CN112052322B

CN112052322B - Intelligent robot conversation strategy generation method based on particle calculation

Info

Publication number: CN112052322B
Application number: CN202010917746.4A
Authority: CN
Inventors: 涂志莹; 田钧锐; 王忠杰; 李楠; 徐晓飞
Original assignee: Harbin Institute of Technology
Current assignee: Harbin Institute of Technology
Priority date: 2020-09-03
Filing date: 2020-09-03
Publication date: 2023-03-21
Anticipated expiration: 2040-09-03
Also published as: CN112052322A; WO2022048164A1

Abstract

The invention discloses an intelligent robot dialogue strategy generation method based on particle computing, which comprises the following steps: s1, constructing a demand pattern library; s2, constructing a required knowledge graph; s3, constructing a particle calculation model; s4, generating a corresponding dialogue strategy based on the particle calculation model; and step S5, finishing the conversation. The method analyzes the user demand data of the demand library by a particle computation fuzzy clustering method so as to construct a proper inquiry path and finally complete the construction of the complete demand of the user. The invention can accurately construct the large-scale requirements of the users and ensure stable requirement accuracy and requirement coverage based on the prior knowledge in the field after few rounds of conversations.

Description

Intelligent robot conversation strategy generation method based on particle calculation

Technical Field

The invention belongs to the technical field of computer services, and relates to a method for quickly constructing a user demand intention.

Background

In the big data age, the service has become one of the most important development trends in the IT world. More and more software services are developed and deployed over the Internet, as well as a large number of virtualized services that connect real-world physical service resources. Services from multiple domains, multiple networks, are aggregated into one large complex service network or ecosystem, which may be referred to as "service internet (IoS)" or "big service. IoS proposes a paradigm in which all content can be used as a service on the Internet. In IoS, an extremely rich mass of services is diverse, distributed, and heterogeneous. By collecting, clustering, and combining these services, service solutions can be generated to meet the needs of the customers.

With the increasing sophistication of a wide variety of services, the needs of users are increasingly becoming increasingly sophisticated. In the standardization process of the service itself, the granularity is gradually reduced, and the combination of the service is naturally generated. Corresponding to the combination of services is a combination of user requirements. Since user demand is typically coarse-grained and cross-border, the service robot must integrate services from multiple domains to meet the demand. Considering that there are too many services available in a service library, a great challenge is how to balance the efficiency of construction and the accuracy of the solution during interaction with the user.

In the different requirements facing users, how to effectively utilize domain prior knowledge and end the construction of the user requirements in a limited turn is a problem worthy of solving for the conversation strategy facing the users. Based on the research background, it can be found that a conversation strategy which faces to the change of user demands and is efficient and ensures higher accuracy can be generated by utilizing a large amount of historical user demands.

Disclosure of Invention

The invention aims to provide an intelligent robot conversation strategy generation method based on particle computing. The invention can accurately construct the large-scale requirements of the users and ensure stable requirement accuracy and requirement coverage based on the prior knowledge in the field after few rounds of conversations.

The purpose of the invention is realized by the following technical scheme:

an intelligent robot dialogue strategy generation method based on particle computing comprises the following steps:

s1, constructing a demand pattern library:

the method comprises the following steps of preprocessing user demand information data to construct a demand pattern library, and specifically comprises the following steps:

(1) Building an intent tree

The intent tree is defined as follows:

ITree＝<G,E>

G＝{goal ₁ ,...goal _i ,...goal _n }

E＝{(goal _i ,goal _j )|goal _i is the parent node of goal _j }

goal _i ＝<intention,{Cons},{OptTarget}>

in the formula, ITree represents an intention tree, G is a set of nodes of the intention tree, E is a set of edges of the intention tree, and the nodes of the intention tree are represented by coarse and comprise two parts of an intention and a constraint set;

(2) Defining demand patterns

The demand pattern is defined as follows:

RP＝<info,{IntentionTree}>

in the formula, RP represents a demand mode and consists of an intention tree fragment set corresponding to RP and info describing information of RP;

