Method and device for model processing, equipment and medium

Abstract

The invention provides a method, a device, equipment and a medium for model processing. In one method, a first knowledge domain of a generic knowledge domain involved in a processing model is determined. A first set of reference queries associated with the first field of knowledge is generated. A first set of reference answers matching the first set of reference queries is obtained from a repository associated with the processing model, the repository relating to at least the first field of knowledge. The processing model is updated with the first set of reference queries and the first set of reference answers. With exemplary implementations of the present disclosure, knowledge reserves in various subdivided knowledge domains can be continuously added to the processing model. In this way, the performance and accuracy of the processing model can be improved, and therefore the query result output by the processing model is more accurate.

Description

Technical Field

[0001] Exemplary implementations of the present disclosure generally relate to machine learning technology, and in particular, to methods, devices, apparatuses, and computer-readable storage media for model processing. Background Art

[0002] Machine learning technology has been widely used to perform a variety of processing tasks. For example, a variety of processing models can be constructed to perform various processing tasks. Queries can be sent to the processing model and query results from the processing model can be received. The processing model should have knowledge of many knowledge fields in order to provide accurate query results for queries in different knowledge fields. However, the processing model cannot cover all knowledge fields, which leads to the fact that in some cases (for example, when involving a certain subdivided knowledge field), the processing model cannot provide accurate query results. At this time, it is expected that a more accurate and comprehensive processing model can be obtained. Summary of the invention

[0003] In a first aspect of the present disclosure, a model processing method is provided. In the method, a first knowledge domain in a general knowledge domain involved in a processing model is determined. A first set of reference queries associated with the first knowledge domain is generated. A first set of reference answers matching the first set of reference queries is obtained from a repository associated with the processing model, the repository at least involving the first knowledge domain. The processing model is updated using the first set of reference queries and the first set of reference answers.

[0004] In a second aspect of the present disclosure, a model processing device is provided. The device includes: a determination module configured to determine a first knowledge domain in the general knowledge domain involved in the processing model; a generation module configured to generate a first set of reference queries associated with the first knowledge domain; an acquisition module configured to acquire a first set of reference answers matching the first set of reference queries from a repository associated with the processing model, the repository at least involving the first knowledge domain; and an update module configured to update the processing model using the first set of reference queries and the first set of reference answers.

[0005] In a third aspect of the present disclosure, an electronic device is provided. The electronic device includes: at least one processing unit; and at least one memory, the at least one memory is coupled to the at least one processing unit and stores instructions for execution by the at least one processing unit, and when the instructions are executed by the at least one processing unit, the electronic device executes the method according to the first aspect of the present disclosure.

[0006] In a fourth aspect of the present disclosure, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the processor implements the method according to the first aspect of the present disclosure.

[0007] It should be understood that the content described in this content section is not intended to limit the key features or important features of the implementation of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easy to understand through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] In the following, in conjunction with the accompanying drawings and with reference to the following detailed description, the above and other features, advantages and aspects of various implementations of the present disclosure will become more apparent. In the accompanying drawings, the same or similar figure labels represent the same or similar elements, wherein:

[0009] Figure 1 A block diagram showing an application environment of a processing model according to an exemplary implementation of the present disclosure;

[0010] Figure 2 A block diagram showing a process of model processing according to some implementations of the present disclosure;

[0011] Figure 3 A block diagram illustrating query types according to some implementations of the present disclosure;

[0012] Figure 4 A block diagram for generating reference samples according to some implementations of the present disclosure is shown;

[0013] Figure 5 A block diagram for generating reference samples associated with a given knowledge domain according to some implementations of the present disclosure is shown;

[0014] Figure 6 A block diagram is shown for generating reference samples associated with various knowledge domains within a general knowledge domain according to some implementations of the present disclosure;

[0015] Figure 7 A block diagram for updating a processing model according to some implementations of the present disclosure is shown;

[0016] Figure 8 A block diagram is shown for processing a user query using an updated processing model according to some implementations of the present disclosure;

[0017] Fig. 9 A flowchart showing a model processing method according to some implementations of the present disclosure is shown;

[0018] Fig.10 A block diagram showing a model processing apparatus according to some implementations of the present disclosure; and

[0019] Fig.11 A block diagram of a device capable of implementing multiple implementations of the present disclosure is shown. DETAILED DESCRIPTION

[0020] The implementation of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although certain implementations of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be interpreted as being limited to the implementations described herein. On the contrary, these implementations are provided to provide a more thorough and complete understanding of the present disclosure. It should be understood that the drawings and implementations of the present disclosure are only for exemplary purposes and are not intended to limit the scope of protection of the present disclosure.

[0021] In the description of the implementation methods of the present disclosure, the term "including" and similar terms should be understood as open inclusion, that is, "including but not limited to". The term "based on" should be understood as "based at least in part on". The term "an implementation" or "the implementation" should be understood as "at least one implementation". The term "some implementations" should be understood as "at least some implementations". The following may also include other explicit and implicit definitions. As used herein, the term "model" can represent the association relationship between various data. For example, the above-mentioned association relationship can be obtained based on a variety of technical solutions currently known and / or to be developed in the future.

