Intention recognition method and device, storage medium and electronic device

Abstract

The embodiment of the invention provides an intention recognition method and device, a storage medium and an electronic device.The method comprises the steps that first interaction data input by a first object included in a target dialogue is determined; inputting the first interaction data into a target network model, and determining a first intention included in the first interaction data by using the target network model; wherein the target network model is obtained by training in the following mode: training an initial model through training data to obtain a first network model; negative feedback data included in the historical dialogue is determined, and the intention of the negative feedback data recognized by the first network model is different from the intention expressed by the object inputting the negative feedback data; and training the first network model through the negative feedback data to obtain a target network model. According to the invention, the problem of inaccurate intention recognition is solved, and the effect of improving the intention recognition accuracy is achieved.

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of communications, and in particular, to a method, device, storage medium and electronic device for identifying intent. Background Art

[0002] In related technologies, as the number of users continues to increase, the amount of consultations users ask when handling business also continues to increase. The efficiency of business handling can be improved through AI automated operation technology. However, the accuracy rate is often low during the interaction between AI and users due to errors in AI’s recognition of user intent.

[0003] From this, we can see that there is a technical problem of inaccurate intent recognition in the relevant technology.

[0004] With respect to the above-mentioned problems existing in the related technologies, no effective solution has been proposed yet. Summary of the invention

[0005] Embodiments of the present invention provide a method, device, storage medium and electronic device for identifying intent, so as to at least solve the problem of inaccurate intent identification existing in the related art.

[0006] According to one embodiment of the present invention, there is provided a method for identifying intent, comprising: determining first interaction data input by a first object included in a target conversation; inputting the first interaction data into a target network model to determine a first intent included in the first interaction data through the target network model; wherein the target network model is trained in the following manner: training an initial model through training data to obtain a first network model; determining negative feedback data included in a historical conversation, wherein the intent of the negative feedback data identified by the first network model is different from the intent expressed by the object that input the negative feedback data; and training the first network model through the negative feedback data to obtain the target network model.

[0007] In an exemplary embodiment, determining the negative feedback data included in the historical conversation includes: using the first network model to identify second interaction data input by a second object in the historical conversation to obtain a first recognition result; when the first recognition result indicates that the first network model has recognized the same intention at least twice in a row, marking the second interaction data as abnormal data; identifying the abnormal data to obtain a second recognition result; when the second recognition result is consistent with the first recognition result, determining the abnormal data as negative feedback data.

[0008] In an exemplary embodiment, determining the negative feedback data included in the historical conversation includes: determining other data included in the historical conversation except abnormal data, wherein the other data is data input by a second object; performing random detection on the other data to obtain a first detection result; when the first detection result is inconsistent with the recognition result of the first network model, determining the other data as the negative feedback data.

[0009] In an exemplary embodiment, determining the negative feedback data included in the historical conversation includes: for any target historical conversation included in the historical conversation, the following operations are performed: when the same response data appears at least twice in a row in the target historical conversation, the question data corresponding to the response data is determined as the negative feedback data.

[0010] In an exemplary embodiment, inputting the first interaction data into a target network model to determine the first intent included in the first interaction data through the target network model includes: inputting the first interaction data into a plurality of sub-models included in the target network model respectively to determine a target score obtained by each sub-model scoring the first interaction data, wherein one sub-model corresponds to one intent, each sub-model corresponds to a different intent, and the target score is used to indicate the association between the first interaction data and the intent corresponding to the sub-model; determining the highest score included in the target score; and when the highest score is greater than a preset threshold, determining the intent corresponding to the highest score as the first intent.

[0011] In an exemplary embodiment, after determining the highest score included in the target score, the method further includes: constructing a target sub-model based on the first interaction data when the highest score is less than or equal to a preset threshold; and adding the target sub-model to the target network model.

[0012] In an exemplary embodiment, after inputting the first interaction data into a target network model to determine the first intent included in the first interaction data through the target network model, the method further includes: determining a reply template corresponding to the first intent; generating reply data based on the reply template; and outputting the reply data.

