CN112948534A

CN112948534A - Interaction method and system for intelligent man-machine conversation and electronic equipment

Info

Publication number: CN112948534A
Application number: CN201911258858.7A
Authority: CN
Inventors: 曹东岩; 周祥生; 屠要峰; 郭斌
Original assignee: ZTE Corp
Current assignee: ZTE Corp
Priority date: 2019-12-10
Filing date: 2019-12-10
Publication date: 2021-06-11

Abstract

The embodiment of the specification discloses an interaction method, an interaction system and electronic equipment for intelligent man-machine conversation, which mainly comprise the following steps: and determining a personalized dialog style matched with the user, accurately predicting intention information expressed by the user message based on a machine learning algorithm, and determining a dialog mode based on the intention information and the historical dialog record, so that personalized reply content is determined according to the determined dialog style, the intention information and the dialog mode and is returned to the user side. Furthermore, the reply content can be accurately matched according to the prediction intention, the reply content can be enriched by adopting a personalized conversation style and even adding emotional colors, and one or more rounds of conversation interaction are executed under the determined conversation mode. Therefore, flexibility and richness of dialog reply are improved, and use viscosity of a user is increased.

Description

Interaction method and system for intelligent man-machine conversation and electronic equipment

Technical Field

The present disclosure relates to the field of computer technologies, and in particular, to an interaction method and system for intelligent human-computer interaction, and an electronic device.

Background

In the field of network operation and customer service, it is essential to capture the key points in the current technological development environment to improve the operation efficiency of enterprises, reduce the cost and improve the customer satisfaction. Among them, intelligent customer service is becoming the main carrier for enterprises to provide services for external users and internal digital transformation.

At present, most of customer service services are not humanized enough, and the reply is rigid and rigid; and the reply content has larger deviation and influences the use experience of the user due to the lack of better context understanding capability and inaccurate semantic recognition.

With the continuous progress of the technology, the user demand is continuously increased. The customer service will no longer satisfy the requirement of a discordant and rigid reply and a content reply in the form of "answer-not-asking", and therefore improvements are needed in the customer service and the man-machine interaction system implemented by the customer service.

Disclosure of Invention

The embodiment of the specification aims to provide an interaction method, an interaction system and electronic equipment for intelligent man-machine conversation, so as to improve semantic recognition accuracy and realize personalized, flexible and diverse man-machine conversation interaction.

In order to solve the above technical problem, the embodiments of the present specification are implemented as follows:

in a first aspect, an interactive system for intelligent human-computer conversation is provided, including:

the system comprises a preprocessing module, a data processing module and a data processing module, wherein the preprocessing module is used for preprocessing received user information, and the user information carries user basic data;

the conversation style matching module is used for determining the conversation style matched with the user message according to the autonomous selection of the user and/or the user basic data;

the deep semantic understanding module is used for carrying out semantic analysis on the preprocessed user message by adopting a deep learning algorithm and identifying intention information expressed by the user message;

the dialogue management module is used for determining a dialogue mode of the user message during the dialogue interaction according to the intention information and the historical dialogue record of the user corresponding to the user message;

the decision-making module is used for determining reply content based on a plurality of confidence degree results generated by the intention information in a preset intention classification model;

and the reply module is used for returning reply content corresponding to the user message to the user based on the conversation mode and by adopting the conversation style.

In a second aspect, an interaction method for intelligent human-computer conversation is provided, including:

preprocessing a received user message, wherein the user message carries user basic data;

determining a conversation style matched with the user message according to the autonomous selection of the user and/or the user basic data;

adopting a deep learning algorithm to carry out semantic analysis on the user message after the preprocessing operation, and identifying intention information expressed by the user message;

determining a conversation mode of the user message during the conversation interaction according to the intention information and a historical conversation record of the user corresponding to the user message;

determining reply content based on a plurality of confidence degree results generated by the intention information in a preset intention classification model;

and returning reply content corresponding to the user message to the user based on the conversation mode and by adopting the conversation style.

In a third aspect, an electronic device is provided, including:

a processor; and

a memory arranged to store computer executable instructions that, when executed, cause the processor to perform the operations of the second aspect.

In a fourth aspect, a computer-readable storage medium is provided, which stores one or more programs that, when executed by an electronic device including a plurality of application programs, cause the electronic device to perform the operations of the second aspect.

According to the technical scheme provided by the embodiment of the specification, the personalized conversation style matched with the user is determined, the intention information expressed by the user message is accurately predicted based on the machine learning algorithm, the conversation mode is determined based on the intention information and the historical conversation record, and therefore the personalized reply content is determined according to the determined conversation style, intention information and conversation mode and is returned to the user side. Furthermore, the method can not only ensure that the reply content is accurately matched by predicting the intention, but also enrich the reply content by adopting the personalized conversation style and even adding emotional colors, and execute one or more rounds of conversation interaction under the determined conversation mode. Therefore, flexibility and richness of dialog reply are improved, and use viscosity of a user is increased.

Drawings

In order to more clearly illustrate the embodiments of the present specification or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below, it is obvious that the drawings in the following description are only some embodiments described in the present specification, and for those skilled in the art, other drawings can be obtained according to the drawings without any creative effort.

Fig. 1 is a diagram of an applicable scene architecture of an interactive system in an embodiment of the present specification, provided by an embodiment of the present specification.

Fig. 2 is a structural diagram of an interactive system of an intelligent human-computer conversation provided by an embodiment of the present specification.

Fig. 3 is a schematic structural diagram of functional units of a preprocessing module according to an embodiment of the present disclosure.