[0022] It is understandable that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and relevant provisions.

[0023] It is understandable that before using the technical solutions disclosed in the embodiments of the present disclosure, the types, scope of use, usage scenarios, etc. of the personal information involved in the present disclosure should be informed to the user and the user's authorization should be obtained in an appropriate manner in accordance with relevant laws and regulations.

[0024] For example, in response to receiving an active request from a user, a prompt message is sent to the user to clearly prompt the user that the operation requested to be performed will require obtaining and using the user's personal information. Thus, the user can autonomously choose whether to provide personal information to software or hardware such as an electronic device, application, server, or storage medium that performs the operation of the technical solution of the present disclosure according to the prompt message.

[0025] As an optional but non-limiting implementation, in response to receiving an active request from the user, the prompt information is sent to the user in a manner such as a pop-up window, in which the prompt information can be presented in text form. In addition, the pop-up window can also carry a selection control for the user to choose "agree" or "disagree" to provide personal information to the electronic device.

[0026] It is understandable that the above notification and the process of obtaining user authorization are merely illustrative and do not constitute a limitation on the implementation of the present disclosure. Other methods that meet relevant laws and regulations may also be applied to the implementation of the present disclosure.

[0027] The term "in response to" as used herein refers to a state in which a corresponding event occurs or a condition is satisfied. It will be understood that the timing of executing a subsequent action executed in response to the event or condition is not necessarily strongly related to the time when the event occurs or the condition is satisfied. For example, in some cases, the subsequent action may be executed immediately when the event occurs or the condition is satisfied; while in other cases, the subsequent action may be executed some time after the event occurs or the condition is satisfied.

[0028] Example Environment

[0029] Machine learning techniques have been widely used to perform a variety of processing tasks. For example, a variety of processing models can be built to perform various processing tasks. Queries can be sent to the processing models and query results can be received from the processing models. See Figure 1 Describes an application environment according to an example implementation of the present disclosure. Figure 1 A block diagram 100 of an application environment of a processing model according to an exemplary implementation of the present disclosure is shown. Figure 1 As shown, a processing model 130 may be pre-built, and a query 110 may be input to the processing model 130. For example, in the context of a language processing model, the query 110 may be, for example, "what is a programming language", etc. At this time, the processing model 130 may provide a query result 120 matching the query 110, for explaining the meaning of a programming language.

[0030] The processing model 130 should have knowledge of many knowledge fields in order to provide accurate query results for queries in different knowledge fields. However, the processing model 130 usually cannot cover all knowledge fields, which leads to the fact that the processing model 130 cannot provide accurate query results in some cases (for example, when it involves a certain subdivided knowledge field). For example, the results output by the processing model 130 may not completely match the input data. For example, when calling a large deep learning model (such as a large language model (LLM) or other models, etc.), the query results may not be accurate.

[0031] Generally speaking, for large pre-trained models, they have been trained with a large amount of data and can answer and solve most knowledge question-answering tasks in the public domain, but their performance in a certain niche (also known as a vertical field) may not be satisfactory. The main reason is usually that no training data for the relevant niche was obtained during the training process. In order to obtain a model that performs better in a certain niche, it is necessary to use the data from the niche to fine-tune the pre-trained model. Therefore, the integration of niche knowledge is crucial, which helps to improve the model's ability to understand and handle problems in a specific field. The fine-tuning process is of great significance, which is reflected in the following aspects.

[0032] Domain expertise and depth: Large pre-trained models that incorporate domain knowledge can demonstrate higher expertise and depth when understanding and addressing specific domain problems. Compared with pre-trained general models, such models can more accurately understand domain terminology, concepts, contextual relationships, etc.

[0033] Improved model performance: By incorporating domain knowledge into the model, the model can better learn domain expertise during the fine-tuning phase, thereby achieving better performance on related tasks in the domain.

[0034] Practical application value: In real-world scenarios, many applications are concentrated in specific niches, and models with niche domain knowledge can better meet the needs of these specific applications. For example, models that incorporate software development domain knowledge can provide more accurate and high-quality results when performing tasks such as terminology queries and error diagnosis.

[0035] Reduce generalization error: The integration of subdivided domain knowledge helps improve the model's domain generalization ability. This is because the model may better understand and handle related problems under the knowledge system of a specific field, thereby reducing generalization error.

[0036] Enhanced model robustness: With the support of segmented domain knowledge, pre-trained models can better cope with some challenging problems. When faced with complex, semantically similar or highly ambiguous tasks, models with domain knowledge will show higher robustness.

[0037] In order to expand the domain knowledge of the pre-trained model, a method has been proposed to manually collect and construct training data in the subdivided knowledge domain and then perform training. However, this technical solution involves a lot of manual labor and is inefficient. At this time, it is expected to train the machine learning model in a more effective way and enable the machine learning model to process queries in various knowledge domains with higher performance and accuracy.

