CN111177338B

CN111177338B - Context-based multi-round dialogue method

Info

Publication number: CN111177338B
Application number: CN201911222528.2A
Authority: CN
Inventors: 孙晓光; 刘为民; 游峰磊
Original assignee: Beijing Borui Tongyun Technology Co ltd
Current assignee: Beijing Borui Tongyun Technology Co ltd
Priority date: 2019-12-03
Filing date: 2019-12-03
Publication date: 2023-07-21
Anticipated expiration: 2039-12-03
Also published as: CN111177338A

Abstract

The invention relates to a multi-round dialogue method based on context, which comprises the following steps: the natural language understanding system receives sentence data input by a current user; identifying scene data, intention information and semantic slot information in the sentence according to the session control data; carrying out semantic matching according to scene data, intention information and state information corresponding to the intention information of the current sentence in the current sentence to obtain semantic matching result data, and sending the semantic matching result data to a natural language processing system; the natural language processing system performs context query according to the semantic matching result data to obtain a context query result; when the number of the contextual query results is multiple, carrying out semantic slot verification on the semantic slot information corresponding to the contextual query results according to the priority in the contextual query results; and obtaining the result data of the current round of session according to the result of the semantic slot verification.

Description

Context-based multi-round dialogue method

Technical Field

The invention relates to the technical field of data processing, in particular to a multi-round dialogue method based on context.

Background

At present, along with popularization of smart phones, updating of communication technology and vigorous development of mobile phone application ecology, people's life is more and more convenient, and information circulation speed is also faster, so that people and mobile phones are more and more closely contacted, and the mobile phones become important tools for people to acquire information.

The mobile phone naturally supports the characteristic of voice and the mature voice recognition and voice synthesis technology at present, so that the voice application has unique advantages in the aspects of direct voice interaction, convenience for user operation and increase of user experience. The current voice is mainly applied to a voice input method, an intelligent robot, an intelligent dialogue system and the like. Intelligent robots and intelligent dialog systems are generally embodied in accomplishing system setup, simple question-answer chat. These applications are struggling in the health field and lack contextual interactions.

Disclosure of Invention

The invention aims at overcoming the defects of the prior art, provides a multi-round dialogue method based on context, combines a natural language understanding system and a natural language processing system in the health field, can enable a user to acquire health knowledge through multi-round dialogue in a voice mode through a context association method, enriches the way of acquiring the health knowledge by the user, fills the blank in the aspect of health voice, and improves the experience of the user.

To achieve the above object, the present invention provides a context-based multi-round dialog method, which includes:

the natural language understanding system receives sentence data input by a current user;

identifying scene data, intention information and semantic slot information in the statement according to the session control data;

carrying out semantic matching according to scene data, intention information and state information corresponding to the intention information of the current sentence in the current sentence to obtain semantic matching result data, and sending the semantic matching result data to a natural language processing system;

the natural language processing system performs context query according to the semantic matching result data to obtain a context query result;

when the context query results are multiple, carrying out semantic slot verification on semantic slot information corresponding to the context query results according to the priority in the context query results;

and obtaining the result data of the current round of session according to the result of the semantic slot verification, and updating the session control data corresponding to the intention information of the current sentence according to the result data of the current round of session.

Preferably, before the natural language understanding system receives sentence data of the current user data, the method further includes:

the natural language understanding system is constructed by a Baxwell model.

Preferably, before the identifying scene data, intention information, and semantic slot information in the sentence according to the session control data, the method further includes:

acquiring the user ID of the current user;

session control data corresponding to the user ID of the current user is acquired.

Preferably, the performing semantic matching according to the scene data, the intention information and the state information corresponding to the intention information of the current sentence in the current sentence, and obtaining the semantic matching result data specifically includes:

determining whether the intention information of the current sentence is the same as the above intention information;

when the intention information of the current sentence is the same as the intention information of the text, performing first semantic matching on the scene data of the current sentence and the context association set information of the current sentence to obtain first semantic matching result data;

and when the intention information of the current sentence is different from the intention information of the text, performing second semantic matching on the scene data of the current sentence and the context association set information of the current sentence to obtain second semantic matching result data.

Preferably, the natural language processing system performs a contextual query according to the matching result data, and the obtaining a contextual query result specifically includes:

acquiring the intention information and the corresponding state information;

and carrying out context query according to the intention information, the corresponding state information and the matching result data of the context to obtain a context query result.

Preferably, when the contextual query results are multiple, the method further comprises:

when the multiple context query results comprise the context query result under the selection scene, verifying the context query result under the selection scene, and determining whether a corresponding matching item exists;

and if the corresponding matching item does not exist, deleting the context query junction in the current selection scene.

