CN113901837A

CN113901837A - Intention understanding method, device, equipment and storage medium

Info

Publication number: CN113901837A
Application number: CN202111214902.1A
Authority: CN
Inventors: 陆晨昱; 张涛; 吴尧; 徐嘉南; 马戈; 章乐
Original assignee: Zebred Network Technology Co Ltd
Current assignee: Zebred Network Technology Co Ltd
Priority date: 2021-10-19
Filing date: 2021-10-19
Publication date: 2022-01-07

Abstract

The invention provides an intention understanding method, an intention understanding device, intention understanding equipment and a storage medium, wherein the intention understanding method comprises the following steps: acquiring target information input by a user; performing intention identification on the target information, and judging whether the identification is successful; when the intention identification fails, acquiring a candidate intention set corresponding to the target information; constructing prompt information based on the candidate intention set and displaying the prompt information; acquiring reply information input by the user aiming at the prompt information; and understanding the reply information, and determining a target intention corresponding to the target information. According to the intention understanding method, the correct intention of the user request which is not understood is obtained in an interactive mode with the user, so that the understanding defect of the dialog system in the actual online use process can be quickly solved, the current requirements of the user are met, and the use experience of the user is improved.

Description

Intention understanding method, device, equipment and storage medium

Technical Field

The invention relates to the technical field of artificial intelligence, in particular to an intention understanding method, device, equipment and storage medium.

Background

The task-based dialog system has wide application in various fields, such as voice assistants on devices such as smart phones and smart speakers, and intelligent customer service in scenes such as telephones and e-commerce platforms. With the development of the technology, especially the breakthrough progress of the deep learning technology based on the neural network, as long as enough training sample data is provided, the task type dialog systems in different fields can achieve good effect, and the user requirements under most conditions are met.

However, in practical applications, regardless of how the effect of the system is optimized, there will always be a proportion of long-tailed user requests that should not be correctly understood by the system. Reasons why these requests are not properly understood may include, but are not limited to: 1. the user request statement contains new entities which do not appear before, such as new song names, new place names and the like; 2. the expression of the user request sentence is not common before, such as the user expressing the intention of "volume turn down" in a comparatively spoken manner "music low point". Therefore, in order to improve the use experience of the user, it is necessary to consider how to deal with the understanding defect occurring during the actual online use of the system and how to make the system continuously and optimally updated, so as to avoid the repeated occurrence of the understanding defect as much as possible.

In the prior art, generally, an off-line iteration mode is used for processing understanding problems occurring on a line, and a system is continuously optimized. However, the period of the scheme optimization in the prior art is generally in cycles, which cannot solve the understanding defect of the online conversation system in real time, so that the current needs of the user cannot be met, and the user experience is poor.

Disclosure of Invention

In view of the above problems in the prior art, an object of the present invention is to provide an intention understanding method, apparatus, device and storage medium, which can quickly solve understanding defects occurring in the actual online use process of a dialog system, meet the current requirements of users, and improve the use experience of users.

In order to solve the above problems, the present invention provides an intention understanding method including:

acquiring target information input by a user;

performing intention identification on the target information, and judging whether the identification is successful;

when the intention identification fails, acquiring a candidate intention set corresponding to the target information;

constructing prompt information based on the candidate intention set and displaying the prompt information;

acquiring reply information input by the user aiming at the prompt information;

and understanding the reply information, and determining a target intention corresponding to the target information.

Further, the performing intent recognition on the target information and determining whether recognition is successful includes:

inputting the target information into a first pre-trained intention recognition model to obtain an intention recognition result corresponding to the target information, wherein the first intention recognition model is determined by training a first preset neural network model by using first sample data obtained in advance;

judging whether the intention identification result is a valid result;

if the intention recognition result is an invalid result, judging that intention recognition fails;

and if the intention identification result is a valid result, judging that the intention identification is successful.

Optionally, when the intention identification fails, the obtaining of the candidate intention set corresponding to the target information includes:

correcting the first intention recognition model to obtain a corrected first intention recognition model;

and inputting the target information into the modified first intention recognition model to obtain a candidate intention set corresponding to the target information.

and inputting the target information into a preset second intention recognition model to obtain a candidate intention set corresponding to the target information.

matching the target information with a pre-stored request statement set, and determining a successfully-matched target request statement set;

acquiring the corresponding intention of each target request statement in the target request statement set;

and generating a candidate intention set corresponding to the target information according to the intention corresponding to each target request statement in the target request statement set.

Further, before constructing and displaying the prompt information based on the candidate intention set, the method further includes:

analyzing the target information, and determining a first intention type corresponding to the target information;

judging whether the target information needs to be identified or not according to the first intention type;

and when the target information needs to be identified, constructing prompt information based on the candidate intention set and displaying the prompt information.

Further, the analyzing the target information and determining the first intention type corresponding to the target information includes:

acquiring current environment information;

inputting the target information and the environment information into a first intention classification model trained in advance to obtain a first intention type corresponding to the target information, wherein the first intention classification model is determined by utilizing third sample data obtained in advance to train a third preset neural network model.