[0038] Overview of Model Processing

[0039] In order to at least partially solve the deficiencies in the prior art, according to an exemplary implementation of the present disclosure, a model processing method is proposed. In summary, the knowledge in the subdivided knowledge field can be used to construct reference samples and then use these reference samples to train the pre-trained processing model 130. In this way, the knowledge reserve of the processing model 130 in each subdivided knowledge field can be further enriched on the basis of the knowledge in the general knowledge field originally possessed by the processing model 130. In this way, the performance and accuracy of the processing model 130 can be improved, so that the query results output by the processing model 130 are more consistent with the query input by the user.

[0040] See also Figure 2 An overview of an exemplary implementation of the present disclosure is described. Figure 2 A block diagram 200 is shown of a process for model processing according to some implementations of the present disclosure. Figure 2 As shown, the processing model 130 may be a pre-trained model and may have various knowledge in the general knowledge domain 210. Specifically, the general knowledge domain 210 may be further subdivided into multiple knowledge domains. For example, the knowledge domain 212 relates to software development, ..., the knowledge domain 214 relates to hardware design, and so on. Similar processing may be performed for each subdivided knowledge domain to obtain reference samples in each knowledge domain. Further, the obtained reference samples may be used to update the processing model 130.

[0041] In the following, only the knowledge domain 212 is used as an example to provide a description. Figure 2 As shown, a knowledge domain 212 (for ease of description, may be referred to as a first knowledge domain) in the general knowledge domain 210 involved in the processing model 130 may be determined. A first set of reference queries 230 associated with the knowledge domain 212 may be generated. It should be understood that a set of reference queries 230 may include one or more reference queries. For example, questions from users in the field of software development may be determined, and these questions may be used to construct reference queries. Alternatively and / or additionally, reference queries may be constructed by data retrieval or the like.

[0042] The reference query may be expressed using natural language, for example, “What is a programming language?” Alternatively and / or additionally, the reference query may have a predetermined format, for example, “Please explain: programming language.” In this case, the portion “Please explain” before the colon indicates the type of action to be performed, and the portion “Programming language” after the colon indicates the object to be explained, and so on.

[0043] Further, a first group of reference answers 240 matching the first group of reference queries 230 may be obtained from the repository 220. Here, the repository 220 is associated with the process model 130, and may include knowledge in one or more knowledge domains in the general knowledge domain 210. Specifically, the repository 220 may at least involve the knowledge domain 212. In this way, answers matching each reference query in the first group of reference queries 230 may be obtained from the repository 220, respectively.

[0044] Assuming that one of the reference queries in the first set of reference queries 230 is “Please explain: programming language”, the results matching the reference query can be searched in the repository 220. For example, “Programming language is…”. Other reference queries in the first set of reference queries 230 can be processed in a similar manner, and corresponding reference answers can be obtained. Further, the processing model 130 can be updated using the first set of reference queries 230 and the first set of reference answers 240.

[0045] By using the exemplary implementation of the present disclosure, knowledge reserves in various subdivided knowledge fields can be continuously added to the processing model 130. In this way, the performance and accuracy of the processing model 130 can be improved, so that the query results output by the processing model 130 are more consistent with the query input by the user.

[0046] Detailed process of model processing

[0047] An overview of an example implementation according to the present disclosure has been described, and more details for determining reference queries and reference answers will be provided below. It should be understood that queries may involve multiple types, in which case corresponding reference queries can be constructed for each query type. First, it is necessary to set what kind of problems the updated processing model 130 is expected to solve. Further, based on this setting, relevant data can be collected and corresponding reference queries can be constructed. Taking the field of software development as an example, it is assumed that the questions that the processing model 130 is expected to answer involve the following:

[0048] Technical query: mainly for the query of various latest technologies in the relevant technical fields, such as concept explanation (for example, what is a programming language, etc.), information retrieval (for example, finding relevant technical documents), knowledge summary (for example, key analysis), etc. Case analysis: for example, for a specific case encountered by a user, a solution to the case is provided according to the corresponding technology. Problem diagnosis: mainly for the corresponding problems encountered by the user during the software development process, reasonable diagnosis is provided, etc. Based on the setting of the above questions, the corresponding answers can be found in the repository 220.

[0049] According to an example implementation of the present disclosure, the query may involve different types. Figure 3 Describes various query types in the context of the present disclosure, the Figure 3 A block diagram 300 of query types according to some implementations of the present disclosure is shown. Figure 3 As shown, the query type 310 may include at least any one of the following: an explanation type 311 , a summary type 312 , a description type 313 , an analysis type 314 , a diagnosis type 315 , a rewrite type 316 , a generation type 317 , and the like.

