Method and device for performing intention recognition on chat content and storage medium

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a chat content intention recognition method and device and a storage medium, and the method comprises the steps: screening pre-selected chat content based on a preset authority parameter, obtaining target chat content, carrying out the intention recognition of the target chat content through an intention recognition funnel matrix, and obtaining the target chat content; the intention recognition funnel matrix comprises a plurality of upper-level intention recognition methods and a plurality of lower-level intention recognition methods which are arranged in a cascading manner in sequence, and a target intention of the target chat content is determined based on intention confirmation information of the target client, the intention confirmation information is a reply of the target client based on the received intention recommendation information, the intention recommendation information comprises at least two pre-selected intentions, and the method for determining the target intention by using the intention recognition funnel matrix and the intention recommendation information effectively improves the intention recognition accuracy.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and provides a method, device and storage medium for intent recognition of chat content. Background Art

[0002] At present, social chat software has become an essential chat tool in people's lives. People can use social chat software for group chats or one-on-one chats, and correspondingly, group chat content and one-on-one chat content will be generated. Especially in the group chat scenario, the group chat content is numerous and complex. Usually, the group chat content will span multiple fields and scenarios. Correspondingly, it brings technical challenges to the intent recognition of chat content, resulting in poor accuracy of intent recognition. Summary of the Invention

[0003] Embodiments of this application provide a method, device and storage medium for intent recognition of chat content to improve the accuracy of intent recognition of chat content.

[0004] The specific technical solutions provided by this application are as follows:

[0005] In a first aspect, embodiments of this application provide a method for intent recognition of chat content, including:

[0006] Screening preselected chat content based on preset permission parameters to obtain target chat content, where the preselected chat content includes group chat content and one-on-one chat content of a target customer;

[0007] Using an intent recognition funnel matrix to perform intent recognition on the target chat content to obtain at least one preselected intent, where the intent recognition funnel matrix includes a plurality of cascaded upper-level intent recognition methods and lower-level intent recognition methods in sequence. If the content of the preselected intent obtained after the upper-level intent recognition method performs intent recognition on the target chat content is empty, then use the lower-level intent recognition method to perform intent recognition on the target chat content;

[0008] Determining the target intent of the target chat content based on the intent confirmation information of the target customer, where the intent confirmation information is the reply of the target customer based on the received intent recommendation information, and the intent recommendation information includes at least two preselected intents.

[0009] Optionally, after using the intent recognition funnel matrix to perform intent recognition on the target chat content to obtain at least one preselected intent, it further includes:

[0010] If the number of preselected intents is one, then determine the preselected intent as the target intent of the target chat content, and execute an intent business strategy based on the target intent.

[0011] Optionally, the preset permission parameters include scenario permissions, domain permissions, time-limit permissions, and role permissions. Based on the preset permission parameters, the preselected chat content is filtered to obtain the target chat content, including:

[0012] For each piece of preselected chat content, perform the following operations:

[0013] Extract and / or combine keywords from the preselected chat content to obtain multiple chat content blocks;

[0014] Use a sliding window to intercept the content of each chat content block to obtain multiple valid chat contents;

[0015] Perform scenario recognition, domain recognition, time-limit recognition, and role recognition on each valid chat content respectively to obtain the target scenario, target domain, target time-limit, and target role;

[0016] If the target scenario does not match the current scenario, delete the preselected chat content; or

[0017] If the target domain does not match the current domain, delete the preselected chat content; or

[0018] If the target time-limit does not match the current time range, delete the preselected chat content; or

[0019] If the target role does not match the current role, delete the preselected chat content;

[0020] Among them, the current scenario, current domain, current time range, and current role are all determined based on the historical associated content of the preselected chat content.

[0021] Optionally, use an intention recognition funnel matrix to perform intention recognition on the target chat content to obtain at least one preselected intention, including:

[0022] Input the target chat content into the intention recognition method at the first level of the intention recognition funnel matrix for intention recognition;

[0023] If the content of the intention recognition result is not empty, use the intention recognition result as at least one preselected intention; or

[0024] If the content of the intention recognition result is empty, input the target chat content into the next-level intention recognition method, and use the next-level intention recognition method to perform intention recognition on the target chat content until the content of the obtained intention recognition result is not empty.

[0025] Optionally, the intention recognition methods include: keyword matching method model, sentence matching method model, supervised model, and unsupervised model. The intention recognition funnel matrix is constructed in the following way:

[0026] Connect the input end of the intention recognition funnel matrix to the input end of the keyword matching method model, and connect the output end of the keyword matching method model to the output end of the intention recognition funnel matrix;

[0027] Connect the output end of the keyword matching method model to the input end of the sentence matching method model, and connect the output end of the sentence matching method model to the output end of the intention recognition funnel matrix;

[0028] Connect the output end of the sentence matching method model to the input end of the supervised model, and connect the output end of the supervised model to the output end of the intention recognition funnel matrix;

[0029] Connect the output end of the supervised model to the input end of the unsupervised model, and connect the output end of the unsupervised model to the output end of the intention recognition funnel matrix.

[0030] Optionally, if the intention recognition method is an unsupervised model, the target chat content is recognized for intention in the following manner:

[0031] Determine an intention association chain based on a pre-established knowledge graph, where the intention association chain includes intention keywords and the parent and child nodes associated with the intention keywords. The intention keywords are the knowledge points corresponding to the target chat content in the knowledge graph, the parent node is the upper-level knowledge point connected to the knowledge point in the knowledge graph, and the child node is the lower-level knowledge point connected to the knowledge point in the knowledge graph;

[0032] Use an unsupervised model with constraints to recognize the intention of the target chat content, where the constraint is the intention association chain.

[0033] Optionally, determine the target intention of the target chat content based on the intention confirmation information of the target customer, including:

[0034] If the number of preselected intentions is at least two, generate intention recommendation information based on the at least two preselected intentions;

[0035] Send the intention recommendation information to the target customer so that the target customer can screen based on the received intention recommendation information and obtain intention confirmation information based on at least one preselected intention selected;

[0036] Determine at least one preselected intention included in the intention confirmation information as the target intention of the target chat content, and execute an intention business strategy based on the target intention.

