Large model-based data expansion method, electronic equipment and storage medium

Through the data expansion method based on the big model, similar sentences are generated and screened, which solves the problem of inefficient filling of similar sentences manually in the customer service quality inspection system, and improves the quality of data updates and the accuracy and scalability of the quality inspection system.

CN119089892APending Publication Date: 2024-12-06QINGDAO HAIER TECH +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411153139.X
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-21
Publication Date
2024-12-06

AI Technical Summary

Technical Problem

The existing customer service quality inspection system relies on manual exhaustive filling of similar sentences, resulting in low efficiency, high cost, low detection accuracy, and poor scalability and maintainability.

Method used

The data expansion method based on the big model is adopted to obtain the session content, determine the corpus information, use the pre-trained big model to generate similar sentences, and filter and update the standard database to improve the data expansion efficiency and quality.

Benefits of technology

It significantly improves the efficiency of data expansion, ensures the quality of data updates, reduces manual intervention, and improves the detection accuracy and scalability of the quality inspection system.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119089892A_ABST
    Figure CN119089892A_ABST
Patent Text Reader

Abstract

The invention discloses a data expansion method based on a large model, electronic equipment and a storage medium, and relates to the technical field of smart home / smart home, and the data expansion method based on the large model comprises the steps: obtaining session content; determining corpus information based on the session content, wherein the corpus information comprises at least one of a standard sentence, a semantic interpretation corresponding to the standard sentence and a semantic tag code; generating a first similar sentence based on the corpus information and a pre-trained large model; screening the first similar sentences to obtain newly added similar sentences corresponding to the standard sentences; and updating the standard database based on the newly added similar sentences. According to the method, the first similar sentences are generated based on the accurate corpus information by means of the generalization ability of a large model, the data expansion efficiency is effectively improved, the generated first similar sentences are screened, the standard database is updated through the newly-added similar sentences obtained through screening, and the data updating quality is guaranteed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of smart home / intelligent family technology, and more specifically, to a data expansion method, electronic device and storage medium based on a large model. Background Art

[0002] The customer service quality inspection system is a powerful support for enterprises to master customer service levels and improve service quality. Its core idea is to detect customer service hotlines and online conversation content to determine whether there are problems such as irrelevant answers, impolite attitudes, and repeated inquiries. The quality inspection system relies on natural language processing (NLP) technologies such as keyword matching to review the content of conversations through pre-defined semantic standard sentences and similar sentences. Among them, for the task of generalizing similar sentences, it currently relies entirely on manual exhaustive filling. This expansion method has many disadvantages. For example, it requires a lot of time and manpower costs, is extremely time-consuming and labor-intensive, and inefficient; secondly, the differences in filling between different quality inspectors may cause consistency problems, which in turn affects the detection accuracy of the quality inspection system; and with the continuous expansion of the standard sentence library, the manual exhaustive method will face the challenge of a surge in workload, which seriously restricts the scalability and maintainability of the system.

[0003] Accordingly, the art needs a new data expansion solution to solve the above problems. Summary of the invention

[0004] In order to overcome the above defects, the present application is proposed to solve or at least partially solve the above technical problems existing in relying on manual exhaustive completion of similar sentences.

[0005] In a first aspect, a data expansion method based on a large model is provided, the method comprising: acquiring conversation content; determining corpus information based on the conversation content, the corpus information comprising a standard sentence, a semantic interpretation corresponding to the standard sentence, and at least one of a semantic label encoding; generating a first similar sentence based on the corpus information and a pre-trained large model; screening the first similar sentence to obtain a newly added similar sentence corresponding to the standard sentence; and updating a standard database based on the newly added similar sentence.

[0006] In a technical solution of the above-mentioned large-model-based data expansion method, the determining of corpus information based on the conversation content includes: preprocessing the conversation content, wherein the preprocessing includes at least one of denoising, word segmentation, extracting standard sentences, and extracting context information; based on the preprocessed conversation content, determining at least one of a standard sentence, a semantic interpretation corresponding to the standard sentence, and a semantic label encoding.