[0050] According to an example implementation of the present disclosure, Figure 3 The query type 310 shown can vary depending on business needs. Figure 3 As shown, the explanation type 311 and the description type 312 are mainly used to inject segmented domain knowledge into the processing model 130, and enable the processing model 130 to utilize the knowledge of these segmented domains during the reasoning process. This query type generally involves enabling the processing model 130 to learn the corresponding domain knowledge through instructions such as definition, explanation, and description. In the case of performing the fine-tuning process using such reference queries and corresponding reference answers, the updated processing model 130 can handle tasks of the explanation and / or description types.

[0051] The summary type 313 is mainly used to guide the processing model 130 to summarize the knowledge of the subdivided domain, and allows users to quickly obtain the desired query results without reading a large number of documents. When the training process is performed using such reference queries and corresponding reference answers, the updated processing model 130 can handle summary-type tasks. The analysis type 314 is mainly through the Chain of Thought (abbreviated as CoT) method, so that the processing model 130 can obtain a certain reasoning ability, and then obtain better reasoning results in specific case analysis. When the fine-tuning process is performed using such reference queries and corresponding reference answers, the updated processing model 130 can handle analysis-type tasks.

[0052] The diagnosis type 315 is mainly used to guide the processing model 130 to find the root cause of the problem through cases, that is, to train the reasoning ability and inductive summary ability of the processing model 130. In the case of performing a fine-tuning process using such reference queries and corresponding reference answers, the updated processing model 130 can handle diagnosis-type tasks. The rewrite type 316 is used to guide the processing model 130 to complete the rewrite in a different style. In the case of performing a fine-tuning process using such reference queries and corresponding reference answers, the updated processing model 130 can handle rewrite-type tasks. The generation type 317 is mainly used to guide the processing model 130 to generate the content required by the user, for example, generate text, images, and so on. In the case of performing a fine-tuning process using such reference queries and corresponding reference answers, the updated processing model 130 can handle generation-type tasks.

[0053] Among the above types, the explanation type and description type are the basic methods of acquiring domain knowledge, the analysis type, summary type and diagnosis type are the basic learning tasks for the model to reason about domain knowledge, and the generation type can be a learning task that can be output to the user based on the learned domain knowledge. Different query types correspond to different types of tasks, and the processing model can continuously strengthen its multi-faceted processing capabilities based on these tasks.

[0054] According to an exemplary implementation of the present disclosure, in each knowledge domain, a target type (e.g., Figure 3 ), and then use the template associated with the target type to generate a first set of reference queries. Different numbers of reference queries can be constructed. For example, hundreds or thousands of reference queries can be constructed, or only a few or dozens of reference queries can be constructed. Each reference query needs to fully incorporate the knowledge of the subdivided field, and the processing model 130 can learn specific processing capabilities through these reference queries and corresponding reference answers.

[0055] According to an example implementation of the present disclosure, a template of a predetermined format may be constructed for each query type, and then a reference query may be generated by filling in the template. For example, the template may have a format as shown in Table 1 below. At this time, the keyword "Explain" at the beginning of the template (or other positions) may indicate the query type. In this way, the type and task type of the generated reference query may be clearly indicated, thereby facilitating the processing model to learn the processing method for each type of reference query.

[0056] Table 1 Example of reference query

[0057] type Reference query Explanation Type Explain… Description Type Describe… Summary Type Summarize… Analysis Type Analyze… Type of diagnosis Trouble-shooting… Override Type Rewrite… Generate Type Generate…

[0058] According to an exemplary implementation of the present disclosure, the template in Table 1 may be called, and specific tasks to be performed may be added at the positions indicated by the ellipsis. For example, a reference query as shown in Table 2 below may be generated.

[0059] Table 2 Example of reference query

[0060]

[0061] It should be understood that although the examples of reference queries are shown in English format above, alternatively and / or additionally, reference queries can be constructed in other natural languages ​​and / or custom formats. Specifically, an explanation-type reference query indicates that the processing model is expected to explain a certain term / terms. For another example, a rewrite-type reference query indicates that the processing model is expected to rewrite the code into C language, and so on.

[0062] According to an example implementation of the present disclosure, the template may indicate the subdivided knowledge domain involved in the reference query. For the domain where the processing model solves the problem, the domain may be divided into two parts, each of which corresponds to a role, and each role may be defined by a corresponding query type. Specifically, for the processing model in the software development field, the following two roles may be designed for the reference query based on the following commands:

[0063] Role 1: You are an assistant answering software development-related questions. Please leverage software development knowledge to solve the following tasks. ###USER:task description###Assistant: (The Chinese meaning of this command is: You are an assistant answering software development-related questions. Please use software development knowledge to solve the following tasks. ###USER:task description###Assistant. USER and Assistant are keywords that define the starting and ending positions of the query.)

[0064] Role 2: You are an assistant answering general questions. Please use related knowledge to solve the following tasks. ###USER:taskdescription###Assistant: (The Chinese meaning of this command is: You are an assistant answering general questions. Please use related knowledge to solve the following tasks. ###USER:taskdescription###Assistant. USER and Assistant are keywords that define the starting and ending positions of the query.)

