A soft service robot supporting social conversation

Abstract

The present invention discloses a soft service robot supporting social dialogue, the soft service robot includes a user cognitive model and a social dialogue model, wherein: the user cognitive model includes three parts: a social user preference model, a user attribute set, and a common sense knowledge base; the social dialogue model includes three parts: a conflict point identification module, a conflict resolution decision module, and a dialogue action generation module. The present invention automatically negotiates with external partners in a natural language manner, and has achieved good experimental results, constructing a dialogue robot framework currently in the field of social dialogue.

Description

Technical Field

[0001] The invention belongs to the technical field of artificial intelligence and relates to a soft service robot supporting social dialogue. Background Art

[0002] With the development of dialogue systems, soft service robots have been widely used in daily life, such as Xiaodu, Tmall Genie, Siri, etc. It can meet the various needs of users, including web search, product recommendation, smart home, etc. Among them, there is a special need, that is, social need. It refers to the negotiation between users and external partners (including family, friends and colleagues) to jointly complete a task. However, with the development of the Internet, the social scope of users has greatly expanded, which has also brought a heavy social burden to users. So far, social burden has not received widespread attention from academia and industry. Summary of the invention

[0003] In order to alleviate the social burden of users, the present invention provides a soft service robot supporting social dialogue.

[0004] The objective of the present invention is achieved through the following technical solutions:

[0005] A soft service robot supporting social dialogue includes a user cognition model and a social dialogue model, wherein:

[0006] The user cognitive model includes three parts: social user preference model, user attribute set, and common sense knowledge base;

[0007] The social user preference model is used to determine the user's preference for the new intent, and then estimate the degree of conflict with the user's preference;

[0008] The user attribute set is used to aggregate preferences of users with similar attributes, and estimate user preferences based on the aggregated preferences;

[0009] The common sense knowledge base is used to model the connections between different areas of user preferences and introduce common sense reasoning into the negotiation process;

[0010] The social dialogue model includes three parts: conflict point identification module, conflict resolution decision module, and dialogue action generation module;

[0011] The conflict point identification module is used to select the most conflicting and important points in the conflict point set based on the user cognitive model;

[0012] The conflict resolution decision module is used to select the intention that best meets the user's preference under the selected conflict point based on the conflict point given by the conflict point identification module and the user cognitive model;

[0013] The dialogue action generation module is used to generate dialogue actions according to the selected conflict points, intentions and dialogue states to simulate the dialogue strategies in the negotiation process.

[0014] Compared with the prior art, the present invention has the following advantages:

[0015] In the absence of the user, the soft service robot completes the negotiation process with the external partner on behalf of the user, and returns the negotiation results to the user, thereby reducing the user's social burden. In order to complete the above scenario, the social soft service robot needs two capabilities. In daily life, people do things according to their preferences (such as buying items, participating in social negotiations). Therefore, in order to make accurate decisions on behalf of users, social soft service robots need to accurately and timely understand the user's preferences and real-time environment (i.e., user cognitive ability). In the process of social negotiation, people actively express their intentions, accept the suggestions of others, judge whether these suggestions are acceptable, and give reasonable suggestions to resolve potential conflicts. Therefore, soft service robots for social scenarios need to complete the above process (i.e., autonomous negotiation capabilities). The present invention verifies that social soft service robots can complete social dialogue scenarios well through the above two capabilities. The present invention automatically negotiates with external partners in a natural language manner, and has achieved good experimental results. It constructs the current dialogue robot framework in the field of social dialogue and expands social functions in the direction of soft service robots. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 The usage scenarios for the present invention;

[0017] Figure 2 This is the overall process diagram of the social soft service robot;

[0018] Figure 3 It is the various modules and algorithms in the social soft service robot framework. DETAILED DESCRIPTION

[0019] The technical solution of the present invention is further described below in conjunction with the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention should be included in the protection scope of the present invention.