Further, the constructing and presenting prompt information based on the candidate intention set comprises:

when the candidate intention set is empty, constructing and displaying guide information, wherein the guide information is used for indicating a user to correct the target information;

when the candidate intention set only comprises one candidate intention, constructing inquiry information according to the candidate intention and displaying the inquiry information, wherein the inquiry information is used for indicating a user to reply confirmation information according to the candidate intention;

when the candidate intention set comprises a plurality of candidate intentions, recommendation information is constructed according to the candidate intentions and displayed, and the recommendation information is used for instructing a user to select a target intention from the candidate intentions.

Optionally, the understanding of the reply information, and the determining of the target intention corresponding to the target information includes:

inputting the reply information into a pre-trained second intention classification model to obtain a second intention type corresponding to the reply information, wherein the second intention classification model is determined by training a fourth preset neural network model by using pre-acquired fourth sample data;

and determining a target intention corresponding to the target information according to the second intention type.

inputting the reply information and the candidate intention set into an intention matching model to obtain a target candidate intention matched with the reply information, wherein the intention matching model is determined by training a fifth preset neural network model by utilizing pre-acquired fifth sample data;

and taking the target candidate intention as a target intention corresponding to the target information.

performing intention identification on the reply information, and judging whether the identification is successful;

when the intention is successfully identified, carrying out repeat detection on the reply information, and judging whether the reply information is repeat information of the target information;

and when the reply information is the repeat information of the target information, acquiring the intention corresponding to the reply information, and taking the intention corresponding to the reply information as the target intention.

Further, the method further comprises:

generating response information aiming at the target information according to the target intention and displaying the response information;

and/or the presence of a gas in the gas,

and correspondingly storing the target information and the target intention.

Another aspect of the present invention provides an intention understanding apparatus, comprising:

the first acquisition module is used for acquiring target information input by a user;

the intention identification module is used for carrying out intention identification on the target information and judging whether the identification is successful;

the second acquisition module is used for acquiring a candidate intention set corresponding to the target information when the intention identification fails;

the information construction module is used for constructing and displaying prompt information based on the candidate intention set;

the third acquisition module is used for acquiring reply information input by the user aiming at the prompt information;

and the reply understanding module is used for understanding the reply information and determining the target intention corresponding to the target information.

Another aspect of the present invention provides an electronic device comprising a processor and a memory, wherein at least one instruction or at least one program is stored in the memory, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the method as intended.

Another aspect of the present invention provides a computer-readable storage medium having at least one instruction or at least one program stored therein, the at least one instruction or the at least one program being loaded and executed by a processor to implement the intent understanding method as described above.

Due to the technical scheme, the invention has the following beneficial effects:

according to the intention understanding method provided by the embodiment of the invention, when the intention identification of the target information input by the user fails (that is, the intention of the target information is not understood), the prompt information is constructed and displayed based on the candidate intention set corresponding to the target information, and the target intention of the target information is finally determined by understanding the reply information input by the user aiming at the prompt information.

Drawings

In order to more clearly illustrate the technical solution of the present invention, the drawings used in the description of the embodiment or the prior art will be briefly described below. It is obvious that the drawings in the following description are only some embodiments of the invention, and that for a person skilled in the art, other drawings can be derived from them without inventive effort.

FIG. 1 is a schematic illustration of an implementation environment provided by an embodiment of the invention;

FIG. 2 is a flow chart of an intent understanding method provided by one embodiment of the present invention;

FIG. 3 is a flow chart of an intent understanding method provided by another embodiment of the present invention;

FIG. 4 is a schematic diagram of an apparatus for understanding purpose provided by an embodiment of the present invention;

FIG. 5 is a schematic diagram of an apparatus for understanding purpose provided by another embodiment of the present invention;

FIG. 6 is a schematic diagram of an apparatus for understanding purpose provided by another embodiment of the present invention;

fig. 7 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.

Detailed Description

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

It should be noted that the terms "first," "second," and the like in the description and claims of the present invention and in the drawings described above are used for distinguishing between similar elements and not necessarily for describing a particular sequential or chronological order. It is to be understood that the data so used is interchangeable under appropriate circumstances such that the embodiments of the invention described herein are capable of operation in sequences other than those illustrated or described herein. Furthermore, the terms "comprises," "comprising," and "having," and any variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, apparatus, article, or device that comprises a list of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, article, or device.

Referring to the specification, fig. 1 is a schematic diagram illustrating an implementation environment provided by an embodiment of the present invention, and as shown in fig. 1, the implementation environment may include at least one terminal device 110 and a server 120. The server 120 and each terminal device 110 may be directly or indirectly connected through wired or wireless communication, which is not limited in this embodiment of the present invention.

The terminal device 110 may include a smart phone, a tablet computer, a notebook computer, a desktop computer, a digital assistant, a smart speaker, a smart wearable device, a vehicle-mounted terminal, a server, and other types of physical devices, and may also include software running in the physical devices, such as an application program, but is not limited thereto. The operating system running on the terminal device 110 may include, but is not limited to, an android system, an IOS system, a linux system, a windows system, and the like.