[0013] In an exemplary embodiment, the method further includes: upon receiving a sparring instruction, determining a business scenario included in the sparring instruction; obtaining a sparring dialogue corresponding to the business scenario; inputting third interaction data included in the sparring dialogue into a dialog box; identifying reply data in the dialog box for the third interaction data; determining a target correlation between the reply data and the third interaction data; when the target correlation is greater than or equal to a second preset threshold, inputting fourth interaction data included in the sparring dialogue into the dialog box; when the target correlation is less than the second preset threshold, adjusting the third interaction data to obtain fifth interaction data, wherein the intent of the fifth interaction data is the same as the intent of the third interaction data, and inputting the fifth interaction data into the dialog box.

[0014] In an exemplary embodiment, after inputting the first interaction data into a target network model to determine a target first intent included in the first interaction data through the target network model, the method further includes: determining business information associated with the first intent; displaying the business information in a target interface; upon receiving a click instruction, sending associated data of the first business information indicated by the click instruction to the target conversation; upon receiving a view instruction, displaying associated data of the second business information indicated by the view instruction in the target interface.

[0015] According to another embodiment of the present invention, there is provided an apparatus for identifying intent, comprising: a first determination module for determining first interaction data input by a first object included in a target conversation; a second determination module for inputting the first interaction data into a target network model to determine a first intent included in the first interaction data through the target network model; wherein the target network model is trained in the following manner: training an initial model through training data to obtain a first network model; determining negative feedback data included in historical conversations, wherein the intent of the negative feedback data identified by the first network model is different from the intent expressed by the object inputting the negative feedback data; and training the first network model through the negative feedback data to obtain the target network model.

[0016] According to yet another embodiment of the present invention, a computer-readable storage medium is provided, in which a computer program is stored, wherein the computer program is configured to execute the steps of any one of the above method embodiments when run.

[0017] According to yet another embodiment of the present invention, there is provided an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0018] According to yet another embodiment of the present invention, a computer program product is provided, including a computer program, and when the computer program is executed by a processor, the steps of the method described in each embodiment of the present application are implemented.

[0019] Through the present invention, the intention of the data input by the object recognized by the first network model obtained after the initial model is trained using the training data in the historical conversation is determined. When the intention recognized by the first network model is different from the intention of the data input by the object, the data input by the object can be determined as negative feedback data. By training the first network model with negative feedback data, a target network model can be obtained. After being trained with negative feedback data, the target network model can accurately determine the first intention of the first interaction data input by the first object in the target conversation. Therefore, the problem of inaccurate intention recognition existing in the related art can be solved, and the effect of improving the accuracy of intention recognition can be achieved. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 is a hardware structure block diagram of a mobile terminal according to a method for identifying intent in an embodiment of the present invention;

[0021] Figure 2 is a flow chart of a method for identifying intent according to an embodiment of the present invention;

[0022] Figure 3 is a schematic diagram of a specific model structure according to an embodiment of the present invention;

[0023] Figure 4 is a schematic diagram of the working principle of an image recognition module according to an embodiment of the present invention;

[0024] Figure 5 is a flow chart of online recall of negative feedback data according to an embodiment of the present invention;

[0025] Figure 6 is a working principle diagram of an ASR speech recognition model according to an embodiment of the present invention;

[0026] Figure 7 is a schematic diagram of an ASR speech recognition model optimization principle according to an embodiment of the present invention;

[0027] Figure 8 is a working principle diagram of target network model optimization according to an embodiment of the present invention;

[0028] Fig. 9is a schematic diagram of a target network model identifying user intent according to an embodiment of the present invention;

[0029] Fig.10 is an architecture diagram of a business-side intelligent platform according to an embodiment of the present invention;

[0030] Fig.11 is a diagram of the client-side intelligent platform architecture according to an embodiment of the present invention;

[0031] Fig.12 is a structural block diagram of an apparatus for identifying intent according to an embodiment of the present invention. DETAILED DESCRIPTION

[0032] Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings and in combination with the embodiments.

[0033] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence.