[0037] In a second aspect, an embodiment of the present application further provides a device for recognizing the intention of chat content, including:

[0038] A screening unit for screening preselected chat content based on preset permission parameters to obtain target chat content, where the preselected chat content includes group chat content and one-on-one chat content of the target customer;

[0039] An identification unit for identifying the intent of the target chat content by using an intent identification funnel matrix to obtain at least one preselected intent, where the intent identification funnel matrix includes a plurality of cascaded upper-level intent identification methods and lower-level intent identification methods in sequence. If the content of the preselected intent obtained by the upper-level intent identification method for identifying the intent of the target chat content is empty, then the lower-level intent identification method is used to identify the intent of the target chat content;

[0040] A determination unit for determining the target intent of the target chat content based on the intent confirmation information of the target customer, where the intent confirmation information is the reply of the target customer based on the received intent recommendation information, and the intent recommendation information includes at least two preselected intents.

[0041] Optionally, after using the intent identification funnel matrix to identify the intent of the target chat content and obtaining at least one preselected intent, it further includes:

[0042] If the number of preselected intents is one, then the preselected intent is determined as the target intent of the target chat content, and the intent business strategy is executed based on the target intent.

[0043] Optionally, the preset permission parameters include scenario permissions, domain permissions, time-limit permissions, and role permissions. The screening unit is used to screen the preselected chat content based on the preset permission parameters to obtain the target chat content, and the screening unit is used for:

[0044] Perform the following operations for each piece of preselected chat content:

[0045] Extract and / or merge keywords from the preselected chat content to obtain multiple chat content blocks;

[0046] Use a sliding window to intercept the content of each chat content block respectively to obtain multiple valid chat contents;

[0047] Perform scenario recognition, domain recognition, time-limit recognition, and role recognition on each valid chat content respectively to obtain the target scenario, target domain, target time-limit, and target role;

[0048] If the target scenario does not match the current scenario, then delete the preselected chat content; or

[0049] If the target domain does not match the current domain, then delete the preselected chat content; or

[0050] If the target time-limit does not match the current time range, then delete the preselected chat content; or

[0051] If the target role does not match the current role, delete the preselected chat content;

[0052] Among them, the current scene, current field, current duration range, and current role are all determined based on the historical associated content of the preselected chat content.

[0053] Optionally, use an intention recognition funnel matrix to perform intention recognition on the target chat content to obtain at least one preselected intention. The recognition unit is used for:

[0054] Input the target chat content into the intention recognition method at the first level of the intention recognition funnel matrix for intention recognition;

[0055] If the content of the intention recognition result is not empty, use the intention recognition result as at least one preselected intention; or

[0056] If the content of the intention recognition result is empty, input the target chat content into the next-level intention recognition method and use the next-level intention recognition method to perform intention recognition on the target chat content until the content of the obtained intention recognition result is not empty.

[0057] Optionally, the intention recognition methods include: keyword matching method model, sentence matching method model, supervised model, and unsupervised model. The intention recognition funnel matrix is constructed in the following way:

[0058] Connect the input end of the intention recognition funnel matrix to the input end of the keyword matching method model, and connect the output end of the keyword matching method model to the output end of the intention recognition funnel matrix;

[0059] Connect the output end of the keyword matching method model to the input end of the sentence matching method model, and connect the output end of the sentence matching method model to the output end of the intention recognition funnel matrix;

[0060] Connect the output end of the sentence matching method model to the input end of the supervised model, and connect the output end of the supervised model to the output end of the intention recognition funnel matrix;

[0061] Connect the output end of the supervised model to the input end of the unsupervised model, and connect the output end of the unsupervised model to the output end of the intention recognition funnel matrix.

[0062] Optionally, if the intention recognition method is an unsupervised model, the target chat content is intention-recognized in the following way:

[0063] Determine an intent association chain based on a pre-established knowledge graph, where the intent association chain includes intent keywords and the parent and child nodes associated with the intent keywords. The intent keywords are the knowledge points corresponding to the target chat content in the knowledge graph, the parent nodes are the upper-level knowledge points connected to the knowledge points in the knowledge graph, and the child nodes are the lower-level knowledge points connected to the knowledge points in the knowledge graph;

[0064] Use an unsupervised model with constraints to perform intent recognition on the target chat content, where the constraint is the intent association chain.

[0065] Optionally, determine the target intent of the target chat content based on the intent confirmation information of the target customer. The determination unit is used for:

[0066] If the number of preselected intents is at least two, generate intent recommendation information based on the at least two preselected intents;

[0067] Send the intent recommendation information to the target customer so that the target customer can screen based on the received intent recommendation information and obtain intent confirmation information based on at least one of the preselected intents;

[0068] Determine at least one of the preselected intents included in the intent confirmation information as the target intent of the target chat content, and execute the intent business strategy based on the target intent.

[0069] In a third aspect, a computing device includes:

[0070] A memory for storing executable instructions;

[0071] A processor for reading and executing the executable instructions stored in the memory to implement the method according to any one of the first aspect.

[0072] In a fourth aspect, a computer-readable storage medium, when the instructions in the storage medium are executed by a processor, enable the processor to execute the method according to any one of the first aspect.

[0073] The beneficial effects of this application are as follows:

[0074] In summary, in the embodiments of the present application, a method, apparatus, and storage medium for intent recognition of chat content are provided. The method includes: screening preselected chat content based on preset permission parameters to obtain target chat content, where the preselected chat content includes group chat content and one-on-one chat content of the target customer; using an intent recognition funnel matrix to perform intent recognition on the target chat content to obtain at least one preselected intent, where the intent recognition funnel matrix includes a plurality of cascaded upper-level intent recognition methods and lower-level intent recognition methods in sequence. If the content of the preselected intent obtained by the upper-level intent recognition method for intent recognition of the target chat content is empty, then the lower-level intent recognition method is used to perform intent recognition on the target chat content; determining the target intent of the target chat content based on the intent confirmation information of the target customer, where the intent confirmation information is the reply of the target customer based on the received intent recommendation information, and the intent recommendation information includes at least two preselected intents. The method of using the intent recognition funnel matrix and the intent recommendation information to determine the target intent effectively improves the accuracy of intent recognition.

[0075] Other features and advantages of the present application will be described in the following specification, and some will become apparent from the specification or be understood by implementing the present application. The objectives and other advantages of the present application can be achieved and obtained through the structures specifically pointed out in the written specification, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The schematic embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation of the present application. In the drawings:

[0077] Figure 1 It is a schematic diagram of the system architecture for intent recognition of chat content in the embodiments of the present application;

[0078] Figure 2 It is a flowchart of a method for intent recognition of chat content in the embodiments of the present application;

[0079] Figure 3 It is a flowchart of screening preselected chat content based on preset permission parameters to obtain target chat content in the embodiments of the present application;

[0080] Figure 4 It is a schematic diagram of the intent recognition funnel matrix in the embodiments of the present application;

[0081] Figure 5 It is a flowchart of using the intent recognition funnel matrix to perform intent recognition on the target chat content to obtain preselected intents in the embodiments of the present application;

[0082] Figure 6 Schematic diagram of the knowledge graph in the embodiment of the present application;

[0083] Figure 7 Schematic flowchart of determining the target intention based on the preselected intention in the embodiment of the present application;

[0084] Figure 8 Schematic flowchart of determining the target intention when the number of preselected intentions is at least two in the embodiment of the present application;

[0085] Figure 9 Schematic diagram of the logical architecture of a device for intent recognition of chat content in the embodiment of the present application;

[0086] Figure 10 Schematic diagram of the entity architecture of a computing device in the embodiment of the present application. Detailed implementation manners

[0087] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are some, but not all, of the embodiments of the technical solutions of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments recorded in this application document without creative efforts shall fall within the scope of protection of the technical solutions of the present application.

[0088] Terms such as "first" and "second" in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and do not necessarily need to describe a specific order or sequence. It should be understood that such used data may be interchanged under appropriate circumstances so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein.

[0089] The preferred embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0090] Refer to Figure 1 As shown, in the embodiment of the present application, the system includes at least one computing device. In the embodiment of the present application, the implementation of the method for intent recognition of chat content is mainly executed on the computing device side, that is, the preselected chat content is input into the computing device for intent recognition, and the specific introduction is as follows.

[0091] Refer to Figure 2 As shown, in the embodiment of the present application, a specific process for intent recognition of chat content is as follows:

[0092] Step 201: Screen the preselected chat content based on preset permission parameters to obtain target chat content, where the preselected chat content includes group chat content and one-on-one chat content of the target customer.

[0093] First of all, it should be noted that the chat content in the embodiments of the present application includes group chat content and one-on-one chat content. The above group chat content is the group where a certain target customer is located (including family groups, enterprise groups, etc.), and the above one-on-one chat content is the one-on-one chat content between a certain target customer and other customers. The above group chat and one-on-one chat can be carried out simultaneously or separately. In the embodiments of the present application, the above group chat content and one-on-one chat content are collectively referred to as preselected chat content.

[0094] Considering that the content of the preselected chat content is large and miscellaneous, it is necessary to screen the preselected chat content during the implementation process to obtain the target chat content that needs to be subject to intent recognition. Exemplarily, the preset permission parameters include scenario permissions, domain permissions, time limit permissions, and role permissions. Screen the preselected chat content based on the preset permission parameters to obtain the target chat content. Refer to Figure 3 As shown, it includes:

[0095] Perform the following operations on each piece of preselected chat content:

[0096] Considering that the preselected chat content is generated one by one, during the implementation process, each time a piece of preselected chat content is generated, the following operation processing is performed on it. When factors such as time limit need to be considered, usually the last piece of preselected chat content is used as the time benchmark. The following specifically introduces how to obtain the target chat content based on the preselected chat content.

[0097] Step 2011: Extract and / or merge keywords from the preselected chat content to obtain multiple chat content blocks.

[0098] In order to facilitate the recognition of the preselected chat content, in one embodiment, when the preselected chat content includes relatively rich chat content, that is, when the sentences of the preselected chat content are relatively long, extract keywords from the preselected chat content to obtain multiple chat content blocks. For example, cut and extract content from the preselected chat content according to the chat scenario, chat domain, chat time limit, and chat role of the target customer in the chat, so as to obtain multiple keywords.

[0099] In another embodiment, when the sentences of the preselected chat content are relatively short, it is also necessary to first merge the first N pieces of preselected chat content associated with this piece of preselected chat content, and then perform the above keyword extraction. The specific value of the above N needs to be set according to the actual usage scenario. The specific merging method can adopt rich text and long text processing, etc.

[0100] For example, when the preselected chat content is "from Beijing to Nanjing", it is necessary to combine the first N pieces of preselected chat content, namely "I want to depart from Beijing", "go to Nanjing", and "please help query ticket information". In this way, after merging the above preselected chat content and then extracting keywords, the multiple chat content blocks obtained include: query ticket information, tickets from Beijing to Nanjing.

[0101] Step 2012: Use a sliding window to intercept the content of each chat content block respectively to obtain multiple valid chat contents.

[0102] Considering that when the content of the preselected chat content is relatively large, the number of multiple chat content blocks obtained is also relatively large. Based on this, it is necessary to streamline the information of the above chat content blocks. During the implementation process, use a sliding window to intercept the content of each chat content block respectively. The specific interception method includes but is not limited to finding the boundaries of two adjacent chat content blocks. Between the two boundaries belonging to the same chat content block, adjust the window length of the sliding window, and use this sliding window to intercept the content of the chat content between the two boundaries, so as to obtain a valid chat content. After intercepting the content of each chat content block, multiple valid chat contents are obtained.

[0103] Step 2013: Perform scene recognition, domain recognition, timeliness recognition, and role recognition on each valid chat content respectively to obtain the target scene, target domain, target timeliness, and target role.

[0104] During the implementation process, after converting the preselected chat content into multiple valid chat contents, perform scene recognition on each valid chat content respectively to obtain the target scene. It should be noted that the above scene mainly refers to the conversation mode in the chat. Common scenes include group chat scenes and one-on-one chat scenes, etc. During the specific implementation process, after determining the target customer, according to the chat identifier and other information of the target customer, determine the identifiers of other customers corresponding to the preselected chat content, and determine whether the target scene is a group chat or a one-on-one chat according to the number of identifiers of other customers.

[0105] In addition, perform domain recognition on each valid chat content to obtain the target domain. During the specific implementation process, support vector machines, logistic regression models, decision trees, etc. can be used to perform domain recognition on each valid chat content, so as to obtain the target domain. It should be supplemented and explained that during the implementation process, it is also possible to directly compare whether the distance value between the semantic vector corresponding to the valid chat content and the semantic vector corresponding to the current domain is greater than the preset distance threshold, and based on this, determine the target domain corresponding to the valid chat content.

[0106] Moreover, perform timeliness recognition on each valid chat content to obtain the target timeliness. In the specific implementation process, use the valid chat content corresponding to the last preselected chat content as the time benchmark, and determine the timeliness between the time of each preselected chat content before the last preselected chat content and the above time benchmark, so as to obtain multiple target timeliness. It should be added that when the preselected chat content carries a time tag, the time interval between the time tags corresponding to the preselected chat content corresponding to two valid chat contents can be directly calculated to obtain multiple target timeliness.

[0107] Moreover, perform role recognition on each valid chat content to obtain the target role. In the specific implementation process, the determination of the target role is directly related to the target scenario. When the target scenario is a group chat scenario, role recognition needs to be performed on each customer involved in the group chat to obtain target role A, target role B, target role C, etc.; when the target scenario is a one-on-one chat scenario, usually only the target customer needs to be recognized for the role to obtain the target role.

[0108] Step 2014: If the target scenario does not match the current scenario, delete the preselected chat content. Or

[0109] If the target field does not match the current field, delete the preselected chat content. Or

[0110] If the target timeliness does not match the current duration range, delete the preselected chat content. Or

[0111] If the target role does not match the current role, delete the preselected chat content.

[0112] Among them, the current scenario, the current field, the current duration range, and the current role are all determined based on the historical associated content of the preselected chat content.

[0113] In the implementation process, in order to screen the preselected chat content according to the scenario permission, field permission, timeliness permission, and role permission, first determine the historical associated content of the preselected chat content. The above historical associated content includes a section of preselected chat content before the current preselected chat content, and the amount of content of the historical associated content is flexibly set according to the actual usage scenario.

[0114] After determining the historical associated content, use the same technical means as the above target scenario, target field, target timeliness, and target role to determine the current scenario, the current field, the current duration range, and the current role. The specific determination process will not be elaborated here one by one.

[0115] After determining the current scenario, compare whether the target scenario matches the current scenario. For example, if the current scenario is a group chat scenario, and the target scenario corresponding to the determined preselected chat content is a one-on-one chat scenario, then delete the preselected chat content.

[0116] After the current domain is determined, the target domain is compared to see if it matches the current domain. For example, if the current domain is the food domain, and the target domain corresponding to the determined pre-selected chat content is the sports domain, then the pre-selected chat content is deleted.

[0117] After determining the current time range, compare whether the target time limit is in line with the current time range. For example, if the current time range is 72 hours, and the target time limit corresponding to the pre-selected chat content is 100 hours ago, then the pre-selected chat content will be deleted.

[0118] After determining the current role, compare whether the target role matches the current role. For example, if the current role is role A in a group chat scenario, and the target role corresponding to the determined pre-selected chat content is role B in a single chat scenario, then the pre-selected chat content is deleted.

[0119] Step 202: Use the intent recognition funnel matrix to perform intent recognition on the target chat content to obtain at least one pre-selected intent, wherein the intent recognition funnel matrix includes multiple upper-level intent recognition methods and lower-level intent recognition methods that are cascaded in sequence. If the content of the pre-selected intent obtained after the upper-level intent recognition method performs intent recognition on the target chat content is empty, the lower-level intent recognition method is used to perform intent recognition on the target chat content.

[0120] First, the intent recognition funnel matrix in the embodiment of the present application is introduced. The intent recognition method includes: a keyword matching method model, a sentence matching method model, a supervised model and an unsupervised model. The intent recognition funnel matrix is ​​constructed in the following way. The internal structure of the constructed intent recognition funnel matrix is ​​shown in FIG. Figure 4 As shown:

[0121] (1) Connect the input end of the intention recognition funnel matrix to the input end of the keyword matching method model, and connect the output end of the keyword matching method model to the output end of the intention recognition funnel matrix.

[0122] In an embodiment of the present application, the intention recognition method at the first level in the intention recognition funnel matrix is ​​a keyword matching method, that is, the input end of the intention recognition funnel matrix is ​​first connected to the keyword matching method model, so that the target chat content is input through the input end of the intention recognition funnel matrix, and then first input into the keyword matching method model for intent recognition.

[0123] If the content of the preselected intention obtained after the keyword matching method model performs intention recognition on the target chat content is not empty, the obtained preselected intention is output through the output end of the intention recognition funnel matrix; if the content of the preselected intention obtained after the keyword matching method model performs intention recognition on the target chat content is empty, the target chat content is continuously input into the intention recognition method at the next level in the intention recognition funnel matrix for intention recognition.

[0124] (2) Connect the output end of the keyword matching method model to the input end of the sentence matching method model, and connect the output end of the sentence matching method model to the output end of the intention recognition funnel matrix.

[0125] In the embodiment of the present application, the intention recognition method at the second level in the intention recognition funnel matrix is the sentence matching method, that is, the output end of the keyword matching method model is connected to the input end of the sentence matching method model. In this way, after the target chat content passes through the keyword matching method model, it is input into the sentence matching method model for intention recognition.

[0126] If the content of the preselected intention obtained after the sentence matching method model performs intention recognition on the target chat content is not empty, the obtained preselected intention is output through the output end of the intention recognition funnel matrix; if the content of the preselected intention obtained after the sentence matching method model performs intention recognition on the target chat content is empty, the target chat content is continuously input into the intention recognition method at the next level in the intention recognition funnel matrix for intention recognition.

[0127] (3) Connect the output end of the sentence matching method model to the input end of the supervised model, and connect the output end of the supervised model to the output end of the intention recognition funnel matrix.

[0128] In the embodiment of the present application, the intention recognition method at the third level in the intention recognition funnel matrix is the supervised model, that is, the output end of the sentence matching method model is connected to the input end of the supervised model. In this way, after the target chat content passes through the keyword matching method model and the sentence matching method model, it is input into the supervised model for intention recognition.

[0129] If the content of the preselected intention obtained after the supervised model performs intention recognition on the target chat content is not empty, the obtained preselected intention is output through the output end of the intention recognition funnel matrix; if the content of the preselected intention obtained after the supervised model performs intention recognition on the target chat content is empty, the target chat content is continuously input into the intention recognition method at the next level in the intention recognition funnel matrix for intention recognition.

[0130] (4) Connect the output of the supervised model to the input of the unsupervised model, and connect the output of the unsupervised model to the output of the intent recognition funnel matrix.

[0131] In an embodiment of the present application, the intent recognition method at the fourth level in the intent recognition funnel matrix is ​​an unsupervised model, that is, the output end of the supervised model is connected to the input end of the unsupervised model, so that the target chat content is input into the unsupervised model for intent recognition after passing through the keyword matching method model, the sentence matching method model and the supervised model.

[0132] If the content of the pre-selected intent obtained after the unsupervised model performs intent recognition on the target chat content is not empty, the pre-selected intent obtained above is output through the output end of the intent recognition funnel matrix; if the content of the pre-selected intent obtained after the unsupervised model performs intent recognition on the target chat content is empty, the target chat content continues to be input into the next level intent recognition method in the intent recognition funnel matrix for intent recognition.

[0133] It should be noted that the above-mentioned intention recognition funnel matrix can also include other intention recognition methods. The connection relationship between the method models corresponding to other intention recognition methods in the intention recognition funnel matrix is ​​the same as the connection method between the above-mentioned keyword matching method model, sentence matching method model, supervised model and unsupervised model, which will not be repeated here one by one.

[0134] Exemplarily, the intent recognition funnel matrix is ​​used to perform intent recognition on the target chat content to obtain at least one pre-selected intent, see Figure 5 As shown, including:

[0135] Step 2021: Input the target chat content into the intent recognition method at the first level of the intent recognition funnel matrix for intent recognition. If the content of the intent recognition result is not empty, execute step 2022; if the content of the intent recognition result is empty, execute step 2023.

[0136] During the implementation process, after constructing the intent recognition funnel matrix, the target chat content is input into the intent recognition method at the first level of the intent recognition funnel matrix for intent recognition, that is, the keyword matching method model is used to perform intent recognition on the target chat content.

[0137] Exemplarily, a keyword template composed of multiple keywords is pre-established in the keyword matching method model. For example, the diet keyword template includes: I want to eat, Yu-Shiang Shredded Pork, Kung Pao Chicken, Cola Chicken Wings; the sports keyword template includes: table tennis, badminton, basketball, football, etc. While establishing the keyword template, a weight value is also pre-assigned to each keyword in the template. During the implementation process, first determine the corresponding keyword template according to the target field of the target chat content. After extracting multiple target keywords from the target chat content, match each target keyword with the keywords in the keyword template and calculate the weight scores. When the total score obtained in this way is higher than the preset score threshold, the intention recognition result is output; when the total score obtained is lower than the preset score threshold, the content of the intention recognition result is set to be empty.

[0138] Step 2022: Use the intention recognition result as at least one preselected intention. Or

[0139] During the implementation process, considering that the content of the target chat content is relatively long, the corresponding intention recognition result may be more than one. In this case, determine whether the content of the intention recognition result is empty. If the judgment result is not empty, use the intention recognition result as the preselected intention; if the judgment result is empty, input the target chat content into the next-level intention recognition method.

[0140] Step 2023: Input the target chat content into the next-level intention recognition method, and use the next-level intention recognition method to perform intention recognition on the target chat content until the content of the obtained intention recognition result is not empty.

[0141] During the implementation process, when the intention recognition result obtained by the previous-level intention recognition method is empty, continue to input the target chat content into the next-level intention recognition method, and use the next-level intention recognition method to perform intention recognition on the target chat content.

[0142] Case 1: The previous-level intention recognition method is the keyword matching method model, and the next-level intention recognition method is the sentence matching method model.

[0143] During the implementation process, use the sentence matching method model to perform intention recognition on the target chat content.

[0144] Exemplarily, a semantic vector composed of multiple sentences is pre-established in the sentence matching method model. During the implementation process, first convert each chat content of the target chat content into a target semantic vector, and compare the vector distance between each target semantic vector and each semantic vector in the model one by one. When the obtained vector distance is less than the preset distance threshold, use the intention recognition result as the preselected intention; when the obtained vector distance is greater than the preset distance threshold, set the content of the intention recognition result to be empty.

[0145] Case 2: The intention recognition method at the upper level is a sentence matching method model, and the intention recognition method at the lower level is a supervised model.

[0146] During the implementation process, the supervised model is used to recognize the intention of the target chat content.

[0147] Exemplarily, the supervised model can be pre-trained in advance with a bert model or other models. During the implementation process, each piece of chat content of the target chat content is input into the supervised model for learning. The output obtained in this way is the intention name and the corresponding confidence level. If the above confidence level is greater than the preset confidence level threshold, the intention recognition result is used as the preselected intention; if the above confidence level is less than the preset confidence level threshold, the content of the intention recognition result is set to be empty.

[0148] Case 3: The intention recognition method at the upper level is a supervised model, and the intention recognition method at the lower level is an unsupervised model.

[0149] During the implementation process, the unsupervised model is used to recognize the intention of the target chat content.

[0150] Exemplarily, the unsupervised model can be pre-trained in advance with a language large model or other models. During the implementation process, each piece of chat content of the target chat content is input into the unsupervised model for learning. The output obtained in this way is the intention name and the corresponding confidence level. If the above confidence level is greater than the preset confidence level threshold, the intention recognition result is used as the preselected intention; if the above confidence level is less than the preset confidence level threshold, the content of the intention recognition result is set to be empty.

[0151] It should be noted that in the embodiments of the present application, in order to make the output of the unsupervised model more accurate, a knowledge graph is also used to add constraint conditions to the unsupervised model. If the intention recognition method is an unsupervised model, the intention of the target chat content is recognized in the following way:

[0152] 1) Determine the intention association chain based on the pre-established knowledge graph. Among them, the intention association chain includes intention keywords and the parent node and child node associated with the intention keyword. The intention keyword is the knowledge point corresponding to the target chat content in the knowledge graph. The parent node is the upper-level knowledge point connected to the knowledge point in the knowledge graph, and the child node is the lower-level knowledge point connected to the knowledge point in the knowledge graph.

[0153] In the embodiments of the present application, multiple knowledge graphs can be pre-established according to historical association content. Refer to Figure 6As shown in the figure, it is a knowledge graph of a common family group chat mode. This knowledge graph includes multiple levels of parent nodes and multiple levels of child nodes. Exemplarily, the first-level knowledge points are established according to the fields of historical associated content: gas payment business, mental health business, diet business, insurance business, and other businesses, etc. The second-level knowledge points are established according to the first-level intentions of historical associated content. For example, the second-level nodes associated with the mental health business include: disease Q&A, emotion Q&A, and health care Q&A, etc. And, the third-level knowledge points are established according to the second-level intentions of historical associated content. For example, the third-level knowledge points associated with disease Q&A include: symptom clarification, disease clarification, drug clarification, and plan clarification, etc. By analogy, the next-level knowledge points can be continuously established in the knowledge graph. It should be noted that for a certain knowledge point, that is, a node, in the knowledge graph, the upper-level knowledge point connected to it is the parent node, and the lower-level knowledge point connected to it is the child node.

[0154] During the implementation process, after extracting the keyword of the target chat content, the intention keyword is obtained. Then, the knowledge point matching the intention keyword is searched in the knowledge graph. Further, according to the found matching knowledge point, the upper-level knowledge point connected to it is found. The upper-level knowledge point is called the parent node. And, according to the found matching knowledge point, the lower-level knowledge point connected to it is found. The lower-level knowledge point is called the child node. Then, the above-mentioned intention keyword, parent node, and child node are determined as an intention association chain.

[0155] 2) Use an unsupervised model with constraints to perform intention recognition on the target chat content, where the constraint condition is the intention association chain.

[0156] After determining the intention association chain, add the above intention association chain to the unsupervised model as the constraint condition of the unsupervised model. During the implementation process, when the target chat content is input into the unsupervised model, use the unsupervised model with constraints to perform intention recognition on the target chat content, and the intention recognition result can be obtained. If the content of the intention recognition result is not empty, then use the intention recognition result as the preselected intention; if the content of the intention recognition result is empty, considering that the unsupervised model is the intention recognition method at the last level in the intention recognition funnel matrix, then set the content of the preselected intention to be empty.

[0157] Step 203: Determine the target intention of the target chat content based on the intention confirmation information of the target customer, where the intention confirmation information is the reply of the target customer based on the received intention recommendation information, and the intention recommendation information includes at least two preselected intentions.

[0158] Considering that the content of the target chat is long and complex, after using the intention recognition funnel matrix to recognize the intention of the target chat content, the number of preselected intentions obtained may be multiple. Based on this, the following two cases are described.

[0159] The first case: The number of preselected intentions is one. After using the intention recognition funnel matrix to recognize the intention of the target chat content and obtaining at least one preselected intention, refer to Figure 7 as shown, it also includes:

[0160] Step 2031: If the number of preselected intentions is one, then determine the preselected intention as the target intention of the target chat content, and execute the intention business strategy based on the target intention.

[0161] During the implementation process, when the number of preselected intentions output by the intention recognition funnel matrix is one, that is, when the intention of the target chat content is clear and unique, determine the above preselected intention as the target intention of the target chat content, and then determine the corresponding intention business strategy according to the target intention, and continue to execute the corresponding intention business strategy. For example, when the target intention is gas payment, continue to execute the corresponding gas payment operation.

[0162] The second case: Step 2032: The number of preselected intentions is at least two. Determine the target intention of the target chat content based on the intention confirmation information of the target customer. Refer to Figure 8 as shown, it includes:

[0163] Step 20321: If the number of preselected intentions is at least two, then generate intention recommendation information based on the at least two preselected intentions.

[0164] During the implementation process, when the number of preselected intentions output by the intention recognition funnel matrix is at least two, that is, when the intention of the target chat content is not clear and not unique, generate intention recommendation information according to the above at least two preselected intentions.

[0165] For example, when the preselected intentions corresponding to the target chat content are gas payment and electronic payment, the generated intention recommendation information is gas payment and / or electronic payment.

[0166] Step 20322: Send the intention recommendation information to the target customer, so that the target customer can screen based on the received intention recommendation information and obtain intention confirmation information based on the at least one preselected intention selected.

[0167] During the implementation process, after generating the intent recommendation information, the above-mentioned intent recommendation information is sent to the target customer. In this way, the target customer can receive the above-mentioned intent recommendation information and perform intent screening based on the received intent recommendation information, so as to screen out at least one preselected intent. Furthermore, an intent confirmation information is generated based on the above-mentioned at least one preselected intent, and the above-mentioned intent confirmation information is sent to the computing device.

[0168] For example, when the intent recommendation information is gas payment and / or electricity payment, the target customer screens out gas payment from the intent recommendation information, and then generates intent confirmation information based on gas payment.

[0169] Step 20323: Determine at least one preselected intent included in the intent confirmation information as the target intent of the target chat content, and execute the intent business strategy based on the target intent.

[0170] During the implementation process, after receiving the intent confirmation information, the intent confirmation information is parsed to obtain at least one preselected intent, and the above-mentioned at least one preselected intent is determined as the target intent of the target chat content. Furthermore, the corresponding intent business strategy is determined according to the target intent, and the corresponding intent business strategy is continuously executed. For example, when the target intent is gas payment, the corresponding gas payment operation can be continuously executed.

[0171] Based on the same inventive concept, refer to Figure 9 As shown, an apparatus for intent recognition of chat content provided in an embodiment of the present application includes:

[0172] A screening unit 901, configured to screen the preselected chat content based on preset permission parameters to obtain target chat content, where the preselected chat content includes group chat content and one-on-one chat content of the target customer;

[0173] An identification unit 902, configured to perform intent recognition on the target chat content by using an intent recognition funnel matrix to obtain at least one preselected intent, where the intent recognition funnel matrix includes a plurality of successively cascaded upper-level intent recognition methods and lower-level intent recognition methods. If the content of the preselected intent obtained by the upper-level intent recognition method for intent recognition of the target chat content is empty, the lower-level intent recognition method is used to perform intent recognition on the target chat content;

[0174] A determination unit 903, configured to determine the target intent of the target chat content based on the intent confirmation information of the target customer, where the intent confirmation information is a reply of the target customer based on the received intent recommendation information, and the intent recommendation information includes at least two preselected intents.