(3) Defining a library of requirements patterns

Defining a collection of a large number of user demand patterns as a demand pattern library;

s2, constructing a required knowledge graph:

collecting a large amount of user requirement data, modeling user requirements by an intention tree structure, associating intentions in an intention tree with entity concepts in a general knowledge graph, associating different intentions through entities in the knowledge graph, and completing the inference of the intentions by means of the inference capability of the general knowledge graph, wherein the specific steps are as follows:

(1) Constructing a general knowledge graph;

(2) When new intention trees are added:

(a) Inserting all intention nodes in the intention tree into the knowledge graph, and adding child relations among corresponding intentions according to the structure of the intention tree;

(b) For each intention node in the current intention tree, extracting keywords of top-k from the intention, searching the entity in the knowledge graph according to the keywords, and establishing a has relation between the corresponding entity and the entity;

s3, constructing a particle calculation model:

through processing the information collected by the demand pattern library and the demand knowledge graph spectrum constructed in the steps S1 and S2, a particle calculation model is constructed based on a Grc algorithm, and the method specifically comprises the following steps:

I. firstly, obtaining a requirement R0 proposed by a user, searching in a requirement pattern library, and entering a free conversation process if the requirement R0 is not matched; after matching is successful, the matching result requirement mode is used as a dialogue basis of the root node intention, and then the intention and the constraint are returned to the user as a whole for requirement confirmation;

II. Acquiring a next round of inquiry sub-intention R1 of the dialogue with the user according to the matched dialogue;

III, if the user does not need or does not have the expression of the sub-intention R1, respectively correspondingly deleting the sub-intention point and constraining the sub-intention point to be random; otherwise, matching the sub-intention R1 in the requirement pattern library, and returning the matching result intention and the constraint and the sub-intention point of the R1 to the user as a whole;

IV, traversing all child intention points under the current father node and repeating III until the intention under the current father node is clear, inquiring the additional demand intention of the corresponding father node for the user, and repeating III until the demands of the user under the corresponding father node are completely proposed without omission;

v, continuously repeating the processes II-IV until all nodes in the conversation basis are traversed, and the user requirement is accurately confirmed at the moment, so that the conversation is finished;

s4, generating a corresponding dialogue strategy based on the particle calculation model:

the method comprises the following steps of processing and integrating constraint attributes of each layer of intention and information of sub-intention points, applying a particle calculation model to obtain inquiry constraints corresponding to the intention, continuously iterating and circularly obtaining inquiry bases of each layer from top to bottom, finally merging results into an inquiry path of a root node intention and using the inquiry path as a dialogue strategy for interacting with a user corresponding to the intention, and specifically comprising the following steps of:

I. collecting and analyzing demand patterns and demand knowledge graph spectrum data, and taking root node intents as analyzed root intents;

II. Constructing a particle computing GrC model by analyzing the constraint and the child nodes corresponding to the root intention, and obtaining an inquiry path corresponding to the intention by analyzing and processing information;

III, traversing the sub-intents of the follow-up intents, and repeating the step II to obtain an inquiry path corresponding to the sub-intents;

IV, all the obtained inquiry paths are connected in series to obtain a robot inquiry strategy corresponding to the follow intention;

step S5, completing the conversation:

continuously interacting with the user based on the conversation strategy generated in the step S4, receiving user information, giving corresponding feedback according to the conversation strategy, completing construction of user requirements, and ending the conversation, wherein the conversation comprises the following specific steps:

and interacting with the user based on the conversation strategy generated by the corresponding intention in the step S4, after the conversation strategy corresponding to the intention is generated, leading the conversation into different classification paths according to different feedbacks after the user is inquired according to the conversation basis, integrating the conversation basis of each layer to form the conversation strategy corresponding to the intention, and finally finishing the conversation after the conversation reaches the bottom layer so as to effectively and quickly finish the conversation with the user.