The server 120 may be an independent physical server, a server cluster or a distributed system formed by a plurality of physical servers, or a cloud server providing basic cloud computing services such as a cloud service, a cloud database, cloud computing, a cloud function, cloud storage, a Network service, cloud communication, a middleware service, a domain name service, a security service, a Content Delivery Network (CDN), a big data and artificial intelligence platform, and the like.

In practical applications, the terminal device 110 may be configured with a dialog system, which may include a speech dialog system, a text input based dialog system, a graphical interface based dialog system, or a multimodal interaction system with speech in combination with an interface, such as a speech assistant, a smart client, etc. The dialogue system can acquire voice information sent by a user through the voice acquisition module and send the acquired voice information to the server 120 for processing, and the dialogue system can also acquire text information input by the user in an interactive interface and send the text information to the server 120 for processing. The server 120 may process the voice information/text information through the intention understanding method provided by the embodiment of the present invention, determine the corresponding target intention, thereby generating corresponding interaction data to be sent to the terminal device 110, so that the dialog system completes an interaction behavior with the user.

It should be noted that fig. 1 is only an example. Those skilled in the art will appreciate that although only 2 terminal devices 110 are shown in fig. 1, this is not a limitation of the embodiments of the present invention and that more or fewer terminal devices 110 may be included than shown.

Referring to the specification, fig. 2 illustrates a flow of an intent understanding method provided by an embodiment of the present invention, which may be applied to the server 120 in fig. 1, and particularly to fig. 2, the method may include the following steps:

s201: target information input by a user is acquired.

In the embodiment of the present invention, the target information may be a voice text corresponding to the voice information input by the user. Specifically, terminal equipment can be configured with a voice conversation system, the voice conversation system can collect voice information input by a user through a voice collection module and send the voice information to a server, the server can adopt a voice recognition technology to carry out recognition processing on the voice information to obtain a voice text corresponding to the voice information, and the voice text is used as the target information. The voice acquisition module can be a sound sensor, a microphone and the like.

In one possible embodiment, the target information may also be text information manually entered by a user through an interactive interface. Specifically, the terminal device may be configured with a dialog system based on a text input/graphical interface, the dialog system based on the text input/graphical interface may acquire text information input by a user and send the text information to a server, and the server may use the received text information as target information.

S203: and performing intention identification on the target information, and judging whether the identification is successful.

In the embodiment of the present invention, natural language understanding technology may be adopted to analyze and understand the target information to determine the intention of the target information.

Specifically, the performing intent recognition on the target information and determining whether recognition is successful may include:

judging whether the intention identification result is a valid result;

Wherein, the valid result refers to the intention label corresponding to the target information, and the invalid result refers to no output result or other recognition results except the intention label.

In the embodiment of the present invention, the first intention recognition model is a model trained in advance for performing intention recognition on a request sentence input by a user. Specifically, first sample data labeled with a corresponding intention may be obtained in advance, and a first preset neural network model may be trained by using the first sample data to obtain the first intention recognition model. The first preset neural network model may include, but is not limited to, a neural network model commonly used in the prior art, and a person skilled in the art may reasonably select a specific neural network model according to needs without limitation in the embodiment of the present invention.

S205: and when the intention identification fails, acquiring a candidate intention set corresponding to the target information.

In the embodiment of the invention, when the intention identification fails, that is, the current intention of the user is not understood, the correct intention representation of the target information which is not understood can be obtained in a mode of further interacting with the user, so that the current requirement of the user is met.

In the embodiment of the invention, under the condition that the current intention of the user is not understood, some possible candidate intentions can be recalled through one or more schemes to be used as guesses of the current intention of the user, and the strategy of interacting with the user is determined according to the candidate intentions.

In one possible embodiment, when the intention identification fails, the obtaining of the candidate intention set corresponding to the target information may include:

Specifically, the modification of the first intention recognition model may include lowering a threshold value of model output so that an intention that cannot be output originally can be output, and taking an intention obtained by lowering the threshold value of model output as a candidate intention of the target information. For example, it is assumed that the rule criterion for the first intention recognition model to output a certain intention is that the model output score of the intention is larger than a set threshold (e.g., 0.5) and can be output, and that the model output score of the intention is larger than 0.3 and equal to or smaller than 0.5 and can be output after the threshold is reduced (e.g., reduced to 0.3).

In some other possible embodiments, the additional output of the first intention recognition model may also be recalled in other ways, as a candidate intention for the target information, for example, an intention that is rejected for recognition by a rule or module may be recalled, and so on.

Wherein the intent to be rejected by a rule or module is: the original model outputs a certain intent, but is rejected by subsequent rules or modules, resulting in the final failure to output the intent. In the event that a candidate intent needs to be obtained, the restrictions of these rejection rules or modules may be lifted, allowing the intent to be output. For example, if no match is found in the knowledge base, which is a kind of rejection rule, the output of the corresponding intention for which no match is found in the knowledge base is generally limited, but in a case where the candidate intention needs to be obtained, the intention that is not output due to this reason may be recalled as the candidate intention.

