A multi-party human-computer interaction reply personification method based on a knowledge graph

Abstract

The application relates to a multi-party human-computer interaction reply personification method based on a knowledge graph, and belongs to the field of human-computer interaction. The method comprises the following steps: S1: multi-party human-computer interaction emotional state evaluation; and S2: knowledge graph water ripple propagation. The method adopts a knowledge graph representation mode to mine similar experiences of multiple parties to find potential interesting dialogue content, similar experience promotes users to share emotional experience and group feelings, additional emotional information is introduced into a model by using a VAD emotional dictionary to be formally represented, a robot can feel emotions of multiple parties, and the emotions are fed back to the multiple parties, so that more suitable emotional and more content-rich replies are generated. Experiments are carried out on multi-party dialogue data sets, and experimental results show that the multi-party human-computer interaction system with emotional temperature and knowledge depth can attract the attention of the rest of the parties and become the focus of the multiple parties.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of human-computer interaction, and relates to a multi-party human-computer interaction reply personification method based on a knowledge graph. BACKGROUND

[0002] Early multi-party human-computer dialogue systems focus on using hand-crafted rules and templates to help robots learn, and identify user intentions through artificially summarized semantic analysis rule templates, such as a multi-party dialogue management game robot based on probability and rules, which mainly uses preset dialogue templates to generate appropriate replies, and a social robot based on hand-crafted rules, which controls dialogue turns to enable the robot to participate in multi-party dialogue. However, most of them have poor experience effects, poor domain transferability, and cannot meet actual entertainment and commercial needs. Thanks to the rapid development of deep learning technology, recent research focuses on accelerating the intelligent learning of robots through deep learning technology to achieve the intelligence of robots in speaker recognition, listener recognition and content reply. In the research on speaker recognition, through the analysis of the AMI and MPR two data sets, the correlation-based feature selection is used to predict the turning change and the next speaker. In the research on listener recognition, the dialogue sentence and the state of the speaker are encoded, the interactive representation of the dialogue sentence and the user is learned, and the listener is displayed and recognized; in the research on content reply, one is to model the multi-party participant dialogue information flow through a graph structure, the purpose of which is to effectively use the dialogue structure to generate replies, and the other is to organize the historical message dialogue of the multi-party dialogue into a tree structure according to the “@” reply relationship, then divide the tree structure into multiple sequence combinations, and then use the dialogue of the branch where the speaker is located to generate a reply.

[0003] Although the above work uses deep learning technology to help robots learn intelligently, it does not introduce a complete knowledge system and rich emotional expression, which will be easily marginalized and cannot become the focus of multi-party. SUMMARY

[0004] Therefore, the present application aims to provide a multi-party human-computer interaction reply personification method based on a knowledge graph, which uses a knowledge graph representation method to mine the similar experiences of multi-parties to find potential interesting dialogue content, and similar experience promotes users to share emotional experience and group feelings, uses a VAD emotion dictionary to introduce additional emotional information into the model to formalize the representation, so that the robot can feel the emotions of multi-parties and feedback the emotions to multi-parties, thereby generating more suitable replies in terms of emotions and more rich replies in terms of content.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0006] A knowledge graph-based method for anthropomorphizing multi-party human-computer interaction responses, comprising the following steps:

[0007] S1: Multi-party human-computer interaction emotional state assessment;

[0008] S2: Knowledge Graph Ripple Propagation;

[0009] Specifically, S1 is:

[0010] Using the NRC_VAD sentiment dictionary from psychology, sentiment is quantified and calculated. Combining the specific meanings of VAD sentiments with TF-IDF technology, the overall sentiment embedding of dialogue statements is calculated, as shown in the following formula:

[0011]

[0012] in, It is the emotional embedding of words. and Here, TF and IDF are the values ​​of the i-th word in sentence u, respectively. TF represents the frequency of a word in a one-party dialogue. The higher the frequency, the more important the word is to that party. IDF measures the probability of a word appearing in all multi-party human-computer dialogues. If a word appears in all multi-party human-computer dialogues, then the word is not important. and The calculation formulas are shown in equations (2) and (3) respectively:

[0013]

[0014] Wherein, c(u i ) is the word u i Word frequency in a one-sided dialogue, s u It refers to the length of the conversation between one party;

[0015]

[0016] Where D is the sum of all input and response sentences in the training dataset, n(u i This indicates that all input and response sentences contain the word "u". i The total number of sentences;