Fig. 4a is a schematic processing flow diagram of a dialog style matching module according to an embodiment of the present disclosure.

Fig. 4b is a flowchart illustrating a process of constructing a user image by the dialog style matching module according to still another embodiment of the present specification.

Fig. 5 is a functional unit structure diagram of a deep semantic understanding module provided in an embodiment of the present specification.

Fig. 6 is a schematic structural diagram of functional units of a dialog management module provided in an embodiment of the present specification.

FIG. 7 is a knowledge base included in a predetermined intent classification model provided by an embodiment of the present description.

Fig. 8 is a schematic flowchart illustrating the cooperation of the dialog management module and the decision module according to an embodiment of the present disclosure.

Fig. 9 is a schematic step diagram of an interaction method of an intelligent human-machine conversation provided by an embodiment of the present specification.

Fig. 10 is a flowchart illustrating an intelligent interactive interaction between an intelligent service robot and a user side according to an embodiment of the present disclosure.

Fig. 11 is a flow chart of the dialog interaction of the personalized service robot in the call-out scenario provided in example 1 of the present specification.

Fig. 12 is a flowchart of the dialogue interaction of the personalized customer service robot in the wechat interaction scenario provided in example 2 of the present specification.

Fig. 13 is a flow chart of the dialog interaction of the personalized customer service robot in the direct human-computer interaction scenario provided in example 3 of the present specification.

Fig. 14 is a flowchart of the dialog interaction of the personalized customer service robot in the web page interaction scenario provided in example 4 of the present specification.

Fig. 15 is a schematic structural diagram of an electronic device provided in an embodiment of the present specification.

Detailed Description

In order to make those skilled in the art better understand the technical solutions in the present specification, the technical solutions in the embodiments of the present specification will be clearly and completely described below with reference to the drawings in the embodiments of the present specification, and it is obvious that the described embodiments are only a part of the embodiments of the present specification, and not all of the embodiments. All other embodiments obtained by a person skilled in the art based on the embodiments in the present specification without any inventive step should fall within the scope of protection of the present specification.

Referring to fig. 1, an applicable scenario architecture diagram of an interactive system in an embodiment of the present specification may be a scenario of a dialog interaction between an interactive system 11 of an intelligent human-computer dialog and an end user side 12. The interactive system 11 may be regarded as a robot or an intelligent customer service robot, and may include a main set of function modules, and a database set and a model set used in cooperation with the function modules. The terminal user side 12 can perform dialogue interaction with the interactive system 11 through various interaction channels such as outgoing call, mobile phone APP, WEB page, entity robot, and the like. In a specific implementation, the interactive system 11 may perform semantic understanding analysis according to the received data information acquired by the data acquisition device 13, and particularly predict a language style by combining with a machine learning model, match a reply content adapted to the conversation style, and feed the reply content back to the terminal user side 12. Therefore, personalized service of conversation interaction is achieved, and conversation interaction experience is improved.

In the following, referring to fig. 2, an interactive system for intelligent human-computer conversation provided by an embodiment of the present specification is described, where the interactive system may include: a preprocessing module 21, a dialogue style matching module 22, a deep semantic understanding module 23, a dialogue management module 24, a decision module 25 and a reply module 26. In fact, the interactive system is not limited to include the above modules, and may also include other functional modules that assist in implementing intelligent man-machine conversation, which are not described one by one here. In addition, the interactive system also comprises various models matched with the internal functional modules.

< pretreatment Module >

The preprocessing module 21 is configured to perform preprocessing operation on a received user message, where the user message carries user basic data.

Referring to fig. 3, the preprocessing module 21 may include at least one or a combination of the following units: a text correction unit 211, a language conversion unit 212, a numeral conversion unit 213, a special word normalization unit 214; wherein,

the text error correction unit 211 is configured to correct a text error in the received user message, for example, a text error existing in a user message input by a user through a keyboard or input by voice. Wherein the corrective action includes at least: homophonic/harmonic correction, dialect correction, fuzzy sound correction, disorder correction and automatic completion.

The language conversion unit 212 is configured to convert the language type and/or the language label in the user message into a language version currently supported by the interactive system. For example, english in a user message input by the user is converted into chinese, or pinyin in the user message is converted into chinese. The language label referred to here may be pinyin, and the language version currently supported by the interactive system may be chinese, english, or other languages.

The digital conversion unit 213 is configured to convert the digits in the user message into a digital version currently supported by the interactive system; for example, a description of a user message containing Chinese numbers such as "one, thirty-one, two-ten-yuan" may be converted to Arabic numbers for international unified use. In fact, the digital version currently supported by the interactive system is not limited to arabic numerals, but can also be chinese numerals.

The special word standardizing unit 214 is configured to unify special words used in a specific field in the user message into standardized words, for example, a set of standard words can be defined uniformly in a specific field such as "300 megabases, 1000M traffic, 50G traffic, 199 yuan package, number 10000" in the telecommunication field, and in a case that there may be multiple expressions.

Through the units contained in the preprocessing module, preprocessing operation in various modes can be carried out on the user message received from the user, so that the language text of the user message is more standard, and subsequent semantic recognition and understanding are facilitated.

< dialogue Style matching Module >

The conversation style matching module 22 is configured to determine a conversation style matched with the user message according to the user's autonomous selection and/or the user basic data.

Optionally, the conversation style matching module 22 is specifically configured to select a conversation style matched with the user from a preset conversation style library according to the autonomous selection of the user and/or the user basic data. During specific implementation, whether a user independently selects a conversation style type before sending a user message is judged, and if the user independently selects the conversation style type, a conversation style matched with the conversation style type independently selected by the user is selected from a preset conversation style library; if not, analyzing the habits and styles of the user based on the basic data of the user, and selecting the conversation style matched with the user from the preset conversation style library.