In another possible embodiment, when the intention identification fails, the obtaining of the candidate intention set corresponding to the target information may include:

Specifically, the second intention recognition model may be determined by training a second preset neural network model using second sample data acquired in advance, or may be an intention recognition model adopted in a natural language understanding system in the prior art. The target information may be input into one or more second intention recognition models, and the intention output by each second intention recognition model is obtained as a candidate intention corresponding to the target information.

Optionally, second sample data labeled with corresponding intents may be obtained in advance, and the one or more second preset neural network models are trained by using the second sample data to obtain one or more second intention recognition models. Each of the second preset neural network models may include, but is not limited to, a common neural network model in the prior art, the second preset neural network model and the first preset neural network model may adopt the same model architecture or different model architectures, and when the second preset neural network model and the first preset neural network model adopt the same model architecture, sample data from different sources may be trained respectively to obtain the second preset neural network model and the first preset neural network model. It should be noted that, in the embodiment of the present invention, what kind of neural network model is specifically adopted is not limited, and those skilled in the art can reasonably select the neural network model according to needs.

Alternatively, one or more second intention recognition models in the natural language understanding system of the related art, which may be trained using completely different model architectures or data sources from the first intention recognition model and thus have a certain complementarity, may be predetermined, and may be capable of outputting the correct intention of the user when the first intention recognition model does not output the intention of the user.

Specifically, when the intention recognition fails, the intention related to the target information stored in the database may also be queried with the retrieval system as a candidate intention. The database may be stored with a plurality of request statement (query) -intention pairs in advance, and the target information may be retrieved and matched with the query in the query-intention pairs stored in the database in a q-q retrieval manner, so as to determine one or more queries that are successfully matched and corresponding intentions, and use the intentions corresponding to the one or more queries that are successfully matched as candidate intentions corresponding to the target information.

In practical application, a search matching model may be trained in advance, the input of the search matching model is the target information and each query stored in the database, the output of the search matching model is the similarity (or matching degree) between the target information and each query, and if the similarity (or matching degree) between a certain query and the target information is higher than a first threshold, it is determined that the query and the target information are successfully matched. The first preset threshold may be reasonably selected according to needs, which is not limited in the embodiment of the present invention.

It should be noted that, in practical applications, one or more of the schemes for obtaining the candidate intention set provided by the several possible embodiments described above may be used to determine the candidate intention corresponding to the target information, and when multiple schemes are used, the multiple schemes may be executed sequentially or in parallel, which is not limited by the embodiment of the present invention.

S207: and constructing prompt information based on the candidate intention set and displaying the prompt information.

In the embodiment of the present invention, when one or more candidate intentions corresponding to the target information are acquired, prompt information may be constructed according to the one or more candidate intentions, so as to prompt a user to determine a target intention from the one or more candidate intentions.

In the embodiment of the present invention, the prompt information may be text information or voice information, and after the prompt information is constructed, the prompt information may be sent to the terminal device for display, when the prompt information is text information, the terminal device may display the text information in an interactive interface, and when the prompt information is voice information, the terminal device may broadcast the voice information.

Specifically, the constructing and presenting the prompt information based on the candidate intention set may include:

In the embodiment of the present invention, when the candidate intention set is empty, guidance information may be constructed and displayed, where the guidance information may be more specific information that suggests that the user tries to change a statement or expression, for example, "do you say that i did not understand, you may also change a statement try," do you say that i did not understand, may express more specific "and the like.

In the embodiment of the present invention, when the candidate intention set only includes one candidate intention, query information is constructed and displayed according to the candidate intention, where the query information may be information inquiring whether an intention expressed by a user is the candidate intention, for example, if the candidate intention is "open window", the query information may be "ask you for a window to be opened".

In the embodiment of the present invention, when the candidate intention set includes a plurality of candidate intentions, recommendation information may be constructed and displayed according to the plurality of candidate intentions, where the recommendation information may be information that recommends the plurality of candidate intentions to the user for the user to select, or information that asks the user what intention the user expresses is, for example, if the candidate intentions include "turn down the volume" and "turn up the volume", the recommendation information may be "ask you to turn down the volume or turn up the volume", "ask you to select to turn down the volume or turn up the volume", and the like.

In one possible embodiment, a preset number of candidate intents from the candidate intent set may also be selected for recommendation to the user, such that the user selects the target intent from the selected preset number of candidate intents. The preset number may be set according to an actual situation, for example, may be set to 3, and the embodiment of the present invention does not limit this.

The embodiment of the invention guesses the intention of the target information input by the user by adopting one or more schemes, recommends the guessed result to the user, can identify the correct intention of the user as much as possible and feed back the correct intention so as to improve the use experience of the user.

In practical applications, since all information input by the user is collected and processed, the information may include information that is not interacted with the dialog system (e.g., chat information of the user with others), and for the information, in case that a valid intention (i.e., no understanding) is obtained, no further interaction may be performed with the user to obtain a corresponding correct intention. Therefore, before further interaction with the user to understand the correct intention of the target information, rejection judgment may be performed, rejection is performed on information that does not interact with the dialog system, and for information that interacts with the dialog system, a policy for interacting with the user may be determined according to the candidate intention set determined in step S205, and finally, the correct intention of the information is determined.