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

1. the invention effectively reduces useless inquiry in user requirement understanding and intention construction without support of service by utilizing a particle calculation (fuzzy clustering) method;

2. according to the method, some requirements (requirement modes) which often appear together and can be aggregated together to form a new coarse granularity are effectively used in the process of constructing the requirement mode library, the efficiency of interaction with a user is effectively improved after the requirement modes are applied, and one round of detailed and complicated inquiry is avoided;

3. the application of the demand-knowledge graph effectively assists the user in demand expansion and service recommendation, and meanwhile, through the demand-knowledge graph, the conversation strategy can effectively predict the demand of the user based on past data instead of pure blank inquiry (does you need other XXX;

4. compared with the existing intelligent robot conversation strategy, the application particle calculation method has the advantages that conversation bases of each round of interaction with the user not only stay in the confirmation of a single attribute type or intention, conversation rounds are effectively reduced, and conversation efficiency is improved;

5. constraint-based analysis enables a dialog strategy to be built on the basis of service resources, and rejection can be represented for user constraints beyond the service range in the process of dialog with users.

Drawings

FIG. 1 is an overall flowchart of an intelligent robot dialogue strategy generation method based on particle computing according to the present invention;

FIG. 2 is an intention tree structure;

FIG. 3 is a practical example of decomposing a requirement into an intent tree;

FIG. 4 is a required knowledge-graph;

FIG. 5 is a schematic diagram of a dialog strategy generation principle;

FIG. 6 is a flow chart of the dialogue strategy generation and particle computation model construction;

FIG. 7 is a diagram illustrating an example of a dialog process interacting with a user;

FIG. 8 is a comparison graph of the principle before and after use of particle calculation;

fig. 9 shows the results of the experiment.

Detailed Description

The technical solution of the present invention is further described below with reference to the accompanying drawings, but not limited thereto, and 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 shall be covered by the protection scope of the present invention.

The invention provides an intelligent robot conversation strategy generation method based on particle computing, which mainly finds out an inquiry path suitable for perfecting user requirements through automatic analysis and processing of prior knowledge by utilizing the prior knowledge in the field obtained from commonness and similarity among a large number of historical user requirements, roughly determines the requirement range of a user through a few pairs of short inquiries of the user, and further determines the complete requirements of the user in subsequent small-scale user interactive conversation fine adjustment. As shown in fig. 1, the method specifically comprises the following steps:

s1, constructing a demand pattern library:

the step is mainly based on the automatic construction of past system demand data of a large number of users. This step requires a structured processing of the user's massive demand data. These requirements tree structures represent some common conventional or relatively stable requirements intent or requirements plan templates with accompanying constraints that can be reused in meeting different customer requirements in some areas. The present invention defines these frequently occurring pieces of service requirements as requirement patterns, which are modular descriptions of user requirements. The demand patterns can meet different user requirements and be reused, and finally, a set of a plurality of demand patterns is defined as a demand pattern library. The method comprises the following specific steps:

1. construction of an intent tree

The intention tree describes the specific needs of an individual. A particular requirement may be split into intent trees, so each person will have several intent trees.

The demand can be expressed in a plurality of different forms, the demand in life is usually described by natural language, such as various voice assistants of Siri, and the description method is close to daily life, is easy to obtain the demand, but also brings challenges to the analysis of the demand. Therefore, in a large service platform, we propose an intention tree, a structured form, to describe the requirements. The intention tree is defined as follows:

ITree＝<G,E>

G＝{goal ₁ ,...goal _i ,...goal _n }

E＝{(goal _i ,goal _j )|goal _i is the parentnode of goal _j }

goal _i ＝<intention,{Cons},{OptTarget}>

in the formula, ITree represents an intention tree, the intention tree represents the containing relationship between different intents by using a tree structure, G is a node set of the intention tree, and E is an edge set of the intention tree. The nodes of the intention tree are represented by the good, which contains two parts of intention and constraint set, and the intention is the description of the user to the specific function requirement and is represented by the natural language.

Constraints express the user's non-functional requirements and restrictions on meeting the demand service, expressed in key-value pairs. The constraints are defined as follows:

Cons＝<Cons _key ,Cons _type ,Cons _value >

in the formula, cons _key Representing the object to be constrained, cons _value Representing specific contents of constraints, cons _type Representing the type of constraint, such as continuous, enumerated, etc. The user's needs need to be analyzed and compared into two parts, intent and constraint. FIG. 2 illustrates an intention tree structure.

The requirement with large granularity can be decomposed into the requirement with finer granularity, and based on the requirement, the intention tree can be formed through continuous decomposition after the user puts forward the requirement.

FIG. 3 is a practical example of decomposing a requirement into an intent tree. The demand of the user travels to Paris, and the demand costs between 1000 and 1500$, the demand is further decomposed into three sub-demands of diet, transportation and accommodation, and the demand requires accommodation for 5 days and costs between 500 and 800 $.

2. Demand patterns

In the real world, many of the user's needs may not be resolved, as these needs already have corresponding solutions. As shown in fig. 3, a large number of travel plans are available in a travel agency, and a user only needs to select the most suitable one from the travel plans, so that the user does not need to care about dining, transportation and other problems. These travel plans are not envisioned by the travel agencies, but rather because more passengers have had similar travel needs, the travel agencies may mine travel plans that fit such needs of the users.

Similarly, in the user-proposed intent tree, there are often some intents that appear together, which may be aggregated together to form a new coarse-grained intent, and we refer to these frequently-appearing intent tree segments as demand patterns. A demand pattern is a modular description of a user's demand, representing a commonality between different demands, and is defined as follows:

RP＝<info,{IntentionTree}>

where RP represents a demand pattern, consisting of its corresponding set of intent tree fragments and info describing its own information. The Info includes information such as domain information, frequency of use, and function description of the demand pattern.

In addition, a demand pattern is not constant, and the demand pattern evolves over time. For example, if a new scenic spot is opened in a certain area, a travel plan whose destination includes the area is likely to be newly added with a service for the scenic spot. Demand patterns also fail and if a demand pattern is rarely used or even not at all for a period of time, then the demand pattern is not necessary. Similarly, if the number of travel services used by a travel agency is too low, the travel services may be off-shelf.

3. Demand pattern library

We define a collection of large user demand patterns as a library of demand patterns.

S2, constructing a demand knowledge graph

The required knowledge graph is constructed on the basis of the general knowledge graph, the intentions in the intention tree are connected with the entity concepts in the general knowledge graph, different intentions are connected through the entities in the knowledge graph, and the inference of the intentions is completed by means of the inference capability of the general knowledge graph. The universal knowledge graph uses the ownthink open-sourced Chinese knowledge graph data (https:// github. Com/ownthink/knowledgedigageGraphData), which is organized in a form of (entity, attribute, value), (entity, relationship, entity) mixture. The demand knowledge-graph uses only this part (entity, relationship, entity).

Nodes of the knowledge-graph are required to contain:

1. the entity: refers to all kinds of entities in the universal knowledge graph;

2. the interaction: the intents are the intents in each node in the intent tree.

Relationships contained in the demand knowledge graph:

1. (entity, relationship, entity) the relationship refers to all types of relationships in the common knowledge graph;

2. (entry, has, entity) has means that the intent contains the entity;

3. (attendance, child, attendance): child refers to a child intent of which an intent is another intent.

The form of the required knowledge graph is shown in FIG. 4, and the construction process is as follows:

1. constructing a general knowledge graph;

2. when new intention trees are added:

(1) Inserting all intention nodes in the intention tree into the knowledge graph, and adding child relations among corresponding intentions according to the structure of the intention tree;

(2) For each intention node in the current intention tree, extracting keywords of top-k from the intention, searching the entity in the knowledge graph according to the keywords, and establishing a has relation between the corresponding entity and the entity.

The intention tree can express the inclusion relation before the disagreement graph of the user requirement, the subordinate relation between the intention and the constraint, and the structured form also has great help for understanding and analyzing the user requirement, but can not analyze the potential relation between different requirements.