Further, for the case that the user selects the conversation style autonomously, when the conversation style matching module 22 selects the conversation style matching the user from the preset conversation style library according to the user basic data, the conversation style matching module is specifically configured to:

constructing a user representation according to the user basic data, wherein the user representation at least comprises: a user fact tag and a user model tag; selecting a conversation style matched with the user from a preset conversation style library according to a label contained in the user portrait;

or,

constructing a user representation according to the user basic data, wherein the user representation at least comprises: a user fact tag and a user model tag; and selecting a conversation style matched with the user from a preset conversation style library according to the label contained in the user portrait and the user message.

The preset conversation style library is constructed in the following way: acquiring user data related to a conversation in a network by using a crawler tool to serve as a training sample; inputting the training sample into a preset language style model for training to obtain different language style models; and the different language style models correspond to different types of conversation styles. In specific implementation, the crawler tool can be used for collecting public data on the Internet, such as data of microblogs, bean bars, Baidu sticking bars and the like, the public data are divided according to the ages and the genders of users, and models in different styles are respectively trained according to the user data of different ages and different genders. In one implementation, each language style model may define a class of customer service robots having a language style corresponding to the language style model.

It should be noted that a crawler or web crawler, also called a web spider, is a web robot for automatically browsing the world wide web. The purpose of the method is to compile network indexes, and web search engines and other websites update their own website contents or indexes to other websites through crawler software. The web crawler can save the pages visited by the web crawler, so that the search engine generates an index afterwards for the user to search.

It should be understood that in the embodiment of the present specification, the preset dialog style library may be pre-established by the interactive system, and is preferably constructed using the dialog style matching module 22.

Alternatively, an implementation may be that after obtaining different language style models, the dialogue style matching module 22 further assigns a character attribute representing the character of the interlocutor to the different language style models based on a text sequence generation technique and a deep neural network migration learning technique, where the character attribute is expressed in the dialogue style. Text sequence generation techniques include, but are not limited to: word-level, phrase-level text repetition techniques, and other sequence generation models, among others. For example, different personage data, such as Taobao, Fangyi, blue spirit, and dialect, are collected to create a customer service robot personality, so that the reply content has emotional colors, is more emotional and more intelligent.

Based on the above analysis, when constructing the preset dialogue style library, the dialogue style matching module 22 not only trains models of different language styles, but also gives different character attributes to the models of different language styles. In the concrete implementation, the customer service robots with different language styles can be obtained through training, and different character colors are given to the customer service robots with different language styles. Therefore, the man-machine interaction is more intelligent, and the conversation content is rich.

In another embodiment, the dialogue style matching module 22 may be a personalized customer service module, which first collects data, collects corpora on the internet, classifies language models of different styles, and after data is cleaned, stores data that may have answers as candidate answers of a search model directly, and after the remaining data is screened, uses the data as personalized language modeling data to perform generative language model reply and general reply processing data. Specifically, referring to fig. 4a, the personalized customer service module may include the following sub-modules:

-a corpus processing module: the system mainly performs work such as corpus collection, cleaning, mining, classification, storage and the like, is a foundation of a personalized customer service module, and provides important data guarantee for subsequent modules.

-an emotion analysis module: the emotion analysis method mainly analyzes emotion states in user input information and utilizes a traditional machine learning algorithm and a deep learning algorithm to conduct emotion analysis on word level, sentence level, target level and other dimensions. Wherein, the target level sentiment analysis considers a specific target, and the target can be an entity and/or an attribute of a certain entity. In the embodiment of the specification, emotion analysis can better understand the state of the user and match personalized reply content in time, and the satisfaction degree of the user is improved. For example, detecting that the user is currently at a low mood, some positive energy or humorous style statements may be replied to strive to improve the user's low mood.

-a content generation module: the method mainly comprises the steps of matching existing content based on a retrieval mode, generating new content based on a generation mode and generating new content after the universal reply is processed in a personalized mode. And generating personalized contents meeting the requirements in the corpus based on the retrieved contents mainly through a semantic matching technology. The semantic matching model is mainly used for carrying out language representation through training a language model and variants thereof and carrying out similarity scoring training by combining a specific matching task. Meanwhile, the powerful training language model can provide powerful guarantee for subsequent content generation. The generation based on the generated content is mainly realized by generating models, such as migration learning, style conversion and the like from the sequence to the sequence model and its variants. The universal reply personalized processing is mainly realized by technical means such as a specific template or rule and the like.

In an embodiment of the present description, referring to FIG. 4b, dialog style matching module 22, when constructing a user representation according to user basic data, may at least include: a fact tag unit and a model tag unit; in addition, the method can further comprise the following steps: an original tag unit and a predicted tag unit; wherein,

the original tag unit is mainly used for processing the acquired user basic data (which can be historical conversation records or user historical data), abstracting the acquired user basic data into a plurality of tags, for example, collecting user natural features such as gender, age and region of a user, and simultaneously performing data such as user interest features (interest preference, system use habits and the like), social features (marital families, social relations and the like), consumption features (product purchase channels, time frequency and purchasing power features) and the like to construct an original tag;

the fact label unit is used for extracting and establishing a user fact label from the original label through a text mining algorithm and a statistical analysis algorithm;

the model label unit is used for training through a statistical analysis algorithm and/or a semantic analysis algorithm to obtain a user model label based on the established user fact label;

and the prediction tag unit is used for predicting the user basic data input by the current user based on the user model tag to obtain a prediction tag.