[0017] A positive input often produces a positive response, making the multi-party human-computer interaction more positive; conversely, a negative input often produces a negative response, making the multi-party human-computer interaction more negative. This dialogue process aligns with the psychological process of actual human-to-human communication. Based on this psychological process, the formula for measuring the similarity of the V-value sentiment between the response and the input is as follows:

[0018]

[0019] Among them, A x A represents the sentiment value in the robot's response. y This represents the sentiment value input from a multi-party dialogue, with the subscript 1 indicating the Manhattan distance;

[0020] When both input and response have a strong desire to dominate, conflict will arise between the dominant party and the robot. The greater the distance between the A and D values ​​of the input and response, the better. The distance between the A and D values ​​of the response and the input is measured by equations (5) and (6), respectively:

[0021]

[0022]

[0023] Combining formulas (4), (5), and (6), the final sentiment score of the robot's candidate responses is defined as follows:

[0024] R e =R v +R A +R D (7)

[0025] Normalizing this, the sentiment affinity of the robot's candidate responses is defined as:

[0026]

[0027] Among them, R min and R max These are the minimum and maximum sentiment scores among the candidate responses;

[0028] S2 is:

[0029] First, topic entities are extracted from the multi-party dialogue content, mapped onto the knowledge graph, and relevant knowledge entities are activated. These knowledge entities are regarded as a set of user interest seeds.

[0030] Secondly, based on the idea of ​​water ripple propagation, it propagates outward along the links in the knowledge graph, and selects the next layer's neighboring entities as propagation entities for the seed entities in the inner layer in turn, forming a multi-party preference distribution for potential seed topics.

[0031] Then, after the seed entity propagates to the H layer along different links, it aggregates the entity information of the current layer and its neighboring entities sequentially from the outside to the inside, starting from the H-1 layer, in order to update the entity representation of the current layer and use it for the next aggregation. The seed entity that aggregates the neighbor information of the H layer is used to represent the final potential interest topic.

[0032] Finally, taking into account the emotional inclinations and interests of various parties, the candidate response with the highest score is selected as the robot's optimal response.

[0033] Optionally, S2 specifically includes:

[0034] First, define the set of multi-party dialogue topic entities obtained in k rounds of dialogue:

[0035] C k ={c k |c k ∈G} (9)

[0036] Where G is a known knowledge graph, c k For multi-party dialogue topic entities, k represents the dialogue round; the knowledge graph triple is represented as G = (h, r, t), i.e., head entity - relation - tail entity, for the acquired multi-party dialogue topic entity set C. k Definition of triplet ripple set:

[0037]

[0038] Where n represents the number of outward propagation layers, such as It is a collection of topics involved in multi-party dialogue. This indicates the first-level related entity that propagates outward from the topic entity in the first round of dialogue;

[0039] Secondly, candidate responses are retrieved from the dialogue corpus, and the response content is sequentially processed using the word2Vec method and the Embedding Average method to obtain the vector representation v∈K of the candidate responses. d Where K is the word embedding matrix and d is the dimension of the word embedding; the vector representation v of the candidate response is coupled with the head entity h. i and relation r i Comparison and calculation of the ripple set after k rounds of multi-party dialogue. Each triple (h) in a first-order related entity i ,r i ,t i The correlation probability P between the sentence feature vector v and the sentence feature vector v i :

[0040]

[0041] Among them, R i ∈K d×d The central relation r of the knowledge graph ripple i The vector representation of H i ∈k d The knowledge graph ripple focuses on the head entity h. iThe vector representation is achieved using the TransD, P knowledge representation method embedded in knowledge graphs. i Viewed as in relation space R i ∈K d×d In the process, the vector representation v of the candidate response and the entity vector representation H i The probability of similarity;

[0042] After obtaining the relevant probabilities, the topics of the multi-party dialogue are then aggregated to form the first-order related tail entity T. i The impact garnered first-order responses from multiple parties in the dialogue. The calculation formula is as follows:

[0043]

[0044] Where T i ∈K d It is the vector representation of the tail entity;

[0045] Then, replace v with Calculate in sequence This process is performed iteratively on the ripple set, so that topics of interest to multiple parties propagate from near to far along the hierarchical relationship of the knowledge graph; the multi-party interest preference vector u is defined as:

[0046]

[0047] In this regard, considering that an excessively large order response would dilute useful information and introduce more noise, the value of H is set to 2;

[0048] Finally, given the multi-party interest preference vector u and the candidate response vector v i Under these conditions, the predictable interest level I of multiple parties in a potential topic is:

[0049] I=σ(u T v) (14)

[0050] Where σ is the sigmoid function, as shown in the formula below:

[0051]

[0052] Combining formulas (8) and (14), the candidate responses are weighted by content interest and emotional friendliness, and normalized as follows:

[0053] y=αR+βI (16)

[0054] Here, α and β are constraint factors, and α+β=1; the range of y is [0,1], and the closer the value is to 1, the more satisfied the parties are with the candidate response.