A generic knowledge-graph refers to a knowledge-graph that contains as many entity concepts and relationships between them as possible. The intention knowledge graph is composed of intention trees, wherein the entity is a demand in the intention trees, and the relationship is an inclusion relationship between coarse-grained intention and fine-grained intention in the intention trees. For each intention in an intention tree generated by a demand, keywords of top-k are extracted from the intention tree, and the intention is related to the keywords, so that the general purpose graph and the intention are closely related. And connecting the obtained knowledge graph with the general knowledge graph and establishing corresponding mapping to construct a required knowledge graph.

Step S3, generating a particle calculation model

The step is based on a demand pattern library and demand knowledge graph spectrum data, and a particle calculation model is constructed through a GrC algorithm. As shown in fig. 5, after the intentions of the user root nodes are obtained, the demand patterns corresponding to the intentions are collected through the demand pattern library and the demand knowledge graph. The modes are merged and analyzed, constraint attributes and sub-nodes (presence service support) of the nodes of each layer are integrated to be used as input of a particle computation model based on a tree structure of a demand mode, and a core algorithm of the model is a GrC algorithm.

Step S4, generating a dialogue strategy based on the particle computation model

As shown in fig. 6, after identifying a root node demand D proposed by a user, collecting all demand patterns corresponding to the root node D and performing the following steps:

I. because D is the demand mode of the root node, the father node of D is not considered; the method comprises the steps of taking the structure of { key1: { value11, value12, \8230; }, key2: { value21, value22, \8230; } as a model input of a particle calculation algorithm, using a classical FCM algorithm as a core algorithm of the particle calculation model, projecting the multidimensional constraint and attribute values required by D in the same plane through a particle calculation process, and performing particle division to generate a plurality of fuzzy particles. Membership boundaries of the ambiguity particles are identified as interrogation paths on the D demand constraint attribute. (for this part, refer to ICWS 2020 User indication Recognition and Recognition Method for switching AI Services.)

After the constraint of the D demand point is processed to obtain the query path, the complete set of child node demands of the D node in all demand patterns is defined as { D1, D2, D3, \8230; }, so for each demand pattern taking D as a root node, the child demands of the D node can be vectorized (a vector consisting of 0 and 1 is generated, 1 represents that the demand pattern includes the child demand point, and 0 represents that the pattern does not include the child intention point). Taking the travel requirement in the data set as an example, the "travel" node includes five seed requirements, such as "hotel", "restaurant", "scenic spot", "subway", "rent", and the like, and from the requirement mode taken as the basis in example 3, the vector corresponding to the travel sub-requirement is [1, 0], and then a vector with a length of 5 can be generated corresponding to each requirement mode. And merging all the sub-demand vectors into an n-m matrix (n is the number of the corresponding demand modes, and m is the length of the full set of the sub-demands) by the system after the sub-demands are vectorized, taking the matrix as primary input of the particle calculation model, and performing fuzzy division through particle calculation again. And analyzing the result, and then synchronizing the result in the step I to obtain the sub-requirement inquiry path corresponding to the step D.

And III, repeating the steps I and II to obtain the inquiry path of the intention Di for the sub-intention Di with the definite demand of 1 in the demand mode in a deep or wide sequence, and then repeating the steps continuously until the inquiry paths of all the associated intention points are determined or the divided grains are small enough to be directly fed back to the user as a return result. The obtained clustering optimal division is used as the basis of inquiry when the current root intention point interacts with the user, after the inquiry turn is finished, the operation is repeated on the sub-nodes of the root intention point, and the inquiry basis corresponding to the sub-intention points is respectively obtained; recursion is continuously performed downwards in the same way until the user requirement is completely constructed; and finally, constructing a conversation strategy according to the combined inquiry of each round and the user intention of the corresponding root intention point.

S5, finishing the interaction with the user based on the conversation strategy

In this step, a dialog strategy generated based on the corresponding intention in step S4 interacts with the user, as shown in fig. 7, after the dialog strategy corresponding to the intention is generated, each parent intention has a corresponding dialog basis, the dialog is guided into different classification paths according to different feedbacks after the user is queried according to the dialog basis, and finally the dialog is ended after the dialog reaches the bottom layer, so that the dialog with the user can be effectively and quickly completed.