[0034] The method embodiments provided in the embodiments of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Taking running on a mobile terminal as an example, Figure 1 FIG. 1 is a hardware structure block diagram of a mobile terminal of an intention recognition method according to an embodiment of the present invention. Figure 1 As shown, the mobile terminal may include one or more ( Figure 1 Only one is shown in the figure) a processor 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA) and a memory 104 for storing data, wherein the mobile terminal may also include a transmission device 106 and an input / output device 108 for communication functions. It can be understood by those skilled in the art that Figure 1 The structure shown is only for illustration and does not limit the structure of the mobile terminal. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations shown.

[0035] The memory 104 can be used to store computer programs, for example, software programs and modules of application software, such as the computer program corresponding to the method for identifying the intention in the embodiment of the present invention. The processor 102 executes various functional applications and data processing by running the computer program stored in the memory 104, that is, to implement the above method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the mobile terminal via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0036] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the mobile terminal. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, referred to as NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0037] In this embodiment, a method for identifying intent is provided. Figure 2 is a flow chart of a method for identifying intent according to an embodiment of the present invention. Figure 2 As shown, the process includes the following steps:

[0038] Step S202, determining first interaction data input by a first object included in the target dialogue;

[0039] Step S204, inputting the first interaction data into a target network model, so as to determine a first intention included in the first interaction data through the target network model;

[0040] The target network model is trained in the following way:

[0041] The initial model is trained using training data to obtain a first network model;

[0042] Determining negative feedback data included in the historical conversation, wherein the intention of the negative feedback data identified by the first network model is different from the intention expressed by the object that input the negative feedback data;

[0043] The first network model is trained by using the negative feedback data to obtain the target network model.

[0044] In the above embodiment, the first interaction data may be data sent by the user (i.e., the above-mentioned first object) to the platform, which may be in text form, image form, or voice form, but is not limited thereto. The first interaction data is input into the target network model, and the target network model may identify the first intention of the first interaction data input by the first object. The initial model may be trained using the original manually annotated data (i.e., the above-mentioned training data) to construct a first network model. The intention of the first network model to recognize the data input by the user in the historical conversation is obtained. When the intention recognized by the first network model is different from the intention expressed by the data input by the user, the data input by the user may be determined as negative feedback data.

[0045] In the above embodiment, the first network model can be an ASR speech recognition model (Automatic Speech Recognition), a TTS speech synthesis model (Text To Speech), an NLP natural language processing (Natural Language Processing), or an OCR text recognition model (Optical Character Recognition), an image recognition module, etc. In the process of model training, a multi-head training strategy can be adopted to integrate the original manually annotated data, positive feedback data, and negative feedback data, wherein the positive feedback data is the data whose intent recognized by the first network model is the same as the intent expressed by the data input by the object. For a specific model structure diagram, please refer to Figure 3 , the production model can be used as the base model, and the original manually labeled data and positive feedback data use the original input layer, hidden layer, and output layer; the negative feedback data uses the original input layer and hidden layer, and the output layer of the original model is copied and used separately. In the model usage stage after obtaining the target network model, only the input layer, hidden layer, and output layer used by the original manually labeled data and positive feedback data can be used. By training the first network model with negative feedback data, an optimized target network model can be obtained, which can achieve more accurate recognition of user input data. In addition, training based on production data generated during platform operation solves the problem of insufficient training data sources and reduces the cost of purchasing original training data.

[0046] In the above embodiment, the image recognition module can scan and register the ID card information through the OCR text recognition model. The working principle diagram of the image recognition module can be referred to Figure 4 ,like Figure 4As shown, the image recognition module captures the ID card photo data, recognizes it through OCR technology and returns the recognized information for dynamic filling. For the recognition of rare Chinese characters, the first network model can be optimized by online recall of negative feedback data after the management system is put into production. The flowchart of online recall of negative feedback data can be referred to Figure 5 ,like Figure 5 As shown in the figure, the OCR model uses the data recalled after production as negative feedback data to optimize the model. Similar to the speech recognition module, both use negative feedback data for multi-head training. Among them, negative feedback data can be data modified by the target object or data that the system does not recognize and reports errors. By effectively formulating data recall rules and strategies, the positive feedback data and negative feedback data accumulated in production can be effectively screened, which can greatly reduce the labeling cost.