[0007] In a technical solution of the above-mentioned large-model-based data expansion method, the first similar sentence is generated based on the corpus information and the pre-trained large model, including: determining whether the semantic label encoding corresponding to the standard sentence has been stored; and generating the first similar sentence based on the determination result and the standard sentence, the semantic interpretation corresponding to the standard sentence, and the pre-trained large model.

[0008] In a technical solution of the above-mentioned large-model-based data expansion method, the first similar sentence is generated based on the judgment result and the standard sentence, the semantic interpretation corresponding to the standard sentence, and the pre-trained large model, including: when the semantic label encoding corresponding to the standard sentence has been stored, obtaining the second similar sentence corresponding to the semantic label encoding; determining the question text based on a preset prompt word template and the standard sentence and the semantic interpretation corresponding to the standard sentence; generating the first similar sentence based on the question text, the second similar sentence and the pre-trained large model.

[0009] In a technical solution of the above-mentioned large-model-based data expansion method, the first similar sentence is generated based on the judgment result and the standard sentence, the semantic interpretation corresponding to the standard sentence, and the pre-trained large model, including: when the semantic label encoding corresponding to the standard sentence is not stored, determining the question text based on a preset prompt word template and the standard sentence, and the semantic interpretation corresponding to the standard sentence; generating the first similar sentence based on the question text and the pre-trained large model.

[0010] In a technical solution of the above-mentioned large model-based data expansion method, the first similar sentences are screened to obtain newly added similar sentences corresponding to the standard sentence, including: determining the semantic similarity between the first similar sentence and the standard sentence; obtaining a first similar sentence whose semantic similarity is greater than a preset threshold as a newly added similar sentence corresponding to the standard sentence; and / or visually displaying the first similar sentence; receiving a screening operation instruction for the first similar sentence; and using the first similar sentence selected by the screening operation instruction as the newly added similar sentence corresponding to the standard sentence.

[0011] In a technical solution of the above-mentioned large-model-based data expansion method, after updating the standard database based on the newly added similar sentences, the method also includes: visually displaying the updated standard database; receiving filtering operation instructions for the newly added similar sentences; and storing the newly added similar sentences selected by the filtering operation instructions as second similar sentences.

[0012] In a technical solution of the above-mentioned large-model-based data expansion method, the method also includes: obtaining operation records for the newly added similar sentences in the standard database, the operation records at least including retention records, modification records and deletion records; based on the operation records, iteratively optimizing the model parameters of the pre-trained large model.

[0013] In a second aspect, an electronic device is provided, comprising at least one processor; and a memory communicatively connected to the at least one processor; wherein a computer program is stored in the memory, and when the computer program is executed by the at least one processor, the method described in any one of the technical solutions of the above-mentioned large model-based data expansion method is implemented.

[0014] In a third aspect, a computer-readable storage medium is provided, in which a plurality of program codes are stored, wherein the program codes are suitable for being loaded and run by a processor to execute the method described in any one of the technical solutions of the above-mentioned large model-based data expansion method.

[0015] The above one or more technical solutions of the present application have at least one or more of the following beneficial effects:

[0016] The present application provides a data expansion method based on a large model, including: obtaining conversation content; determining corpus information based on the conversation content, the corpus information including at least one of a standard sentence, a semantic interpretation corresponding to the standard sentence, and a semantic tag encoding; generating a first similar sentence based on the corpus information and a pre-trained large model; screening the first similar sentence to obtain a newly added similar sentence corresponding to the standard sentence; and updating a standard database based on the newly added similar sentences. The present application determines accurate corpus information based on the conversation content, and with the help of the generalization ability of the large model, uses the corpus information to generate the first similar sentence, which significantly improves the efficiency of data expansion, and by screening the generated first similar sentence, the standard database is updated using the newly added similar sentences obtained by screening, thereby ensuring the quality of data update. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present application and, together with the description, serve to explain the principles of the present application.

[0018] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0019] Figure 1It is a flowchart of the main steps of a data expansion method based on a large model according to an embodiment of the present application;

[0020] Figure 2 is a flow chart of the application of a large model-based data expansion method according to an embodiment of the present application;

[0021] Figure 3 It is a schematic diagram of the main structure of an electronic device according to an embodiment of the present application. Reference numerals:

[0022] 11: processor; 12: memory. DETAILED DESCRIPTION

[0023] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0024] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0025] In the enterprise customer service quality inspection system, the quality inspection task is mainly to review the conversation content between the customer service and the user to determine whether there are any situations that have a negative impact on the user experience and service quality, including service failure. The existing workflow includes:

[0026] Collection and analysis of conversation content: Customer service quality inspection personnel regularly collect conversation content between customer service and users.