In a possible embodiment, with reference to fig. 3 of the specification, before constructing and presenting the prompt information based on the candidate intent set, the method may further include:

s206: analyzing the target information, and determining a first intention type corresponding to the target information; judging whether the target information needs to be identified or not according to the first intention type; and when the target information needs to be identified, constructing prompt information based on the candidate intention set and displaying the prompt information.

The first intention type is used for indicating whether the target information is information of interaction between a user and a dialog system. When the target information can be determined to be the interaction information between the user and the dialog system according to the first intention type, the target information can be determined to need to be identified, when the target information can be determined to be not the interaction information between the user and the dialog system according to the first intention type, the target information is rejected, and the processing process aiming at the target information is finished.

Specifically, the analyzing the target information and determining the first intention type corresponding to the target information may include:

acquiring current environment information;

Wherein the input of the first intention classification model comprises the target information and the environment information, the environment information may comprise historical input information, action behaviors of a user (such as operating a physical button or a touch screen interface), information displayed to the user by the dialog system, and the like, and the output of the first intention classification model is a first intention type, and the first intention type is used for indicating whether the target information is information for the user to interact with the dialog system.

Specifically, third sample data labeled with a corresponding intention type may be obtained in advance, and a third preset neural network model may be trained by using the third sample data to obtain the first intention classification model. The third preset neural network model may include, but is not limited to, a neural network model commonly used in the prior art, and may be, for example, a two-class model based on a pre-trained transformer, and the like. It should be noted that, in the embodiment of the present invention, what kind of neural network model is specifically adopted is not limited, and those skilled in the art can reasonably select the neural network model according to needs.

In some possible embodiments, the specific implementation of step S205 to step S207 may also be implemented by training an end-to-end generative model, and the target information and the current environmental information are input into the generative model, so that the prompt information for asking questions or guiding may be obtained and displayed.

S209: and acquiring reply information input by the user aiming at the prompt information.

In the embodiment of the present invention, after the terminal device displays the prompt information, the terminal device may acquire the reply voice information input by the user for the prompt information through a voice acquisition module disposed in the terminal device and send the reply voice information to the server, and the server may perform recognition processing on the reply voice information by using a voice recognition technology to obtain a reply voice text corresponding to the reply voice information, and use the reply voice text as the reply information.

In a possible embodiment, after the terminal device displays the prompt message, a reply text message manually input by the user through an interactive interface can be acquired and sent to the server, and the server can use the received reply text message as the reply message.

Specifically, the reply information may be an intention confirmation sentence, an intention selection sentence, an intention negation sentence of the user, correction information for the target information, or the like.

S211: and understanding the reply information, and determining a target intention corresponding to the target information.

In the embodiment of the present invention, the reply message may include three types of messages: positive feedback information, negative feedback information, and other feedback information. The positive feedback information is feedback information which can be learned, and comprises information that a user expresses positive intention aiming at the inquiry information, information that the user selects one candidate intention in the candidate intentions aiming at the recommendation information, and information that the user reforms one-pass intention aiming at the guide information. The negative feedback information is feedback that learning is not possible, and includes information that the user has expressed a negative intention for the query information and information that the user has expressed a negative intention for all the candidates for the recommendation information. The other feedback information is information that the user does not manage the prompt information and expresses other intentions.

In the embodiment of the invention, the reply information can be understood through various schemes to determine the type of the reply information, and further determine the correct intention corresponding to the target information input by the user.

In one possible embodiment, the user may reply to the reminder by positive or negative means. Accordingly, the understanding of the reply information, and the determining of the target intention corresponding to the target information may include:

Wherein the input of the second intention classification model comprises the reply information, the output of the second intention classification model is a second intention type, the second intention type is used for indicating the intention type expressed by the reply information, and comprises a positive intention, a negative intention and other intentions, the reply information can be determined to be positive feedback information, negative feedback information or other feedback information according to the second intention type and the information type of the prompt information, and further, the target intention corresponding to the target information can be determined.

In the embodiment of the present invention, fourth sample data labeled with a corresponding intention type may be obtained in advance, and a fourth preset neural network model is trained by using the fourth sample data to obtain the second intention classification model. The fourth preset neural network model may include, but is not limited to, a neural network model commonly used in the prior art, and a person skilled in the art may reasonably select a specific neural network model according to needs without limitation in the embodiment of the present invention.

Specifically, when the prompt information is query information, if the second intention type corresponding to the reply information is a positive intention, it may be determined that the reply information is positive feedback information, and further it may be determined that the candidate intention is a correct intention corresponding to the target information; if the second intention type corresponding to the reply information is negative intention, determining that the reply information is negative feedback information, and further determining that the candidate intention is not correct intention corresponding to the target information; and if the second intention type corresponding to the reply information is other intentions, the correct intention corresponding to the target information cannot be determined.