The algorithm models involved in the various label units include, but are not limited to: EM algorithms, GM algorithms, classification algorithms such as decision trees, logistic regression algorithms, random forest algorithms, deep learning algorithms, and the like; the corresponding tag types in the various types of tag units can be referred to as an example in fig. 4.

< deep semantic understanding Module >

The deep semantic understanding module 23 is configured to perform semantic parsing on the user message after the preprocessing operation by using a deep learning algorithm, and identify intention information expressed by the user message.

Referring to fig. 5, the deep semantic understanding module 23 specifically includes: a domain identifying unit 231, an intention identifying unit 232, and a semantic parsing unit 233; wherein,

the domain identification unit 231 is configured to identify information of a service domain where the user message is located by using a rule algorithm, a machine learning algorithm, and a deep learning algorithm; in other words, the domain identification unit 231 is mainly used to identify the domain range where the user intends to be located, where the domain range or the service domain information includes: domain-specific business knowledge, domain-specific industry knowledge, general encyclopedia knowledge, daily chatting, and the like. The field identification mainly adopts the matching of rules combined with a machine learning algorithm and a deep learning algorithm. The rule model and the machine learning model mainly solve the problem of cold start of the system, for example, a new product is applied on line, the collected available information is not much, and at the moment, the field can be identified by combining the rule model and the machine learning model; the deep learning algorithm solves the problem that the field recognition accuracy is improved after the system runs for a period of time.

The intention identifying unit 232 is configured to perform text classification on the user message by using a rule algorithm, a machine learning algorithm, and a deep learning algorithm, so as to identify intention information expressed by the user message in a service field corresponding to the service field information; the intention identification unit is mainly used for identifying the exact intention of the user, for example, the user uses the question-answering system to perform traffic package inquiry in the telecommunication field, wherein the telecommunication field is the result of identification analysis, and the traffic inquiry is specific intention information needing to be identified. The intention recognition algorithm is generally the same as the field recognition algorithm, and mainly uses the rule + machine learning + deep learning algorithm to classify texts, which is not described herein.

The semantic parsing unit 233 is configured to identify semantic slot information related to the intention information by using a named entity identification algorithm and a deep learning algorithm. The semantic parsing unit is mainly used for identifying necessary and optional semantic slot information related to intentions, for example, in the intention of 'traffic query', the necessary and optional information may be 'user mobile phone number', 'query year and month'; the optional information may be information such as "user zone". Semantic parsing is mainly performed by adopting a named entity recognition algorithm such as rule extraction and conditional random field CRF and a deep learning model.

The deep semantic understanding module is used for identifying the final intention of the user, constantly making clear and reasoning on the intention by combining the context data model and the field data model, completing the operations of intention completion, intention classification, intention transfer and the like, and accordingly analyzing more accurate and real intention information.

< dialogue management Module >

The dialogue management module 24 is configured to determine a dialogue mode of the user message during the dialogue interaction according to the intention information and a historical dialogue record of the user corresponding to the user message.

It should be understood that the dialog interactions referred to in the embodiments of the present description may be one or more rounds of interactions, that is, the intelligent customer service robot may perform one dialog with the user or may perform multiple rounds of reply dialogs. The method is mainly used for judging when to perform omission recovery, reference resolution, dialog state clarification, interactive question reversal and the like.

Optionally, as shown in fig. 6, the dialog management module 24 specifically includes: a dialog state tracking unit 241 and a dialog action selection unit 242; wherein,

the dialog state tracking unit 241 is configured to determine a current user state based on the intention information and the semantic slot information identified by the deep semantic understanding module in combination with the historical dialog record of the user. Specifically, the dialog state tracking unit 241 is responsible for receiving the slot position and the intention information after the natural language understanding module performs language identification, and obtains the user state at the current time by integrating the slot position, the intention information, and the historical state information of the user in the current turn.

The dialog action selection unit 242 is configured to determine a dialog action to be currently executed based on a preset algorithm model and a preset intention classification model. Specifically, the dialogue action selection unit may determine which action to perform jointly using a model such as a reinforcement learning model, a machine learning and deep learning model, and a knowledge base of a back support. Wherein the preset algorithm model at least comprises: a reinforcement learning model, a machine learning model and a deep learning model; referring to fig. 7, the preset intention classification model at least includes: domain-specific atlas database, general encyclopedia database, domain-specific FAQ database, chat FAQ database, rules database, and dictionary database. Through the preset intention classification models, under the condition that the problems of the user are highly repeated, intention information analyzed directly based on the user information can be matched from the intention classification models, manpower is released to a great extent, and labor cost is saved. Wherein the specific domain may be a telecommunications domain or other domain.

In other words, as shown in fig. 8, after the user message is analyzed, the slot and the intention information are obtained, it is confirmed whether the user intention is complete, the omission restoration is performed if the intention is missing but the omission restoration condition is satisfied, the interaction is performed to inquire more information if the intention is missing and the omission restoration condition is not satisfied, and the intention clarification is performed if the information collection is complete (the slot information is complete) but it is still not determined whether the user intention is the final intention. And the dialogue management module applies a reinforcement learning algorithm and obtains the current optimal action through the reward function maximization, namely the reward is a positive value if the interaction result is satisfied through user feedback, and the reward replied by the current system is set to be a negative value if the user provides the system with unsatisfied results, so that the iterative optimization is continuously performed.

The decision module 25 is configured to determine the reply content based on a plurality of confidence level results generated by the intention information in a preset intention classification model. In specific implementation, the decision module is configured to sort a plurality of confidence level results generated by the intention information in a preset intention classification model, and determine reply content based on a prediction result of the preset intention classification model with a confidence level greater than a threshold.