[0047] Through the present invention, the intention of the data input by the object recognized by the first network model obtained after the initial model is trained using the training data in the historical conversation is determined. When the intention recognized by the first network model is different from the intention of the data input by the object, the data input by the object can be determined as negative feedback data. By training the first network model with negative feedback data, a target network model can be obtained. After being trained with negative feedback data, the target network model can accurately determine the first intention of the first interaction data input by the first object in the target conversation. Therefore, the problem of inaccurate intention recognition existing in the related art can be solved, and the effect of improving the accuracy of intention recognition can be achieved.

[0048] Optionally, the execution subject of the above steps may be a CRM (Customer Relationship Management) customer management system, or any module with identification capability such as a customer management module.

[0049] In an exemplary embodiment, determining the negative feedback data included in the historical conversation includes: using the first network model to identify second interaction data input by a second object in the historical conversation to obtain a first recognition result; when the first recognition result indicates that the first network model has recognized the same intention at least twice in a row, marking the second interaction data as abnormal data; identifying the abnormal data to obtain a second recognition result; when the second recognition result is consistent with the first recognition result, determining the abnormal data as negative feedback data.

[0050] In the above embodiment, the speech recognition model ASR can be used to recognize the second interaction data input by the user (i.e., the above second object) in the historical conversation to obtain the first recognition result. The working principle of the ASR speech recognition model can be referred to Figure 6 , Figure 6is a working principle diagram of an ASR speech recognition model according to an embodiment of the present invention, such as Figure 6 As shown, the speech recognition model can convert audio data into text, extract features from the input audio, and determine the text with the highest matching rate with the audio data as the recognition result by decoding the search text sequence.

[0051] In the above embodiment, the ASR speech recognition model can define the standard for audio data feature extraction as 13-dimensional MCFF (Mel-frequency cepstral coefficients, based on multi-scale convolutional neural network feature fusion) features and first- and second-order differences and total energy, and can map each frame of audio signal to a 40-dimensional feature vector. The acoustic model that constitutes the decoder can map the audio sequence to an acoustic unit sequence, the dictionary can convert the acoustic unit sequence into a text sequence, and the language model provides a posteriori probability to correct the text sequence output by the dictionary. Among them, the dictionary and language model can set proper nouns according to different dialogue scenarios, and can increase and update the weights of proper nouns and professional terms in real time.

[0052] In the above embodiment, when the output result of the model does not correctly identify the user's intention, the user usually continues to input data for query, and when the recognition result output at this time is the same as the previous time, it means that the model has not recognized the user's intention. Figure 7 As shown, in the process of the first network model identifying the full amount of data (i.e., the second interaction data mentioned above), the second interaction data corresponding to the same intention identified twice in a row and the second interaction data without any recognition result by the first network model can be determined as suspected negative feedback data (i.e., the abnormal data mentioned above). A second recognition result can be obtained by performing a second recognition on the abnormal data (i.e., the re-examination described in the figure). When the second recognition result is consistent with the first recognition result, the abnormal data can be determined as negative feedback data.

[0053] In an exemplary embodiment, determining the negative feedback data included in the historical conversation includes: determining other data included in the historical conversation except abnormal data, wherein the other data is data input by a second object; performing random detection on the other data to obtain a first detection result; when the first detection result is inconsistent with the recognition result of the first network model, determining the other data as the negative feedback data.

[0054] In the above embodiment, the other data except the abnormal data in the full amount of data input by the second object in the historical conversation can be randomly detected to obtain a detection result (i.e., the above first detection result). When the first detection result is inconsistent with the recognition result of the first network model, the other data can be determined as negative feedback data.

[0055] In an exemplary embodiment, determining the negative feedback data included in the historical conversation includes: for any target historical conversation included in the historical conversation, the following operations are performed: when the same response data appears at least twice in a row in the target historical conversation, the question data corresponding to the response data is determined as the negative feedback data.

[0056] In the above embodiment, any target historical dialogue can be understood as the dialogue process of any object in the historical dialogue, such as Figure 8 As shown, all data in the production process are automatically marked. For the matched data, when the first network model gives the same response to at least two input data of the same object (i.e., the question data corresponding to the above-mentioned response data), that is, the same response data appears at least twice in a row in a certain target historical dialogue, it can be said that the first network model did not correctly identify the object's intention, and the object's input data can be determined as negative feedback data. Random detection can be performed on the matched data, and the misidentified data can be screened in a timely manner to ensure the accuracy of the existing result recognition.