Specifically, when the prompt information is recommendation information, if the second intention type corresponding to the reply information is negative intention, it may be determined that the reply information is negative feedback information, and none of the candidate intentions is a correct intention corresponding to the target information; if the second intention type corresponding to the reply information is positive intention or other intention, the correct intention corresponding to the target information cannot be determined.

In another possible embodiment, the user may reply to the prompt message by repeating the candidate intentions in the prompt message to select a candidate intention, in which case understanding may be achieved by replying to intention matching. Accordingly, the understanding of the reply information, and the determining of the target intention corresponding to the target information may include:

Wherein the intention matching model may match the reply information with the candidate intentions in the candidate intention set one by one, and output a result of a match hit (if any). According to the output result of the intention matching model, whether the reply information is positive feedback information, negative feedback information or other feedback information can be determined, and further, the target intention corresponding to the target information can be determined.

In the embodiment of the present invention, fifth sample data labeled with a matching relationship may be obtained in advance, and a fifth preset neural network model is trained by using the fifth sample data to obtain the intention matching model. The fifth preset neural network model may include, but is not limited to, a neural network model commonly used in the prior art, and a person skilled in the art may reasonably select a specific neural network model according to needs without limitation in the embodiment of the present invention.

Specifically, when the prompt information is query information, if the intention matching model has an output result, it may be determined that the reply information is positive feedback information, and it may be further determined that the candidate intention is a correct intention corresponding to the target information; if the intention matching model does not output a result, the information type of the reply information cannot be determined, and the correct intention corresponding to the target information cannot be determined.

Specifically, when the prompt information is recommendation information, if the intention matching model has an output result, it may be determined that the reply information is positive feedback information, and further it may be determined that a target candidate intention in the output result is a correct intention corresponding to the target information; if the intention matching model does not output a result, the information type of the reply information cannot be determined, and the correct intention corresponding to the target information cannot be determined.

In another possible embodiment, the user may reply to the prompt message by adjusting or refining the expression and repeating the intention to feed back the intention of the target message, and at this time, the understanding may be realized by repeating the detection. Accordingly, the understanding of the reply information, and the determining of the target intention corresponding to the target information may include:

In the embodiment of the present invention, the reply information may be subjected to intent recognition by using the first intent recognition model to obtain a corresponding intent recognition result, when the intent recognition result is an invalid result, it is determined that the intent recognition is failed, and when the intent recognition result is a valid result, it is determined that the intent recognition is successful. And the valid result refers to an intention label corresponding to the reply information, and the invalid result refers to no output result or other recognition results except the intention label.

In the embodiment of the present invention, no matter what type of the prompt information is, when the reply information is successfully identified (i.e., the intention of the reply information can be understood), it may be determined whether the reply information is the repeat information of the target information, and if so, it may be determined that the reply information is positive feedback information, and further, it may be determined that the intention corresponding to the reply information is the correct intention corresponding to the target information; if not, the information type of the reply information cannot be determined, and the correct intention corresponding to the target information cannot be determined. When the reply information is failed to be identified, the correct intention corresponding to the target information cannot be determined.

In practical application, a sixth preset neural network model may be trained by using the retest detection sample data obtained in advance, so as to obtain a retest detection model for retest detection of the reply information. The input of the repeat detection model is reply information and target information input by a user, the output of the repeat detection model is a result of whether the semantics of the reply information and the target information are the same, if the semantics of the reply information and the target information are the same, the reply information can be determined to be the repeat information of the target information, otherwise, the reply information can be determined not to be the repeat information of the target information. The sixth preset neural network model may include, but is not limited to, a neural network model commonly used in the prior art, for example, a neural network model based on a pre-training model, and the like.

It should be noted that, in practical applications, one or more of the schemes provided by the above several possible embodiments for understanding the reply information may be selected according to the information type of the prompt information to understand the reply information, and when multiple schemes are adopted, the multiple schemes may be executed sequentially or in parallel, which is not limited in this embodiment of the present invention.

In some possible embodiments, the specific implementation scheme of step S211 may also be implemented by training an end-to-end intention understanding model that is understood in combination with the dialog history of the user and the system, and the target intention corresponding to the target information may be obtained by inputting the reply information and the dialog history information into the intention understanding model.

In summary, according to the intention understanding method of the embodiment of the present invention, when the intention recognition of the target information input by the user fails (that is, the intention of the target information is not understood), the prompt information is constructed and displayed based on the candidate intention set corresponding to the target information, and the target intention of the target information is finally determined by understanding the reply information input by the user with respect to the prompt information.

In one possible embodiment, the method may further include:

and/or the presence of a gas in the gas,

and correspondingly storing the target information and the target intention.

Specifically, after the target intention corresponding to the target information is acquired, response information for the target information may be generated according to the target intention, and the response information is sent to the terminal device, so that a dialog system in the terminal device executes corresponding interactive operation according to the response information. The response information may include reply information for the target information, and may also include an interactive operation for a target execution object. When the response message is a reply message, the dialog system may display the reply message, for example, display the reply message in an interactive interface, or broadcast the reply message by voice. When the response information is the interactive operation for the target execution object, the dialog system may generate an operation instruction for executing the interactive operation and send the operation instruction to the target execution object, and the target execution object may execute the interactive operation in response to the operation instruction to obtain an execution result.