And the reply module 26 is configured to return reply content corresponding to the user message to the user based on the conversation mode and by using the conversation style. That is, the reply module performs secondary processing on the reply content determined by the decision module according to the session mode determined by the session management module and the session style determined by the session style module to generate reply contents with different language styles, and the reply contents may also show emotional colors, so that the individualized reply contents are returned to the terminal user side.

Through the technical scheme, the personalized conversation style matched with the user can be determined, the intention information expressed by the user information is accurately predicted based on the machine learning algorithm, the conversation mode is determined based on the intention information and the historical conversation record, and therefore the personalized reply content is determined according to the determined conversation style, the intention information and the conversation mode and is returned to the user side. Furthermore, the method can not only ensure that the reply content is accurately matched by predicting the intention, but also enrich the reply content by adopting the personalized conversation style and even adding emotional colors, and execute one or more rounds of conversation interaction under the determined conversation mode. Therefore, flexibility and richness of dialog reply are improved, and use viscosity of a user is increased.

Referring to fig. 9, a schematic diagram illustrating steps of an interaction method for an intelligent human-machine conversation provided in an embodiment of the present specification, where the method may include the following steps:

step 302: and preprocessing the received user message, wherein the user message carries user basic data.

Optionally, the preprocessing operation is performed on the received user message, which specifically includes one or a combination of the following operations:

text error correction operation: correcting the text errors in the received user message, wherein the correcting operation at least comprises the following steps: homophonic/harmonic correction, dialect correction, fuzzy sound correction, disorder correction and automatic completion;

language conversion operation: converting the language type and/or language label in the user message into a language version currently supported by the interactive system;

digital conversion operation: converting the numbers in the user message into a digital version currently supported by the interactive system;

special word standardization operation: and unifying special words used in a specific field in the user message into a standardized expression.

Step 304: and determining a conversation style matched with the user message according to the autonomous selection of the user and/or the user basic data.

Optionally, in step 304, when determining a dialog style matched with the user message according to the user's autonomous selection and/or the user basic data, specifically performing: and selecting a conversation style matched with the user from a preset conversation style library according to the autonomous selection of the user and/or the basic data of the user.

Further, the dialog style may be determined in two specific ways:

the first method is as follows:

the second method comprises the following steps:

In an embodiment of the present specification, the preset dialog style library is constructed by: acquiring user data related to a conversation in a network by using a crawler tool to serve as a training sample; inputting the training sample into a preset language style model for training to obtain different language style models; and the different language style models correspond to different types of conversation styles.

In fact, after obtaining different language style models, the different language style models can be given character attributes representing characters of an interlocutor based on a text sequence generation technology and a deep neural network transfer learning technology, wherein the character attributes are expressed in the conversation style.

It should be appreciated that the constructing a user representation from the user base data includes at least: establishing a user fact label through a text mining algorithm and a statistical analysis algorithm; and training by a statistical analysis algorithm and/or a semantic analysis algorithm based on the established user fact label to obtain a user model label.

Step 306: and performing semantic analysis on the preprocessed user message by adopting a deep learning algorithm, and identifying intention information expressed by the user message.

Optionally, in step 306, when performing semantic parsing on the preprocessed user message by using a deep learning algorithm to identify intention information expressed by the user message, the method may specifically perform:

step one, identifying the information of the service field where the user message is located by adopting a rule algorithm, a machine learning algorithm and a deep learning algorithm; the service domain information at least comprises: domain specific business knowledge, domain specific industry knowledge, general encyclopedia knowledge, and daily chat content.

And secondly, text classification is carried out on the user message by adopting a rule algorithm, a machine learning algorithm and a deep learning algorithm so as to identify intention information expressed by the user message in a service field corresponding to the service field information.

And thirdly, recognizing semantic slot information related to the intention information by adopting a named entity recognition algorithm and a deep learning algorithm.

Step 308: and determining a conversation mode of the user message during the conversation interaction according to the intention information and the historical conversation record of the user corresponding to the user message.

Optionally, in step 308, when determining a dialog mode of the user message during the dialog interaction according to the intention information and a historical dialog record of the user corresponding to the user message, the method may specifically perform: determining the current user state based on the intention information and the semantic slot information and combined with the historical dialogue record of the user; and determining the current conversation action to be executed based on the preset algorithm model and the preset knowledge base model. It should be understood that after learning the user state and the session actions to be performed, it is equivalent to determining the dialog mode; wherein the session action comprises: omitting recovery, referring to resolution, dialog state clarification, interactive question reversal, etc. The preset algorithm model at least comprises: a reinforcement learning model, a machine learning model and a deep learning model; the preset knowledge base model at least comprises: domain-specific atlas database, general encyclopedia database, domain-specific FAQ database, chat FAQ database, rules database, and dictionary database.

Step 310: and determining reply content based on a plurality of confidence degree results generated by the intention information in a preset intention classification model.

The scheme can be implemented by sorting a plurality of confidence level results generated by the intention information in a preset intention classification model, and determining reply content based on a predicted result of the preset intention classification model with the confidence level larger than a threshold value.

Step 312: and returning reply content corresponding to the user message to the user based on the conversation mode and by adopting the conversation style.

The specific implementation of each step shown in fig. 9 may refer to the interactive system scheme provided in the first embodiment, which is not described herein again.

The following description will be given by taking an example in which the interactive system is a personalized customer service robot, and the personalized customer service robot can realize diversified interactive interaction.

Referring to fig. 10, a flowchart of the intelligent service robot performing intelligent dialogue interaction with the user side is shown.

Firstly, a user side sends a user message to a personalized customer service robot in a voice or text input mode, the personalized customer service robot receives the user message, and respectively realizes semantic understanding, style matching, session management, decision reply and other operations according to a self-configured function module, and then content packaging is carried out on the determined reply content, the conversation mode and the conversation style, and the reply content, the conversation mode and the conversation style are output to the user side in a text or voice mode.

Example 1: telephone outbound scenario

Referring to fig. 11, in this scenario, a user performs a dialogue interaction with the personalized customer service robot through a telephone, and the system defaults that the personalized customer service robot has collected historical interaction information in advance through a telephone outbound mode to establish various models and knowledge bases; correspondingly, the personalized robot mainly executes the following steps in the conversation interaction process:

step 402: receiving a telephone voice message sent by a user: what traffic package is there now?

It should be appreciated that in this step 402, the received voice message carries a user grounding message to facilitate creation of a user representation.

Step 404: and the personalized customer service robot is established based on the voice message and has a user portrait matching with a stable style.

Step 406: and (4) intention classification: a domain of telecommunications map.

Step 408: and (3) session management: no interaction is required and the intention is clear.

Step 410: and (4) decision content: the existing flow package is as follows: 19 yuan package, 59 yuan package and 99 yuan package.

Step 412: and (3) individual reply: mr. good, the existing flow package XXXX recommends that you use a 99 yuan national unlimited flow package according to your basic situation.

Therefore, according to the voice message and the basic information of the user carried by the voice message, the personalized reply content which accords with the style and the current emotion of the user is determined in a matching mode, and the user can directly reply without interaction.

Example 2: WeChat interaction scenario

Referring to fig. 12, in this scenario, a user performs dialogue interaction with the personalized customer service robot through WeChat, and the personalized customer service robot defaults to previously collect historical interaction information to establish various models and knowledge bases; correspondingly, the personalized robot mainly executes the following steps in the conversation interaction process:

step 502: receiving a WeChat voice message sent by a user: which sending voice call?

It should be appreciated that in this step 502, the received voice message carries a user grounding message to facilitate creation of a user representation.

Step 504: and the personalized customer service robot is established based on the voice message, and the user portrait is matched with the lovely style.

Step 506: and (4) intention classification: a domain of telecommunications map.

Step 508: and (3) session management: the recovery needs to be omitted and the recovered entity is a traffic package.

Step 510: and (4) decision content: the 19 yuan package is given without voice, the 59 yuan package is given for the call duration of 100 minutes, and the 99 yuan package is given for the call duration of 300 minutes.

Step 512: and (3) individual reply: miss, just help you inquire about the next, at present we give a 59 yuan package for 100 minutes of call duration, a 99 yuan package for 300 minutes of call duration, see what you have without bell?

Therefore, the reply content with the lovely style can be matched according to the user message, the intention expressed by the user message is not clear, the recovery needs to be omitted, and the reply content can be directly matched and determined for the user after the recovery.

Example 3: direct human-computer interaction scenario

Referring to fig. 13, in this scenario, a user directly performs dialogue interaction with the personalized customer service robot, and the system defaults that the personalized customer service robot has previously collected historical interaction information to establish various models and knowledge bases; correspondingly, the personalized robot mainly executes the following steps in the conversation interaction process:

step 602: receiving a text message sent by a user: how much you are?

It should be appreciated that in this step 602, the received voice message carries a user grounding message to facilitate creation of a user representation.

Step 604: and the user portrait established based on the text message is matched with the personalized customer service robot with a lovely style.

Step 606: and (4) intention classification: chat to the FAQ.

Step 608: and (3) session management: the intention is clear.

Step 610: and (4) decision content: i am 5 years old this year.

Step 612: and (3) individual reply: thanks to your interest, i am 5 years old today, not very lovely?

In this example, the customer service robot with the lovely style can be determined to reply by analyzing and matching according to the user message, that is, the reply content is lovely and friendly and is rich in infectivity.

Example 4: web page interaction scenario

Referring to fig. 14, in this scenario, a user performs a dialogue interaction with the personalized customer service robot through a web page, and the personalized customer service robot defaults to have previously collected historical interaction information to establish various models and knowledge bases; correspondingly, the personalized robot mainly executes the following steps in the conversation interaction process:

step 702: receiving a text message sent by a user: which sending voice call?

It should be appreciated that in this step 702, the received voice message carries a user grounding message to facilitate creation of a user representation.

Step 704: and the user portrait established based on the text message is matched with the personalized customer service robot with a lovely style.

Step 706: and (4) intention classification: a domain of telecommunications map.

Step 708: and (3) session management: the recovery is omitted and the recovered entity is a flow package.

Step 710: and (4) decision content: the 19 yuan package is given without voice, the 59 yuan package is given for the call duration of 100 minutes, and the 99 yuan package is given for the call duration of 300 minutes.

Step 712: and (3) individual reply: miss, just help you inquire about the next, at present we give a 59 yuan package for 100 minutes of call duration, a 99 yuan package for 300 minutes of call duration, see what you have without bell?

In the above example 4, the user communicates with the intelligent customer service through the web page without collecting user information, the system pushes the customer service robot matching the user by default, and the user has a clear intention of inputting information and can directly match answers.

According to the 4 examples, the personalized customer service robot can perform conversation interaction with the user side in various interaction modes, and can determine reply contents matched with the language style and the emotional color of the user based on the received user information, so that personalized reply service is improved, and the use viscosity of the user is increased; by the scheme of the specification, the user intention expressed by the user message can be accurately analyzed, and the accuracy of the reply content is further improved.

Fig. 15 is a schematic structural diagram of an electronic device according to an embodiment of the present specification. Referring to fig. 15, at a hardware level, the electronic device includes a processor, and optionally further includes an internal bus, a network interface, and a memory. The Memory may include a Memory, such as a Random-Access Memory (RAM), and may further include a non-volatile Memory, such as at least 1 disk Memory. Of course, the electronic device may also include hardware required for other services.

The processor, the network interface, and the memory may be connected to each other via an internal bus, which may be an ISA (Industry Standard Architecture) bus, a PCI (Peripheral Component Interconnect) bus, an EISA (Extended Industry Standard Architecture) bus, or the like. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of illustration, only one double-headed arrow is shown in FIG. 15, but that does not indicate only one bus or one type of bus.

And the memory is used for storing programs. In particular, the program may include program code comprising computer operating instructions. The memory may include both memory and non-volatile storage and provides instructions and data to the processor.

The processor reads the corresponding computer program from the nonvolatile memory into the memory and then runs the computer program to form the shared resource access control device on the logic level. The processor is used for executing the program stored in the memory and is specifically used for executing the following operations:

The method performed by the interactive system of intelligent human-machine conversation disclosed in the embodiments shown in the figures of the present specification can be applied to or implemented by a processor. The processor may be an integrated circuit chip having signal processing capabilities. In implementation, the steps of the above method may be performed by integrated logic circuits of hardware in a processor or instructions in the form of software. The Processor may be a general-purpose Processor, including a Central Processing Unit (CPU), a Network Processor (NP), and the like; but also Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field Programmable Gate Arrays (FPGAs) or other Programmable logic devices, discrete Gate or transistor logic devices, discrete hardware components. The various methods, steps and logic blocks disclosed in the embodiments of the present specification may be implemented or performed. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like. The steps of a method disclosed in connection with the embodiments of the present specification may be embodied directly in a hardware decoding processor, or in a combination of hardware and software modules in the decoding processor. The software module may be located in ram, flash memory, rom, prom, or eprom, registers, etc. storage media as is well known in the art. The storage medium is located in a memory, and a processor reads information in the memory and completes the steps of the method in combination with hardware of the processor.

The electronic device can also execute the method in the attached drawings and realize the functions of the interactive system of the intelligent man-machine conversation in the embodiment shown in the attached drawings, and the embodiment of the specification is not described again here.

Of course, besides the software implementation, the electronic device of the embodiment of the present disclosure does not exclude other implementations, such as a logic device or a combination of software and hardware, and the like, that is, the execution subject of the following processing flow is not limited to each logic unit, and may also be hardware or a logic device.

Embodiments of the present specification also propose a computer-readable storage medium storing one or more programs, the one or more programs comprising instructions, which when executed by a portable electronic device comprising a plurality of application programs, are capable of causing the portable electronic device to perform the method of the embodiments shown in the drawings, and in particular to perform the method of:

In short, the above description is only a preferred embodiment of the present disclosure, and is not intended to limit the scope of the present disclosure. Any modification, equivalent replacement, improvement and the like made within the spirit and principle of the present specification shall be included in the protection scope of the present specification.

The systems, devices, modules or units illustrated in the above embodiments may be implemented by a computer chip or an entity, or by a product with certain functions. One typical implementation device is a computer. In particular, the computer may be, for example, a personal computer, a laptop computer, a cellular telephone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

Computer-readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), Static Random Access Memory (SRAM), Dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), Digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium that can be used to store information that can be accessed by a computing device. As defined herein, a computer readable medium does not include a transitory computer readable medium such as a modulated data signal and a carrier wave.

It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.

The embodiments in the present specification are described in a progressive manner, and the same and similar parts among the embodiments are referred to each other, and each embodiment focuses on the differences from the other embodiments. In particular, for the system embodiment, since it is substantially similar to the method embodiment, the description is simple, and for the relevant points, reference may be made to the partial description of the method embodiment.

Claims

1. An interactive system for intelligent man-machine conversation, comprising:

2. The interactive system of claim 1, wherein the preprocessing module specifically comprises one or a combination of the following units: the system comprises a text error correction unit, a language conversion unit, a digital conversion unit and a special word standardization unit; wherein,

the text error correction unit is configured to correct a text error in the received user message, where the correction operation at least includes: homophonic/harmonic correction, dialect correction, fuzzy sound correction, disorder correction and automatic completion;

the language conversion unit is used for converting the language type and/or the language label in the user message into a language version currently supported by the interactive system;

the digital conversion unit is used for converting the digits in the user message into a digital version currently supported by the interactive system;

the special word standardizing unit is used for unifying the special words used in the specific field in the user message into standardized terms.

3. The interactive system of claim 1, wherein the deep semantic understanding module specifically comprises: the system comprises a field identification unit, an intention identification unit and a semantic parsing unit; wherein,

the domain identification unit is used for identifying the information of the service domain where the user message is located by adopting a rule algorithm, a machine learning algorithm and a deep learning algorithm;

the intention identification unit is used for carrying out text classification on the user message by adopting a rule algorithm, a machine learning algorithm and a deep learning algorithm so as to identify intention information expressed by the user message in a service field corresponding to the service field information;

and the semantic parsing unit is used for identifying semantic slot information related to the intention information by adopting a named entity identification algorithm and a deep learning algorithm.

4. The interactive system of claim 3, wherein the business segment information includes at least: domain specific business knowledge, domain specific industry knowledge, general encyclopedia knowledge, and daily chat content.

