceptron, and obtain the prior probability distribution of the relationship between entities; Recommended reply module: Generates recommended replies by reconstructing the tail entities related to the head entity in the specific subgraph of the conversation context.

9. A computer device, characterized in that: include: A processor, a memory, and a bus, wherein the memory stores machine-readable instructions executable by the processor. When the computer device is running, the processor and the memory communicate via the bus. When the machine-readable instructions are executed by the processor, the steps of the conversational recommendation method for incomplete knowledge graphs as described in any one of claims 1 to 7 are performed.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, executes the steps of the conversational recommendation method for incomplete knowledge graphs as described in any one of claims 1 to 7.

Images