According to the embodiment of the invention, after the correct intention of the user request which is not understood before is learned through feedback in interaction with the user, the response information of the learned user request can be given in time, the requirement of the user is met, the user can sense the learning of a dialog system, and the user experience is better.

Specifically, after a target intention corresponding to the target information is acquired, a query may be generated as the target information, and a query-intention pair with an intention as the target intention may be stored in a database of a retrieval system. And adding the learned query-intention pairs into a database of a retrieval system, and then matching hit requests in the database through the retrieval system if the same or similar unintelligible user requests occur again, so that the hit requests can be directly validated, and the intentions of the hit query-intention pairs are output.

Specifically, after the target intention corresponding to the target information is acquired, new entity knowledge (such as song names, place names, etc., if any) can be extracted from the query-intention pairs generated by the target information and the target intention and stored in a knowledge base. And then if the same entity appears in the user request again, the correct recognition and understanding can be carried out by means of the knowledge base.

Specifically, after the target intention corresponding to the target information is acquired, a training sample pair may also be generated by using the target information and the target intention, and stored in a training data pool of an intention recognition model. The defects of the original model can be optimized in a targeted manner during the next optimization training of the intention recognition model, so that the updated model can correctly understand the user requests which are the same as or similar to the sample pairs.

It should be noted that other schemes may also be adopted in the embodiment of the present invention to correspondingly store and validate the target information and the target intent, for example, a query-intent pair generated by the target information and the target intent may be validated by adding an online white list, so that the user request that is the same or similar to the query-intent pair can be correctly understood later, an online learning manner may also be adopted to perform optimization training on an intent recognition model, and a newly learned sample pair is directly transmitted to the model for online learning and online updating.

According to the embodiment of the invention, the user request and the intention which are fed back and learned in interaction with the user are correspondingly stored in the retrieval system and added into the online white list, or the new entity knowledge extracted from the user request and the intention are stored in the knowledge base, or the new entity knowledge is correspondingly stored in the training sample pool so as to carry out optimization training on the original intention recognition model in time, so that the understanding capability of the dialogue system can be rapidly improved, and the learning and use can be realized, namely, if the user expresses the request sentence which is the same as or similar to the target information again, the user can correctly understand the request sentence without interaction and learning again.

Reference is made to the description accompanying fig. 4, which illustrates an example of an apparatus 400 for understanding the structure provided by an embodiment of the present invention. As shown in fig. 4, the apparatus 400 may include:

a first obtaining module 410, configured to obtain target information input by a user;

an intention identifying module 420, configured to perform intention identification on the target information, and determine whether the identification is successful;

a second obtaining module 430, configured to obtain a candidate intent set corresponding to the target information when intent recognition fails;

an information construction module 440, configured to construct and display prompt information based on the candidate intention set;

a third obtaining module 450, configured to obtain reply information input by the user for the prompt message;

and a reply understanding module 460, configured to understand the reply information and determine a target intention corresponding to the target information.

In one possible embodiment, as shown in fig. 5, the apparatus 400 may further include:

a rejection judging module 470, configured to analyze the target information and determine a first intention type corresponding to the target information; judging whether the target information needs to be identified or not according to the first intention type; and when the target information needs to be identified, constructing prompt information based on the candidate intention set and displaying the prompt information.

Alternatively, the second obtaining module 430, the information constructing module 440, and the rejection judging module 470 may be integrated into a question policy module, and the question policy module implements the functions of the above modules.

Alternatively, as shown in fig. 6, the intention recognition module 420, the question policy module, and the reply understanding module 460 may be integrated into a natural language understanding module, and the functions of the modules are implemented by the natural language understanding module.

In one possible embodiment, as shown in fig. 6, the apparatus 400 may further include a dialog management module for generating response information for the target information according to the target intention.

In one possible embodiment, as shown in fig. 6, the apparatus 400 may further include a knowledge storage module for correspondingly storing the target information and the target intention.

It should be noted that, when the apparatus provided in the foregoing embodiment implements the functions thereof, only the division of the functional modules is illustrated, and in practical applications, the functions may be distributed by different functional modules according to needs, that is, the internal structure of the apparatus may be divided into different functional modules to implement all or part of the functions described above. In addition, the apparatus provided in the above embodiments and the corresponding method embodiments belong to the same concept, and specific implementation processes thereof are detailed in the corresponding method embodiments and are not described herein again.

An embodiment of the present invention further provides an electronic device, which includes a processor and a memory, where at least one instruction or at least one program is stored in the memory, and the at least one instruction or the at least one program is loaded by the processor and executed to implement the intent understanding method provided by the above method embodiment.

The memory may be used to store software programs and modules, and the processor may execute various functional applications and data processing by operating the software programs and modules stored in the memory. The memory can mainly comprise a program storage area and a data storage area, wherein the program storage area can store an operating system, application programs needed by functions and the like; the storage data area may store data created according to use of the apparatus, and the like. Further, the memory may include high speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid state storage device. Accordingly, the memory may also include a memory controller to provide the processor access to the memory.