[0057] In an exemplary embodiment, inputting the first interaction data into a target network model to determine the first intent included in the first interaction data through the target network model includes: inputting the first interaction data into a plurality of sub-models included in the target network model respectively to determine a target score obtained by each sub-model scoring the first interaction data, wherein one sub-model corresponds to one intent, each sub-model corresponds to a different intent, and the target score is used to indicate the association between the first interaction data and the intent corresponding to the sub-model; determining the highest score included in the target score; and when the highest score is greater than a preset threshold, determining the intent corresponding to the highest score as the first intent.

[0058] In the above embodiment, the online data processing results can be extracted regularly. The principle diagram of the target network model identifying the user intention can be found in Fig. 9 ,like Fig. 9 As shown, the target network model can be a FAQ (Frequently Asked Questions) model, and a FAQ database can be constructed based on the questioning method and answer of each user's question. Multiple FAQ models will identify the data input by the user (i.e., the first interaction data mentioned above) to obtain multiple target scores. Among them, one FAQ model corresponds to one intent, and the target score can illustrate the correlation between the first interaction data and the intent represented in each FAQ model.

[0059] In the above embodiment, the highest score in the target score is determined, which can indicate that the intention of the first interaction data is most highly correlated with the intention of this FAQ model. When the highest score reaches a preset threshold, the intention of the FAQ model corresponding to the highest score can be determined as the first intention, wherein the preset threshold can be 60, 75, 90, etc., but is not limited thereto.

[0060] In an exemplary embodiment, after determining the highest score included in the target score, the method further includes: constructing a target sub-model based on the first interaction data when the highest score is less than or equal to a preset threshold; and adding the target sub-model to the target network model.

[0061] In the above embodiment, if Fig. 9 As shown, when the highest score is less than or equal to the preset threshold (i.e., unmatched data), it can be determined whether the unmatched first interaction data and the negative feedback data belong to the existing FAQ model. The data belonging to the existing FAQ list can be used to enrich the database, and the data that does not belong to the existing FAQ list can be collected and the relevant parts of the feedback can be unified to build a new FAQ model (i.e., the above-mentioned target sub-model) and added to the target network model.

[0062] In an exemplary embodiment, after inputting the first interaction data into a target network model to determine the first intent included in the first interaction data through the target network model, the method further includes: determining a reply template corresponding to the first intent; generating reply data based on the reply template; and outputting the reply data.

[0063] In the above embodiment, when the first intention in the first interaction data is determined, a reply template matching the first intention can be determined according to the FAQ database, and reply data is generated by the reply template and input into the target dialogue.

[0064] In an exemplary embodiment, the method further includes: upon receiving a sparring instruction, determining a business scenario included in the sparring instruction; obtaining a sparring dialogue corresponding to the business scenario; inputting third interaction data included in the sparring dialogue into a dialog box; identifying reply data in the dialog box for the third interaction data; determining a target correlation between the reply data and the third interaction data; when the target correlation is greater than or equal to a second preset threshold, inputting fourth interaction data included in the sparring dialogue into the dialog box; when the target correlation is less than a second preset threshold, adjusting the third interaction data to obtain fifth interaction data, wherein the intent of the fifth interaction data is the same as the intent of the third interaction data, and inputting the fifth interaction data into the dialog box.

[0065] In the above embodiment, the intelligent training module in the CRM customer relationship management system can provide a training function, such as Fig.10 As shown, Fig.10 It is an architecture diagram of the business-side intelligent platform according to an embodiment of the present invention. When receiving a training instruction, different business scenarios can be selected according to demand, wherein the business scenario can be a complaint scenario, a consultation scenario, or an after-sales scenario, etc. The intelligent training module can obtain the training dialogue corresponding to the selected business scenario, and input the third interaction data included in the training dialogue into the dialog box. By identifying the reply data for the third interaction data in the dialog box, the target correlation between the reply data and the third interaction data can be determined. When the target correlation is greater than or equal to the second preset threshold, it means that the training result is relatively accurate, and the interaction of the current data can be ended, and the interaction of the next data (i.e., the fourth interaction data mentioned above) can be carried out; when the target correlation is less than the second preset threshold, the third interaction data is adjusted to the fifth interaction data with the same intention as the third interaction data to continue the interaction.