Preferably, according to the priority in the context query result, verifying the semantic slot information corresponding to the context query result specifically includes:

verifying semantic slot information corresponding to the contextual query result according to the priority in the contextual query result, and determining whether the semantic slot information corresponding to the contextual query result corresponds to the intention information of the current sentence;

and deleting the current context query result when the semantic slot information corresponding to the context query result does not correspond to the intention information of the current sentence.

Preferably, the obtaining the result data of the current round of session according to the result of the semantic slot verification specifically includes:

generating display output data according to the result of the semantic slot verification;

receiving a corresponding result of a user input by a current user according to the display output data;

and obtaining the result data of the current round of session according to the corresponding result of the user.

According to the context-based multi-round dialogue method provided by the embodiment of the invention, a natural language understanding system and a natural language processing system in the health field are combined, and a user can acquire health knowledge through multi-round dialogue in a voice mode by the context association method, so that the way of acquiring the health knowledge by the user is enriched, the blank of the health voice aspect is filled, and meanwhile, the experience of the user is improved.

Drawings

Fig. 1 is a flowchart of a context-based multi-round dialog method according to an embodiment of the present invention.

Detailed Description

The technical scheme of the invention is further described in detail through the drawings and the embodiments.

The context-based multi-round dialogue method provided by the embodiment of the invention is used for realizing more accurate dialogue results through a context-related method in a multi-round dialogue state. For a better understanding of the content of the present application, the following first explains the definition of key terms referred to in the present application:

session control data: in computers, and in particular in network applications, session control is also denoted Session. The session control data is used to store attributes and configuration information required for a particular user session. Thus, when a user jumps between requests of an application, the variables stored in the session control data will not be lost, but will be present throughout the user session. It should be noted that, each session control data corresponds to a session control data ID, so as to distinguish different session control data.

Natural language understanding system (Natural Language Understanding, NLU): for parsing the user's request text.

Natural language processing system (Natural Language Processing, NLP): the system is also called NLP dialogue system, which is used for processing the semantic analysis result of NLU system, filtering, verifying data, inquiring data, generating processing result, and finally responding the processing result to the client.

A round of conversation: and starting from the initiation of a request from a user to the NLP dialogue system and generating a response result from the NLP dialogue system to the user.

The intention is: the NLU system performs semantic analysis on the text of the user request, and those areas of the result that are indicated as being processable by the NLP dialog system are called intents.

Semantic slot: the specific concepts extracted from sentences by NLU systems are variables that are complemented in order for the user's intent to translate into user-specific instructions. For example, a semantic slot of "disease" would be expected for "health" to indicate the name of the disease. If the user says what symptom the cold has, the semantic slot of "disease" is "cold" through NLU system analysis; if the user says "what symptoms are in headache," the semantic slot of "disease" is "headache.

As shown in fig. 1, a multi-round dialogue method based on context provided by an embodiment of the present invention includes the following steps:

step 101, a natural language understanding system receives sentence data input by a current user;

specifically, a natural language understanding system (Natural Language Understanding, NLU) receives sentence data that is currently input by a user. Sentence data input by a user can be understood as a request text of the user to be parsed. The sentence data may be in the form of voice or text.

Step 102, identifying scene data, intention information and semantic slot information in the statement data according to session control data;

specifically, the NLU system is a system constructed by a barker pattern (BNF) model, and a syntax file for scene division is loaded in the NLU system. The grammar file for the barker's pattern to describe the speech recognition is loaded through the NLU system and then compiled into the recognition network. And the NLU system carries out path matching on text input by a user or text after voice recognition on a recognition network, and finally recognizes and proposes scene data, intention information and semantic slot information in sentence data.

It should be noted that, in the NLU system, the results of recognizing the same sentence data may be different from one another. As session control data corresponding to different users may be different. Therefore, before implementing this step, it is also necessary to acquire the user ID of the current user and to control the session data according to the user ID of the current user and the session control data corresponding to the ID, so as to ensure the accuracy of the session control data.

In a specific example, grammar files in the NLU system are written according to intention and scene, and the scene data that can be identified include: "entry" means "context free"; "more" means "the context is still used by the context, but the semantic slots need to be replaced"; "slot" means "the context is related to the semantic slot, the semantic slot information needs to be added"; "confirm" means "related to confirmation hereinafter"; "selection" means "the following is related to selection". For example, under "healthy" intent, the "entry" scene data configured file is "health. Bnf" used to represent a "healthy" intent context-free grammar rule set, and the "confirm" scene data configured file is "health confirm. Bnf" used to represent a health intent context-representation confirmation statement grammar rule set; the healthselect. Bnf is used to represent the set of grammar rules that the context identification makes the selection in the healthy intent.