[0065] Role 1 specifies a processing model to answer questions related to the software development field (i.e., a segmented knowledge field), and uses knowledge related to software development to answer them. Role 2 specifies a processing model to answer general questions (i.e., a general knowledge field), and any relevant knowledge can be used. Specifically, "###USER" and "###Assistant" can be used as the start and end identifiers for each round of query dialogue. After setting the roles, specific reference queries can be defined under each role as shown in Table 2 above.

[0066] According to an example implementation of the present disclosure, after a set of reference queries within a certain knowledge domain has been determined, reference answers matching these reference queries can be searched in the repository 220. In the context of the present disclosure, the specific format of the repository 220 is not limited, but the repository 220 may include a variety of storage methods currently known and / or to be developed in the future. For example, the public or internal knowledge bases, documents, tables and other contents related to the subdivided knowledge domain can be collected and formatted to obtain the original data set. It should be understood that the various knowledge bases, documents, tables and other contents collected here are collected with authorization. Various knowledge bases can be used, for example, to retrieve or classify the content related to the subdivided knowledge domain from the professional knowledge base in the public domain and format it, and so on. Furthermore, vertical data sets can be constructed using documents, concepts and other information formed within the business, and so on.

[0067] According to an example implementation of the present disclosure, public data sets and knowledge bases can be used. Many open source data sets and knowledge bases are specifically targeted at specific fields, such as technology, medicine, finance, and law. These resources can serve as a starting point for collecting domain knowledge. For example, various current knowledge bases contain a large amount of domain knowledge that can be used to build domain-specific models. In the field of software development, a large amount of information such as terms, concepts, cases, and development system knowledge can be collected through the Internet. Based on this information, the required data set can be further constructed.

[0068] According to an example implementation of the present disclosure, text mining and topic modeling can be performed to extract desired information. Data is collected from sources such as domain-related question-and-answer communities, and through text mining and topic modeling, key concepts, terms, and associations in the text can be automatically extracted, thereby expanding the knowledge scope of the domain.

[0069] According to an example implementation of the present disclosure, protocol knowledge and help documents can be analyzed to extract the desired information. Each field has corresponding protocols, such as various open source protocols in the field of software development, etc. Each open source software has corresponding responsibilities and obligations, as well as corresponding help documents. These protocols and documents serve as sources of model fine-tuning data.

[0070] According to an example implementation of the present disclosure, manual question-and-answer data can be used to extract the desired information. Specifically, domain-related question-and-answer datasets can be collected, which contain a large amount of knowledge from domain experts and practitioners and can be used to extract and integrate domain information. For example, in case analysis, there will be a large amount of manual case analysis data, which can be collected to construct the final dataset.

[0071] According to an exemplary implementation of the present disclosure, translation alignment between different natural languages ​​can be performed. For example, if there is less data on certain knowledge in a certain language, it can be expanded to other languages ​​through translation alignment, thereby increasing the diversity of domain knowledge.

[0072] By using the exemplary implementation of the present disclosure, knowledge in multiple fields can be collected in a more comprehensive and rich manner, so that the processing model 130 can grasp more abundant knowledge and provide more correct answers.

[0073] According to an example implementation of the present disclosure, in order to prevent unstable factors from appearing in the training data, it should be ensured that the answers to the same and / or similar reference queries are consistent. At this time, the similarity between the reference queries in a set of reference queries can be compared. For example, the distance between the reference queries can be determined based on text analysis and / or semantic analysis. If it is determined that the similarity between the first reference query and the second reference query in the first set of reference queries meets a predetermined condition (for example, the similarity is greater than or equal to 90% or other thresholds), any one of the first reference query and the second reference query can be removed from the first set of reference queries.

[0074] By using the example implementation of the present disclosure, it is possible to avoid the situation where different reference answers are retrieved for the same and / or similar reference queries. In other words, for multiple same and / or similar reference queries, only one reference query is retained. This achieves pre-filtering of reference queries, thereby ensuring that the knowledge learned by the processing model 130 is stable and consistent from the source of the data.

[0075] According to an example implementation of the present disclosure, a reference sample for training the processing model 130 may be constructed. Specifically, in the process of updating the processing model, for a target reference query in the first group of reference queries, a target reference sample for updating the processing model is created using the target reference query and a target answer in the first group of reference answers that matches the target reference query. Further, the processing model is updated using the target reference sample. In this way, reference samples may be continuously constructed in a supervised manner, thereby generating richer and more accurate training data for training the processing model. According to an example implementation of the present disclosure, updating the processing model may include updating various parameters of the processing model, for example, relevant parameters of each node of the processing model, such as weights, etc. may be updated.