Example (b):

in the embodiment, data sources are CrossWOZ, https:// github. Com/thu-coai/CrossWOZ, a corresponding demand mode library and a demand knowledge graph are constructed after a data set is preprocessed, data in the demand mode library and the demand knowledge graph are used as particle calculation models to be input, a corresponding intention conversation strategy is obtained through calculation of the particle calculation models, and finally interaction with a user is carried out based on the conversation strategy.

The experimental result is shown in fig. 9, after the requirement mode library, the requirement knowledge graph and the particle calculation model are applied, consumption and cost required for completing the conversation are effectively reduced, conversation efficiency for understanding the intention of the user is improved, the satisfaction of the user is improved, and the situation that the satisfaction of the user is reduced.

As shown in FIG. 8, the system may save query rounds on demand constraints by binding demand points of constraints in demand patterns as feedback to the user. Compared with a fine-grained conversation strategy based on a requirement pattern library and a requirement knowledge graph and focusing on each constraint attribute of each requirement, the pruning strategy based on the grain calculation model enlarges the granularity of a bottom-layer framework on which the whole conversation depends, reduces the fine-grained stub part in the user requirement on the whole and improves the conversation efficiency.

Claims

1. An intelligent robot dialogue strategy generation method based on particle computing is characterized by comprising the following steps:

s1, constructing a demand pattern library:

the method comprises the steps of constructing a demand pattern library by preprocessing user demand information data;

s2, constructing a required knowledge graph:

collecting a large amount of user requirement data, modeling user requirements by an intention tree structure, associating intentions in an intention tree with entity concepts in a general knowledge graph, associating different intentions through entities in the knowledge graph, and completing the inference of the intentions by virtue of the inference capability of the general knowledge graph;

s3, constructing a particle calculation model:

the method comprises the following steps of processing and integrating constraint attributes of each layer of intention and information of sub-intention points, applying a particle calculation model to obtain inquiry constraint corresponding to the intention, continuously iterating and circulating to obtain inquiry bases of each layer from top to bottom, finally combining results into an inquiry path of a root node intention and using the inquiry path as a dialogue strategy for interacting with a user corresponding to the intention, and the specific steps are as follows:

III, traversing the sub-intents following the intents, and repeating the step II to obtain an inquiry path corresponding to the sub-intents;

step S5, completing the conversation:

and (4) continuously interacting with the user based on the conversation strategy generated in the step (S4), receiving user information, giving corresponding feedback according to the conversation strategy, completing construction of user requirements, and ending the conversation.

2. The intelligent robot dialogue strategy generation method based on particle computing according to claim 1, wherein the specific steps of the step S1 are as follows:

(1) Building an intent tree

The intent tree is defined as follows:

ITree＝<G,E>

G＝{goal ₁ ,…goal _i ,…goal _n }

E＝{(goal _i ,goal _j )∣goal _i is the parent node of goal _j }

goal _i ＝<intention,{Cons}>

in the formula, ITree represents an intention tree, G is a node set of the intention tree, E is an edge set of the intention tree, and the nodes of the intention tree are represented by coarse and comprise two parts, namely intention intersection and constraint set { Cons };

(2) Defining demand patterns

The demand pattern is defined as follows:

RP＝<info,{IntentionTree}>

(3) Defining a library of demand patterns

A collection of a large number of user demand patterns is defined as a library of demand patterns.

3. The intelligent robot dialogue strategy generation method based on particle computing according to claim 1, wherein the specific steps of the step S2 are as follows:

(1) Constructing a general knowledge graph;

(2) When new intention trees are added:

(b) For each intention node in the current intention tree, extracting keywords of top-k from the intention, searching the entity in the knowledge graph according to the keywords, and establishing a has relation between the corresponding entity and the entity.

4. The intelligent robot dialogue strategy generation method based on particle computing according to claim 1, wherein the specific steps of the step S5 are as follows: