Dialogue method and electronic device based on domain pre-training model

Abstract

The present application provides a conversation method and electronic device based on a domain pre-training model. The method includes: obtaining a business model constructed based on a business tag set and a domain pre-training model constructed based on a domain tag set, obtaining a first score of a user question classified under a business tag and a second score of a domain tag through the business model and the domain pre-training model; obtaining a corrected score of the user question classified under a business tag based on the first score of the user question classified under a business tag and the second score of the domain tag; determining a target business tag corresponding to the user question based on the corrected score of the user question classified under each business tag, and outputting a standard answer corresponding to the target business tag. This solution reduces the amount of expansion of conversation business corpus and reduces operating costs by integrating the domain pre-training model method.

Description

Technical Field

[0001] The present application relates to the field of intelligent question-answering technology, and in particular to a dialogue method and electronic device based on a domain pre-training model. Background Art

[0002] Dialogue systems based on algorithmic models require experts to sort out dialogue business corpus. Traditional methods require experts to combine business knowledge and linguistic knowledge and use templates, word order changes, sentence changes and other methods to manually expand the corpus. This requires a lot of manpower costs and expert capabilities, and the sorting work of different experts is difficult to ensure stability and uniformity.

[0003] To reduce operating costs and reduce the workload of sorting out conversational business corpora, customers generally directly use pre-trained models provided by manufacturers. Specifically, manufacturers use neural network training to integrate their own accumulated domain corpora into domain pre-trained models and provide customers with domain pre-trained model services. This reduces the cost of corpus expansion and protects domain data resources. However, many times, customers customize a set of labeling systems based on their own business. Some labels will completely overlap with the labeling system of the manufacturer's domain pre-trained model, while others may have similar semantics but different expressions. Therefore, it is not possible to directly put the manufacturer's domain pre-trained model into use. Summary of the Invention

[0004] The embodiment of the present application provides a dialogue method based on a domain pre-training model to reduce the amount of expansion of dialogue business corpus and reduce operating costs.

[0005] The present application provides a conversation method based on a domain pre-training model, including:

[0006] Obtain a business model built based on a business tag set and a domain pre-training model built based on a domain tag set;

[0007] Classify the received user questions using the business model and the domain pre-training model respectively, and obtain a first score for the user question being classified under any business tag in the business tag set and a second score for the user question being classified under any domain tag in the domain tag set;

[0008] Determining a correction score for the user question classified under each of the business tags according to a first score for the user question classified under any business tag in the business tag set and a second score for the user question classified under any domain tag in the domain tag set;

[0009] According to the correction score of each business tag categorized by the user question, a target business tag corresponding to the user question is determined, and a standard answer corresponding to the target business tag is output.

[0010] In one embodiment, determining the correction score of the user question classified under each of the business tags based on a first score of the user question classified under any business tag in the business tag set and a second score of the user question classified under any domain tag in the domain tag set includes:

[0011] For each business tag, filtering out a target domain tag corresponding to the business tag from the domain tag set;

[0012] If the target domain label is obtained through screening, the first score of the business label is corrected according to the second score of the target domain label to obtain a corrected score for classifying the user question under the business label.

[0013] In one embodiment, determining the correction score of the user question classified under each of the business tags based on a first score of the user question classified under any business tag in the business tag set and a second score of the user question classified under any domain tag in the domain tag set includes:

[0014] For each business tag, filtering out a target domain tag corresponding to the business tag from the domain tag set;

[0015] If the target domain tag does not exist, the first score of the user question classified as the business tag is used as a correction score for classifying the user question as the business tag.

[0016] In one embodiment, for each service tag, filtering out a target domain tag corresponding to the service tag from the domain tag set includes:

[0017] For each business tag, a domain tag having a text similarity with the business tag greater than a first threshold is screened out from the domain tag set as the target domain tag.

[0018] In one embodiment, for each service tag, filtering out a target domain tag corresponding to the service tag from the domain tag set includes:

[0019] According to the second score of each domain tag classified by the user question, delete the domain tags whose second score is lower than a second threshold in the domain tag set;

[0020] For each business tag, a target domain tag corresponding to the business tag is screened out from the remaining domain tags. The target domain tag is a domain tag whose text similarity with the business tag is greater than a first threshold.

[0021] In one embodiment, for each service tag, filtering out a target domain tag corresponding to the service tag from the domain tag set includes:

[0022] For each domain tag, calculate the text similarity between the domain tag and each business tag, and delete all domain tags whose text similarity is less than a preset similarity;

[0023] For each business tag, a target domain tag corresponding to the business tag is screened out from the remaining domain tags. The target domain tag is a domain tag whose text similarity with the business tag is greater than a first threshold.

[0024] In one embodiment, for each service tag, filtering out a target domain tag corresponding to the service tag from the domain tag set includes:

[0025] According to the second score of each domain tag categorized by the user question, domain tags having the second score lower than a second threshold are deleted to obtain a first domain tag list;

[0026] Calculating the text similarity between the domain tag and each business tag in the first domain tag list, deleting all domain tags whose text similarity is less than a preset similarity, and obtaining a second domain tag list;

[0027] For each business tag, a target domain tag corresponding to the business tag is screened out from the second domain tag list. The target domain tag is a domain tag whose text similarity with the business tag is greater than a first threshold.

[0028] In one embodiment, the correcting the first score of the service tag according to the second score of the target domain tag to obtain a corrected score for the user question classified under the service tag includes:

[0029] According to the second score of the target domain label, a weighted sum is performed on the first score of the corresponding business label and the second score of the target domain label to obtain a correction score for classifying the user question under the business label.

[0030] In one embodiment, determining a target business tag corresponding to the user question based on the correction score of each business tag categorized by the user question, and outputting a standard answer configured corresponding to the target business tag includes:

[0031] According to the correction scores of each business tag classified by the user question, the business tag with the largest correction score is determined as the target business tag, and the standard answer corresponding to the target business tag is output.

[0032] An embodiment of the present application further provides an electronic device, comprising:

[0033] processor;

[0034] a memory for storing processor-executable instructions;

[0035] The processor is configured to execute the above-mentioned dialogue method based on the domain pre-training model.

[0036] An embodiment of the present application also provides a computer-readable storage medium, which stores a computer program. The computer program can be executed by a processor to complete the above-mentioned dialogue method based on the domain pre-training model.

[0037] The technical solution provided by the above embodiments of the present application can directly utilize the domain data accumulated by others through the domain pre-training model, which helps users reduce the amount of corpus expansion and reduce operating costs compared to expanding the dialogue corpus from scratch; according to the score of the domain label, the correction score of the business label is obtained, and the integration of the domain pre-training model and the business model of the inconsistent label system is realized. Compared with directly using the domain pre-training model, the flexibility of the customized label system is realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following briefly introduces the drawings required for use in the embodiments of the present application.

[0039] Figure 1 is a structural diagram of an electronic device provided in an embodiment of the present application;

[0040] Figure 2 This is a flowchart of a conversation method based on a domain pre-training model provided in an embodiment of the present application;

[0041] Figure 3 This is a schematic diagram of the principle of the dialogue method based on the domain pre-training model provided in an embodiment of the present application;

[0042] Figure 4 yes Figure 2 Detailed flowchart of step S230 in the corresponding embodiment;

[0043] Figure 5 yes Figure 4 Detailed flowchart of step S231 in the corresponding embodiment;

[0044] Figure 6 This is a schematic diagram of the principle of a dialogue method based on a domain pre-training model provided in one embodiment of the present application;

[0045] Figure 7 Another embodiment provides Figure 4Detailed flowchart of step S231 in the corresponding embodiment;

[0046] Figure 8 Yet another embodiment provides Figure 4 Detailed flowchart of step S231 in the corresponding embodiment;

[0047] Figure 9 This is a block diagram of a conversation device based on a domain pre-training model provided in an embodiment of the present application. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0049] Similar reference numerals and letters denote similar items in the following drawings. Therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings. At the same time, in the description of this application, the terms "first", "second", etc. are only used to distinguish the description and should not be understood as indicating or implying relative importance.

[0050] Figure 1 1 is a schematic diagram of the structure of the electronic device provided in the embodiment of the present application. The electronic device 100 can be used to execute the dialogue method based on the domain pre-training model provided in the embodiment of the present application. Figure 1 As shown, the electronic device 100 includes: one or more processors 102 and one or more memories 104 storing processor-executable instructions. The processor 102 is configured to execute the conversation method based on the domain pre-training model provided in the following embodiments of the present application.

[0051] The processor 102 can be a gateway, a smart terminal, or a device including a central processing unit (CPU), a graphics processing unit (GPU), or other forms of processing units with data processing capabilities and / or instruction execution capabilities. It can process data of other components in the electronic device 100 and control other components in the electronic device 100 to perform desired functions.

[0052] The memory 104 may include one or more computer program products, and the computer program products may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. The non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 102 may execute the program instructions to implement the dialogue method based on the domain pre-training model described below. Various applications and various data may also be stored in the computer-readable storage medium, such as various data used and / or generated by the application.

[0053] In one embodiment, Figure 1 The electronic device 100 may further include an input device 106, an output device 108, and a data acquisition device 110. These components are interconnected via a bus system 112 and / or other forms of connection mechanisms (not shown). Figure 1 The components and structure of the electronic device 100 shown are merely exemplary and non-limiting. The electronic device 100 may also have other components and structures as needed.

[0054] The input device 106 can be a device used by a user to input instructions, and can include one or more of a keyboard, a mouse, a microphone, and a touch screen. The output device 108 can output various information (e.g., images or sounds) to the outside (e.g., a user), and can include one or more of a display, a speaker, etc. The data acquisition device 110 can capture images of an object and store the captured images in the memory 104 for use by other components. For example, the data acquisition device 110 can be a camera.

[0055] In one embodiment, the various components in the example electronic device 100 for implementing the dialogue method based on the domain pre-training model of the embodiment of the present application can be integrated or dispersed, such as integrating the processor 102, memory 104, input device 106 and output device 108 into one, and separating the data acquisition device 110.

[0056] In one embodiment, the example electronic device 100 for implementing the dialogue method based on the domain pre-training model of the embodiment of the present application can be implemented as an intelligent terminal such as a smart phone, a tablet computer, a desktop computer, a smart watch, a vehicle-mounted device, a server, etc.

[0057] Figure 2 This is a flow chart of the dialogue method based on the domain pre-training model provided in the embodiment of the present application. Figure 2 As shown, the method includes the following steps S210 to S240.

[0058] Step S210: Acquire a business model constructed based on the business tag set and a domain pre-training model constructed based on the domain tag set.

[0059] Both the business model and domain pre-training are classification models, but they are trained using different labeling systems. The business model is trained based on a business label set, while the domain pre-training model is trained based on a domain label set. The business label set and the domain label set may not contain exactly the same labels. For example, the business label set may contain labels that are not present in the domain label set, the business label set and the domain label set may have partially overlapping labels, or the business label set and the domain label set may contain labels with inconsistent semantics.

[0060] To distinguish them, tags in a business tag set can be referred to as business tags, and tags in a domain tag set can be referred to as domain tags. Tags can be thought of as standard questions. For example, a business tag set contains 10 standard questions in the banking sector, while a domain tag set contains 8 standard questions. These 10 standard questions in the business tag set are not identical to the 8 standard questions in the domain tag set.

[0061] It should be noted that a business tag set is a set of tags customized by the user based on their business needs. A domain tag set is an externally provided tag set that covers the same domain but has inappropriate content. Therefore, you cannot directly apply a domain pre-trained model built based on an externally provided domain tag set.

[0062] Step S220: Classify the received user question using the business model and the domain pre-training model respectively to obtain a first score for the user question being classified under any business tag in the business tag set and a second score for the user question being classified under any domain tag in the domain tag set.

[0063] For example, a user question might be "How long does it take to open a credit card account?" This user question serves as input to both the business model and the domain pre-trained model. The business model classifies the user question and outputs a score for each business tag. This score, called the first score, represents the likelihood that the user question falls under that business tag. Assuming the business tag set contains 10 business tags, 10 scores are output, each corresponding to the likelihood that the user question falls under each of the 10 business tags.

[0064] Similarly, the domain pre-trained model classifies user questions and outputs a score for each domain label. This score, for differentiation, is called the second score. The score represents the likelihood that the user question falls under that domain label. Assuming the domain label set contains eight domain labels, eight scores are output, each corresponding to the likelihood that the user question falls under each of these eight domain labels.

[0065] like Figure 3 As shown, the business label set includes 7 business labels (i.e., standard questions 1.1 to 1.7), and the domain label set includes 9 domain labels (i.e., standard questions 2.1 to 2.9). Through the business module, 7 first scores corresponding to standard questions 1.1 to 1.7 can be obtained, and through the domain pre-training model, 9 second scores corresponding to standard questions 2.1 to 2.9 can be obtained.

[0066] Step S230: Determine a correction score for the user question classified under each of the business tags according to a first score for the user question classified under any business tag in the business tag set and a second score for the user question classified under any domain tag in the domain tag set.

[0067] In one embodiment, if Figure 4 As shown, step S230 specifically includes: step S231: for each business tag, filtering out a target domain tag corresponding to the business tag from the domain tag set. step S232: if the target domain tag is obtained through filtering, correcting the first score of the business tag based on the second score of the target domain tag to obtain a corrected score for the user question categorized under the business tag.

[0068] In one embodiment, the target domain tag may be a domain tag in the domain tag set whose text similarity with the business tag is greater than a first threshold. The calculation method of the text similarity can be implemented with reference to the existing technology. Figure 3 As shown, the text similarity between Standard Question 2.5 and Standard Question 1.6 is 0.9, which is greater than the first threshold of 0.8. Therefore, Standard Question 2.5 is the target domain label whose text similarity with Standard Question 1.6 is greater than the first threshold. Since the second score of Standard Question 2.5 is 0.85 and the first score of Standard Question 1.6 is 0.80, the first score can be corrected to obtain a corrected score of 0.825 for Standard Question 1.6 (i.e., the average of 0.8 and 0.85).

[0069] It should be noted that for any business tag, there may be no target domain tag, or one or more target domain tags with a textual similarity greater than a first threshold. If no target domain tag exists, no correction is performed on the first score of the business tag. If one or more target domain tags exist, the first score of the business tag is corrected based on the second score of the target domain tag. The corrected first score is referred to as the corrected score.

[0070] In one embodiment, assuming that there are one or more target domain tags, the first score of the business tag and the second score of the target domain tag can be weighted and summed according to the second score of the target domain tag to obtain a correction score for the user question classified under the business tag.

[0071] For example, if the text similarity between business tag 5 and domain tag 8 is greater than the first threshold, domain tag 8 is the target domain tag for business tag 5. Assuming the second score of domain tag 8 is 0.9 and the first score of business tag 5 is 0.8, the corrected score for business tag 5 can be (0.8 + 0.9) / 2 = 0.85. This means that the weights of the target domain tag and the business tag are both 50%.

[0072] Among them, the weight of the target domain label can be related to the size of the text similarity with the business label. Assuming that there is more than one target domain label similar to the business label, the higher the text similarity between the target domain label and the business label, the greater the corresponding weight. For example, the weight of the business label can be 50%. Assuming that domain label 6, domain label 7, and domain label 8 are all target domain labels whose text similarity with the business label is greater than the first threshold, and the text similarity with the business label is 60%, 70%, and 80% respectively, the weights of domain labels 6, 7, and 8 can be increased in turn, and the total is equal to 50%.

[0073] In other embodiments, if the target domain tag does not exist, the first score of the user question classified under the business tag may be used as a correction score for classifying the user question under the business tag.

[0074] For example, if there are no domain tags in the domain tag set whose text similarity with the business tag exceeds the first threshold, the target domain tag corresponding to the business tag cannot be filtered out. Therefore, the first score of the user question classified under the business tag is not corrected. The correction score is the first score of the user question classified under the business tag.

[0075] Step S240: determining the target business tag corresponding to the user question based on the correction score of each business tag categorized by the user question, and outputting the standard answer configured corresponding to the target business tag.

[0076] The target business tag refers to the corrected scores assigned to each business tag by the user's question. The business tag that most closely matches the user's question is selected for differentiation. The standard answer corresponding to each business tag can be pre-configured and stored in the database.

[0077] In one embodiment, based on the correction scores of each service tag categorized by the user question, the service tag with the largest correction score may be determined as the target service tag, and the standard answer corresponding to the target service tag may be output.

[0078] In other words, the business tag with the largest correction score can be considered as the standard question closest to the user's question. The standard answer corresponding to the standard question can be used as the answer to the user's question, so that the standard answer can be output to conduct an intelligent dialogue with the user. Figure 3 As shown, after correction, the first score of standard question 1.6 is corrected to 0.825. At this time, standard question 1.6 has the highest score, and the standard answer corresponding to standard question 1.6 can be output as the response to the user's question.

[0079] In one embodiment, if Figure 5 As shown, the above step S231 specifically includes: step S510-step S520.

[0080] Step S510 : According to the second score of each domain tag categorized by the user question, delete the domain tags in the domain tag set whose second score is lower than a second threshold.

[0081] Step S520 : For each business tag, a target domain tag corresponding to the business tag is screened out from the remaining domain tags. The target domain tag is a domain tag whose text similarity with the business tag is greater than a first threshold.

[0082] For example, the second threshold can be 0.8, and the domain tags with a second score lower than 0.8 are deleted, so that only the domain tags close to the user's question are retained. Then, the domain tags whose text similarity with the business tag is greater than the first threshold can be screened out from the remaining domain tags and used as the target domain tags, so as to improve the accuracy of the correction score when the first score of the business tag is corrected later.

[0083] In one embodiment, assuming that after deleting domain tags with a second score below the second threshold, there are no remaining domain tags, or no domain tags with a text similarity greater than the first threshold, it can be assumed that the target domain tag does not exist. The correction score is the first score of the business tag. Based on the first score of each business tag assigned to the user question, the target business tag corresponding to the user question is determined, and the standard answer corresponding to the target business tag is output. In this case, the target business tag can be the business tag with the highest first score.

[0084] like Figure 6 As shown, the second scores corresponding to each domain label (standard questions 2.1 to 2.9) are all lower than the second threshold (for example, 0.8). Therefore, after deleting the domain labels whose second scores are lower than the second threshold, there are no remaining domain labels. At this time, the first score of standard question 1.2 is 0.86, which is the highest. The standard answer corresponding to standard question 1.2 can be directly output.

[0085] In another embodiment, if Figure 7 As shown, the above step S231 specifically includes: step S710-step S720.

[0086] Step S710 : For each domain tag, calculate the text similarity between the domain tag and each business tag, and delete all domain tags in the domain tag set whose text similarity is less than a preset similarity.

[0087] Step S720 : For each business tag, a target domain tag corresponding to the business tag is screened out from the remaining domain tags. The target domain tag is a domain tag whose text similarity with the business tag is greater than a first threshold.

[0088] It should be noted that if the textual similarity between a domain tag and all business tags is less than a preset similarity, this indicates that the domain tag differs significantly from the user's business tag and does not meet the user's business needs. Therefore, the domain tag can be deleted in advance. Then, from the remaining domain tags, a domain tag with a textual similarity greater than a first threshold can be selected as the target domain tag. Based on the second score of the target domain tag, the first score of the business tag can be corrected to improve the accuracy of the corrected score.

[0089] Similarly, assuming that after deleting domain tags whose textual similarity with all business tags is less than a preset similarity, there are no remaining domain tags, or no domain tags whose textual similarity exceeds a first threshold, it can be assumed that the target domain tag does not exist. The correction score is then the first score of the business tag. Based on the first score of each business tag assigned to the user question, the target business tag corresponding to the user question is determined, and the standard answer corresponding to the target business tag is output. In this case, the target business tag can be the business tag with the highest first score.

[0090] In other embodiments, Figure 8 As shown, the above step S231 specifically includes: step S810-step S830.

[0091] Step S810: According to the second score of each domain tag categorized by the user question, domain tags with the second score lower than a second threshold are deleted to obtain a first domain tag list.

[0092] The first domain tag list refers to a set of remaining domain tags after deleting domain tags with second scores lower than the second threshold from the domain tag set.

[0093] Step S820: Calculate the text similarity between the domain tag and each business tag in the first domain tag list, delete all domain tags whose text similarity is less than a preset similarity, and obtain a second domain tag list.

[0094] Step S830: For each business tag, filter out a target domain tag corresponding to the business tag from the second domain tag list, where the target domain tag is a domain tag whose text similarity with the business tag is greater than a first threshold.

[0095] The second domain tag list is the set of domain tags remaining after removing domain tags whose textual similarity with all business tags is less than a preset similarity from the first domain tag list. After removing domain tags whose textual similarity with all business tags is less than the preset similarity, target domain tags whose textual similarity with business tags is greater than a first threshold are found from the second domain tag list.

[0096] Similarly, if the first domain tag list is empty or the second domain tag list is empty or there is no domain tag with a text similarity greater than the first threshold, it can be considered that there is no target domain tag. The correction score is the first score of the business tag. The target business tag corresponding to the user question is determined directly based on the first score of each business tag classified by the user question, and the standard answer corresponding to the target business tag is output.

[0097] The technical solution provided by the above embodiments of the present application can directly utilize the domain data accumulated by others through the domain pre-training model, which helps users reduce the amount of corpus expansion and operating costs compared to expanding the dialogue corpus from scratch; through the text similarity method, the domain pre-training model and business model of the inconsistent label system are integrated, which realizes the flexibility of the customized label system compared to directly using the domain pre-training model.

[0098] The following is an embodiment of the device of the present application, which can be used to execute the embodiment of the conversation method based on the domain pre-training model of the present application. For details not disclosed in the embodiment of the device of the present application, please refer to the embodiment of the conversation method based on the domain pre-training model of the present application.

[0099] Figure 9 This is a block diagram of a conversation device based on a domain pre-training model provided in an embodiment of the present application.

[0100] like Figure 9 As shown, the device includes:

[0101] Model acquisition module 910, used to acquire a business model built based on a business tag set and a domain pre-training model built based on a domain tag set;

[0102] A score calculation module 920 is configured to classify the received user question using the business model and the domain pre-trained model, respectively, to obtain a first score for the user question when it is classified under any business tag in the business tag set and a second score for the user question when it is classified under any domain tag in the domain tag set;

[0103] A score correction module 930 is configured to determine a correction score for the user question classified under each of the business tags according to a first score for the user question classified under any business tag in the business tag set and a second score for the user question classified under any field tag in the field tag set;

[0104] The answer output module 940 is configured to determine the target business tag corresponding to the user question based on the correction score of each business tag categorized by the user question, and output the standard answer configured corresponding to the target business tag.

[0105] The implementation process of the functions and effects of each module in the above-mentioned device is specifically described in the implementation process of the corresponding steps in the above-mentioned dialogue method based on the domain pre-training model, and will not be repeated here.

[0106] In several embodiments provided in this application, the disclosed devices and methods may also be implemented in other ways. The device embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions, and operations of the devices, methods, and computer program products according to multiple embodiments of the present application. In this regard, each box in the flowchart or block diagram may represent a module, a program segment, or a portion of code, and the module, program segment, or a portion of code contains one or more executable instructions for implementing the specified logical functions. In some alternative implementations, the functions marked in the boxes may also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes may actually be executed substantially in parallel, or they may sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart, may be implemented using a dedicated hardware-based system that performs the specified functions or actions, or may be implemented using a combination of dedicated hardware and computer instructions.

[0107] In addition, the functional modules in each embodiment of the present application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0108] If the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a number of instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the various embodiments of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

Claims

1. A dialogue method based on a domain pre-training model, characterized in that: include: Obtain a business model built based on a business tag set and a domain pre-training model built based on a domain tag set; Classify the received user questions using the business model and the domain pre-training model respectively, and obtain a first score for the user question being classified under any business tag in the business tag set and a second score for the user question being classified under any domain tag in the domain tag set; Determining a correction score for the user question classified under each of the business tags according to a first score for the user question classified under any business tag in the business tag set and a second score for the user question classified under any domain tag in the domain tag set; According to the correction score of each business tag categorized by the user question, a target business tag corresponding to the user question is determined, and a standard answer corresponding to the target business tag is output.

2. The method according to claim 1, characterized in that Determining, based on a first score of the user question being classified under any business tag in the business tag set and a second score of the user question being classified under any domain tag in the domain tag set, a correction score for the user question being classified under each business tag includes: For each business tag, filtering out the target domain tag corresponding to the business tag from the domain tag set; If the target domain label is obtained through screening, the first score of the business label is corrected according to the second score of the target domain label to obtain a corrected score for classifying the user question under the business label.

3. The method according to claim 1, characterized in that Determining, based on a first score of the user question being classified under any business tag in the business tag set and a second score of the user question being classified under any domain tag in the domain tag set, a correction score for the user question being classified under each business tag includes: For each business tag, filtering out the target domain tag corresponding to the business tag from the domain tag set; If the target domain tag does not exist, the first score of the user question classified as the business tag is used as a correction score for classifying the user question as the business tag.

4. The method according to claim 2 or 3, characterized in that The step of filtering out, for each service tag, a target domain tag corresponding to the service tag from the domain tag set includes: For each business tag, a domain tag having a text similarity with the business tag greater than a first threshold is screened out from the domain tag set as the target domain tag.

5. The method according to claim 2 or 3, characterized in that The step of filtering out, for each service tag, a target domain tag corresponding to the service tag from the domain tag set includes: According to the second score of each domain tag classified by the user question, delete the domain tags whose second score is lower than a second threshold in the domain tag set; For each business tag, a target domain tag corresponding to the business tag is screened out from the remaining domain tags. The target domain tag is a domain tag whose text similarity with the business tag is greater than a first threshold.

6. The method according to claim 2 or 3, characterized in that The step of filtering out, for each service tag, a target domain tag corresponding to the service tag from the domain tag set includes: For each domain tag, calculate the text similarity between the domain tag and each business tag, and delete all domain tags in the domain tag set whose text similarity is less than a preset similarity; For each business tag, a target domain tag corresponding to the business tag is screened out from the remaining domain tags. The target domain tag is a domain tag whose text similarity with the business tag is greater than a first threshold.

7. The method according to claim 2 or 3, characterized in that The step of filtering out, for each service tag, a target domain tag corresponding to the service tag from the domain tag set includes: According to the second score of each domain tag categorized by the user question, domain tags having the second score lower than a second threshold are deleted to obtain a first domain tag list; Calculating the text similarity between the domain tag and each business tag in the first domain tag list, deleting all domain tags whose text similarity is less than a preset similarity, and obtaining a second domain tag list; For each business tag, a target domain tag corresponding to the business tag is screened out from the second domain tag list. The target domain tag is a domain tag whose text similarity with the business tag is greater than a first threshold.

8. The method according to claim 2, characterized in that The correcting the first score of the business tag according to the second score of the target domain tag to obtain a corrected score of the user question classified under the business tag includes: According to the second score of the target domain label, a weighted sum is performed on the first score of the corresponding business label and the second score of the target domain label to obtain a correction score for classifying the user question under the business label.

9. The method according to claim 1, characterized in that The step of determining a target business tag corresponding to the user question based on the correction score of each business tag categorized by the user question, and outputting a standard answer configured corresponding to the target business tag, includes: According to the correction scores of each business tag classified by the user question, the business tag with the largest correction score is determined as the target business tag, and the standard answer corresponding to the target business tag is output.

10. An electronic device, characterized in that: The electronic device comprises: processor; a memory for storing processor-executable instructions; The processor is configured to execute the dialogue method based on the domain pre-training model according to any one of claims 1 to 9.

Images