[0175] Optionally, after performing intention recognition on the target chat content using the intention recognition funnel matrix and obtaining at least one preselected intention, it further includes:

[0176] If the number of preselected intentions is one, determine the preselected intention as the target intention of the target chat content, and execute the intention business strategy based on the target intention.

[0177] Optionally, the preset permission parameters include scenario permissions, domain permissions, time limit permissions, and role permissions. Filter the preselected chat content based on the preset permission parameters to obtain the target chat content. The filtering unit 901 is used for:

[0178] Perform the following operations on each preselected chat content:

[0179] Extract and / or merge keywords from the preselected chat content to obtain multiple chat content blocks;

[0180] Use a sliding window to intercept the content of each chat content block respectively to obtain multiple valid chat contents;

[0181] Perform scenario recognition, domain recognition, time limit recognition, and role recognition on each valid chat content respectively to obtain the target scenario, target domain, target time limit, and target role;

[0182] If the target scenario does not match the current scenario, delete the preselected chat content; or

[0183] If the target domain does not match the current domain, delete the preselected chat content; or

[0184] If the target time limit does not match the current time range, delete the preselected chat content; or

[0185] If the target role does not match the current role, delete the preselected chat content;

[0186] Among them, the current scenario, current domain, current time range, and current role are all determined based on the historical associated content of the preselected chat content.

[0187] Optionally, perform intention recognition on the target chat content using the intention recognition funnel matrix to obtain at least one preselected intention. The recognition unit 902 is used for:

[0188] Input the target chat content into the intention recognition method at the first level of the intention recognition funnel matrix for intention recognition;

[0189] If the content of the intention recognition result is not empty, use the intention recognition result as at least one preselected intention; or

[0190] If the content of the intent recognition result is empty, the target chat content is input into the next-level intent recognition method, and the next-level intent recognition method is used to perform intent recognition on the target chat content until the obtained intent recognition result is not empty.

[0191] Optionally, the intent recognition method includes: keyword matching method model, sentence matching method model, supervised model, and unsupervised model. The intent recognition funnel matrix is constructed in the following way:

[0192] Connect the input end of the intent recognition funnel matrix to the input end of the keyword matching method model, and connect the output end of the keyword matching method model to the output end of the intent recognition funnel matrix;

[0193] Connect the output end of the keyword matching method model to the input end of the sentence matching method model, and connect the output end of the sentence matching method model to the output end of the intent recognition funnel matrix;

[0194] Connect the output end of the sentence matching method model to the input end of the supervised model, and connect the output end of the supervised model to the output end of the intent recognition funnel matrix;

[0195] Connect the output end of the supervised model to the input end of the unsupervised model, and connect the output end of the unsupervised model to the output end of the intent recognition funnel matrix.

[0196] Optionally, if the intent recognition method is an unsupervised model, the intent recognition of the target chat content is performed in the following way:

[0197] Determine the intent association chain based on the pre-established knowledge graph. Among them, the intent association chain includes intent keywords and the parent nodes and child nodes associated with the intent keywords. The intent keywords are the knowledge points corresponding to the target chat content in the knowledge graph, the parent nodes are the upper-level knowledge points connected to the knowledge points in the knowledge graph, and the child nodes are the lower-level knowledge points connected to the knowledge points in the knowledge graph;

[0198] Use the unsupervised model with constraints to perform intent recognition on the target chat content, where the constraint condition is the intent association chain.

[0199] Optionally, determine the target intent of the target chat content based on the intent confirmation information of the target customer. The determination unit 903 is used for:

[0200] If the number of preselected intents is at least two, generate intent recommendation information based on the at least two preselected intents;

[0201] Send the intent recommendation information to the target customer so that the target customer can screen based on the received intent recommendation information and obtain intent confirmation information based on at least one preselected intent selected;

[0202] Determine at least one preselected intent included in the intent confirmation information as the target intent of the target chat content, and execute an intent business strategy based on the target intent.

[0203] Based on the same inventive concept, refer to Figure 10 As shown, an embodiment of the present application provides a computing device, including: a memory 1001 for storing executable instructions; a processor 1002 for reading and executing the executable instructions stored in the memory, and executing any one of the methods in the first aspect above.

[0204] Based on the same inventive concept, an embodiment of the present application provides a computer-readable storage medium, when the instructions in the storage medium are executed by a processor, enabling the processor to execute the method described in any one of the above first aspects.

[0205] In summary, in the embodiment of the present application, a method, device, and storage medium for intent recognition of chat content are provided. The method includes: screening preselected chat content based on preset permission parameters to obtain target chat content, where the preselected chat content includes group chat content and one-on-one chat content of a target customer; using an intent recognition funnel matrix to perform intent recognition on the target chat content to obtain at least one preselected intent, where the intent recognition funnel matrix includes a plurality of successively cascaded upper-level intent recognition methods and lower-level intent recognition methods. If the content of the preselected intent obtained after the upper-level intent recognition method performs intent recognition on the target chat content is empty, then use the lower-level intent recognition method to perform intent recognition on the target chat content; determine the target intent of the target chat content based on the intent confirmation information of the target customer, where the intent confirmation information is a reply from the target customer based on the received intent recommendation information, and the intent recommendation information includes at least two preselected intents. The method of using the intent recognition funnel matrix and intent recommendation information to determine the target intent effectively improves the accuracy of intent recognition.

[0206] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system, or a computer program product system. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product system implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0207] This application is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to the application. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to produce a machine, such that the instructions executed by the processor of the computer or other programmable data processing device produce means for implementing the functions specified in one flow Figure 1 one flow or more flows and / or blocks Figure 1 one block or more blocks.

[0208] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means that implement the functions specified in one flow Figure 1 one flow or more flows and / or blocks Figure 1 one block or more blocks.