[0055] The beneficial effects of this invention are as follows:

[0056] Experiments were conducted on a multi-party dialogue dataset. The results show that a multi-party human-computer interaction system that combines emotional warmth and knowledge depth can attract the attention of other parties and become the focus of attention.

[0057] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0058] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0059] Figure 1 The knowledge graph is a water ripple propagation process; (a) is a dialogue topic entity; (b) is a dialogue topic first-order related entity; (c) is a dialogue topic second-order related entity; (d) is a dialogue topic n-order related entity. Detailed Implementation

[0060] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0061] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0062] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0063] This invention mainly consists of two parts: multi-party human-computer interaction emotional state assessment and knowledge graph water ripple propagation.

[0064] (1) Multi-party human-computer interaction emotion state assessment

[0065] To enable robots to perceive and express emotions, this invention utilizes the NRC_VAD sentiment dictionary, widely adopted in psychology and available in multiple languages, for better quantification and calculation of emotions. The NRC_VAD sentiment dictionary is a VAD sentiment dictionary containing nearly 20,000 words. VAD includes three sentiment dimensions: Valence, Arousal, and Dominance, with values ​​ranging from 0 (weakest) to 1 (strongest). Valence represents the degree of negativity or positiveness of a word, Arousal represents the degree of passivity or activeness, and Dominance represents the degree of submissiveness or dominance. To achieve amicable multi-party communication, the overall sentiment embedding of dialogue sentences is calculated by combining the specific meanings of VAD emotions with TF-IDF (term frequency-inverse document frequency) technology, as shown in the following formula:

[0066]

[0067] in, It is the emotional embedding of words. and Here, TF and IDF values ​​are the i-th word in sentence u, respectively. TF represents the frequency of a word in a one-party dialogue; a higher frequency indicates greater importance of the word to that party. IDF measures the probability of a word appearing in all multi-party human-computer dialogues; if a word appears in all multi-party human-computer dialogues, then that word may not be very important. and The calculation formulas are shown in equations (2) and (3) respectively:

[0068]

[0069] Wherein, c(u i ) is the word u i Word frequency in a one-sided dialogue, s u It refers to the length of the dialogue content on one side.

[0070]

[0071] Where D is the sum of all input and response sentences in the training dataset, n(u i This indicates that all input and response sentences contain the word "u". i The total number of sentences.

[0072] A positive input often produces a positive response, making the human-computer interaction more positive; conversely, a negative input often produces a negative response, making the human-computer interaction more negative. This dialogue process aligns with the psychological processes involved in actual human communication. Based on this psychological process, the formula for measuring the similarity of the V-value sentiment between the response and the input is as follows:

[0073]

[0074] Among them, A x A represents the sentiment value in the robot's response. y This represents the sentiment value input from the multi-party dialogue, with the subscript 1 indicating the Manhattan distance.

[0075] When both input and response exhibit strong dominance, conflict arises between the dominant party and the robot. Therefore, to ensure harmonious multi-party dialogue, the greater the distance between the A and D values ​​of the input and response, the better. Based on the above process, the distance between the A and D values ​​of the response and the input is measured as shown in equations (5) and (6), respectively:

[0076]

[0077]

[0078] Combining formulas (4), (5), and (6), the final sentiment score of the robot's candidate responses is defined as follows:

[0079] R e =R v +R A +R D (7)

[0080] Normalizing this, the sentiment affinity of the robot's candidate responses is defined as:

[0081]

[0082] Among them, R min and R max These are the minimum and maximum sentiment scores among the candidate responses.