Step 103, carrying out semantic matching according to scene data, intention information in the current sentence and state information corresponding to the intention information of the current sentence to obtain semantic matching result data;

specifically, the NLU system determines whether the intention information of the current sentence is identical to the intention information of the above sentence according to the grammar file. When the intention information of the current sentence is the same as the intention information, carrying out first semantic matching on the scene data of the current sentence and the state information corresponding to the intention information of the current sentence to obtain first semantic matching result data; when the intention information of the current sentence is different from the intention information, performing second semantic matching on the scene data of the current sentence and the state information corresponding to the intention information of the current sentence to obtain second semantic matching result data. This process can be understood as: NLU systems perform a selective path matching process on all recognition networks within the grammar file by the user's above intent and state.

In a specific example, the first semantic matching process includes:

if the current scene data is "entry" or "more", performing semantic matching; if the current scene data is a slot scene and the above state information is a state of waiting for a semantic slot reply, carrying out semantic matching; if the current scene data is "confirm" and the above state information is "wait for confirmation" state, then carrying out semantic matching; if the current scene data is "selection" and the above state is the information "wait for selection" state, then semantic matching is performed.

In a specific example, the second semantic matching process includes:

if the current scene data is "entry", semantic matching is performed; if the current scene data is "selection" and the above state information is "wait for selection" state, then performing semantic matching; other scene data do not perform semantic matching.

104, transmitting semantic matching result data to a natural language processing system;

specifically, the NLU system sends the semantic matching result data to a natural language processing system (Natural Language Processing, NLP) system.

Step 105, the natural language processing system performs context query according to the semantic matching result data to obtain a context query result;

specifically, the NLP system acquires the intention information and the corresponding state information of the context, and performs context query according to the intention information and the corresponding state information of the context and the matching result data to obtain a context query result. This process can be understood as: when a user asks, the NLP system calls the last conversation intention and the responded state stored by the user in the NLP system, assigns values to different state bits according to the states, and finally calculates the context value. Through the process, the purpose of selectively matching the reply of the user by the NLU system can be achieved by transmitting the intention of the last request of the user and calculating the acquired context value to the NLU.

In a specific example, if the current user does not have a dialogue, the intention information and the corresponding state information of the context are both null, and the context query result is also null; if the intention information of the upper and the state information of the upper are expressed as 'waiting for the reply of the semantic slot', the context query result is expressed as 'waiting for the reply of the semantic slot' state bit; if the above intention information and the above status information are expressed as "wait for semantic slot confirmation or wait for result confirmation", the context query result is expressed as "user confirmation status bit"; if the above intent information and the above status information are expressed as "wait for selection confirmation", the context query result is expressed as "wait for selection status bit"; if the above intent information and the above status information are expressed as "save last intent", the contextual query result is expressed as "infinite status bit".

Step 106, when the context query results are multiple, carrying out semantic slot verification on the semantic slot information corresponding to the context query results according to the priority in the context query results;

specifically, when the plurality of context query results include the context query result in the selection scene, the NLP system verifies the context query result in the selection scene to determine whether a corresponding matching item exists. If no corresponding matching item exists, the NLP system deletes the context query junction under the current selection scene.

And the NLP system verifies the semantic slot information corresponding to the context query result according to the priority in the context query result, and determines whether the semantic slot information corresponding to the context query result corresponds to the intention information of the current sentence. And deleting the current context query result when the semantic slot information corresponding to the context query result does not correspond to the intention information of the current sentence.

In a specific example, if the user first inputs a question of "which symptoms the cold will have," after the NLP system makes an answer to the question, the user outputs a question of "beijing, the second question is grammatically indistinguishable from the expression" that hypertension. The matched semantic slot value is Beijing, the Beijing is used as a disease name, the fact that the semantic slot value is invalid can be judged through a disease verification interface, a user says that the user is not a question of healthy intention, and possibly is a question of other intention, and the matching result is filtered out.

Step 107, obtaining the result data of the current round of session according to the result of the semantic slot verification, and updating session control data corresponding to the intention information of the current sentence according to the result data of the current round of session;

specifically, the NLP system generates display output data according to the result of the semantic slot verification, receives the corresponding result of the user input by the current user according to the display output data, finally obtains the result data of the current round of session according to the corresponding result of the user, and updates session control data corresponding to the intention information of the current sentence according to the result data of the current round of session.

The present solution is described below in two specific examples.

Example 1: the current session is ended, but the user performs the session again based on the previous session, so that the intention of the user of the last session is associated.