[0066] In an exemplary embodiment, after inputting the first interaction data into a target network model to determine a target first intent included in the first interaction data through the target network model, the method further includes: determining business information associated with the first intent; displaying the business information in a target interface; upon receiving a click instruction, sending associated data of the first business information indicated by the click instruction to the target conversation; upon receiving a view instruction, displaying associated data of the second business information indicated by the view instruction in the target interface.

[0067] In the above embodiment, after identifying the first intention of the target object, the intelligent assistant module can determine all business information related to the first intention and display all relevant business information in the target interface, wherein the relevant business information may include the intention of the target object, keyword information, etc. When the management system receives a click instruction, it can send the associated data of the business information indicated by the click instruction (i.e., the above-mentioned first business information) to the target dialogue; when receiving a view instruction, it can display the associated data of the second business information indicated by the view instruction in the target interface.

[0068] In an exemplary embodiment, the client-side intelligent platform CRM customer relationship management system architecture diagram can be found in Fig.11 ,like Fig.11As shown in the figure, the CRM customer relationship management system can be understood as a customer information management platform, including modules such as identity information, business processing information, and customer tags. It can trigger the response module to execute the business process, and can also collect information reported by each module, including: text customer service module, CC call center, voiceprint recognition module, image recognition module, PS recognition module, and face recognition module. Among them, the text customer service module can interact with the target object in the form of text and update the data of the target object in a timely manner; the CC call center includes active inbound calls from the target object, such as IVR (Interactive Voice Response) navigation, and active outbound calls from the management system, such as voice for collection, return visits, and marketing, etc., and can establish voice calls with the target object through ASR, TTS, NLP and other technologies to build models; the voiceprint recognition module, face recognition module, image recognition module, and PS recognition module can detect the authenticity of the target user's ID card information during the information registration process. If it is identified as forged, the suspicious label of the target object can be recorded in the customer management relationship system; the face recognition module can also be used to confirm the identity of the target user when operating the interface to handle business.

[0069] In an exemplary embodiment, the CRM customer relationship management system in the business side intelligent platform may also include a quality inspection module, such as Fig.10 As shown, the quality inspection module can save the text and voice conversation records of the target object, and can check and evaluate the quality of service provided to customers in the past through key items such as voice intonation and the use of polite language.

[0070] Through the description of the above implementation methods, those skilled in the art can clearly understand that the method according to the above embodiment can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0071] In the present embodiment, a device for identifying intention is also provided, and the device is used to implement the above-mentioned embodiments and preferred embodiments, and the descriptions that have been made will not be repeated. As used below, the term "module" can implement a combination of software and / or hardware of a predetermined function. Although the device described in the following embodiments is preferably implemented in software, the implementation of hardware, or a combination of software and hardware is also possible and conceived.

[0072] Fig.12 is a structural block diagram of an apparatus for identifying intent according to an embodiment of the present invention, such as Fig.12 As shown, the device includes

[0073] A first determining module 1202, configured to determine first interaction data input by a first object included in a target dialogue;

[0074] A second determination module 1204, configured to input the first interaction data into a target network model, so as to determine a first intent included in the first interaction data through the target network model;

[0075] The target network model is trained in the following way:

[0076] The initial model is trained using training data to obtain a first network model;

[0077] Determining negative feedback data included in the historical conversation, wherein the intention of the negative feedback data identified by the first network model is different from the intention expressed by the object that input the negative feedback data;

[0078] The first network model is trained by using the negative feedback data to obtain the target network model.

[0079] In an exemplary embodiment, the device can implement the determination of negative feedback data included in the historical conversation in the following manner: using the first network model to identify second interaction data input by a second object in the historical conversation to obtain a first recognition result; when the first recognition result indicates that the first network model has recognized the same intention at least twice in a row, marking the second interaction data as abnormal data; identifying the abnormal data to obtain a second recognition result; when the second recognition result is consistent with the first recognition result, determining the abnormal data as negative feedback data.