[0083] (2) Knowledge Graph Ripple Propagation

[0084] In multi-party human-computer interaction, appropriately introducing new topics can often be eye-catching and attract attention. To characterize the interests and preferences of multiple parties and identify potential topics of interest, the following steps are taken: First, topic entities are extracted from the multi-party dialogue content and mapped onto a knowledge graph. Relevant knowledge entities are activated, and these entities are considered as a set of user interest seeds. Second, based on the ripple propagation concept, the knowledge graph is continuously propagated outward along links, sequentially selecting neighboring entities of the next layer from the seed entities in the inner layer as propagation entities. This forms a distribution of preferences for potential seed topics among multiple parties, facilitating a smooth transition to the topics that may be discussed. Then, after the seed entities propagate along different links to layer H, they are aggregated from the outside inward, starting from layer H-1, with the entities in this layer and their neighboring entity information. This aggregated entity representation is used for the next aggregation, and the seed entities with aggregated neighboring information from layer H represent the final potential topics of interest. Finally, considering the emotional tendencies and interests of multiple parties, the candidate response with the highest score is selected as the robot's optimal response.

[0085] like Figure 1 The diagram illustrates the ripple propagation process of a knowledge graph. (a) represents the dialogue topic entity; (b) represents the first-order related entity of the dialogue topic; (c) represents the second-order related entity of the dialogue topic; and (d) represents the nth-order related entity of the dialogue topic. The topic information of multi-party dialogues is mapped onto the knowledge graph, much like a stone thrown into a pond, and the phenomenon is as follows... Figure 1 As shown in (a), ripples spread across a calm water surface, forming concentric circles that represent the distribution of preferences for candidate topics. When multiple ripples meet and overlap, the amplitude in certain specific areas increases, resulting in enhanced interference. Figure 1 As shown in (b), the darker areas indicate a higher intensity of multi-party interest, which can be used as potential interest content to guide users. However, as the ripples spread further and further, the amplitude of the ripples also weakens. As the association order increases, the association between the central topic entity and the surrounding topic entities gradually weakens, and the intensity of multi-party potential interest topics also gradually decreases.

[0086] First, define the set of multi-party dialogue topic entities obtained in k rounds of dialogue:

[0087] C k ={c k |c k ∈G} (9)

[0088] Where G is a known knowledge graph, c k Let C represent the multi-party dialogue topic entities, and k represent the dialogue round. A knowledge graph triple is represented as G = (h, r, t), i.e., head entity - relation - tail entity, for the acquired multi-party dialogue topic entity set C. k Definition of triplet ripple set:

[0089]

[0090] Where n represents the number of outward propagation layers, such as It is a collection of topics involved in multi-party dialogue. This indicates the first-level related entity that propagates outward from the topic entity in the first round of dialogue.

[0091] Secondly, candidate responses are retrieved from the dialogue corpus, and the response content is sequentially processed using the word2Vec method and the Embedding Average method to obtain the vector representation v∈K of the candidate responses. d Where K is the word embedding matrix and d is the dimension of the word embeddings. The vector representation v of the candidate response is then compared with the head entity h. i and relation r i By comparison, the ripple set S can be calculated after k rounds of multi-party dialogue. k Each triple (h) in a first-order related entity of 1 i ,r i ,t i The correlation probability P between the sentence feature vector v and the sentence feature vector v i :

[0092]

[0093] Among them, R i ∈K d×d The central relation r of the knowledge graph ripple i The vector representation of H i ∈k d The knowledge graph ripple focuses on the head entity h. i The vector representation is achieved using the TransD, P knowledge representation method embedded in knowledge graphs. i It can be viewed as being in the relation space R i ∈K d×d In the process, the vector representation v of the candidate response and the entity vector representation H i The similarity probability.

[0094] After obtaining the relevant probabilities, the topics of the multi-party dialogue are then aggregated to form the first-order related tail entity T. i The impact garnered first-order responses from multiple parties in the dialogue. The calculation formula is as follows:

[0095]

[0096] Where T i ∈K d It is the vector representation of the tail entity.

[0097] Then, replace v with Calculate in sequence This process is performed iteratively on the ripple set, allowing topics of interest to propagate outwards along the hierarchical relationships of the knowledge graph. In summary, the multi-party interest preference vector u can be defined as:

[0098]

[0099] In this regard, considering that an excessively large order response would dilute useful information and introduce more noise, the value of H is set to 2.