The user asks "what is hypertension to eat? "

The NLP system answers "hypertension requires a light food, and periodic checks are performed.

The user asks "that catch cold? "

NLP system answers "cold needs more water, attention rest.

For the first time a user asks, the NLU system only matches grammars in the "health.bnf" file because of no context value, and the matching result is that the user intends to be "health", and the semantic slots have two: one name is "disease" value is "hypertension", the other name is "type" value is "eating", the NLP dialogue system inquires about hypertension and inquires about eating, after responding the inquiry result to the user, the intention of the user is "health" and the semantic slot of the session is recorded, and the state of the session is the ending state.

When the user asks the cold woolen again, the NLU system analyzes the intention of the user, and because the intention of the user is "health", the state bit of the NLU system query has infinite query, the NLU system can match the question of the user for two files of "health. Bnf" and "health. Bnf" represented by the scene data of "entry" and "more", and as a result, a matching result can be obtained in the file of "health. Bnf", and the obtained user intention information is "health", the semantic slot one: the name "disease" value is "cold". When the NLP system processes, the intention information context semantic of "health" has the highest priority, and the "health" processing is performed, because only the disease name lacks the "type", the value of "type" recorded by the last session can be used at this time, so that the NLP system queries about the cold in terms of eating, the NLP system queries about the cold, records the user intention as "health" after responding to the query result, records the semantic slot of the session at this time, and the session state is the end state.

Example 2: after the user is analyzed by the session, the processing of the semantic slot finds an uncertain semantic slot value, which requires the user to confirm:

the user asks "what is i want to check that he is not sleeping? "

NLP system answers, "insomnia meets the symptoms you want to query, do you want to see the detailed description? "

The user asks "Yes"

NLP system answers "insomnia generally refers to a subjective experience of a patient that is not satisfied with sleep time and/or quality and affects daytime social functions, including difficulty falling asleep, frequent wakefulness, and/or early morning wakefulness.

The first time the user asks, the NLU system analyzes the user intention as: "health", semantic slots have two: the "symptom" value is "drowsiness", and the "action" value is "query". The NLP system inquires that the symptom is an unconscious disease, and only one result is found: the user may ask for insomnia, in which case, the NLP system will respond to the user with a confirmation question, record the intention of the user as "health", record the semantic slot to be confirmed as "break", and the session state is to wait for the user to confirm the semantic slot state.

When the user answers affirmatively, the NLP system sets an infinite state bit and a confirmation state bit, the NLU matches the user questions with the "health. Bnf" and "health. Bnf" files indicated by the "entry" and "confirm" scenes, and a matching result is obtained from the "health. Bnf".

Those of skill would further appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that the various illustrative elements and steps are described above generally in terms of function in order to clearly illustrate the interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied in hardware, in a software module executed by a user terminal, or in a combination of the two. The software modules may be disposed in Random Access Memory (RAM), memory, read Only Memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

The foregoing description of the embodiments has been provided for the purpose of illustrating the general principles of the invention, and is not meant to limit the scope of the invention, but to limit the invention to the particular embodiments, and any modifications, equivalents, improvements, etc. that fall within the spirit and principles of the invention are intended to be included within the scope of the invention.

Claims

1. A method of context-based multi-round dialog, the method comprising:

carrying out semantic matching according to scene data, intention information in the current sentence and state information corresponding to the intention information of the previous sentence to obtain semantic matching result data, and sending the semantic matching result data to a natural language processing system;

2. The context-based multi-round dialog method of claim 1, wherein before the natural language understanding system receives sentence data of current user data, the method further comprises:

the natural language understanding system is constructed by a Baxwell model.

3. The context-based multi-round dialog method of claim 1, wherein prior to the identifying scene data, intent information, and semantic slot information in the statement from session control data, the method further comprises:

acquiring the user ID of the current user;

4. The context-based multi-round dialogue method according to claim 1, wherein the performing semantic matching according to the scene data, the intention information in the current sentence and the state information corresponding to the intention information of the previous sentence, to obtain the semantic matching result data specifically includes:

5. The context-based multi-round dialog method of claim 1, wherein the natural language processing system performs a context query according to the matching result data, and the obtaining a context query result specifically includes:

acquiring the intention information and the corresponding state information;

6. The context-based multi-round dialog method of claim 1, wherein when the context query result is a plurality, the method further comprises:

7. The context-based multi-round dialog method according to claim 1, wherein verifying the semantic slot information corresponding to the context query result according to the priority in the context query result is specifically:

8. The context-based multi-round dialog method according to claim 1, wherein the obtaining the result data of the current round of dialog according to the result of the semantic slot verification is specifically: