CN110232108B - Man-machine conversation method and conversation system - Google Patents

Abstract

This application relates to artificial intelligence technology, and provides a method for man-machine dialogue, including: the dialogue system acquires the first user sentence input by the user, wherein the first user sentence indicates the first user behavior; The above-mentioned first user behavior is an abnormal behavior; in the case that the abnormal behavior is a behavior that can be recognized by the dialog system but cannot be executed, identify the reason why the abnormal behavior cannot be executed, and generate a first reply statement to reply to the abnormal behavior, wherein , the first reply sentence is used to prompt the reason why the first user behavior cannot be executed; or, in the case that the abnormal behavior is a behavior that cannot be recognized by the dialog system, identify the user intention according to the first user sentence, and generate a reply to the abnormal behavior The second reply sentence, wherein the second reply sentence is used to prompt that the dialog system does not support the user's intention.

Description

Man-machine dialogue method and dialogue system

technical field

The present application relates to the field of artificial intelligence, in particular to a man-machine dialogue method and a dialogue system.

Background technique

Currently, there are two types of dialogue systems: task-based dialogue systems and chat-type dialogue systems. Among them, the task-based dialogue refers to the dialogue conducted to complete a specific task. In task-based dialogue, the dialogue system needs to understand the user's intention according to the user's sentence, extract the information in the user's sentence, and complete the task specified by the user according to the extracted information.

However, those skilled in the art have found through long-term research that the robustness of the dialog system in dealing with abnormal behaviors is very poor.

Contents of the invention

The present application provides a man-machine dialogue method and a dialogue system, which can clearly prompt the user why the dialogue system cannot perform abnormal behaviors, thereby improving the robustness of the dialogue system in handling abnormal behaviors.

In the first aspect, a method for man-machine dialogue is provided, including:

The dialog system acquires a first user sentence input by a user indicating a first user behavior. The dialogue system determines that the first user behavior is an abnormal behavior through a behavior recognition model according to the first user statement. Wherein, the abnormal behavior is a behavior that the dialog system cannot handle normally. In the case that the abnormal behavior is a behavior that can be recognized by the dialog system but cannot be executed, the dialog system identifies a reason why the abnormal behavior cannot be executed, and generates a first reply sentence for replying to the abnormal behavior. Wherein, the first reply statement is used to prompt the reason why the first user action cannot be executed. Or, in the case that the abnormal behavior is a behavior that cannot be recognized by the dialog system, the dialog system recognizes the user's intention according to the first user sentence, and generates a second reply sentence for replying to the abnormal behavior. Wherein, the second reply sentence is used to prompt that the dialogue system does not support the user intention.

In the above solution, the user clearly knows the reason why the dialogue system cannot complete the user's behavior from the reply sentence output by the dialogue system, and adjusts the direction of the dialogue in time to make the dialogue proceed smoothly.

In some possible designs, the dialog system can generate a first reply statement to reply to the abnormal behavior by the following method:

The dialogue system searches for a corresponding first template according to the reason why the abnormal behavior cannot be executed. Wherein, the first template includes the first slot associated with the first user behavior. The dialogue system acquires first slot information for filling the first slot. The dialogue system fills the first slot information into the first slot to obtain the first reply sentence.

In some possible designs, the dialog system looks for a second template that includes a second slot. The dialogue system acquires a first associated user behavior of the first user behavior, wherein the first associated user behavior is a user behavior whose degree of correlation with the first user behavior is greater than a threshold. The dialog system fills the first associated user behavior into the second slot to obtain a third reply sentence, wherein the third reply sentence is used to prompt the dialog system to perform the first associated user behavior.

In the above solution, the dialogue system can clearly point out the user behaviors that the dialogue system can support related to the previous user statement according to the previous user statement, so it can well guide the direction of the user's dialogue, so that the dialogue can be carried out more clearly .

In some possible designs, the dialogue system can generate a second reply sentence for replying to the abnormal behavior through the following method: the dialogue system searches for a third template, wherein the third template contains a third slot; The user intends to fill in the third slot to obtain the second reply statement.

In some possible designs, the dialog system looks for a second template that includes a second slot. The dialog system acquires a second user sentence input by the user before the first user sentence. Wherein, the second user sentence indicates a second user behavior. The dialogue system obtains a second associated user behavior of the second user behavior. Wherein, the second associated user behavior is a user behavior whose degree of correlation with the second user behavior is greater than a threshold. The dialogue system fills the second associated user behavior into the second slot to obtain a fourth reply sentence. Wherein, the fourth reply sentence is used to prompt the dialogue system to perform the second associated user behavior.

In the above solution, the dialogue system can clearly point out the user behaviors that can be supported by the dialogue system related to the current user statement according to the current user statement. Therefore, it can well guide the direction of the user's dialogue, so that the dialogue can be carried out more clearly.

In some possible designs, the dialogue system recognizes the user's intention according to the first user sentence in the following manner:

The dialogue system inputs the first user sentence into a user intention recognition model to obtain the user intention, wherein the user intention recognition model is obtained by training through known user sentences and known user intentions. Wherein, the user intention recognition model is a classifier, and the classifier is any one of a K-nearest neighbor classifier, a support vector machine, a decision tree, and a random forest.

In a second aspect, a dialogue system is provided, including an input module, a determination module and a reply module.

The input module is used to acquire a first user sentence input by a user, wherein the first user sentence indicates a first user behavior.

The determining module is configured to determine that the first user behavior is an abnormal behavior through a behavior recognition model according to the first user statement, wherein the abnormal behavior is a behavior that the dialogue system cannot handle normally.

The reply module is used to identify the reason why the abnormal behavior cannot be executed when the abnormal behavior is a behavior that can be recognized by the dialog system but cannot be executed, and generate a first response to the abnormal behavior. A reply sentence, wherein the first reply sentence is used to prompt the reason why the first user action cannot be executed. or,

The reply module is also used to identify the user's intention according to the first user sentence and generate a second reply sentence for replying to the abnormal behavior when the abnormal behavior is a behavior that cannot be recognized by the dialog system , wherein the second reply sentence is used to prompt that the dialogue system does not support the user intention.

In some possible designs, the reply module is further configured to search for a corresponding first template according to the reason why the abnormal behavior cannot be executed, where the first template includes the first template associated with the first user behavior. Slot: Obtaining first slot information for filling the first slot through a slot identification model; filling the first slot information into the first slot to obtain the first reply statement.

In some possible designs, the reply module is also used to search for a second template, where the second template includes a second slot; acquire the first associated user behavior of the first user behavior, where the The first associated user behavior is a user behavior whose correlation with the first user behavior is greater than a threshold; fill the first associated user behavior into the second slot to obtain a third reply statement, wherein the first associated user behavior The three reply sentences are used to prompt the dialogue system to perform the first associated user behavior.

In some possible designs, the reply module is also used to search for a third template, where the third template includes a third slot; fill the user intention into the third slot to obtain the The second reply statement.

In some possible designs, the reply module is also used to search for a second template, where the second template includes a second slot; acquire a second user sentence input by the user, where the second user sentence is in Before the first user statement, the second user statement indicates a second user behavior; obtain a second associated user behavior of the second user behavior, wherein the second associated user behavior is related to the second user behavior A user behavior whose correlation degree of user behavior is greater than a threshold value; filling the second associated user behavior into the second slot to obtain a fourth reply sentence, wherein the fourth reply sentence is used to prompt the dialog system An executable behavior related to the second user behavior.

In some possible designs, the reply module is further configured to input the first user sentence into a user intention recognition model, so as to obtain the user intention, wherein the user intention recognition model is based on known user sentences and existing learned from user intent training. Wherein, the user intention recognition model is a classifier, and the classifier is any one of a K-nearest neighbor classifier, a support vector machine, a decision tree, and a random forest.

In a third aspect, a computer program product is provided. When the computer program product is read and executed by a computing device, the method described in any one of the first aspect will be executed.

In a fourth aspect, there is provided a computer non-transitory storage medium, including instructions, and when the instructions are run on a computing device, the computing device is made to execute the method according to any one of the first aspect.

According to a fifth aspect, an intelligent terminal is provided, including a processor and a memory, and the processor executes codes in the memory to execute the method described in any one of the first aspect.

Description of drawings

In order to more clearly illustrate the technical solutions in the embodiment of the present application or the background art, the following will describe the drawings that need to be used in the embodiment of the present application or the background art.

FIG. 1 is a schematic diagram of an application scenario of a task-based dialogue involved in the present application;

Fig. 2 is a schematic diagram of the task-type dialogue of the abnormal behavior involved in the first case of the present application;

Fig. 3 is a schematic diagram of the task-type dialogue of the abnormal behavior involved in the second situation of the present application;

Fig. 4 is a schematic diagram of using the first reply sentence to reply to the abnormal behavior in the first case involved in this application;

Fig. 5 is a schematic diagram of using the first reply sentence and the third reply sentence to reply to the abnormal behavior in the first case involved in this application;

Fig. 6 is a schematic diagram of using the second reply sentence to reply to the abnormal behavior in the second case involved in this application;

Fig. 7 is a schematic diagram of using the second reply sentence and the fourth reply sentence to reply to the abnormal behavior in the second case involved in this application;

FIG. 8 is a schematic structural diagram of a computing device provided by the present application;

FIG. 9 is a schematic structural diagram of a computing cluster provided by the present application;

Fig. 10 is a schematic structural diagram of a dialogue system provided by the present application.

Detailed ways

The terms used in the embodiments of the present application are only used to explain specific embodiments of the present invention, and are not intended to limit the present invention.

Referring to FIG. 1 , FIG. 1 is a schematic diagram of an application scenario of a task-based dialogue involved in the present application. In this application scenario, the user conducts a task-based dialogue with the dialogue system. As shown in Figure 1, the right side of Figure 1 is the user sentence input by the user, and the left side of Figure 1 is the reply sentence output by the dialog system according to the user sentence. First, the user enters the user sentence "Book an air ticket to Beijing." Then, the dialog system outputs a reply sentence "Okay, I found a flight ticket from Shenzhen to Beijing at 8 o'clock tomorrow for you. Do you need to book it?". Then, the user enters the user sentence "Book it for me." The dialogue system outputs a reply sentence "Okay, I have booked a flight ticket for you. Is there anything else I can help you with?". The user then inputs the user sentence "No, thank you". It should be understood that the application scenario shown in FIG. 1 is only an example of one of the task-type dialogues. In practical applications, the task-type dialogue may also be a dialogue about making a phone call, a dialogue about querying a geographic location, or a dialogue about ordering food delivery. , conversations about asking about the weather, conversations about booking a hotel, etc., are not specifically limited here.

In task-based dialogs, user actions are usually included in user sentences entered by users. Wherein, the user behavior is a behavior in which the user makes a request to the dialogue system. Take the user sentence "book an air ticket to Beijing" as an example. This user sentence puts forward a request to the dialogue system to book an air ticket. Therefore, this user sentence includes the user behavior "book a flight ticket". It should be understood that the above example of user behavior is only an example. In other embodiments, user behavior can also be "calling", "querying location", "ordering food", "asking about weather" and "booking a hotel". "Etc., no specific limitation is made here.

The user behavior may be recognized by the dialogue system inputting user sentences into the behavior recognition model. In a specific embodiment, the behavior recognition model may be a classifier, for example, a K-nearest neighbor classifier, a support vector machine, a decision tree, a random forest, etc., which are not specifically limited here. The action recognition model can classify according to the user actions supported by the dialogue system. For example, the user behaviors supported by the dialogue system include "booking air tickets", "making phone calls", "ordering food delivery", "asking about the weather" and "booking hotels", then the classification of the behavior recognition model includes "booking air tickets", " Call," "Order Food," "Ask the Weather," and "Book a Hotel." The behavior recognition model can determine that it should be classified into the category of "booking an air ticket" according to the user sentence "book a flight ticket to Beijing" input by the user, thereby identifying that the user behavior contained in the user sentence is "booking a flight ticket".

User statements can also include slot information. Wherein, the slot information is a value used to fill slots related to user behavior. Taking the user behavior "booking a flight ticket" as an example, the slots related to "booking a flight ticket" include departure time, departure location, and destination point. Corresponding to the above-mentioned slots related to "booking an air ticket", in the user sentence "book an air ticket from Shenzhen to Beijing at ten o'clock tomorrow", the slot information includes departure time = "ten o'clock tomorrow", departure location = "Shenzhen" and Destination = "Beijing" and so on. The dialogue system needs to fill up all slots related to user actions before performing user actions. For example, the dialog system needs to fill up the departure time, departure location, and destination of the slots related to "booking a flight ticket" before it can perform the user behavior of "booking a flight ticket".

The dialogue system can input user sentences into the slot identification model to identify slot information. In a specific embodiment, the slot recognition model may be a traditional machine learning model, such as a conditional random field model, a circular convolution network, or a dictionary search method, etc., which are not specifically limited here. The slot recognition model can extract key information provided in user dialogues. For example, the slot types of ticket booking include "departure place" and "destination place", and the slot recognition model needs to extract the information of "departure place" and "destination place". The slot recognition model is based on the user sentence input by the user "I want to book a flight ticket from Beijing to Shanghai", and the recognition results are "departure: Beijing" and "destination: Shanghai", thus providing slot information to the dialogue system.

It can be understood that the above-mentioned behavior recognition model and slot recognition model may be two different models, or may be integrated into the same model, which is not specifically limited here.

User behavior can be divided into two categories: normal behavior and abnormal behavior. Wherein, the dialog system distinguishes whether the user behavior is a normal behavior or an abnormal behavior based on whether the user behavior is supported by the dialog system and whether the user behavior can be executed. If the user behavior is a behavior supported by the dialog system, and the user behavior can be executed, then the user behavior is a normal behavior. On the contrary, the user behavior is an abnormal behavior. Among them, the abnormal behavior may include two types: first, the abnormal behavior is a user behavior that is supported by the dialogue system but cannot be executed. Second, abnormal behavior is behavior that is not supported by the dialogue system.

The dialogue system can judge whether the user behavior is a behavior supported by the dialogue system in the following way: Continuing to take the classifier as an example, suppose the classifier divides the user behavior into "booking a ticket", "calling", "ordering food", "asking Weather” and “Hotel Booking” and many other categories. Then, if the user statement entered by the user is "ordered a high-speed rail ticket", the classifier cannot classify it into any of the above categories. At this time, the dialogue system can judge that the user's behavior does not belong to the behavior supported by the dialogue system. On the contrary, the dialog system can judge that the user behavior belongs to the behavior supported by the dialog system.

The dialog system can judge whether the user action can be executed in the following ways. Take the user sentence "book an air ticket from Shenzhen to Beijing at ten o'clock tomorrow" as an example, the user behavior is "book an air ticket", and the slot information includes departure time = "ten o'clock tomorrow", departure location = "Shenzhen" and destination = " Beijing". However, after the dialogue system searches, it only finds the flight from Shenzhen to Beijing at 8:00 tomorrow, that is, the slot information of the user sentence cannot be matched with the search results, so the dialogue system judges that the user's behavior cannot be executed, and vice versa , the dialog system judges that the user action can be executed.

When the user's behavior is normal, the dialogue system normally outputs reply sentences, as shown in Figure 1 for details, and will not be described here.

When the user's behavior is abnormal, the dialog system will output reply sentences such as "I don't understand" and "Can't answer your question". Corresponding to the above two abnormal behaviors, the reply sentences output by the dialogue system under these two abnormal behaviors will be introduced in detail below.

In the first case, abnormal behaviors are behaviors that are supported by the dialog system, but cannot be executed. As shown in Figure 2, the right side of Figure 2 is the user sentence input by the user, and the left side of Figure 2 is the reply sentence of the dialogue system. First, the user enters the user sentence "Book a flight ticket to Beijing tomorrow." Then, the dialog system determines that the user behavior "booking a flight ticket" is a behavior supported by the dialog system. According to the user behavior, it is found that the flight from Shenzhen to Beijing at 8:00 am tomorrow Air ticket, and output the reply sentence "Okay, I found the air ticket from Shenzhen to Beijing at 8 o'clock tomorrow morning for you, do you need to make a reservation". The user continues to input the user sentence "help me change an air ticket at ten o'clock". The dialogue system determines that the user behavior "booking an airline ticket" is a behavior supported by the dialogue system. However, the dialogue system finds that the slot information departure time = "10 o'clock tomorrow" cannot match the search result, and the dialogue system outputs a reply sentence "Sorry, I don't seem to have I understand." The user cannot determine the reason why the dialogue system cannot complete the user's behavior, and does not know how to continue the dialogue, resulting in the failure of the dialogue.

In the second case, abnormal behavior is behavior that is not supported by the dialog system. As shown in Figure 3, the right side of Figure 3 is the user sentence input by the user, and the left side of Figure 3 is the reply sentence of the dialogue system. First, the user enters the user sentence "book a plane ticket to Beijing tomorrow", and then, the dialog system determines that the behavior of "booking a plane ticket" belongs to the behavior supported by the user, and finds out the plane from Shenzhen to Beijing at 8:00 tomorrow morning based on this user behavior ticket, and output the reply sentence "Okay, I found the plane ticket from Shenzhen to Beijing at 8:00 tomorrow morning for you, do you need to make a reservation". The user continues to input the user sentence "Let's see if there is a high-speed rail ticket coming back the day after tomorrow", but the dialogue system determines that it is not a behavior supported by the dialogue system, so it outputs the reply sentence "Sorry, I don't seem to understand it". The user cannot determine the reason why the dialogue system cannot complete the user's behavior, and does not know how to continue the dialogue, resulting in the failure of the dialogue.

In the above method, when the user's behavior is abnormal, the dialogue system will simply output reply sentences such as "I don't understand" and "Can't answer your question", which makes the user confused and unable to determine whether the dialogue system cannot complete the user's behavior. cause, the dialogue cannot continue, and eventually the dialogue fails.

For two different abnormal behaviors, the dialog system can handle them in two different ways, which will be described in detail below.

For the first abnormal behavior, the dialog system can directly prompt the user why the behavior cannot be executed. As shown in Figure 4, the right side of Figure 4 is the user sentence input by the user, and the left side of Figure 4 is the reply sentence of the dialogue system. First, the user enters the user sentence "book a plane ticket to Beijing tomorrow." Then, the dialogue system determines that the user behavior "booking a plane ticket" is a behavior supported by the dialogue system. Air ticket, and output the reply sentence "Okay, I found the air ticket from Shenzhen to Beijing at 8:00 tomorrow morning for you, do you need to make a reservation?" The user continues to input the user sentence "help me change an air ticket at ten o'clock". The dialogue system determines that the user behavior "booking an airline ticket" is a behavior supported by the dialogue system. However, the dialogue system finds that the slot information departure time = "10 o'clock tomorrow" cannot match the search result, and the dialogue system outputs "Sorry, this day is from Shenzhen to Beijing There is only this one flight" reply statement. The user clearly knows the reason why the user behavior of "booking a flight ticket" cannot be executed from the reply sentence output by the dialog system, so the user continues to input the user sentence "Okay, just this one". The dialog system determines that the user behavior of "booking a flight ticket" is a behavior supported by the dialog system, executes the user behavior of "booking a flight ticket", and outputs "OK, the flight has been booked for you".

Comparing Figure 2 and Figure 4, it can be found that in the method corresponding to Figure 2, when the dialogue system cannot perform the user behavior corresponding to "help me change a ten o'clock air ticket", it will output "Sorry, I don't seem to have Do you understand?" caused the user to be confused and unable to determine the reason why the dialogue system could not complete the user's behavior, resulting in the inability to continue the dialogue. When the user behaves corresponding to "an air ticket at ten o'clock", the reply sentence of "Sorry, there is only one flight from Shenzhen to Beijing that day" will be output. The user can clearly know from the reply sentence output by the dialogue system The reason why the user behavior cannot be completed, adjust the direction of the dialogue in time, and the dialogue proceeds smoothly.

In a specific implementation, the first reply sentence output by the dialogue system "Sorry, there is only one flight from Shenzhen to Beijing on this day" can be generated in the following way: The dialogue system obtains the corresponding first template "Sorry , there is only one flight from {departure point} to {destination point} on this day", then, the dialog system obtains the slot information destination point as "Beijing" from the user sentence "book a plane ticket to Beijing tomorrow", And the starting point of the slot information obtained from the global positioning system is "Shenzhen" (the starting point is the current location by default). Afterwards, the dialogue system fills the departure point and destination point into the {departure point} and {destination point} in the first template, thus obtaining "Sorry, there is only one flight from Shenzhen to Beijing on that day".

In a specific implementation manner, the dialog system may acquire the corresponding first template in the following manner: the dialog system obtains the user behavior "booking an airline ticket", queries the airline ticket database, and determines that the reason why the user behavior cannot be executed is that no matching user is found. The required flight information, so as to get the system behavior of "notify that there is no ticket". According to the category of the "notification without ticket" system behavior, the dialogue system finds the corresponding first template from multiple modules, wherein the correspondence between multiple system behaviors and multiple templates may be preset.

In addition to directly prompting the reason why the current user behavior cannot be executed, the dialog system can also prompt other user behaviors supported by the dialog system. For example, the dialog system can prompt user behaviors that are more relevant to the current user behavior. As shown in Figure 5, the right side of Figure 5 is the user sentence input by the user, and the left side of Figure 5 is the reply sentence of the dialogue system. First, the user enters the user sentence "book a plane ticket to Beijing tomorrow." Then, the dialogue system determines that the user behavior "booking a plane ticket" is a behavior supported by the dialogue system. Air ticket, and output the reply sentence "Okay, I found the air ticket from Shenzhen to Beijing at 8:00 tomorrow morning for you, do you need to make a reservation?" The user continues to input the user sentence "help me change an air ticket at ten o'clock". The dialogue system determines that the user behavior "booking an airline ticket" is a behavior supported by the dialogue system. However, the dialogue system finds that the slot information departure time = "10 o'clock tomorrow" cannot match the search result, and the dialogue system outputs "Sorry, this day is from Shenzhen to Beijing In addition to the first reply statement of "there is only this one flight", the third reply statement of "you can continue to book air tickets or check the recommended ticket information" is also output. From the first reply sentence output by the dialog system, the user clearly knows the reason why the user behavior of "booking a flight ticket" cannot be performed, and also clearly knows other user behaviors supported by the dialog system through the third reply sentence. Therefore, the user continues to input the user sentence "well, just this one". The dialog system determines that the user behavior of "booking a flight ticket" is a behavior supported by the dialog system, executes the user behavior of "booking a flight ticket", and outputs "OK, the flight has been booked for you".

Here, the dialogue system clearly points out the user behaviors supported by the dialogue system, so it can well guide the direction of the user's dialogue, so that the dialogue can be carried out more clearly.

In a specific implementation manner, the third reply sentence "You can continue to book tickets or query recommended ticket information" output by the dialogue system may be generated in the following manner: the dialogue system obtains the corresponding second template "You can continue to {recommended User Behavior 1} or {Recommended User Behavior 2}", and then, the dialogue system finds the corresponding user behavior in the user behavior list according to the user behavior "booking a flight ticket" corresponding to the user sentence "help me change a plane ticket at ten o'clock". The user behaviors of "booking air tickets" and "providing air ticket information" and "querying recommended air ticket information" are more relevant. Afterwards, the dialogue system fills in the {recommended user behavior 1} and {recommended user behavior 2} in the second template with "provide air ticket information" and "query recommended air ticket information", so as to get the third reply sentence "You can continue to book air tickets Or check the recommended air ticket information".

In a specific implementation manner, the user behavior list can be preset in the dialogue system. Wherein, the user behavior list includes multiple user behaviors that the dialogue system can support. After the user behavior list is set, the dialogue system can keep the user behavior list unchanged, and can also delete, add, and update the user behavior list during use. In a specific embodiment, the user behavior list can be as shown in Table 1:

Table 1 User Behavior List

serial number user behavior 1 book a flight 2 Provide ticket information 3 Query recommended air ticket information 4 Book a hotel … …

It should be understood that the above user behavior list is only an example. In other embodiments, the user behavior list may also include more or less user behaviors, and the user behaviors in the user behavior list may also be the same as those in Table 1. User behaviors are partly the same or completely different, which should not be construed as specific limitations here.

In a specific implementation, the dialogue system can obtain the corresponding second template in the following manner: the dialogue system obtains the user behavior "booking an airline ticket", queries the airline ticket database, and finds no flight information that meets the user's needs, thus obtaining " Prompt user behavior" system behavior. According to the category of the "Prompt User Behavior" system behavior, the dialog system finds the corresponding second template from the multiple templates, wherein the correspondence between the multiple system behaviors and the multiple templates may be preset.

In a specific implementation, the dialogue system finds in the list of user behaviors that it is highly correlated with "booking an air ticket" based on the user behavior "booking an air ticket" corresponding to the user sentence "help me change an air ticket at ten o'clock". The user behaviors of "providing air ticket information" and "querying recommended air ticket information" can be implemented as follows: classify user behaviors, and set the correlation of the same type of user behaviors to be relatively high, and the correlation of different types of user behaviors set to a lower degree. Then, according to the behavior of the target user, a user behavior with a relatively high correlation is selected from the user behavior list for recommendation. Taking the list of user behaviors shown in Table 1 as an example, "booking air tickets", "providing air ticket information" and "querying recommended air ticket information" are user behaviors that belong to the category of "air tickets", while the user behaviors of "booking hotels" are not User behaviors belonging to the category of "air tickets", therefore, the relevance of "booking air tickets", "providing air ticket information" and "querying recommended air ticket information" can be set relatively high, and the relevance of "booking hotels" can be set relatively low. According to the behavior of the target user as "booking air tickets", the dialogue system can select highly relevant "provide air ticket information" and "query recommended air ticket information" from the user behavior list for recommendation.

The foregoing FIG. 4 and related examples are only examples and should not be construed as specific limitations. In other implementation manners, they may also be dialogues in other scenarios, which are not specifically limited here.

For the second type of abnormal behavior, the dialogue system can identify the user's intention and prompt that the function corresponding to the user's intention is not supported. As shown in Figure 6, the right side of Figure 6 is the user sentence input by the user, and the left side of Figure 6 is the reply sentence of the dialogue system. First, the user enters the user sentence "book a plane ticket to Beijing tomorrow", and then the dialogue system determines that the user behavior "booking a plane ticket" is a user behavior supported by the dialogue system, and finds out the plane from Shenzhen to Beijing at 8:00 tomorrow morning based on the user behavior ticket, and output the reply sentence "Okay, I found the plane ticket from Shenzhen to Beijing at 8:00 tomorrow morning for you, do you need to make a reservation". The user continues to input the user sentence "first check whether there is a high-speed rail ticket coming back the day after tomorrow". The dialogue system determines that it does not belong to the user behavior supported by the dialogue system, so the dialogue system recognizes the user's intention as "booking a high-speed rail ticket" according to the user sentence, and outputs "Sorry, the function of booking a high-speed rail ticket is not supported for now". The user clearly knows from the reply sentence output by the dialog system that "booking a high-speed rail ticket" is a function not supported by the dialog system, so the user enters the user sentence "Then book an air ticket at 8:00 tomorrow." The dialog system determines that the user behavior "booking a flight ticket" is a user behavior supported by the dialog system, executes the user behavior of "booking a flight ticket", and outputs "OK, the flight has been booked for you".

Comparing Figure 3 and Figure 6, it can be found that in the method corresponding to Figure 3, when the dialogue system encounters a user behavior that cannot be supported by the dialogue system such as "booking a high-speed rail ticket", it will output "Sorry, I don't seem to understand The reply sentence of "What?" caused the user to be confused and unable to determine the reason why the dialogue system could not complete the user's behavior, resulting in the inability to continue the dialogue. When the user behavior cannot be supported by the dialogue system, the dialogue system will recognize the user's intention according to the user sentence, and output the second reply sentence "Sorry, the function of ordering high-speed rail tickets is not supported at the moment". The user clearly understands from the reply sentence output by the dialogue system Knowing that the dialogue system does not support the function of ordering high-speed rail tickets, I adjusted the direction of the dialogue in time, and the dialogue went smoothly.

In a specific implementation manner, the second reply sentence output by the dialog system "Sorry, the function of ordering high-speed rail tickets is not supported for now" may be generated in the following manner: the dialog system obtains the corresponding third template "Sorry, not supported for now { User intent} function”, and then, the dialogue system recognizes the user’s intention as “booking a high-speed rail ticket” based on the user sentence “First check if there is a high-speed rail ticket coming back the day after tomorrow”. Afterwards, the dialogue system fills in the third template with the user's intention of "booking high-speed rail tickets", and thus obtains "Sorry, the function of booking high-speed rail tickets is not supported for now".

In a specific implementation, the dialogue system can obtain the corresponding third template in the following manner: the dialogue system finds that it is impossible to classify the corresponding user behavior according to the user sentence "first check if there is a high-speed rail ticket coming back the day after tomorrow". Any user behavior supported by the dialog system, so it is determined that the user behavior is not supported by the dialog system, and thus the system behavior of "Notification does not support the function" is obtained. According to the category of the "notification does not support the function" system behavior, the dialog system finds the corresponding third template from the multiple templates, wherein the correspondence between the multiple system behaviors and the multiple templates may be preset.

In a specific implementation manner, the dialog system may realize the recognition of the user intention according to the user sentence in the following manner: the dialog system may input the user sentence into the intent recognition model to recognize the user intent. In a specific embodiment, the intent recognition model may be a classifier, for example, a K-nearest neighbor classifier, a support vector machine, a decision tree, a random forest, etc., which are not specifically limited here. The behavior recognition model can classify user intentions into categories related to dialogue system support scenarios and commonly used in daily life, such as ordering high-speed rail tickets, sending messages, opening applications, searching for food, taking taxis, and so on. According to the user sentence input by the user, the intention recognition model can determine whether there is a high-speed rail ticket coming back the day after tomorrow. It should be classified into the category of "booking a high-speed rail ticket", thereby identifying that the user intention contained in the user sentence is "booking a high-speed rail ticket".

In a specific implementation manner, the intention recognition model may be obtained by training the dialog system through a large number of known user sentences and a large number of known user intentions. The following takes the deep neural network as an example to describe the training process of the intent recognition model in detail.

来描述：从物理层面深度神经网络中的每一层的工作可以理解为通过五种对输入空间(输入向量的集合)的操作，完成输入空间到输出空间的变换(即矩阵的行空间到列空间)，这五种操作包括：1、升维/降维；2、放大/缩小；3、旋转；4、平移；5、“弯曲”。其中1、2、3的操作由

完成，4的操作由+b完成，5的操作则由a()来实现。这里之所以用“空间”二字来表述是因为被分类的对象并不是单个事物，而是一类事物，空间是指这类事物所有个体的集合。其中，W是权重向量，该向量中的每一个值表示该层神经网络中的一个神经元的权重值。该向量W决定着上文所述的输入空间到输出空间的空间变换，即每一层的权重W控制着如何变换空间。训练深度神经网络的目的，也就是最终得到训练好的神经网络的所有层的权重矩阵(由很多层的向量W形成的权重矩阵)。因此，神经网络的训练过程本质上就是学习控制空间变换的方式，更具体的就是学习权重矩阵。The essence of training the intent recognition model is to determine the weight matrix in the deep neural network through the training process. Specifically: the work of each layer in a deep neural network can be expressed mathematically

To describe: From the physical level, the work of each layer in the deep neural network can be understood as completing the transformation from the input space to the output space (that is, the row space of the matrix to the column space) through five operations on the input space (a collection of input vectors). Space), these five operations include: 1. Dimension up/down; 2. Zoom in/out; 3. Rotate; 4. Translation; 5. "Bending". Among them, the operations of 1, 2, and 3 are performed by

Complete, the operation of 4 is completed by +b, and the operation of 5 is realized by a(). The reason why the word "space" is used here is because the classified object is not a single thing, but a kind of thing, and space refers to the collection of all individuals of this kind of thing. Wherein, W is a weight vector, and each value in the vector represents the weight value of a neuron in this layer of neural network. The vector W determines the space transformation from the input space to the output space described above, that is, the weight W of each layer controls how to transform the space. The purpose of training the deep neural network is to finally obtain the weight matrix of all layers of the trained neural network (the weight matrix formed by the vector W of many layers). Therefore, the training process of the neural network is essentially to learn the way to control the spatial transformation, and more specifically, to learn the weight matrix.

The process of training the intent recognition model is the process of continuously reducing the difference between the predicted value and the target value of the deep neural network by updating the weights. Because it is hoped that the output of the deep neural network is as close as possible to the value you really want to predict, you can compare the predicted value of the current network with the target value you really want, and then update each layer of neural network according to the difference between the two. The weight vector of the network (of course, there is usually an initialization process before the first update, which is to pre-configure parameters for each layer in the deep neural network). For example, if the network's predicted value is high, adjust the weight vector to make it The prediction is lower, and it is constantly adjusted until the deep neural network can predict the desired target value. Therefore, it is necessary to pre-define "how to compare the difference between the predicted value and the target value", which is the loss function (loss function) or objective function (objective function), which are used to measure the difference between the predicted value and the target value. equation. Among them, taking the loss function as an example, the higher the output value (loss) of the loss function, the greater the difference. Then the training of the deep neural network becomes a process of reducing the loss as much as possible. Specifically, a known user sentence is input into the deep neural network, and a predicted value can be obtained through forward propagation. Then, by comparing the known user intent corresponding to the known user sentence, the difference value between the two is obtained, and the weight vector W of each layer of the neural network can be updated by performing backpropagation according to the difference value. Continuously use a large number of known user sentences and their corresponding large number of known user intentions to adjust the weight vector W, until the deep neural network can predict the real desired target value.

In a specific implementation, a large number of known user sentences used in training the deep neural network can be manually input into the dialogue system, can be obtained by the dialogue system itself from the historical dialogue between the user and the dialogue system, or can be obtained through self-production (self-play) way. Among them, the self-production method is to manually design a dialogue system that can realize task-based dialogue logic, and a user simulator that can cover as much state information as possible. Allow the user simulator to automatically interact with the dialogue engine to generate corresponding dialogue data (including user sentences). A large number of known user intentions used in training deep neural networks can be obtained by manually labeling a large number of known user sentences.