5. The interactive system of claim 3 or 4, wherein the session management module specifically comprises: a dialogue state tracking unit and a dialogue action selection unit; wherein,

the dialog state tracking unit is used for determining the current user state based on the intention information and the semantic slot information identified by the deep semantic understanding module and combined with the historical dialog record of the user;

and the dialogue action selection unit is used for determining the current conversation action to be executed based on the preset algorithm model and the preset intention classification model.

6. The interactive system according to claim 5, characterized in that said preset algorithmic model comprises at least: a reinforcement learning model, a machine learning model and a deep learning model;

the preset intention classification model at least comprises: specific [ Tonya1] domain atlas database, general encyclopedia atlas database, specific domain FAQ database [ Tonya2], chatting FAQ database, rules database, and dictionary database.

7. The interactive system of claim 1, wherein the dialog style matching module is specifically configured to select a dialog style matching the user from a preset dialog style library according to the user's autonomous selection and/or the user basic data.

8. The interactive system of claim 7, wherein the dialog style matching module, when selecting the dialog style matching the user from a preset dialog style library based on the user base data, is specifically configured to:

constructing a user representation according to the user basic data, wherein the user representation at least comprises: a user fact tag and a user model tag;

selecting a conversation style matched with the user from a preset conversation style library according to a label contained in the user portrait;

or,

and selecting a conversation style matched with the user from a preset conversation style library according to the label contained in the user portrait and the user message.

9. The interactive system of claim 8, wherein the library of preset dialog styles is constructed by:

acquiring user data related to a conversation in a network by using a crawler tool to serve as a training sample;

inputting the training sample into a preset language style model for training to obtain different language style models; and the different language style models correspond to different types of conversation styles.

10. The interactive system of claim 9, after obtaining the different linguistic models, further comprising:

and assigning character attributes representing characters of the interlocutors to the different language style models based on a text sequence generation technology and a deep neural network transfer learning technology, wherein the character attributes are represented in the dialogue style.

11. The interactive system of claim 9, wherein the dialog style matching module comprises at least:

the fact label unit is used for establishing a user fact label through a text mining algorithm and a statistical analysis algorithm;

and the model label unit is used for training through a statistical analysis algorithm and/or a semantic analysis algorithm to obtain a user model label based on the established user fact label.

12. The interactive system of claim 6,

the decision module is specifically configured to rank a plurality of confidence level results generated by the intention information in a preset intention classification model, and determine reply content based on a prediction result of the preset intention classification model whose confidence level is greater than a threshold value.

13. An interaction method of intelligent man-machine conversation is characterized by comprising the following steps:

14. The interactive method according to claim 13, wherein the preprocessing operation is performed on the received user message, and specifically includes one or a combination of the following operations:

15. The interaction method according to claim 13, wherein performing semantic parsing on the user message after the preprocessing operation by using a deep learning algorithm to identify intention information expressed by the user message specifically comprises:

identifying the information of the service field where the user message is located by adopting a rule algorithm, a machine learning algorithm and a deep learning algorithm;

adopting a rule algorithm, a machine learning algorithm and a deep learning algorithm to classify the text of the user message so as to identify intention information expressed by the user message in a service field corresponding to the service field information;

and identifying semantic slot information related to the intention information by adopting a named entity identification algorithm and a deep learning algorithm.

16. The interactive method of claim 15, wherein the business domain information comprises at least: domain specific business knowledge, domain specific industry knowledge, general encyclopedia knowledge, and daily chat content.

17. The interaction method according to claim 15 or 16, wherein determining a dialog mode of the user message during the current dialog interaction according to the intention information and a history dialog record of the user corresponding to the user message specifically includes:

determining the current user state based on the intention information and the semantic slot information and combined with the historical dialogue record of the user;

and determining the current conversation action to be executed based on the preset algorithm model and the preset knowledge base model.

18. The interactive method according to claim 17, characterized in that said preset algorithmic model comprises at least: a reinforcement learning model, a machine learning model and a deep learning model;

the preset knowledge base model at least comprises: domain-specific atlas database, general encyclopedia database, domain-specific FAQ database, chat FAQ database, rules database, and dictionary database.

19. The interaction method according to claim 13, wherein determining a dialog style matching the user message according to the user's autonomous selection and/or the user basic data specifically comprises:

and selecting a conversation style matched with the user from a preset conversation style library according to the autonomous selection of the user and/or the basic data of the user.

20. The interaction method according to claim 13, wherein selecting a dialog style matching the user from a preset dialog style library according to the user's autonomous selection and/or the user basic data specifically comprises:

or,

21. The interaction method of claim 20, wherein the library of preset dialog styles is constructed by:

22. The interactive method of claim 21, after deriving the different linguistic models, the method further comprising:

23. The interactive method of claim 21, wherein said constructing a user representation from said user base data comprises at least:

establishing a user fact label through a text mining algorithm and a statistical analysis algorithm;

and training by a statistical analysis algorithm and/or a semantic analysis algorithm based on the established user fact label to obtain a user model label.

24. The interaction method according to claim 18, wherein determining the reply content based on the plurality of confidence level results generated by the intention information in a preset intention classification model comprises:

and sequencing a plurality of confidence level results generated by the intention information in a preset intention classification model, and determining reply content based on a prediction result of the preset intention classification model with the confidence level being greater than a threshold value.

25. An electronic device, comprising:

a processor; and

a memory arranged to store computer executable instructions that, when executed, cause the processor to perform the steps of the intelligent human machine dialog interaction method of any of claims 13-24.

26. A computer readable storage medium storing one or more programs which, when executed by an electronic device including a plurality of application programs, cause the electronic device to perform the steps of the interaction method of intelligent human-machine conversation as claimed in any one of claims 13 to 24.