[0020] like Figure 2As shown, the social soft service robot needs to have two capabilities: user cognition and autonomous negotiation. The present invention proposes a social soft service robot framework: 1. Because the social soft service robot needs to represent the user and negotiate with other users on the basis of cognition of the user, the social soft service robot needs to have a certain degree of cognition of the user. Therefore, the social soft service robot framework includes a user cognition model, which models the user from three perspectives: user social preferences, user attributes, and common sense reasoning. 2. The social soft service robot framework also includes a social dialogue model to support the autonomous negotiation ability of the social soft service robot. The social dialogue framework includes three parts: a conflict point identification module, which is responsible for selecting the points with the greatest degree of conflict and the most important (that is, the user is most concerned about); a conflict resolution decision module, which is responsible for selecting the intention that best meets the user's preferences under the selected conflict point; and a dialogue action generation module, which is responsible for selecting dialogue actions (for example, proposing intentions, agreeing to the intentions of external partners, and requesting recommendations). Usage scenarios such as Figure 1 shown.

[0021] 1. User Cognitive Model

[0022] Social soft service robots negotiate on behalf of users and need to have the ability to recognize users. Current user modeling research focuses on specific dimensions of users, rather than modeling users from a holistic perspective. The present invention models users from three perspectives: user social preferences, user attributes, and social networks. Social soft service robots also need reasoning capabilities to support explicit and explainable decisions. Therefore, the user cognition model contains three parts: a social user preference model, a user attribute set, and a common sense knowledge base. Among them:

[0023] The social user preference model is used to model the changes in user preferences over time. When a new intention appears, the social soft service robot determines the user's preference for the new intention based on the social user preference model, and then estimates the degree of conflict with the user's preference.

[0024] The user attribute set is based on the intuition that similar preferences exist between similar users according to the central theory of the recommendation system. The user attribute set can aggregate the preferences of users with similar attributes and estimate user preferences based on the aggregated preferences; there are usually relationships between different fields, such as sports and running, travel and booking air tickets, etc.

[0025] In order to model the connection between different fields in user preferences, the present invention introduces a common sense reasoning process in the negotiation process, while maintaining a common sense knowledge base in the user cognitive model.

[0026] 2. Social Conversation Model

[0027] like Figure 3As shown, the social dialogue model includes three parts: a conflict point identification module, a conflict resolution decision module, and a dialogue action generation module, wherein: the conflict point identification module selects the point with the largest and most important conflict based on the user cognitive model; the conflict resolution decision module constructs a candidate intention set under the conflict point based on the given conflict point, and then selects the intention acceptable to both parties; the dialogue action generation module generates dialogue actions according to the selected conflict point, intention and dialogue state to simulate the dialogue strategy in the negotiation process. In addition to the above modules, the present invention also introduces a reasoning method. This method solves the following problem: the new intention proposed by the external partner is not included in the user cognitive model, making it impossible for the social soft service robot to judge.

[0028] (1) Conflict point identification

[0029] During the conversation, the social soft service robot needs to determine whether the discussion on the current conflict point is over. Once the negotiation is completed and the external partner does not raise a new conflict point, the social soft service robot will select a new conflict point. i The conflict point set is expressed as follows: S = {point1, point2,…, point n The conflict point identification module needs to select the conflict point with the maximum conflict degree and importance from the conflict point set. The conflict degree can be determined based on the social user preference model and the intention tree. The conflict function is defined as: score conf =f(SUP,point i ), SUP represents the user cognitive model, iTree t is an intention tree, representing the current state of the conversation, from which the conflict point set S is extracted. The importance can be determined based on the social user preference model. For any conflict point, the importance function can be defined as: score imp =h(SUP). The conflict point with the maximum conflict degree and importance is expressed as follows:

[0030]

[0031] We also considered another problem: the user had only mentioned exercise to the social soft service robot before, but the external partner mentioned running in this negotiation. According to human thinking habits, running and exercise are closely related. However, since the user had not mentioned running to the social soft service robot before, the social soft service robot could not judge whether the user would accept this proposal, which is obviously unreasonable. Therefore, the present invention proposes a reasoning method and integrates it into the conflict point identification module and the conflict resolution decision module.

[0032] The present invention is divided into four parts to construct a conflict point identification module: 1. Seed set selection, 2. Preference estimation, 3. Reasoning path construction, 4. Conflict point scoring. The specific steps are as follows:

[0033] The first step is seed set selection. The previous step of reasoning is to find similar fields, and then the reasoning path can be constructed. Seed set selection completes the function of finding related fields. Specifically, the present invention uses WordNet adjacent words to calculate the word set similarity between two domains: Assuming the original intention is I i , the target intention is I j , while I i The synonym set is Nerbor (I i ), I j The synonym set is Nerbor (I j ), and then the similarity calculation formula is summarized as follows: where dis i and dis j They represent Nerbor(I i ) and Neibor(I j ). Finally, the candidate intents are sorted according to Similarity, and the top-k domains are selected as the candidate domain set, where k is a constant, and the value of the present invention is 3.

[0034] The second step is preference estimation. After obtaining the candidate intent set, preference estimation uses Glove to encode different domain nodes. i , to get the corresponding representation emb(e i ), and apply such representation to the calculation of scoring functions and the construction of reasoning paths.

[0035] The third step is to construct the reasoning path. In order to obtain the relationship between the two fields and enhance the interpretability of the model, the reasoning path construction constructs the reasoning path between the start and end nodes. Specifically, let e s and e o are the starting nodes respectively. For any node e i , let its adjacent node set be N(e i ), the reasoning path is e s →e1→…→e k ,E={e o ,e1,…,e k}, then, the formula for getting the next node is as follows:

[0036]

[0037] As the reasoning path expands, it gradually approaches the target node and eventually forms a complete reasoning path.

[0038] The fourth step is scoring conflict points. This part scores candidate conflict points based on the reasoning path and node representation. If the domain or slot value already exists in the user cognitive model, the corresponding weight is directly used. In the case where the domain exists but the slot does not exist, similar domains can be obtained through reasoning. If similar slot values ​​exist in similar domains, the conflict point identification module directly uses the weight of the slot value as the score. The algorithm of the conflict point identification module is shown in Table 1.

[0039] Table 1

[0040]

[0041] (2) Conflict resolution decision

[0042] The conflict resolution decision module selects a set of candidate intents under the selected conflict point and then scores the candidate intents. If the intent proposed by the external partner meets the user's preference, the module returns this intent as the intention of the social soft service robot; otherwise, the module will look for an intent that can satisfy both user preferences and maximize user benefits, and then submit it to the external partner. The construction of the conflict resolution decision module is divided into three parts: intent construction, consistency level calculation, and suggestion generation. The specific steps are as follows:

[0043] The first step is to build intention. At the selected conflict point i In the next step, the intention construction part constructs a candidate intention set that meets the user's preferences based on the user cognitive model. For example, if the user used to eat between 17:00 and 18:00 in the evening, then the intention set is {17:00, 18:00}. If an external partner also proposes an intention, then this intention should also be added to the candidate intention set, and then the candidate intention set is further narrowed down based on the similarity between the candidate intention and the other party's intention. Let intention u Represents the intention proposed by the external partner. First, for the conflict point selected by the conflict point identification module i , which can satisfy the conflict point i The candidate intention set is U. There is intention i ∈U∩{intention u =U s As the candidate intention set. Let U a As the set of intentions that can satisfy both users, the candidate intention selection function is q, and we have: q(U)≤U a .

[0044] The second step is consistency level calculation. Based on the candidate intent set, the consistency level calculation is scored according to the weight of the corresponding intent.a , the value function f selects the intent that best meets the user's needs, and p is used to measure whether the selected intent is acceptable to external partners:

[0045]

[0046] The third step is suggestion generation. Since the candidate intent set and the corresponding scores have been obtained in the first and second steps, when the social soft service robot cannot accept the intent proposed by the external partner, the suggestion generation will select the intent with the highest score from the candidate intent set as the suggestion. The scoring function of this part is similar to that of the consistency level calculation part.

[0047] (3) Dialogue Action Generation Decision

[0048] The dialogue action generation module is similar to the intention prediction module in task-based dialogue. However, there are some differences: traditional task-based dialogue considers how to understand the needs of users and guide them to express their needs. However, in negotiation scenarios, some special dialogue actions appear. In the negotiation process, there are different behaviors (such as cooperative behavior, competitive behavior), desires, and restriction levels. For example, when the external partner disagrees to go out on Sunday, the user forces him to agree by insisting on his opinion, which is a typical competitive behavior. Dialogue strategies can be implemented through different combinations of dialogue actions; researchers must focus on these strategies to conduct effective negotiations and maximize the interests of users. The construction of the dialogue action generation module is divided into two parts: goal achievement detection and dialogue action prediction.

[0049] The first step is goal achievement detection. When both parties agree on the current intention and do not raise new conflict points, the social soft service robot enters the conflict point selection module to select the next conflict point; otherwise, it enters the conflict resolution decision module. The goal achievement detection determines whether the conflict has been resolved based on the conversation history H. For each {a i ,(d i ,s i,m ,u i,m )}, the present invention encodes it into a unified representation: R i =emb({a i ,(d i ,s i,m ,u i,m )}), where a i represents the user action in round i, d i represents the area discussed in round i, s i,m Indicates the field d i The attributes m,u in i,m Represents attribute s i,mThe value of round i. Specifically, because the number of dialogue actions is limited, the present invention randomly generates representations for all dialogue actions. The domain and slot are encoded by Glove. For the slot value v, when it is a string type, it is directly encoded by Glove. When it is an integer type, the integer value is written to the first position of the representation, and the remaining positions are filled with 0. Finally, all the representations are connected to obtain the representation R of this round of dialogue i The formula is as follows:

[0050] {G 1，u , G 2，u , …, G t，u}=LSTM({R 1，u , R 2，u , …, R t，u})

[0051] goal=softmax(w1(RELU(w2*G t，u +b1))+b2)

[0052] The second step is dialogue action prediction. The dialogue action prediction determines the social soft service robot action of this round based on the generated intention and the output of the goal achievement detection part. This part of the present invention is implemented in a rule-based manner.

[0053] (4) Dialogue action design

[0054] The goal of non-cooperative negotiation is to maximize the interests of the user and minimize the interests of the other party, and there is no overlap in the interests of both parties. Therefore, these dialogue behaviors express a strong competitive tendency. In the scenario of the present invention, the interests of both parties are not completely conflicting. The social soft service robot must express its intentions rather than simply guiding the external partners to express their intentions. Therefore, the present invention designs ten dialogue actions to represent the competitive and cooperative behaviors that may occur during the negotiation process. In each round, the external partners and the social soft service robots will take corresponding actions to express their intentions explicitly or implicitly. Table 2 is the ten dialogue actions and corresponding examples.

[0055] Table 2

[0056]

Claims

1. A soft service robot supporting social dialogue, characterized in that The soft service robot includes a user cognition model and a social dialogue model, wherein: The user cognitive model includes three parts: social user preference model, user attribute set, and common sense knowledge base; The social user preference model is used to determine the user's preference for the new intent, and then estimate the degree of conflict with the user's preference; The user attribute set is used to aggregate preferences of users with similar attributes, and estimate user preferences based on the aggregated preferences; The common sense knowledge base is used to model the connections between different areas of user preferences and introduce common sense reasoning into the negotiation process; The social dialogue model includes three parts: conflict point identification module, conflict resolution decision module, and dialogue action generation module; The conflict point identification module is used to select the most conflicting and important points in the conflict point set based on the user cognitive model; The conflict resolution decision module is used to select the intention that best meets the user's preference under the selected conflict point based on the conflict point given by the conflict point identification module and the user cognitive model; The dialogue action generation module is used to generate dialogue actions according to the selected conflict points, intentions and dialogue states to simulate the dialogue strategies in the negotiation process.

2. The soft service robot supporting social dialogue according to claim 1, characterized in that The conflict point set is expressed as follows: s = {point1, point2, ..., point n }, the conflict point identification module selects the conflict point with the maximum conflict degree and importance in the conflict point set i The conflict degree is determined based on the social user preference model and the intention tree. The conflict function is defined as: score conf =f(SUP,point i ), SUP represents the user cognitive model; the importance is determined according to the social user preference model. For any conflict point, the importance function is defined as: score imp = h(SUP), the conflict point with the maximum conflict degree and importance is expressed as follows:

3. The soft service robot supporting social dialogue according to claim 1 or 2, characterized in that The steps for constructing the conflict point identification module are as follows: The first step is seed set selection: use WordNet adjacent words to calculate the word set similarity between the two domains: Assuming the original intent is I i , the target intention is I j , I i The synonym set is Neibor (I i ), I j The synonym set is Neibor (I j ), the similarity calculation formula is as follows: where dis i and dis j Respectively represent Neibor(I i ) and Neibor(I j ), the candidate intents are sorted according to Similarity, and the top-k fields are selected as the candidate field set; The second step is preference estimation: After obtaining the candidate intent set, preference estimation uses Glove to encode different domain nodes. i , to get the corresponding representation emb(e i ), and apply such representation to score function calculation and reasoning path construction; The third step is to construct the reasoning path: Let e s and e o are the starting nodes respectively, for any node e i , let its adjacent node set be N(e i ), the reasoning path is e s →e1→…→e k ,E={e o ,e1,…,e k }, then, the formula for getting the next node is as follows: As the reasoning path expands, it gradually approaches the target node and eventually forms a complete reasoning path; The fourth step is conflict point scoring: score the candidate conflict points according to the reasoning path and node representation. If the domain or slot value already exists in the user cognitive model, the corresponding weight is directly used. In the case where the domain exists but the slot does not exist, the similar domain is obtained through reasoning. If similar slot values ​​exist in similar domains, the conflict point identification module directly uses the weight of the slot value as the score.

4. The soft service robot supporting social dialogue according to claim 1, characterized in that The conflict resolution decision module selects a candidate intention set under the selected conflict point, and then scores the candidate intention set. If the intention proposed by the external partner meets the user's preference, the module returns this intention as the intention of the social soft service robot; Otherwise, the module will look for an intention that satisfies both user preferences and maximizes user benefits, and then submit it to the external partner.

5. The soft service robot supporting social dialogue according to claim 1 or 4, characterized in that The steps for constructing the conflict resolution decision module are as follows: The first step is to build intention: at the selected conflict point i In this case, we construct a candidate intent set that meets the user's preferences based on the user cognitive model. If an external partner also proposes an intent, we also add this intent to the candidate intent set. Then, we further narrow the candidate intent set based on the similarity between the candidate intent and the other party's intent. Let intention u Representing the intention proposed by the external partner, firstly, for the conflict point selected by the conflict point identification module i , which can satisfy the conflict point i The candidate intention set is U, and there is intention i ∈U∩{intention u =U s As the candidate intention set, let U a As the set of intentions that can satisfy both users, the candidate intention selection function is q, and we have: q(U)≤U a ; The second step is consistency level calculation: Based on the candidate intent set, the consistency level calculation is scored according to the weight of the corresponding intent. a , the value function f selects the intent that best meets the user's needs, and p is used to measure whether the selected intent is acceptable to external partners: The third step is suggestion generation: in the first and second steps, the candidate intent set and the corresponding scores have been obtained. When the social soft service robot cannot accept the intent proposed by the external partner, the suggestion generation will select the intent with the highest score from the candidate intent set as the suggestion.

6. The soft service robot supporting social dialogue according to claim 1, characterized in that The steps for constructing the dialogue action generation module are as follows: The first step is goal achievement detection: when both parties agree on the current intention and do not raise new conflict points, the social soft service robot enters the conflict point selection module and selects the next conflict point; Otherwise, enter the conflict resolution decision module; Goal achievement detection determines whether the conflict has been resolved based on the dialogue history H; for each {a i ,(d i ,s i,m ,u i,m )}, and encode it into a unified representation: R i =emb({a i ,(d i ,s i,m ,u i,m )}), where a i represents the user action in round i, d i represents the area discussed in round i, s i,m Indicates the field d i The attributes m,u in i,m Represents attribute s i,m The value of round i; randomly generate representations for all dialogue actions, the domain and slot are encoded by Glove, for the slot value v, when it is a string type, it is directly encoded by Glove, when it is an integer type, the integer value is written to the first position of the representation, and the remaining positions are filled with 0, finally, all the representations are connected to get the representation G of this round of dialogue i , the formula is as follows: {G 1,u ,G 2,u ,…,G t,u }=LSTM({R 1,u ,R 2,u ,…,R t,u }) goal=softmax(w1(RELU(w2*G t,u +b1))+b2) The second step is dialogue action prediction: dialogue action prediction determines the actions of the social soft service robot in this round based on the generated intention and the output of the goal achievement detection part.

Images