[0100] Finally, given the multi-party interest preference vector u and the candidate response vector v i Under these conditions, the predictable interest level I of multiple parties in a potential topic is:

[0101] I=σ(u T v) (14)

[0102] Where σ is the sigmoid function, as shown in the formula below:

[0103]

[0104] Combining formulas (8) and (14), the candidate responses are weighted by content interest and emotional friendliness, and normalized as follows:

[0105] y=αR+βI (16)

[0106] Here, α and β are constraint factors, and α + β = 1. The value range of y is [0, 1], and the closer the value is to 1, the more satisfied the parties are with the candidate response.

[0107] 1. Introduction to the Experiment

[0108] To more intuitively evaluate and analyze the performance of the proposed model, a text-based multi-party chatbot was built using the Chatterbot and wxpy libraries in Python. First, the input content was mapped onto a knowledge graph for entity extraction and linking. Second, multi-party sentiment analysis was performed based on the VAD sentiment dictionary. Finally, considering both interest preference and sentiment friendliness, the candidate response with the highest score was selected as the chatbot's output.

[0109] 2. Experimental Data

[0110] This dataset mainly comes from Chinese online forums such as Douban Groups and Baidu Tieba. Each dialogue statement includes the speaker ID, posting time, and dialogue content. The total number of dialogue rounds is 88,619, and the total number of speakers is 53,581. Each dialogue has a minimum of 3 rounds and a minimum of 3 speakers. 80% of the dataset is used as the training set, and 20% is used as the test set.

[0111] 3. Evaluation Indicators

[0112] The accuracy of the model proposed in this invention is automatically evaluated using two internationally recognized metrics: Mean Reciprocal Rank (MRR) and Mean Average Precision (MAP).

[0113] In addition to the automatic evaluation method, this invention also employs a manual evaluation method to further verify the effectiveness of the proposed model in multi-party interaction processes. Thirty-three volunteers were invited through multiple channels to conduct multi-party human-computer interaction experiments under different models, and the number and duration of interactions were evaluated. The volunteers were randomly divided into eleven groups of three people each, including 21 males and 12 females.

[0114] 4. Experimental Results and Analysis

[0115] This invention uses two metrics, MRR and MAP, for objective evaluation. Taking into account the time complexity of retrieval and the response quality of candidate content, the number of candidate response sets is n=10. The results are shown in Table 1.

[0116] Table 1. Automatic evaluation results of different comparison models

[0117]

[0118] As shown in Table 1, compared with the other four models, the model proposed in this invention achieved more satisfactory results. This is mainly because the proposed model considers interpersonal communication when ranking candidate responses, requiring both emotional warmth and knowledge depth. Constraints on response content were imposed from the perspectives of interest preference and emotional friendliness. Among them, the CATD, GSN, and MSA-CI models all achieved better results compared to the ChatterBot model. The main reason is that the CATD model considers the topic preferences of multiple parties, resulting in a higher matching degree between candidate topics and context topics, thus improving the relevance of the content response; the GSN model considers the conversational information flow of multiple parties, making the content response more coherent by predicting the response relationships between multiple parties; and the MSA-CI model considers the personalized information of multiple parties, making the content response more interesting. These three models respectively consider the topic preferences, conversation structure, and personalized information of multiple parties to improve model performance. The CATD and GSN models are quite similar in results, indicating that the topic preferences and conversation structure of multiple parties improve the quality of candidate content responses to a certain extent. MIDS and the model of this invention performed well, indicating that the analysis of personalized information and dialogue psychology can significantly improve the quality of responses to candidate content. The ChatterBot model scored the lowest, mainly because its poor contextual dependence led to many meaningless responses.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for anthropomorphizing multi-party human-computer interaction responses based on knowledge graphs, characterized in that: The method includes the following steps: S1: Multi-party human-computer interaction emotional state assessment; S2: Knowledge Graph Ripple Propagation; Specifically, S1 is: Using the NRC_VAD sentiment dictionary from psychology, sentiment is quantified and calculated. Combining the specific meanings of VAD sentiments with TF-IDF technology, the overall sentiment embedding of dialogue statements is calculated, as shown in the following formula: (1) in, It is the emotional embedding of words. and Sentences The Middle The TF and IDF values ​​of a word are used to measure the frequency of a word in a one-party dialogue. The higher the frequency, the more important the word is to that party. The IDF measures the probability of a word appearing in all multi-party human-computer dialogues. If a word appears in all multi-party human-computer dialogues, then the word is not important. and The calculation formulas are shown in equations (2) and (3) respectively: (2) in, For words Word frequency in a conversation between two parties It refers to the length of the conversation between one party; (3) in, It is the sum of all input and response sentences in the training dataset. This indicates that all input and response sentences contain the word. The total number of sentences; A positive input often produces a positive response, making the multi-party human-computer interaction more positive; conversely, a negative input often produces a negative response, making the multi-party human-computer interaction more negative. This dialogue process aligns with the psychological process of actual human-to-human communication. Based on this psychological process, the formula for measuring the similarity of the V-value sentiment between the response and the input is as follows: (4) in, This indicates the sentiment value in the robot's response. This represents the sentiment value input from a multi-party dialogue, with the subscript 1 indicating the Manhattan distance; When both input and response have a strong desire to dominate, conflict will arise between the dominant party and the robot. The greater the distance between the A and D values ​​of the input and response, the better. The distance between the A and D values ​​of the response and the input is measured by equations (5) and (6), respectively: (5) (6) Combining formulas (4), (5), and (6), the final sentiment score of the robot's candidate responses is defined as follows: (7) Normalizing this, the sentiment affinity of the robot's candidate responses is defined as: (8) in, and These are the minimum and maximum sentiment scores among the candidate responses; S2 is: First, topic entities are extracted from the multi-party dialogue content, mapped onto the knowledge graph, and relevant knowledge entities are activated. These knowledge entities are regarded as a set of user interest seeds. Secondly, based on the idea of ​​water ripple propagation, it propagates outward along the links in the knowledge graph, and selects the next layer's neighboring entities as propagation entities for the seed entities in the inner layer in turn, forming a multi-party preference distribution for potential seed topics. Then, the seed entity propagates along different links to After the first layer, proceed from the outside inwards. Each layer begins by aggregating the entity information of its current entity with that of its neighboring entities, updating the entity representation of that layer, and using this aggregation for the next aggregation. Seed entities in the layer neighbor information are used to characterize the final potential topics of interest; Finally, taking into account the emotional inclinations and interests of various parties, the candidate response with the highest score is selected as the robot's optimal response.

2. The method for anthropomorphizing multi-party human-computer interaction responses based on knowledge graphs according to claim 1, characterized in that: Specifically, S2 is: First, define The set of multi-party dialogue topic entities obtained in a turn-based dialogue: (9) in, Given a known knowledge graph, For the subject of multi-party dialogue, Represented as dialogue turns; knowledge graph triples are represented as That is, head entity-relationship-tail entity, which refers to the set of multi-party dialogue topic entities that have been acquired. Definition of triplet ripple set: (10) in, The number of outward propagation layers is represented, where It is a collection of topics involved in multi-party dialogue. This indicates the first-level related entity that propagates outward from the topic entity in the first round of dialogue; Secondly, candidate responses are retrieved from the dialogue corpus, and the response content is sequentially processed using the word2Vec method and the EmbeddingAverage vector mean method to obtain vector representations of the candidate responses. ,in For word embedding matrix, The dimension of the word embedding; representing the vector of the candidate response. With head entity and relationships Comparison, calculation Multi-party dialogue, ripples gather Each triple in a first-order related entity With sentence feature vectors Association probability : (11) in, Representing the central relationships of a knowledge graph ripple Vector representation, Represents the head entity of the knowledge graph ripple center. The vector representation is obtained by using TransD, a knowledge representation method that embeds knowledge graphs. Viewed in relational space In the vector representation of candidate responses With entity vector representation The probability of similarity; After obtaining the relevant probabilities, the topics of the multi-party dialogue are then aggregated to form first-order related tail entities. The impact garnered first-order responses from multiple parties in the dialogue. The calculation formula is as follows: (12) in It is the vector representation of the tail entity; Then, Replace with Calculate in sequence This process is performed iteratively on the ripple set, so topics of interest to multiple parties propagate from near to far along the hierarchical relationship of the knowledge graph; multi-party interest preference vector Defined as: (13) In particular, considering that an excessively high order response can dilute useful information and introduce more noise, The value of is 2; Finally, given a multi-party interest preference vector and candidate response vectors Under these conditions, the interest of multiple parties in potential topics can be predicted. for: (14) in It is the sigmoid function, and the formula is shown below: (15) Combining formulas (8) and (14), the candidate responses are weighted by content interest and emotional friendliness, and normalized as follows: (16) in, and As a constraint factor, and having ; The range of values ​​is The closer the value is to 1, the more satisfied the parties are with the candidate response.

Images