The method embodiments provided by the embodiments of the present invention may be executed in a terminal, a server, or a similar computing device, that is, the electronic device may include a terminal, a server, or a similar computing device. Taking the example of running on a server, as shown in fig. 7, it shows a schematic structural diagram of a server implementing the intent understanding method provided by the embodiment of the present invention. The server 700 may vary significantly due to configuration or performance, and may include one or more Central Processing Units (CPUs) 710 (e.g., one or more processors) and memory 730, one or more storage media 720 (e.g., one or more mass storage devices) storing applications 723 or data 722. Memory 730 and storage medium 720 may be, among other things, transient storage or persistent storage. The program stored in the storage medium 720 may include one or more modules, each of which may include a series of instruction operations for the server. Still further, central processor 710 may be configured to communicate with storage medium 720 and execute a series of instruction operations in storage medium 720 on server 700. The server 700 may also include one or more power supplies 760, one or more wired or wireless network interfaces 750, one or more input-output interfaces 740, and/or one or more operating systems 721, such as Windows Server, Mac OS XTM, UnixTM, LinuxTM, FreeBSDTM, etc.

The input/output interface 740 may be used to receive or transmit data via a network. Specific examples of the network described above may include a wireless network provided by a communication provider of the server 700. In one example, the input/output Interface 740 includes a Network adapter (NIC) that can be connected to other Network devices through a base station to communicate with the internet. In one example, the input/output interface 740 may be a Radio Frequency (RF) module for communicating with the internet in a wireless manner, and the wireless communication may use any communication standard or protocol, including but not limited to Global System for Mobile communication (GSM), General Packet Radio Service (GPRS), Code Division Multiple Access (CDMA), Wideband Code Division Multiple Access (WCDMA), Long Term Evolution (LTE), email, Short Message Service (SMS), and the like.

Those skilled in the art will appreciate that the configuration shown in fig. 7 is merely illustrative and that server 700 may include more or fewer components than shown in fig. 7 or have a different configuration than shown in fig. 7.

An embodiment of the present invention further provides a computer-readable storage medium, which can be disposed in an electronic device to store at least one instruction or at least one program for implementing an intention understanding method, where the at least one instruction or the at least one program is loaded and executed by the processor to implement the intention understanding method provided by the above-mentioned method embodiment.

Optionally, in an embodiment of the present invention, the storage medium may include, but is not limited to: a U-disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a removable hard disk, a magnetic or optical disk, and other various media capable of storing program codes.

An embodiment of the invention also provides a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The processor of the computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, causing the computer device to perform the intent understanding method provided in the various alternative embodiments described above.

It should be noted that: the precedence order of the above embodiments of the present invention is only for description, and does not represent the merits of the embodiments. And specific embodiments thereof have been described above. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims may be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing may also be possible or may be advantageous.

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

It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program instructing relevant hardware, where the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.

The above description is only for the purpose of illustrating the preferred embodiments of the present invention and is not to be construed as limiting the invention, and any modifications, equivalents, improvements and the like that fall within the spirit and principle of the present invention are intended to be included therein.

Claims

1. An intent understanding method, comprising:

acquiring target information input by a user;

acquiring reply information input by the user aiming at the prompt information;

2. The method of claim 1, wherein the performing intent recognition on the target information and determining whether recognition is successful comprises:

judging whether the intention identification result is a valid result;

3. The method according to claim 2, wherein when the intention identification fails, the obtaining of the candidate intention set corresponding to the target information comprises:

4. The method according to claim 1, wherein when the intention identification fails, the obtaining of the candidate intention set corresponding to the target information comprises:

5. The method according to claim 1, wherein when the intention identification fails, the obtaining of the candidate intention set corresponding to the target information comprises:

6. The method of claim 1, wherein before constructing and presenting the hint information based on the set of candidate ideas, further comprising:

7. The method of claim 6, wherein analyzing the target information and determining the first intent type corresponding to the target information comprises:

acquiring current environment information;

8. The method of claim 1, wherein the constructing and presenting hints based on the set of candidate ideas comprises:

9. The method according to claim 1, wherein the understanding of the reply information, and the determining of the target intention corresponding to the target information comprises:

10. The method according to claim 1, wherein the understanding of the reply information, and the determining of the target intention corresponding to the target information comprises:

11. The method according to claim 1, wherein the understanding of the reply information, and the determining of the target intention corresponding to the target information comprises:

12. The method of claim 1, further comprising:

and/or the presence of a gas in the gas,

and correspondingly storing the target information and the target intention.

13. An intent understanding apparatus, comprising:

14. An electronic device, comprising a processor and a memory, wherein at least one instruction or at least one program is stored in the memory, and the at least one instruction or the at least one program is loaded and executed by the processor to implement the intent understanding method of any of claims 1-12.

15. A computer-readable storage medium, in which at least one instruction or at least one program is stored, which is loaded and executed by a processor to implement the intent understanding method of any of claims 1-12.