In addition to directly prompting functions that do not support user intentions, the dialogue system can also prompt other user behaviors supported by the dialogue system. For example, the dialogue system can prompt user behaviors that are more relevant to the current user behavior. As shown in Figure 7, the right side of Figure 7 is the user sentence input by the user, and the left side of Figure 7 is the reply sentence of the dialogue system. First, the user enters the user sentence "book a plane ticket to Beijing tomorrow", and then the dialogue system determines that the user behavior "booking a plane ticket" is a user behavior supported by the dialogue system, and finds out the plane from Shenzhen to Beijing at 8:00 tomorrow morning based on the user behavior ticket, and output the reply sentence "Okay, I found the plane ticket from Shenzhen to Beijing at 8:00 tomorrow morning for you, do you need to make a reservation". The user continues to input the user sentence "first check whether there is a high-speed rail ticket coming back the day after tomorrow". The dialogue system determines that it does not belong to the user behavior supported by the dialogue system. Therefore, the dialogue system recognizes the user's intention as "booking a high-speed rail ticket" according to the user sentence, and outputs a reply sentence of "Sorry, the function of booking a high-speed rail ticket is not supported at the moment", and also outputs " You can continue to book air tickets or check recommended air ticket information" reply statement. From the reply sentence output by the dialogue system, the user clearly knows that "booking high-speed rail tickets" is a function not supported by the dialogue system, and also clearly knows other user behaviors supported by the dialogue system. Therefore, the user continues to input the user sentence "then book an air ticket at eight o'clock tomorrow". The dialog system determines that the user behavior "booking a flight ticket" is a user behavior supported by the dialog system, executes the user behavior of "booking a flight ticket", and outputs "OK, the flight has been booked for you".

In a specific implementation manner, the fourth reply sentence "You can continue to book tickets or query recommended ticket information" output by the dialogue system may be generated in the following manner: the dialogue system obtains the corresponding second template "You can continue to {recommended User Behavior 1} or {Recommended User Behavior 2}", and then, the dialogue system finds the user behavior "booking an air ticket" corresponding to the user sentence "booking a plane ticket to Beijing tomorrow" in the user behavior list. The user behaviors of "booking air tickets" and "providing air ticket information" and "querying recommended air ticket information" are more relevant. Afterwards, the dialog system fills in the {recommended user behavior 1} and {recommended user behavior 2} in the second template with "provide air ticket information" and "query recommended air ticket information", so as to get the fourth reply sentence "You can continue to book air tickets Or check the recommended air ticket information".

The above-mentioned FIG. 6 , FIG. 7 and related examples are just examples and should not be construed as specific limitations. In other implementation manners, they may also be dialogues in other scenarios, which are not specifically limited here.

The abnormal behavior reply method provided by the present application can clearly prompt the user why the dialog system cannot execute the abnormal behavior, thereby improving the robustness of the dialog system in handling abnormal behaviors. The abnormal behavior reply method of the embodiment of the present application includes the following steps:

S101: Acquire a first user sentence input by a user.

In a specific implementation manner, the first user sentence input by the user includes the first user behavior. Wherein, the user behavior is a behavior in which the user makes a first request to the dialogue system. Taking the first user sentence "book a flight ticket to Beijing" as an example, the first user sentence puts forward the first request of "booking a flight ticket" to the dialogue system. Air tickets". It should be understood that the above example of the first user behavior is only an example. In other embodiments, the first user behavior may also be "calling", "querying geographical location", "ordering takeaway", "asking about the weather" And "book a hotel", etc., which are not specifically limited here.

In a specific implementation manner, the first user behavior may be recognized by the dialog system inputting the first user sentence into the behavior recognition model. Wherein, the behavior recognition model may be a classifier, for example, a K-nearest neighbor classifier, a support vector machine, a decision tree, a random forest, etc., which are not specifically limited here. The action recognition model can classify according to the user actions supported by the dialogue system. For example, the user behaviors supported by the dialogue system include "booking air tickets", "making phone calls", "ordering food delivery", "asking about the weather" and "booking hotels", then the classification of the behavior recognition model includes "booking air tickets", " Call," "Order Food," "Ask the Weather," and "Book a Hotel." The behavior recognition model can determine that it should be classified into the category of "booking an air ticket" according to the user sentence "book a flight ticket to Beijing" input by the user, thereby identifying that the user behavior contained in the user sentence is "booking a flight ticket".

In a specific implementation manner, the first user sentence may further include slot information. Wherein, the slot information is a value used to fill a slot related to the first user behavior. Taking the first user behavior "booking a flight ticket" as an example, the slots related to "booking a flight ticket" include departure time, departure location, and destination point. Corresponding to the above-mentioned slots related to "booking air tickets", in the first user sentence "booking an air ticket from Shenzhen to Beijing at ten o'clock tomorrow", the slot information includes departure time = "ten o'clock tomorrow", departure location = "Shenzhen " and destination = "Beijing" and so on. The dialogue system needs to fill up all slots related to user actions before performing user actions. For example, the dialog system needs to fill up the departure time, departure location, and destination of the slots related to "booking a flight ticket" before it can perform the user behavior of "booking a flight ticket".

In a specific implementation manner, the slot information may be obtained by the dialogue system inputting user sentences into the slot recognition model. Wherein, the slot recognition model may be a traditional machine learning model, such as a conditional random field model, a circular convolution network, or a dictionary search method, etc., which are not specifically limited here. The slot recognition model can extract key information provided in user dialogues. For example, the slot types of ticket booking include "departure place" and "destination place", and the slot recognition model needs to extract the information of "departure place" and "destination place". The slot recognition model is based on the user sentence input by the user "I want to book a flight ticket from Beijing to Shanghai", and the recognition results are "departure: Beijing" and "destination: Shanghai", thus providing slot information to the dialogue system.

S102: Determine that there is an abnormal behavior according to the first user sentence, where the abnormal behavior is a behavior that the dialog system cannot handle normally.

In a specific implementation manner, abnormal behaviors may include two types: first, abnormal behaviors are user behaviors that are supported by the dialog system but cannot be executed. Second, abnormal behavior is behavior that is not supported by the dialogue system.

In a specific implementation, the dialogue system can judge whether the first user behavior is a behavior supported by the dialogue system in the following manner: continue to take the classifier as an example, assuming that the classifier divides the user behavior into "booking an airline ticket", "playing phone calls”, “order takeout”, “ask weather” and “book hotel” and many other categories. Then, if the user statement entered by the user is "ordered a high-speed rail ticket", the classifier cannot classify it into any of the above categories. At this time, the dialogue system can judge that the user's behavior does not belong to the behavior supported by the dialogue system. On the contrary, the dialog system may determine that the first user behavior is a behavior supported by the dialog system.

In a specific implementation manner, the dialogue system can determine whether the user action can be executed in the following manner. Take the user sentence "book an air ticket from Shenzhen to Beijing at ten o'clock tomorrow" as an example, the user behavior is "book an air ticket", and the slot information includes departure time = "ten o'clock tomorrow", departure location = "Shenzhen" and destination = " Beijing". However, after the dialogue system searches, it only finds the flight from Shenzhen to Beijing at 8:00 tomorrow, that is, the slot information of the user sentence cannot be matched with the search results, so the dialogue system judges that the user's behavior cannot be executed, and vice versa , the dialog system judges that the user action can be executed.

S103: In the case that the abnormal behavior is a behavior that can be recognized by the dialog system but cannot be executed, identify the reason why the abnormal behavior cannot be executed, and generate a first reply sentence for replying to the abnormal behavior .

In a specific implementation manner, the first reply statement is used to prompt the reason why the first user action cannot be executed. Taking the first reply sentence output by the dialog system "Sorry, there is only one flight from Shenzhen to Beijing on this day" as an example, the first reply sentence can be generated in the following way: the dialog system obtains the corresponding first template "Sorry , there is only one flight from {departure point} to {destination point} on this day", then, the dialog system obtains the slot information destination point as "Beijing" from the user sentence "book a plane ticket to Beijing tomorrow", And the starting point of the slot information obtained from the global positioning system is "Shenzhen" (the starting point is the current location by default). Afterwards, the dialogue system fills the departure point and destination point into the {departure point} and {destination point} in the first template, thus obtaining "Sorry, there is only one flight from Shenzhen to Beijing on that day".

In a specific implementation manner, the first template includes the first slot associated with the first user behavior. Taking the first template "Sorry, there is only one flight from {departure point} to {destination point} on this day" as an example, the first slot includes {departure point} and {destination point}.

In a specific implementation manner, the dialogue system may obtain the corresponding first template in the following manner: the dialogue system searches for the corresponding first template according to the reason why the abnormal behavior cannot be executed, wherein the first template Including the first slot associated with the first user behavior; acquiring first slot information used to fill the first slot; filling the first slot information into the first slot, Obtain the first reply statement. For example, the dialogue system obtains the user behavior "booking a flight ticket", queries the flight ticket database, and determines that the reason why the user behavior cannot be executed is that no flight information that meets the user's needs can be found, thus obtaining the system behavior of "notifying that there is no ticket". According to the category of the "notification without ticket" system behavior, the dialogue system finds the corresponding first template from multiple modules, wherein the correspondence between multiple system behaviors and multiple templates may be preset.

In a specific implementation manner, the dialog system may also output a third reply sentence, wherein the third reply sentence is used to prompt the dialog system to perform an action. The dialogue system searches for a second template, wherein the second template includes a second slot; obtains a first associated user behavior of the first user behavior, wherein the first associated user behavior is related to the first user behavior A user behavior whose behavior correlation is greater than a threshold; fill the first associated user behavior into the second slot, so as to obtain a third reply sentence. Taking the third reply sentence output by the dialog system as "You can continue to book tickets or query recommended flight ticket information" as an example, the third reply sentence can be generated in the following way: the dialog system obtains the corresponding second template "You can continue{ Recommend user behavior 1} or {recommended user behavior 2}", and then, the dialogue system "books an air ticket" according to the user behavior corresponding to the second user sentence before the first user sentence "Help me change an air ticket at ten o'clock", In the user behavior list, it is found that the user behaviors "providing airline ticket information" and "querying recommended airline ticket information" are relatively highly correlated with the user behavior "booking airline tickets". Afterwards, the dialogue system fills in the {recommended user behavior 1} and {recommended user behavior 2} in the second template with "provide air ticket information" and "query recommended air ticket information", so as to get the third reply sentence "You can continue to book air tickets Or check the recommended air ticket information".

In a specific implementation manner, the user behavior list can be preset in the dialogue system. Wherein, the user behavior list includes multiple user behaviors that the dialogue system can support. After the user behavior list is set, the dialogue system can keep the user behavior list unchanged, and can also delete, add, and update the user behavior list during use.

In a specific embodiment, the second template includes a second slot. Taking the second template "You can continue with {recommended user behavior 1} or {recommended user behavior 2}" as an example, the second slot includes: {recommended user behavior 1} and {recommended user behavior 2}".

In a specific implementation, the dialogue system can obtain the corresponding second template in the following manner: the dialogue system obtains the user behavior "booking an airline ticket", queries the airline ticket database, and finds no flight information that meets the user's needs, thus obtaining " Prompt user behavior" system behavior. According to the category of the "Prompt User Behavior" system behavior, the dialog system finds the corresponding second template from the multiple templates, wherein the correspondence between the multiple system behaviors and the multiple templates may be preset.

In a specific implementation, the dialogue system finds in the list of user behaviors that it is highly correlated with "booking an air ticket" based on the user behavior "booking an air ticket" corresponding to the user sentence "help me change an air ticket at ten o'clock". The user behaviors of "providing air ticket information" and "querying recommended air ticket information" can be implemented as follows: classify user behaviors, and set the correlation of the same type of user behaviors to be relatively high, and the correlation of different types of user behaviors set to a lower degree. Then, according to the behavior of the target user, a user behavior with a relatively high correlation is selected from the user behavior list for recommendation. Taking the list of user behaviors shown in Table 1 as an example, "booking air tickets", "providing air ticket information" and "querying recommended air ticket information" are user behaviors that belong to the category of "air tickets", while the user behaviors of "booking hotels" are not User behaviors belonging to the category of "air tickets", therefore, the relevance of "booking air tickets", "providing air ticket information" and "querying recommended air ticket information" can be set relatively high, and the relevance of "booking hotels" can be set relatively low. According to the behavior of the target user as "booking air tickets", the dialogue system can select highly relevant "provide air ticket information" and "query recommended air ticket information" from the user behavior list for recommendation.

S104: In the case that the abnormal behavior is a behavior that cannot be recognized by the dialogue system, identify user intentions according to the first user statement, and generate a second reply statement for replying to the abnormal behavior.

In a specific implementation manner, the second reply sentence is used to prompt that the dialogue system does not support the user intention. Taking the second reply sentence output by the dialogue system as "Sorry, the function of ordering high-speed rail tickets is not supported at the moment" as an example, the second reply sentence can be generated by the following method: find the third template, and fill in the user intention into the The third slot, so as to get the second reply statement. Taking the third template as "Sorry, the {user intent} function is not supported at this time" as an example, the dialogue system recognizes the user intention as "booking a high-speed rail ticket" based on the user sentence "First check if there is a high-speed rail ticket coming back the day after tomorrow". Afterwards, the dialogue system fills in the third template with the user's intention of "booking high-speed rail tickets", and thus obtains "Sorry, the function of booking high-speed rail tickets is not supported for now".

In a specific implementation, the dialogue system can obtain the corresponding third template in the following manner: the dialogue system finds that it is impossible to classify the corresponding user behavior according to the user sentence "first check if there is a high-speed rail ticket coming back the day after tomorrow". Any user behavior supported by the dialog system, so it is determined that the user behavior is not supported by the dialog system, and thus the system behavior of "Notification does not support the function" is obtained. According to the category of the "notification does not support the function" system behavior, the dialog system finds the corresponding third template from the multiple templates, wherein the correspondence between the multiple system behaviors and the multiple templates may be preset.

In a specific implementation manner, the dialog system may realize the recognition of the user intention according to the user sentence in the following manner: the dialog system may input the user sentence into the intent recognition model to recognize the user intent. Wherein, the intent recognition model may be a classifier, for example, a K-nearest neighbor classifier, a support vector machine, a decision tree, a random forest, etc., which are not specifically limited here. The intent recognition model can classify user intent into categories related to dialogue system support scenarios and commonly used in daily life, such as ordering high-speed rail tickets, sending messages, opening applications, searching for food, taking taxis, and so on. According to the user sentence input by the user, the intention recognition model can determine whether there is a high-speed rail ticket coming back the day after tomorrow. It should be classified into the category of "booking a high-speed rail ticket", thereby identifying that the user intention contained in the user sentence is "booking a high-speed rail ticket".

In a specific implementation manner, the intention recognition model may be obtained by training the dialog system through a large number of known user sentences and a large number of known user intentions. The following takes the deep neural network as an example to describe the training process of the intent recognition model in detail.

来描述：从物理层面深度神经网络中的每一层的工作可以理解为通过五种对输入空间(输入向量的集合)的操作，完成输入空间到输出空间的变换(即矩阵的行空间到列空间)，这五种操作包括：1、升维/降维；2、放大/缩小；3、旋转；4、平移；5、“弯曲”。其中1、2、3的操作由

完成，4的操作由+b完成，5的操作则由a()来实现。这里之所以用“空间”二字来表述是因为被分类的对象并不是单个事物，而是一类事物，空间是指这类事物所有个体的集合。其中，W是权重向量，该向量中的每一个值表示该层神经网络中的一个神经元的权重值。该向量W决定着上文所述的输入空间到输出空间的空间变换，即每一层的权重W控制着如何变换空间。训练深度神经网络的目的，也就是最终得到训练好的神经网络的所有层的权重矩阵(由很多层的向量W形成的权重矩阵)。因此，神经网络的训练过程本质上就是学习控制空间变换的方式，更具体的就是学习权重矩阵。The essence of training the intent recognition model is to determine the weight matrix in the deep neural network through the training process. Specifically: the work of each layer in a deep neural network can be expressed mathematically

To describe: From the physical level, the work of each layer in the deep neural network can be understood as completing the transformation from the input space to the output space (that is, the row space of the matrix to the column space) through five operations on the input space (a collection of input vectors). Space), these five operations include: 1. Dimension up/down; 2. Zoom in/out; 3. Rotate; 4. Translation; 5. "Bending". Among them, the operations of 1, 2, and 3 are performed by

Complete, the operation of 4 is completed by +b, and the operation of 5 is realized by a(). The reason why the word "space" is used here is because the classified object is not a single thing, but a kind of thing, and space refers to the collection of all individuals of this kind of thing. Wherein, W is a weight vector, and each value in the vector represents the weight value of a neuron in this layer of neural network. The vector W determines the space transformation from the input space to the output space described above, that is, the weight W of each layer controls how to transform the space. The purpose of training the deep neural network is to finally obtain the weight matrix of all layers of the trained neural network (the weight matrix formed by the vector W of many layers). Therefore, the training process of the neural network is essentially to learn the way to control the spatial transformation, and more specifically, to learn the weight matrix.

The process of training the intent recognition model is the process of continuously reducing the difference between the predicted value and the target value of the deep neural network by updating the weights. Because it is hoped that the output of the deep neural network is as close as possible to the value you really want to predict, you can compare the predicted value of the current network with the target value you really want, and then update each layer of neural network according to the difference between the two. The weight vector of the network (of course, there is usually an initialization process before the first update, which is to pre-configure parameters for each layer in the deep neural network). For example, if the network's predicted value is high, adjust the weight vector to make it The prediction is lower, and it is constantly adjusted until the deep neural network can predict the desired target value. Therefore, it is necessary to pre-define "how to compare the difference between the predicted value and the target value", which is the loss function (loss function) or objective function (objective function), which are used to measure the difference between the predicted value and the target value. equation. Among them, taking the loss function as an example, the higher the output value (loss) of the loss function, the greater the difference. Then the training of the deep neural network becomes a process of reducing the loss as much as possible. Specifically, a known user sentence is input into the deep neural network, and a predicted value can be obtained through forward propagation. Then, by comparing the known user intent corresponding to the known user sentence, the difference value between the two is obtained, and the weight vector W of each layer of the neural network can be updated by performing backpropagation according to the difference value. Continuously use a large number of known user sentences and their corresponding large number of known user intentions to adjust the weight vector W, until the deep neural network can predict the real desired target value.

In a specific implementation, a large number of known user sentences used in training the deep neural network can be manually input into the dialogue system, can be obtained by the dialogue system itself from the historical dialogue between the user and the dialogue system, or can be obtained through self-production (self-play) way. Among them, the self-production method is to manually design a dialogue system that can realize task-based dialogue logic, and a user simulator that can cover as much state information as possible. Allow the user simulator to automatically interact with the dialogue engine to generate corresponding dialogue data (including user sentences). A large number of known user intentions used in training deep neural networks can be obtained by manually labeling a large number of known user sentences.

Referring to FIG. 8 , FIG. 8 is a schematic structural diagram of a dialogue system provided by the present application. The dialogue system provided by the embodiment of the present application includes: an input module 110 , a determination module 120 and a reply module 130 .

The input module 110 is configured to acquire a first user statement input by a user, wherein the first user statement indicates a first user behavior;

The determination module 120 is configured to determine the first user behavior as an abnormal behavior through a behavior recognition model according to the first user statement, wherein the abnormal behavior is a behavior that the dialogue system cannot handle normally;

The reply module 130 is used to identify the reason why the abnormal behavior cannot be executed when the abnormal behavior is a behavior that can be recognized by the dialog system but cannot be executed, and generate a response to the abnormal behavior. A first reply sentence, wherein the first reply sentence is used to prompt the reason why the first user action cannot be executed; or

The reply module 130 is further configured to identify user intentions according to the first user statement and generate a second reply to reply to the abnormal behavior when the abnormal behavior is a behavior that cannot be recognized by the dialog system statement, wherein the second reply statement is used to prompt that the dialogue system does not support the user intention.

For simplicity, this embodiment does not describe how the first reply sentence, the second reply sentence, etc. are generated. For details, please refer to FIGS. 1-5 and related descriptions. This embodiment also does not introduce how the third reply sentence and the fourth reply sentence are generated, please refer to FIG. 6 , FIG. 7 and related descriptions for details.

Referring to FIG. 9 , FIG. 9 is a schematic structural diagram of a computing cluster provided by the present application. The dialogue system shown in FIG. 8 can be implemented in a computing cluster as shown in FIG. 9 . The computing cluster includes at least one computing node 210 and at least one storage node 220 .

Compute node 210 includes one or more processors 211 , communication interface 212 and memory 213 . Wherein, the processor 211 , the communication interface 212 and the memory 213 may be connected through a bus 224 .

The processor 211 includes one or more general-purpose processors, wherein the general-purpose processor may be any type of device capable of processing electronic instructions, including a central processing unit (Central Processing Unit, CPU), a microprocessor, a microcontroller, a mainframe Processor, controller and ASIC (Application Specific Integrated Circuit, application specific integrated circuit) and so on. It can be a dedicated processor used only by compute node 210 or can be shared with other compute nodes 210 . The processor 211 executes various types of digitally stored instructions, such as software or firmware programs stored in the memory 213, which enable the computing node 210 to provide a wide variety of services. For example, the processor 211 can execute programs or process data to perform at least part of the methods discussed herein. The input module, the determination module and the reply module as shown in FIG. 8 may run in the processor 211 .

The communication interface 212 may be a wired interface (such as an Ethernet interface) for communicating with other computing nodes or users.

The memory 213 may include a volatile memory (Volatile Memory), such as a random access memory (Random Access Memory, RAM); the memory may also include a non-volatile memory (Non-VolatileMemory), such as a read-only memory (Read-Only Memory). , ROM), flash memory (Flash Memory), hard disk (Hard Disk Drive, HDD) or solid-state disk (Solid-State Drive, SSD) storage may also include a combination of the above-mentioned types of storage.

The storage node 220 includes one or more processors 221 , a communication interface 222 and a memory 223 . Wherein, the processor 221 , the communication interface 222 and the memory 223 may be connected through a bus 224 .

Processor 221 includes one or more general-purpose processors, wherein a general-purpose processor can be any type of device capable of processing electronic instructions, including CPUs, microprocessors, microcontrollers, main processors, controllers, and ASICs, etc. . It can be a dedicated processor for storage node 220 only or can be shared with other storage nodes 220 . The processor 221 executes various types of digital storage instructions, such as software or firmware programs stored in the memory 223, which enable the storage node 220 to provide a wide variety of services. For example, the processor 221 can execute programs or process data to perform at least part of the methods discussed herein.

The communication interface 222 may be a wired interface (such as an Ethernet interface) for communicating with other computing devices or users.

Storage nodes 220 include one or more storage controllers 221 and storage arrays 225 . Wherein, the storage controller 221 and the storage array 225 may be connected through a bus 226 .

The storage controller 221 includes one or more general-purpose processors, wherein the general-purpose processors may be any type of device capable of processing electronic instructions, including CPUs, microprocessors, microcontrollers, main processors, controllers, and ASICs, etc. wait. It can be a dedicated processor only for a single storage node 220 or can be shared with compute nodes or other storage nodes 220 . It can be understood that, in this embodiment, each storage node includes a storage controller, and in other embodiments, multiple storage nodes may share one storage controller, which is not specifically limited here.

Memory array 225 may include multiple memories. The memory may be a non-volatile memory, such as ROM, flash memory, HDD or SSD. The memory may also include a combination of the above-mentioned types of memory. For example, the storage array may be composed of multiple HDDs or multiple SDDs, or the storage array may be composed of HDDs and SDDs. Among them, with the assistance of the storage controller 221, multiple storages are combined in different ways to form a storage group, thereby providing higher storage performance than a single storage and providing data backup technology. Optionally, memory array 225 may include one or more data centers. Multiple data centers may be set at the same location, or at different locations, which is not specifically limited here. The memory array 225 may store program codes as well as databases. Wherein, the program code includes obtaining the module code, determining the module code and replying the module code. The database includes: behavior recognition model, slot recognition model and intent recognition model.

Wherein, the computing node 211 is used to execute the following steps by calling the program code in the storage node 213:

Acquiring a first user statement input by a user, wherein the first user statement indicates a first user behavior;

Determining that the first user behavior is an abnormal behavior through a behavior recognition model according to the first user statement, wherein the abnormal behavior is a behavior that the dialog system cannot handle normally;

In the case that the abnormal behavior is a behavior that can be recognized by the dialog system but cannot be executed, identify the reason why the abnormal behavior cannot be executed, and generate a first reply statement for replying to the abnormal behavior, wherein , the first reply statement is used to prompt the reason why the first user action cannot be executed; or

In the case that the abnormal behavior is a behavior that cannot be recognized by the dialog system, identify the user's intention according to the first user sentence, and generate a second reply sentence for replying to the abnormal behavior, wherein the second The reply sentence is used to prompt that the dialogue system does not support the user intention.

For simplicity, this embodiment does not describe how the first reply sentence, the second reply sentence, etc. are generated. For details, please refer to FIGS. 1-5 and related descriptions. This embodiment also does not introduce how the third reply sentence and the fourth reply sentence are generated, please refer to FIG. 6 , FIG. 7 and related descriptions for details.

Referring to FIG. 10 , FIG. 10 is a schematic structural diagram of a computing device provided in the present application. The dialog system shown in FIG. 8 can be implemented in a computing device as shown in FIG. 10 . In a specific embodiment, the computing device may be a smart terminal, and the computing device in this embodiment includes one or more processors 311 , a communication interface 312 and a memory 313 . Wherein, the processor 311 , the communication interface 312 and the memory 313 may be connected through a bus 324 .