[0080] In an exemplary embodiment, the device can implement the determination of negative feedback data included in the historical conversation in the following manner: determine other data included in the historical conversation except abnormal data, wherein the other data is data input by the second object; perform random detection on the other data to obtain a first detection result; when the first detection result is inconsistent with the recognition result of the first network model, determine the other data as the negative feedback data.

[0081] In an exemplary embodiment, the device can implement the determination of negative feedback data included in the historical conversation in the following manner: for any target historical conversation included in the historical conversation, the following operations are performed: when the same response data appears at least twice in a row in the target historical conversation, the question data corresponding to the response data is determined as the negative feedback data.

[0082] In an exemplary embodiment, the device can input the first interaction data into a target network model in the following manner to determine the first intent included in the first interaction data through the target network model: input the first interaction data into multiple sub-models included in the target network model respectively to determine a target score obtained by each sub-model scoring the first interaction data, wherein one sub-model corresponds to one intent, each sub-model corresponds to a different intent, and the target score is used to indicate the correlation between the first interaction data and the intent corresponding to the sub-model; determine the highest score included in the target score; and when the highest score is greater than a preset threshold, determine the intent corresponding to the highest score as the first intent.

[0083] In an exemplary embodiment, the device can be used to, after determining the highest score included in the target score: when the highest score is less than or equal to a preset threshold, build a target sub-model based on the first interaction data; and add the target sub-model to the target network model.

[0084] In an exemplary embodiment, the device can also be used to, after inputting the first interaction data into a target network model to determine the first intent included in the first interaction data through the target network model: determine a reply template corresponding to the first intent; generate reply data based on the reply template; and output the reply data.

[0085] In an exemplary embodiment, the device can also be used to: upon receiving a sparring instruction, determine the business scenario included in the sparring instruction; obtain a sparring dialogue corresponding to the business scenario; input the third interaction data included in the sparring dialogue into a dialog box; identify reply data for the third interaction data in the dialog box; determine a target correlation between the reply data and the third interaction data; when the target correlation is greater than or equal to a second preset threshold, input the fourth interaction data included in the sparring dialogue into the dialog box; when the target correlation is less than the second preset threshold, adjust the third interaction data to obtain fifth interaction data, wherein the intention of the fifth interaction data is the same as the intention of the third interaction data, and input the fifth interaction data into the dialog box.

[0086] In an exemplary embodiment, the device can also be used to, after inputting the first interaction data into a target network model to determine the target intent included in the first interaction data through the target network model: determine the business information associated with the first intent; display the business information in the target interface; when receiving a click instruction, send the associated data of the first business information indicated by the click instruction to the target dialogue; when receiving a view instruction, display the associated data of the second business information indicated by the view instruction in the target interface.

[0087] It should be noted that the above modules can be implemented by software or hardware. For the latter, it can be implemented in the following ways, but not limited to: the above modules are all located in the same processor; or the above modules are located in different processors in any combination.

[0088] An embodiment of the present invention further provides a computer-readable storage medium, in which a computer program is stored, wherein the computer program is configured to execute the steps of any of the above method embodiments when running.

[0089] In an exemplary embodiment, the computer-readable storage medium may include, but is not limited to, various media that can store computer programs, such as a USB flash drive, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk.

[0090] An embodiment of the present invention further provides an electronic device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor is configured to run the computer program to execute the steps in any one of the above method embodiments.

[0091] In an exemplary embodiment, the electronic device may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0092] An embodiment of the present invention further provides a computer program product, including a computer program, which implements the steps of the method in each embodiment of the present application when the computer program is executed by a processor.

[0093] For specific examples in this embodiment, reference may be made to the examples described in the above embodiments and exemplary implementation modes, and this embodiment will not be described in detail herein.

[0094] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0095] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for identifying intention, characterized in that: include: Determine first interaction data input by a first object included in the target conversation; Inputting the first interaction data into a target network model to determine a first intent included in the first interaction data through the target network model; The target network model is trained in the following way: The initial model is trained using training data to obtain a first network model; Determining negative feedback data included in the historical conversation, wherein the intention of the negative feedback data identified by the first network model is different from the intention expressed by the object that input the negative feedback data; The first network model is trained by using the negative feedback data to obtain the target network model.

2. The method according to claim 1, characterized in that The determining of the negative feedback data included in the historical conversation includes: Using the first network model to recognize second interaction data input by a second object in the historical conversation, to obtain a first recognition result; If the first recognition result indicates that the first network model recognizes the same intent at least twice in a row, marking the second interaction data as abnormal data; Identifying the abnormal data to obtain a second identification result; When the second recognition result is consistent with the first recognition result, the abnormal data is determined as negative feedback data.

3. The method according to claim 1, characterized in that The negative feedback data included in the determination of the historical dialogue includes: Determine other data included in the historical conversation except the abnormal data, wherein the other data is data input by the second object; Performing random detection on the other data to obtain a first detection result; When the first detection result is inconsistent with the recognition result of the first network model, the other data is determined as the negative feedback data.

4. The method according to claim 1, characterized in that: The negative feedback data included in the determination of the historical dialogue includes: For any target historical dialogue included in the historical dialogues, the following operations are performed: when the same answer data appears at least twice in succession in the target historical dialogue, the question data corresponding to the answer data is determined as the negative feedback data.

5. The method according to claim 1, characterized in that Inputting the first interaction data into a target network model to determine, through the target network model, a first intent included in the first interaction data includes: Inputting the first interaction data into a plurality of sub-models included in the target network model respectively to determine a target score obtained by scoring the first interaction data by each sub-model, wherein one sub-model corresponds to one intent, each sub-model corresponds to a different intent, and the target score is used to indicate the degree of association between the first interaction data and the intent corresponding to the sub-model; determining a maximum score included in the target score; When the highest score is greater than a preset threshold, the intention corresponding to the highest score is determined as the first intention.

6. The method according to claim 5, characterized in that After determining the highest score included in the target score, the method further includes: When the highest score is less than or equal to a preset threshold, constructing a target sub-model based on the first interaction data; The target sub-model is added to the target network model.

7. The method according to claim 1, characterized in that After inputting the first interaction data into a target network model to determine a first intent included in the first interaction data through the target network model, the method further includes: Determine a response template corresponding to the first intention; generating reply data based on the reply template; The reply data is output.

8. The method according to claim 1, characterized in that The method further comprises: In the case of receiving a training instruction, determining a business scenario included in the training instruction; Obtain a training dialogue corresponding to the business scenario; Inputting the third interaction data included in the sparring dialogue into the dialog box; identifying reply data in the dialog box to the third interaction data; determining a target correlation between the reply data and the third interaction data; In a case where the target relevance is greater than or equal to a second preset threshold, inputting fourth interaction data included in the sparring dialogue into the dialog box; When the target association degree is less than the second preset threshold, the third interaction data is adjusted to obtain fifth interaction data, wherein the intention of the fifth interaction data is the same as the intention of the third interaction data, and the fifth interaction data is input into the dialog box.

9. The method according to claim 1, characterized in that: After inputting the first interaction data into a target network model to determine a target intent first intent included in the first interaction data through the target network model, the method further includes: Determining business information associated with the first intention; Displaying the business information in the target interface; When receiving the click instruction, sending the associated data of the first business information indicated by the click instruction to the target dialogue; When the viewing instruction is received, the associated data of the second business information indicated by the viewing instruction is displayed in the target interface.

10. An intention recognition device, characterized in that: include: A first determining module, used to determine first interaction data input by a first object included in a target dialogue; a second determination module, configured to input the first interaction data into a target network model, so as to determine a first intent included in the first interaction data through the target network model; The target network model is trained in the following way: The initial model is trained using training data to obtain a first network model; Determining negative feedback data included in the historical conversation, wherein the intention of the negative feedback data identified by the first network model is different from the intention expressed by the object that input the negative feedback data; The first network model is trained by using the negative feedback data to obtain the target network model.

11. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 9 when executed.

12. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 9.

13. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method described in any one of claims 1 to 9 are implemented.

Images