[0027] Determination of semantic standard sentences: Quality inspectors extract a series of representative semantic standard sentences from the conversation content based on business requirements and quality inspection standards. These standard sentences are used to identify the performance of customer service during the service process.

[0028] Manual process of similar sentence generation: Quality inspectors manually list sentences with semantic similarities to these standard sentences. This requires quality inspectors to manually fill in a large number of similar sentences that may appear based on their own experience and understanding. For example, for the standard sentence "Please wait a moment, I will help you check", quality inspectors may fill in similar sentences such as "Wait a moment, I will check your information", "Please wait a moment, I am checking", etc.

[0029] Application of the quality inspection system: Input standard sentences and similar sentences into the quality inspection system, use keyword matching and other technologies to make semantic judgments, and identify service failures in conversations.

[0030] The following are common quality inspection items in the customer service quality inspection system and their corresponding standard sentences and similar sentences: Table 1

[0031] These similar sentences based on the generalization of standard sentences currently rely entirely on manual exhaustive filling. This expansion method has many disadvantages. For example, it requires a lot of time and manpower costs, is extremely time-consuming and labor-intensive, and inefficient. Secondly, the differences in filling between different quality inspectors may cause consistency problems, which in turn affects the detection accuracy of the quality inspection system. Moreover, with the continuous expansion of the standard sentence library, the manual exhaustive method will face the challenge of a surge in workload, which seriously restricts the scalability and maintainability of the system.

[0032] To this end, the present application provides a data expansion method based on a large model, including: obtaining conversation content; determining corpus information based on the conversation content, the corpus information including at least one of a standard sentence, a semantic interpretation corresponding to the standard sentence, and a semantic label encoding; generating a first similar sentence based on the corpus information and a pre-trained large model; screening the first similar sentence to obtain a newly added similar sentence corresponding to the standard sentence; and updating a standard database based on the newly added similar sentences. The present application determines accurate corpus information based on the conversation content, and uses the corpus information with the generalization ability of the large model to generate a first similar sentence, which significantly improves the efficiency of data expansion, and by screening the generated first similar sentences and updating the standard database with the newly added similar sentences obtained by screening, the quality of data update is guaranteed.

[0033] See attached Figure 1 , Figure 1 FIG. 1 is a flow chart of the main steps of a data expansion method based on a large model according to an embodiment of the present application. Figure 1 As shown, the large model-based data expansion method in the embodiment of the present application mainly includes the following steps S101 to S105.

[0034] Step S101: Acquire the conversation content.

[0035] In this embodiment, the conversation content refers to the conversation content between the customer service and the user, such as hotline recordings and online chat records.

[0036] Step S102: determining corpus information based on the conversation content, wherein the corpus information includes at least one of a standard sentence, a semantic interpretation corresponding to the standard sentence, and a semantic tag encoding.

[0037] In this embodiment, the standard sentence refers to the standard phraseology sentence commonly used by customer service in actual conversations. Semantic interpretation refers to the interpretation of the semantics of the standard sentence, explaining which quality inspection scenario the standard sentence is for. Semantic label encoding refers to the encoding used to identify the standard sentence and its corresponding semantic interpretation.

[0038] Step S103: Generate a first similar sentence based on the corpus information and the pre-trained large model.

[0039] In this embodiment, the pre-trained large model can be a pre-trained large-scale natural language generation model, such as a GPT model. A first similar sentence similar to the standard sentence is generated through the determined corpus information and the pre-trained large model.

[0040] Step S104: screening the first similar sentences to obtain newly added similar sentences corresponding to the standard sentences.

[0041] In this embodiment, the first similar sentences are screened using a preset screening method, and the first similar sentences with higher similarity are selected as newly added similar sentences.

[0042] Step S105: updating the standard database based on the newly added similar sentences.

[0043] Based on the method described in steps S101 to S105 above, the present application determines accurate corpus information according to the conversation content, and with the help of the generalization ability of the large model, uses the corpus information to generate the first similar sentences, which significantly improves the efficiency of data expansion. By screening the generated first similar sentences and using the newly added similar sentences obtained by screening to update the standard database, the quality of data update is guaranteed.

[0044] The above steps S101 to S105 are further explained below.

[0045] With respect to step S101, the conversation content is obtained.

[0046] Specifically, collect the conversation content between customer service and users, such as hotline recordings and online chat records.

[0047] With respect to step S102, in one embodiment, determining the corpus information based on the conversation content includes: preprocessing the conversation content, wherein the preprocessing includes at least one of denoising, word segmentation, extracting standard sentences, and extracting context information; and determining at least one of a standard sentence, a semantic interpretation corresponding to the standard sentence, and a semantic tag encoding based on the preprocessed conversation content.

[0048] Specifically, the collected conversation content is preprocessed, and the preprocessing may include denoising and word segmentation of the conversation content, extracting standard sentences from the conversation content, and extracting context information of the standard sentences.

[0049] De-noising refers to removing unnecessary information such as background noise and interference sounds that may be contained in hotline recordings, or deleting idle chat and repeated information that may be contained in online chat records.

[0050] For the collected hotline recordings, this embodiment also includes performing text conversion processing on the hotline recordings to obtain text information corresponding to the hotline recordings, and using the text information corresponding to the hotline recordings as the conversation content.

[0051] Word segmentation refers to dividing the continuous text in the conversation content into word sequences through word segmentation technology.

[0052] Extracting standard sentences refers to identifying the content of a conversation and extracting standard sentences that are commonly used in actual conversations. Natural language processing (NLP) technology, machine learning technology, rule matching, or manual intervention can be used to determine the standard sentences.

[0053] Extracting context information refers to obtaining context information related to the standard sentence, such as the initiator, receiver, time, topic, context, etc. of the conversation, which helps to understand the semantics of the standard sentence more accurately.

[0054] Based on the preprocessed conversation content, including word sequences, standard sentences, and contextual information, the conversation content is semantically analyzed and understood through a pre-trained model to determine the semantic interpretation of the standard sentence. The pre-trained model can identify key information in the conversation content and understand the quality inspection scenario it targets.

[0055] According to the semantic interpretation of the standard sentence and the predefined semantic tag set, the semantic tag of the standard sentence is determined, and the semantic tag is converted into a coding form, which can be a number, a string, or a structured data (such as JSON, XML, etc.). Alternatively, the semantic tag coding of the standard sentence is determined by manual labeling. The semantic tag coding can identify the semantic meaning of the standard sentence. One semantic tag coding corresponds to a standard sentence and the first similar sentence, the second similar sentence, and the newly added similar sentence corresponding to the standard sentence.

[0056] The above is the description of step S101 and step S102. The following is a further description of step S103.

[0057] For step S103, in one embodiment, generating a first similar sentence based on the corpus information and the pre-trained large model includes: determining whether the semantic label encoding corresponding to the standard sentence has been stored; generating a first similar sentence based on the determination result and the standard sentence, the semantic interpretation corresponding to the standard sentence, and the pre-trained large model.

[0058] Specifically, the pre-trained large model has a corresponding cache module, which is specifically used to store the second similar sentences obtained after screening the first similar sentences generated by the pre-trained large model, and the semantic label encoding corresponding to the second similar sentences.

[0059] Use the semantic label encoding corresponding to the standard sentence as the query key to find out whether there is a match in the cache module of the pre-trained large model. If the same semantic label encoding exists in the cache module, it means that the standard sentence has been used to generate a similar sentence before.

[0060] If the same semantic label encoding does not exist in the cache module, it means that this is the first time that the standard sentence is used to generate a similar sentence.

[0061] Based on the judgment result and the standard sentence, the semantic interpretation of the standard sentence and the pre-trained large model, a first similar sentence corresponding to the standard sentence is generated. The first similar sentence can be multiple sentences similar to the standard sentence. The first similar sentence and the standard sentence have the same semantic label encoding.

[0062] In one embodiment, the generating of the first similar sentence based on the judgment result and the standard sentence, the semantic interpretation corresponding to the standard sentence, and the pre-trained large model includes: obtaining the second similar sentence corresponding to the semantic label code when the semantic label code corresponding to the standard sentence has been stored; determining the question text based on a preset prompt word template and the standard sentence, and the semantic interpretation corresponding to the standard sentence; generating the first similar sentence based on the question text, the second similar sentence, and the pre-trained large model.

[0063] Specifically, when the cache module contains the same semantic label code as that corresponding to the standard sentence, a second similar sentence corresponding to the semantic label code stored in the cache module is obtained. The second similar sentence can serve as additional supplementary corpus or reference sample to help the pre-trained large model better understand the semantics of the standard sentence, so as to enhance the quality and relevance of the generated first similar sentence.

[0064] The question text is determined based on the preset prompt word template, the standard sentence, and the semantic interpretation of the standard sentence. For example, the standard sentence is "Hello, how can I help you?", and the semantic interpretation of the standard sentence is "no polite words are used, no honorifics such as "you" or "hello" are used".

[0065] The question text generated based on the preset prompt word template, standard sentence, and the semantic interpretation corresponding to the standard sentence can be "You are an expert in semantic generalization, and are good at effectively understanding the "standard sentence" I input and its "semantic interpretation" to generate similar sentences. The standard sentence is {Hello, how can I help you?}, and the semantic interpretation is {no polite words are used, and no honorifics such as "you, hello" are used}. Please combine the "standard sentence" and "semantic interpretation" to help me generate 5 similar sentences."

[0066] Then, the question text and the second similar sentence are provided to the pre-trained large model. The question text prompts the model what to do, that is, specific instructions for generating similar sentences. The second similar sentence provides a generation example of the standard sentence, which helps the model better understand the semantics of the standard sentence and generate a more appropriate first similar sentence based on it.

[0067] For example, the standard sentence is "Hello, how can I help you?" and a second similar sentence is "Hello, how can I help you?" and the question text is "Please combine the "standard sentence" and "semantic interpretation" to help me generate 5 similar sentences", then this information can be input into the pre-trained large model together, and the pre-trained large model will generate 5 new sentences with similar meanings to the standard sentence but different wording, such as: "Hello, how can I help you?", "Hello, what can I do for you?", "Hello, what kind of help do you need?", "Hello, do you have any questions that need to be answered?", "Hello, is there anything I can do for you?".

[0068] In one embodiment, the generating of the first similar sentence based on the judgment result and the standard sentence, the semantic interpretation corresponding to the standard sentence, and the pre-trained large model includes: when the semantic label encoding corresponding to the standard sentence is not stored, determining the question text based on the preset prompt word template and the standard sentence, and the semantic interpretation corresponding to the standard sentence; generating the first similar sentence based on the question text and the pre-trained large model.

[0069] Specifically, when the cache module does not contain the same semantic label code as the semantic label code corresponding to the standard sentence, the question text is generated only based on the preset prompt word template and the standard sentence, and the semantic interpretation corresponding to the standard sentence, and the generated question text is input into the pre-trained large model. The pre-trained large model generates multiple first similar sentences that are semantically similar to the standard sentence according to the content of the question text.

[0070] The above is the description of step S103 , and the following is a further description of step S104 .

[0071] With respect to step S104, in one embodiment, the screening of the first similar sentences to obtain a newly added similar sentence corresponding to the standard sentence includes: determining the semantic similarity between the first similar sentence and the standard sentence; obtaining a first similar sentence having a semantic similarity greater than a preset threshold as a newly added similar sentence corresponding to the standard sentence; and / or visually displaying the first similar sentence; receiving a screening operation instruction for the first similar sentence; and using the first similar sentence selected by the screening operation instruction as a newly added similar sentence corresponding to the standard sentence.

[0072] Specifically, similarity calculation techniques, such as cosine similarity, Jaccard similarity or more complex neural network models, are used to calculate the semantic similarity between each first similar sentence and the standard sentence. Based on the calculated similarity scores, the first similar sentences with semantic similarity greater than a preset threshold are retained as new similar sentences corresponding to the standard sentence.

[0073] Alternatively, an interactive interface is generated based on multiple first similar sentences to help users visually view these first similar sentences and filter the first similar sentences. Then, the user's filtering operation instructions for the first similar sentences are obtained, and according to the user's filtering operation instructions, the selected first similar sentences are used as new similar sentences of the standard sentences.

[0074] In addition, the similarity calculation technology can be used to first filter the first similar sentences to obtain the first similar sentences whose similarity is greater than a preset threshold, and then these first similar sentences with higher similarity are displayed to the user in a visual manner, and the user's filtering operation instructions for these first similar sentences filtered by the similarity calculation technology are obtained. According to the user's filtering operation, the selected first similar sentences are used as new similar sentences of the standard sentence.

[0075] The above is the description of step S104 , and the following is a further description of step S105 .

[0076] With respect to step S105 , the standard database is updated based on the newly added similar sentences.

[0077] Specifically, the structure of the standard database may include information such as quality inspection items, standard sentences, similar sentences, and semantic tag codes. In the standard database, the semantic tag codes of the newly added similar sentences are used to locate the standard sentences that match the meaning of the newly added similar sentences, and the newly added similar sentences are added to the similar sentence part of the corresponding standard sentences. After the standard database is updated, more similar sentences are added, so that the standard database can cover more expressions and variants, better support customer service quality inspection work, and improve the comprehensiveness and accuracy of quality inspection.

[0078] In one embodiment, after updating the standard database based on the newly added similar sentences, the method further includes: visually displaying the updated standard database; receiving filtering operation instructions for the newly added similar sentences; and storing the newly added similar sentences selected by the filtering operation instructions as second similar sentences.

[0079] Specifically, after the standard database is updated, the updated standard database is visualized, and the user's filtering operation instructions for newly added similar sentences in the standard database are obtained. A number of newly added similar sentences (such as 5) with the most recent database time are added to the selected newly added similar sentences as second similar sentences, and stored in the cache module of the pre-trained large model. These second similar sentences are used as supplementary corpus when the user generates the first similar sentences again based on the same semantic label encoding.

[0080] In one embodiment, the method also includes: obtaining operation records for the newly added similar sentences in the standard database, the operation records at least including retention records, modification records and deletion records; based on the operation records, iteratively optimizing the model parameters of the pre-trained large model.

[0081] Specifically, various operation records for adding similar sentences to the standard database are collected, including but not limited to retaining records, modifying records, and deleting records. Based on these operation records, the effect of the pre-trained large model in generating similar sentences is analyzed. According to the analysis results, the pre-trained large model is iteratively optimized to continuously improve the quality of similar sentence generation.

[0082] For example, retaining records are defined as positive data (+1), deleting records are defined as negative data (-1), and modifying records are positioned as neutral data (0). Different weights are assigned to each type of record based on actual conditions. For example, retaining records have a higher weight, modifying records have a lower weight, and deleting records have a lower weight than modifying records.

[0083] Use statistical methods to analyze the quantitative ratio of positive data, negative data, and neutral data, and analyze the specific reasons for negative data and neutral data, such as whether they were deleted or modified due to grammatical errors, inaccurate semantics, inconsistency with the context, etc.

[0084] Based on the analysis results, determine where the pre-trained large model needs to be improved and adjust the model parameters, such as changing the way of word embedding, adjusting the weight of the attention mechanism, changing the loss function, etc. In this way, user feedback can be continuously collected and effectively used to continuously optimize the performance of the pre-trained large model and continuously improve the quality of similar sentence generation.

[0085] See attached Figure 2 , Figure 2It is a flow chart of the application of a data expansion method based on a large model according to an embodiment of the present application. In this embodiment, the data expansion method based on a large model is applied to a quality inspection system and a large model cloud service.

[0086] Specifically, the quality inspection system collects conversation content including hotline recordings and online chat records, and preprocesses the conversation content, including at least one of denoising, word segmentation, extracting standard sentences, and extracting context information. Then, based on the processed conversation content, the standard sentences, the semantic interpretations and semantic label codes corresponding to the standard sentences are determined, and the standard sentences, the semantic interpretations and semantic label codes of the standard sentences are sent to the large model cloud service.

[0087] The large model cloud service determines whether the semantic label code of the standard sentence has been stored in the cache module of the pre-trained large model. If so, the corresponding second similar sentence in the cache module is obtained according to the semantic label code, and the second similar sentence is provided to the pre-trained large model. The question text is generated according to the standard sentence, the semantic interpretation of the standard sentence and the preset prompt word template, and is input into the pre-trained large model. If not, the question text is directly generated based on the standard sentence, the semantic interpretation of the standard sentence and the preset prompt word template, and is input into the pre-trained large model.

[0088] The pre-trained large model generates first similar sentences, determines the semantic similarity between each first similar sentence and the standard sentence, eliminates the first similar sentences with low semantic similarity, obtains the first similar sentences with semantic similarity greater than a preset threshold, and uses them as new similar sentences corresponding to the standard sentence, and adds the new similar sentences to the standard database.

[0089] After the standard database is updated, the standard database is visualized, and the user's filtering operation instructions for the newly added similar sentences in the standard database are obtained. The newly added similar sentences with the most recent addition time are selected as the second similar sentences and sent back to the cache module of the pre-trained large model for storage.

[0090] Collect various operation records for adding similar sentences to the standard database, including but not limited to retaining records, modifying records, and deleting records. Based on these operation records, analyze the effect of the pre-trained large model in generating similar sentences. According to the analysis results, iteratively optimize the pre-trained large model to improve the quality of similar sentence generation.

[0091] It should be pointed out that although the various steps in the above embodiments are described in a specific order, those skilled in the art can understand that in order to achieve the effect of the present application, different steps do not have to be executed in such an order. They can be executed simultaneously (in parallel) or in other orders. These adjusted schemes are equivalent to the technical schemes described in this application, and therefore will also fall within the scope of protection of this application.

[0092] It is understood by those skilled in the art that all or part of the processes in the method for implementing the above-mentioned embodiment of the present application can also be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable storage medium may include: any entity or device, medium, U disk, mobile hard disk, disk, optical disk, computer memory, read-only memory, random access memory, electrical carrier signal, telecommunication signal and software distribution medium, etc. that can carry the computer program code.

[0093] Another aspect of the present application provides an electronic device.

[0094] See attached Figure 3 , Figure 3 exemplarily shows that the processor 11 and the memory 12 are connected in communication via a bus. In an embodiment of an electronic device according to the present application, the electronic device may include at least one processor; and a memory connected in communication with the at least one processor; wherein a computer program is stored in the memory, and when the computer program is executed by the at least one processor, the data expansion method based on the large model described in any of the above embodiments is implemented.

[0095] Another aspect of the present application also provides a computer-readable storage medium.

[0096] In an embodiment of a computer-readable storage medium according to the present application, the computer-readable storage medium may be configured to store a program for executing the large-model-based data expansion method of the above-mentioned method embodiment, and the program may be loaded and run by a processor to implement the above-mentioned large-model-based data expansion method. For ease of explanation, only the parts related to the embodiment of the present application are shown. For specific technical details not disclosed, please refer to the method section of the embodiment of the present application. The computer-readable storage medium may be a storage device formed by various electronic devices. Optionally, the computer-readable storage medium in the embodiment of the present application is a non-temporary computer-readable storage medium.

[0097] Further, it should be understood that the method of the present application is widely used in smart home (Smart Home), smart home, smart home device ecology, smart residential (Intelligence House) ecology and other whole-house intelligent digital control application scenarios. The above-mentioned network may include but is not limited to at least one of the following: wired network, wireless network. The above-mentioned wired network may include but is not limited to at least one of the following: wide area network, metropolitan area network, local area network, and the above-mentioned wireless network may include but is not limited to at least one of the following: WI FI (Wireless Fidelity, Wireless Fidelity), Bluetooth. Terminal devices may not be limited to PCs, mobile phones, tablet computers, smart air conditioners, smart range hoods, smart refrigerators, smart ovens, smart stoves, smart washing machines, smart water heaters, smart washing equipment, smart dishwashers, smart projection equipment, smart TVs, smart clothes drying racks, smart curtains, smart audio and video, smart sockets, smart speakers, smart speakers, smart fresh air equipment, smart kitchen and bathroom equipment, smart bathroom equipment, smart sweeping robots, smart window cleaning robots, smart mopping robots, smart air purification equipment, smart steamers, smart microwave ovens, smart kitchen treasures, smart purifiers, smart water dispensers, smart door locks, etc.

[0098] So far, the technical solution of the present application has been described in conjunction with an embodiment shown in the accompanying drawings, but it is easy for those skilled in the art to understand that the protection scope of the present application is obviously not limited to these specific embodiments. Without departing from the principles of the present application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will fall within the protection scope of the present application.

Claims

1. A data expansion method based on large models, It is characterized in that The method comprises: Get the session content; Determine corpus information based on the conversation content, the corpus information including at least one of a standard sentence, a semantic interpretation corresponding to the standard sentence, and a semantic tag encoding; Based on the corpus information and the pre-trained large model, generate a first similar sentence; Screening the first similar sentences to obtain newly added similar sentences corresponding to the standard sentences; The standard database is updated based on the newly added similar sentences.

2. The data expansion method based on a large model according to claim 1, It is characterized in that The determining of corpus information based on the conversation content includes: Preprocessing the conversation content, wherein the preprocessing includes at least one of denoising, word segmentation, extracting standard sentences, and extracting context information; Based on the preprocessed conversation content, at least one of a standard sentence, a semantic interpretation corresponding to the standard sentence, and a semantic label encoding is determined.

3. The data expansion method based on a large model according to claim 1, It is characterized in that The generating a first similar sentence based on the corpus information and the pre-trained large model comprises: Determining whether the semantic tag code corresponding to the standard sentence has been stored; Based on the judgment result and the standard sentence, the semantic interpretation corresponding to the standard sentence, and the pre-trained large model, a first similar sentence is generated.

4. The data expansion method based on a large model according to claim 3, It is characterized in that The generating a first similar sentence based on the judgment result and the standard sentence, the semantic interpretation corresponding to the standard sentence, and the pre-trained large model includes: When the semantic label code corresponding to the standard sentence has been stored, obtaining a second similar sentence corresponding to the semantic label code; Determine the question text based on the preset prompt word template and the standard sentence and the semantic interpretation corresponding to the standard sentence; Based on the question text, the second similar sentence and the pre-trained large model, a first similar sentence is generated.

5. The data expansion method based on a large model according to claim 3, It is characterized in that The generating a first similar sentence based on the judgment result and the standard sentence, the semantic interpretation corresponding to the standard sentence, and the pre-trained large model includes: In the case where the semantic tag code corresponding to the standard sentence is not stored, determining the question text based on a preset prompt word template and the standard sentence and the semantic interpretation corresponding to the standard sentence; Based on the question text and the pre-trained large model, a first similar sentence is generated.

6. The data expansion method based on a large model according to claim 1, It is characterized in that The screening of the first similar sentences to obtain the newly added similar sentences corresponding to the standard sentences includes: Determining the semantic similarity between the first similar sentence and the standard sentence; Acquire a first similar sentence whose semantic similarity is greater than a preset threshold as a newly added similar sentence corresponding to the standard sentence; and / or, Visually displaying the first similar sentence; receiving a filtering operation instruction for the first similar sentence; The first similar sentence selected by the screening operation instruction is used as a newly added similar sentence corresponding to the standard sentence.

7. The data expansion method based on a large model according to claim 1, It is characterized in that After updating the standard database based on the newly added similar sentences, the method further includes: Visually displaying the updated standard database; receiving a filtering operation instruction for the newly added similar sentences; The newly added similar sentence selected by the screening operation instruction is stored as the second similar sentence.

8. The data expansion method based on a large model according to any one of claims 1 to 7, It is characterized in that The method further comprises: Obtaining an operation record for the newly added similar sentence in the standard database, wherein the operation record at least includes a retention record, a modification record, and a deletion record; Based on the operation record, the model parameters of the pre-trained large model are iteratively optimized.

9. An electronic device comprising at least one processor and at least one memory, wherein the memory is adapted to store a plurality of program codes, It is characterized in that The program code is suitable for being loaded and run by the processor to execute the large model-based data expansion method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a plurality of program codes stored therein, It is characterized in that The program code is suitable for being loaded and run by a processor to execute the large model-based data expansion method according to any one of claims 1 to 8.