[0209] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one flow Figure 1 one flow or more flows and / or blocks Figure 1 one block or more blocks.

[0210] Obviously, those skilled in the art can make various changes and modifications to this application without departing from the spirit and scope of this application. Thus, if these modifications and variations of this application fall within the scope of the claims of this application and their equivalent technologies, this application is also intended to include these changes and modifications.

Claims

1. A method for identifying the intent of chat content, characterized in that: The method comprises: The pre-selected chat content is screened based on preset permission parameters to obtain target chat content, wherein the pre-selected chat content includes group chat content and single chat content of the target customer; Performing intent recognition on the target chat content using an intent recognition funnel matrix to obtain at least one pre-selected intent, wherein the intent recognition funnel matrix includes a plurality of upper-level intent recognition methods and lower-level intent recognition methods that are sequentially cascaded, and if the content of the pre-selected intent obtained after the upper-level intent recognition method performs intent recognition on the target chat content is empty, then performing intent recognition on the target chat content using the lower-level intent recognition method; The target intent of the target chat content is determined based on the intent confirmation information of the target customer, wherein the intent confirmation information is the reply of the target customer based on the received intent recommendation information, and the intent recommendation information includes at least two of the pre-selected intentions.

2. The method according to claim 1, characterized in that After the intent recognition funnel matrix is ​​used to perform intent recognition on the target chat content and at least one pre-selected intent is obtained, the method further includes: If the number of the pre-selected intents is one, the pre-selected intent is determined as the target intent of the target chat content, and the intent business policy is executed based on the target intent.

3. The method according to claim 1, characterized in that The preset permission parameters include scene permission, domain permission, time permission and role permission. The pre-selected chat content is screened based on the preset permission parameters to obtain the target chat content, including: For each of the pre-selected chat contents, perform the following operations: Extracting and / or merging keywords from the pre-selected chat content to obtain a plurality of chat content blocks; Using a sliding window to intercept the content of each chat content block respectively, to obtain multiple valid chat contents; Respectively perform scene recognition, field recognition, time recognition and role recognition on each of the effective chat contents to obtain a target scene, a target field, a target time recognition and a target role; If the target scenario does not match the current scenario, deleting the pre-selected chat content; or If the target domain does not match the current domain, deleting the pre-selected chat content; or If the target time limit does not meet the current time range, deleting the pre-selected chat content; or If the target role does not match the current role, deleting the pre-selected chat content; Among them, the current scene, the current field, the current time range and the current role are all determined based on the historical related content of the pre-selected chat content.

4. The method according to claim 1, characterized in that The using the intention recognition funnel matrix to perform intent recognition on the target chat content to obtain at least one pre-selected intent includes: Inputting the target chat content into the intention recognition method located at the first level of the intention recognition funnel matrix for intention recognition; If the content of the intent recognition result is not empty, taking the intent recognition result as at least one of the pre-selected intents; or If the content of the intent recognition result is empty, the target chat content is input into the next level intent recognition method, and the next level intent recognition method is used to perform intent recognition on the target chat content until the content of the intent recognition result is not empty.

5. The method according to claim 1, characterized in that The intention recognition method includes: a keyword matching method model, a sentence matching method model, a supervised model and an unsupervised model, and the intention recognition funnel matrix is ​​constructed in the following manner: Connecting the input end of the intention identification funnel matrix to the input end of the keyword matching method model, and connecting the output end of the keyword matching method model to the output end of the intention identification funnel matrix; Connecting the output end of the keyword matching method model to the input end of the sentence matching method model, and connecting the output end of the sentence matching method model to the output end of the intention recognition funnel matrix; Connecting the output end of the sentence matching method model to the input end of the supervised model, and connecting the output end of the supervised model to the output end of the intention recognition funnel matrix; The output end of the supervised model is connected to the input end of the unsupervised model, and the output end of the unsupervised model is connected to the output end of the intent recognition funnel matrix.

6. The method according to claim 1, characterized in that If the intent recognition method is an unsupervised model, the intent recognition of the target chat content is performed in the following manner: Determine an intention association chain based on a pre-established knowledge graph, wherein the intention association chain includes an intention keyword and a parent node and a child node associated with the intention keyword, the intention keyword is a knowledge point corresponding to the target chat content in the knowledge graph, the parent node is an upper-level knowledge point connected to the knowledge point in the knowledge graph, and the child node is a lower-level knowledge point connected to the knowledge point in the knowledge graph; An unsupervised model with constraints is used to identify the intent of the target chat content, wherein the constraints are the intent association chain.

7. The method according to any one of claims 1 to 6, characterized in that: The determining the target intention of the target chat content based on the intention confirmation information of the target customer includes: If the number of the pre-selected intentions is at least two, generating the intention recommendation information based on at least two of the pre-selected intentions; Sending the intention recommendation information to the target customer, so that the target customer performs screening based on the received intention recommendation information and obtains the intention confirmation information based on at least one pre-selected intention screened out; One of the preselected intents included in the intent confirmation information is determined as the target intent of the target chat content, and an intent business policy is executed based on the target intent.

8. A device for identifying the intention of chat content, characterized in that: include: A screening unit, configured to screen the pre-selected chat content based on a preset permission parameter to obtain target chat content, wherein the pre-selected chat content includes group chat content and single chat content of the target customer; an identification unit, configured to perform intent identification on the target chat content using an intent identification funnel matrix to obtain at least one pre-selected intent, wherein the intent identification funnel matrix includes a plurality of upper-level intent identification methods and lower-level intent identification methods that are sequentially cascaded, and if the content of the pre-selected intent obtained after the upper-level intent identification method performs intent identification on the target chat content is empty, then the lower-level intent identification method is used to perform intent identification on the target chat content; A determination unit is used to determine the target intent of the target chat content based on the intention confirmation information of the target customer, wherein the intention confirmation information is the reply of the target customer based on the received intention recommendation information, and the intention recommendation information includes at least two of the pre-selected intentions.

9. A computing device, characterized in that include: A memory for storing executable instructions; A processor, configured to read and execute the executable instructions stored in the memory to implement the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: When the instructions in the storage medium are executed by a processor, the processor is enabled to execute the method according to any one of claims 1 to 7.

Images