The processor 311 includes one or more general-purpose processors, wherein the general-purpose processor can be any type of device capable of processing electronic instructions, including a central processing unit (Central Processing Unit, CPU), a microprocessor, a microcontroller, a mainframe Processor, controller and ASIC (Application Specific Integrated Circuit, application specific integrated circuit) and so on. Processor 311 executes various types of digitally stored instructions, such as software or firmware programs stored in memory 313, which enable the computing device to provide a wide variety of services. For example, the processor 311 can execute programs or process data to perform at least part of the methods discussed herein. The input module, the determination module and the reply module as shown in FIG. 8 may run in the processor 311 .

The communication interface 312 may be a wired interface (such as an Ethernet interface) for communicating with other computing nodes or users.

The memory 313 may include a volatile memory (Volatile Memory), such as a Random Access Memory (Random Access Memory, RAM); the memory may also include a non-volatile memory (Non-VolatileMemory), such as a Read-Only Memory (Read-Only Memory). , ROM), flash memory (Flash Memory), hard disk (Hard Disk Drive, HDD) or solid-state disk (Solid-State Drive, SSD) storage may also include a combination of the above-mentioned types of storage. The memory 313 can store program codes and program data. Wherein, the program code includes obtaining the module code, determining the module code and replying the module code. The database includes: behavior recognition model, slot recognition model and intent recognition model.

Wherein, the processor 311 is used to perform the following steps by calling the program code in the memory 313:

Acquiring a first user statement input by a user, wherein the first user statement indicates a first user behavior;

Determining that the first user behavior is an abnormal behavior through a behavior recognition model according to the first user statement, wherein the abnormal behavior is a behavior that the dialog system cannot handle normally;

In the case that the abnormal behavior is a behavior that can be recognized by the dialog system but cannot be executed, identify the reason why the abnormal behavior cannot be executed, and generate a first reply statement for replying to the abnormal behavior, wherein , the first reply statement is used to prompt the reason why the first user action cannot be executed; or

In the case that the abnormal behavior is a behavior that cannot be recognized by the dialog system, identify the user's intention according to the first user sentence, and generate a second reply sentence for replying to the abnormal behavior, wherein the second The reply sentence is used to prompt that the dialogue system does not support the user intention.

For simplicity, this embodiment does not describe how the first reply sentence, the second reply sentence, etc. are generated. For details, please refer to FIGS. 1-5 and related descriptions. This embodiment also does not introduce how the third reply sentence and the fourth reply sentence are generated, please refer to FIG. 6 , FIG. 7 and related descriptions for details.

In the above embodiments, all or part of them may be implemented by software, hardware, firmware or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present application will be generated in whole or in part. The computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices. The computer instructions may be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website, computer, server or data center Transmission to another website site, computer, server, or data center by wired (eg, coaxial cable, optical fiber, DSL) or wireless (eg, infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that can be accessed by a computer, or a data storage device such as a server or a data center integrated with one or more available media. The available medium may be a magnetic medium (for example, a floppy disk, a storage disk, a magnetic tape), an optical medium (for example, DVD), or a semiconductor medium (for example, a Solid State Disk (SSD)).

Claims (16)

1. A method for human-computer interaction, comprising:

the dialogue system acquires a first user statement input by a user, wherein the first user statement indicates a first user behavior; determining the first user behavior as an abnormal behavior through a behavior recognition model according to the first user statement, wherein the abnormal behavior is a behavior which cannot be normally processed by the dialog system;

in the case that the abnormal behavior is a behavior which can be recognized by the dialog system but cannot be executed, identifying a reason why the abnormal behavior cannot be executed, and generating a first reply statement which replies to the abnormal behavior, wherein the first reply statement is used for prompting a reason why the first user behavior cannot be executed; or

And under the condition that the abnormal behavior cannot be recognized by the dialogue system, recognizing user intention according to the first user statement, and generating a second reply statement for replying the abnormal behavior, wherein the second reply statement is used for prompting that the dialogue system does not support the user intention.

2. The method of claim 1, wherein generating a first reply statement that replies to the anomalous behavior comprises:

searching a corresponding first template according to the reason that the abnormal behavior cannot be executed, wherein the first template comprises a first slot position associated with the first user behavior;

acquiring first slot position information used for filling the first slot position;

and filling the first slot position information into the first slot position to obtain the first reply statement.

3. The method of claim 2, further comprising:

searching a second template, wherein the second template comprises a second slot position;

acquiring a first associated user behavior of the first user behavior, wherein the first associated user behavior is a user behavior with a correlation degree with the first user behavior larger than a threshold value;

and filling the first associated user behavior into the second slot to obtain a third reply statement, wherein the third reply statement is used for prompting that the dialog system can execute the first associated user behavior.

4. The method of claim 1, wherein generating a second reply statement that replies to the anomalous behavior comprises:

searching a third template, wherein the third template comprises a third slot position;

and filling the third slot with the user intention to obtain the second reply statement.

5. The method of claim 4, further comprising:

searching a second template, wherein the second template comprises a second slot position;

acquiring a second user statement input by a user, wherein the second user statement is before the first user statement, and the second user statement indicates a second user behavior;

obtaining a second associated user behavior of the second user behavior, wherein the second associated user behavior is a user behavior with a correlation degree with the second user behavior larger than a threshold value;

and filling the second associated user behavior into the second slot to obtain a fourth reply statement, wherein the fourth reply statement is used for prompting that the dialog system can execute the second associated user behavior.

6. The method of any one of claims 1 to 5, wherein identifying a user intent from the first user statement comprises:

inputting the first user sentence into a user intention recognition model so as to obtain the user intention, wherein the user intention recognition model is obtained by training known user sentences and known user intentions.

7. The method of claim 6, wherein the user intent recognition model is a classifier, and wherein the classifier is any one of a K-nearest neighbors classifier, a support vector machine, a decision tree, and a random forest.

8. A dialogue system is characterized by comprising an input module, a determination module and a reply module,

the input module is used for acquiring a first user statement input by a user, wherein the first user statement indicates a first user behavior;

the determining module is used for determining the first user behavior as an abnormal behavior through a behavior recognition model according to the first user statement, wherein the abnormal behavior is a behavior which cannot be normally processed by the dialog system;

the reply module is used for identifying the reason why the abnormal behavior cannot be executed when the abnormal behavior is the behavior which can be identified but cannot be executed by the dialogue system, and generating a first reply statement for replying to the abnormal behavior, wherein the first reply statement is used for prompting the reason why the first user behavior cannot be executed; or alternatively

The reply module is further configured to, when the abnormal behavior is a behavior that cannot be recognized by the dialog system, recognize a user intention according to the first user statement, and generate a second reply statement that replies to the abnormal behavior, where the second reply statement is used to prompt that the dialog system does not support the user intention.

9. The system of claim 8,

the reply module is further used for searching a corresponding first template according to the reason that the abnormal behavior cannot be executed, wherein the first template comprises a first slot position associated with the first user behavior; acquiring first slot position information for filling the first slot position through a slot position identification model; and filling the first slot position information into the first slot position to obtain the first reply statement.

10. The system of claim 9,

the reply module is further used for searching a second template, wherein the second template comprises a second slot position; obtaining a first associated user behavior of the first user behavior, wherein the first associated user behavior is a user behavior with a correlation degree with the first user behavior larger than a threshold value; and filling the first associated user behavior into the second slot to obtain a third reply statement, wherein the third reply statement is used for prompting that the dialog system can execute the first associated user behavior.

11. The system of claim 8,

the reply module is further used for searching a third template, wherein the third template comprises a third slot position; and filling the third slot with the user intention to obtain the second reply statement.

12. The system of claim 11,

the reply module is further used for searching a second template, wherein the second template comprises a second slot position; acquiring a second user statement input by a user, wherein the second user statement is before the first user statement, and the second user statement indicates a second user behavior; obtaining a second associated user behavior of the second user behavior, wherein the second associated user behavior is a user behavior whose correlation degree with the second user behavior is greater than a threshold value; and filling the second associated user behavior into the second slot to obtain a fourth reply statement, wherein the fourth reply statement is used for prompting the behavior which can be executed by the dialog system and is related to the second user behavior.

13. The system of any one of claims 8 to 12,

the reply module is further used for inputting the first user statement into a user intention recognition model so as to obtain the user intention, wherein the user intention recognition model is obtained by training through known user statements and known user intentions.

14. The system of claim 13, wherein the user intent recognition model is a classifier, and wherein the classifier is any one of a K-neighbor classifier, a support vector machine, a decision tree, and a random forest.

15. A computer non-transitory storage medium including instructions that, when executed on a computing device, cause the computing device to perform the method of any of claims 1 to 7.

16. An intelligent terminal comprising a processor and a memory, the processor executing code in the memory to perform the method of any of claims 1 to 7.

Images