[0076] Figure 4 A block diagram 400 for generating reference samples according to some implementations of the present disclosure is shown. Figure 4 As shown, a set of reference queries 450 may be created based on the above-described method. The set of reference queries 450 may include multiple reference queries 410, 420, ..., and 430. Specifically, a reference answer matching each reference query may be searched in the repository 220, thereby obtaining a set of reference answers 460. For example, the reference answer 412 matches the reference query 410, the reference answer 422 matches the reference query 420, ..., the reference answer 432 matches the reference query 430, and so on.

[0077] At this time, reference sample 440 may be constructed using reference query 410 and reference answer 412. Similarly, reference sample 420 and reference answer 422 may be constructed, ..., reference sample 430 and reference answer 432 may be constructed. In this way, multiple reference samples may be generated in a simple and effective manner, thereby providing richer training data for processing model 130.

[0078] The above describes the process of generating reference samples of a specific query type in a specific knowledge domain. Other query types in the knowledge domain can be processed in a similar manner to generate a large number of reference samples of multiple types. Figure 5 Describe in more detail the Figure 5A block diagram 500 is shown for generating reference samples associated with a given knowledge domain according to some implementations of the present disclosure. Figure 5 As shown, the knowledge domain 212 is related to software development, and in the knowledge domain 212, multiple query types 510, ..., 520, etc. may be involved. For example, the query type 510 may be a description type, the query type 520 may be a generation type, and so on.

[0079] A set of reference queries can be constructed for each query type, and the repository 220 can be used to obtain reference answers that match each reference query. Figure 4 A set of reference samples is generated in the manner shown in Figure 5 For query type 510, a set of reference queries 512 may be generated, and a set of reference answers 514 matching the set of reference queries 512 may be obtained from repository 220. Further, a corresponding set of reference samples 516 may be generated based on the set of reference queries 512 and the set of reference answers 514.

[0080] According to an example implementation of the present disclosure, similar processing may be performed for other query types. Specifically, another target type among the multiple types may be determined. Further, another template associated with the other target type may be used to generate a portion of the first set of reference queries associated with the other target type.

[0081] Figure 5 The right branch in shows the processing process for another query type. For query type 520, similar processing can be performed. For example, a set of reference queries 522 can be generated, and a set of reference answers 524 matching the set of reference queries 522 can be obtained from the repository 220. Further, a corresponding set of reference samples 526 can be generated based on the set of reference queries 522 and the set of reference answers 524. In this way, corresponding reference samples can be generated for different query types, thereby providing richer training data for the fine-tuning process.

[0082] According to an example implementation of the present disclosure, in order to ensure that the fine-tuning process is performed based on each query type in a relatively balanced manner, the difference between the number of reference samples related to each query type may be specified to be not too large. Assuming that the number of reference samples of a certain query type is too large (e.g., 1000), overfitting may occur for the query type, and reference samples of other query types with fewer reference samples (e.g., 10) may be ignored.

[0083] Since the number of reference samples is the same as the number of reference queries, that is, in the process of constructing the reference queries, it can be specified that the difference between the number of parts associated with the target type in the first set of reference queries and the number of parts associated with another target type in the first set of reference queries satisfies the predetermined difference condition. In this way, the fine-tuning process can be performed in a more balanced manner, thereby ensuring that the updated processing model 130 can answer questions of various query types more comprehensively.

[0084] According to an example implementation of the present disclosure, it is possible to specify Figure 5 The number of the set of reference queries 512 in is the same as or similar to the number of the set of reference queries 522. For example, the predetermined difference condition may specify: the difference is not higher than 10, or the difference is not higher than 1%, and so on.

[0085] Further, a set of reference samples 516 and 526 may be used to train the processing model 130. Using the example implementation of the present disclosure, a corresponding set of reference samples may be generated for each query type in each knowledge domain. In this way, more abundant reference samples may be generated, thereby providing a rich knowledge base for training the processing model 130.

[0086] According to an example implementation of the present disclosure, each subdivided knowledge domain in the general knowledge domain can be processed in a similar manner. Specifically, a second knowledge domain in the general knowledge domain can be determined. A second set of reference queries associated with the second knowledge domain can be generated. A second set of reference answers matching the second set of reference queries can be obtained from a repository. The processing model can be updated using the second set of reference queries and the second set of reference answers.

[0087] Figure 6 A block diagram 600 is shown for generating reference samples associated with various knowledge domains within a general knowledge domain according to some implementations of the present disclosure. Figure 6 As shown, corresponding reference samples can be generated in each subdivided knowledge domain in the general knowledge domain 210. Specifically, in the knowledge domain 610, a group of reference samples 612 can be generated; ...; and in the knowledge domain 620, a group of reference samples 622 can be generated.

[0088] It should be understood that the number of user questions in each sub-field is different. In order to make the ability of the processing model 130 to answer questions in each sub-field match the user's needs, the number of reference samples (i.e., the number of reference queries) in each sub-field can be determined based on the ratio between the number of user questions in each sub-field. At this time, the number of the first group of reference queries in the first knowledge field and the number of the second group of reference queries in the second knowledge field can be determined based on the number of user queries belonging to the first knowledge field and the number of user queries belonging to the second knowledge field.

[0089] like Figure 6 As shown, the ratio between the number of a set of reference samples 612 and the number of a set of reference samples 622 is determined based on the number of user queries in the knowledge domain 610 and the number of user queries in the knowledge domain 620. Using the example implementation of the present disclosure, sufficient reference samples can be provided for the fine-tuning process in a manner consistent with user needs. At this time, the processing model 130 updated with these reference samples can grasp the knowledge of each subdivided field, thereby providing a more accurate answer.

[0090] It should be understood that although the above only shows examples of improving the processing model in the software development field, hardware design field and other sub-fields by way of example. In the context of the present disclosure, the technical solution described above can be implemented in any other sub-field. For example, a variety of knowledge in the knowledge fields of science and technology, finance, law, music, sports, advertising, communication, etc. can be collected, and the reference queries and reference answers in each field can be used to construct corresponding reference samples, and then the processing model 130 can be updated in each knowledge field.

[0091] After the reference samples in each subdivided knowledge domain have been obtained, the processing model 130 can be updated using these reference samples. For example, after the processing model 130 has been pre-trained, the processing model 130 can be directly trained using each set of reference samples obtained. It should be understood that during the training process, the knowledge in the subsequent reference samples may affect the accumulation of knowledge in the previous reference samples by the processing model 130. For example, the processing model 130 may "forget" the knowledge in the general domain that was previously learned. At this time, the reference samples in the subdivided knowledge domain and the reference samples in the general knowledge domain can be mixed.

[0092] According to an example implementation of the present disclosure, in the process of updating the processing model, at least one general reference sample for updating the processing model can be obtained, and the general reference sample includes a reference query and a reference answer in the general knowledge domain. Further, the processing model can be updated using the target reference sample and the at least one general reference sample. Using the example implementation of the present disclosure, the situation where the subdivided domain knowledge covers the general domain knowledge can be alleviated.

[0093] According to an example implementation of the present disclosure, the fine-tuning process may be performed in a batch mode. Figure 7 A block diagram 700 for updating a processing model according to some implementations of the present disclosure is shown. Figure 7 As shown, fine-tuning can be performed in multiple batches, and a batch 710 can include reference samples in two different fields: one or more reference samples 720 in a subdivided knowledge field, and one or more reference samples 730 in a general knowledge field. In this way, reference samples in different fields can act together in the fine-tuning process and complete the update of the processing model 130. At this point, the updated processing model 130 will have knowledge in both the subdivided knowledge field and the general knowledge field.

[0094] According to an exemplary implementation of the present disclosure, the ratio between the number of target reference samples in the subdivided knowledge domain and the number of at least one general reference sample may satisfy a predetermined ratio condition. Figure 7 The ratio between the number of reference samples 720 and the number of reference samples 730 in the example may satisfy a predetermined ratio condition. Generally speaking, users may be more inclined to ask questions in the general knowledge field and less likely to ask questions in the subdivided knowledge field. At this time, the ratio condition may be set according to the ratio of the number of questions in each field.

[0095] For example, the ratio between the number of reference samples 720 and the number of reference samples 730 may be set to 2:8. Alternatively and / or additionally, the ratio between the number of reference samples 720 and the number of reference samples 730 may be set to 1.5:8.5, or other ratios. In this way, it is possible to ensure that the processing model 130 can master knowledge in various fields in a manner that better meets user needs.

[0096] According to an exemplary implementation of the present disclosure, a fine-tuning process may be performed using the reference samples obtained based on various methods currently known and / or to be developed in the future, so as to update the processing model 130. The updated processing model 130 may further be used to serve user queries. Specifically, if a user query targeting the processing model is received, the updated processing model may be used to obtain query results matching the user query. Figure 8 Provide more information, the Figure 8 A block diagram 800 is shown for processing a user query utilizing an updated processing model in accordance with some implementations of the present disclosure.

[0097] like Figure 8 As shown, the processing model 130 can interact with the user in a question-and-answer manner in page 810. Specifically, the user can submit a user query 820, for example, a query of “What is a programming language?” At this time, the updated processing model 130 can provide a query result 830 matching the user query 820. At this time, compared with before fine-tuning, the processing model 130 can provide more knowledge about programming languages ​​and answer the question in a more detailed and accurate manner.

[0098] According to an example implementation of the present disclosure, it is proposed to construct a fine-tuning dataset for a segmented field, and to use the fine-tuning dataset to update the processing model. By collecting data from each segmented field, reference samples and reference answers within the corresponding segmented field can be determined, and then suitable reference samples can be constructed. Furthermore, by balancing the number of reference samples related to the query type within each segmented field, and balancing the number of reference samples within each segmented field, a high-quality, scalable fine-tuning dataset can be constructed. In this way, the knowledge reserve of the processing model related to the segmented field can be greatly improved, and relevant questions in each segmented field can be answered in a more detailed and accurate manner.

[0099] Example Process

[0100] Fig. 9 A flow chart of a model processing method 900 according to some implementations of the present disclosure is shown. At box 910, a first knowledge domain in the general knowledge domains involved in the processing model is determined. At box 920, a first set of reference queries associated with the first knowledge domain is generated. At box 930, a first set of reference answers matching the first set of reference queries is obtained from a repository associated with the processing model, the repository at least involving the first knowledge domain. At box 940, the processing model is updated using the first set of reference queries and the first set of reference answers.

[0101] According to an example implementation of the present disclosure, generating a first set of reference queries includes: determining a target type among multiple types of queries input to a processing model in a first knowledge domain; and generating the first set of reference queries using a template associated with the target type.

[0102] According to an exemplary implementation of the present disclosure, the template indicates at least any one of the following: a first knowledge domain, and a target type.

[0103] According to an example implementation of the present disclosure, the method 900 further includes: in response to determining that the similarity between the first reference query and the second reference query in the first set of reference queries satisfies a predetermined condition, removing any one of the first reference query and the second reference query from the first set of reference queries.

[0104] According to an example implementation of the present disclosure, updating a processing model includes: for a target reference query in a first group of reference queries, creating a target reference sample for updating the processing model using the target reference query and a target answer in a first group of reference answers that matches the target reference query; and updating the processing model using the target reference sample.

[0105] According to an example implementation of the present disclosure, updating a processing model includes: obtaining at least one general reference sample for updating the processing model, the general reference sample including a reference query and a reference answer in a general knowledge domain; and updating the processing model using a target reference sample and at least one general reference sample.

[0106] According to an example implementation of the present disclosure, a ratio between the number of target reference samples and the number of at least one common reference sample satisfies a predetermined ratio condition.

[0107] According to an example implementation of the present disclosure, generating the first set of reference queries further includes: determining another target type among the multiple types; and generating a portion of the first set of reference queries associated with the other target type using another template associated with the other target type.

[0108] According to an example implementation of the present disclosure, a difference between the number of parts associated with a target type in the first set of reference queries and the number of parts associated with another target type in the first set of reference queries satisfies a predetermined difference condition.

[0109] According to an example implementation of the present disclosure, the method 900 further includes: determining a second knowledge domain in the general knowledge domain; generating a second set of reference queries associated with the second knowledge domain; obtaining a second set of reference answers matching the second set of reference queries from a repository; and updating the processing model using the second set of reference queries and the second set of reference answers.

[0110] According to an example implementation of the present disclosure, the number of the first set of reference queries and the number of the second set of reference queries are determined based on the number of user queries belonging to the first knowledge domain and the number of user queries belonging to the second knowledge domain.

[0111] According to an exemplary implementation of the present disclosure, the plurality of types include at least any one of the following: an explanation type, a summary type, a description type, an analysis type, a diagnosis type, a rewrite type, and a generation type.

[0112] According to an example implementation of the present disclosure, the method 900 further includes: in response to receiving a user query for the processing model, using the updated processing model to obtain a query result matching the user query.

[0113] Example devices and equipment

[0114] Fig.10 A block diagram of a model processing device 1000 according to some implementations of the present disclosure is shown. The device 1000 includes: a determination module 1010, configured to determine a first knowledge domain in the general knowledge domains involved in the processing model; a generation module 1020, configured to generate a first set of reference queries associated with the first knowledge domain; an acquisition module 1030, configured to acquire a first set of reference answers matching the first set of reference queries from a repository associated with the processing model, the repository at least involving the first knowledge domain; and an update module 1040, configured to update the processing model using the first set of reference queries and the first set of reference answers.

[0115] According to an example implementation of the present disclosure, the generation module 1020 includes: a type determination module configured to determine, in a first knowledge domain, a target type from multiple types of queries input to a processing model; and a query generation module configured to generate a first set of reference queries using a template associated with the target type.

[0116] According to an exemplary implementation of the present disclosure, the template indicates at least any one of the following: a first knowledge domain, and a target type.

[0117] According to an example implementation of the present disclosure, the apparatus 1000 further includes: a removal module configured to remove any one of the first reference query and the second reference query from the first group of reference queries in response to determining that the similarity between the first reference query and the second reference query in the first group of reference queries satisfies a predetermined condition.

[0118] According to an example implementation of the present disclosure, the update module 1040 includes: a creation module configured to create a target reference sample for updating the processing model for a target reference query in a first group of reference queries, using the target reference query and a target answer in a first group of reference answers that matches the target reference query; and a model update module configured to update the processing model using the target reference sample.

[0119] According to an example implementation of the present disclosure, the model updating module includes: a general sample acquisition module, configured to acquire at least one general reference sample for updating the processing model, the general reference sample including a reference query and a reference answer in a general knowledge domain; and a combined updating module, configured to update the processing model using a target reference sample and at least one general reference sample.

[0120] According to an example implementation of the present disclosure, a ratio between the number of target reference samples and the number of at least one common reference sample satisfies a predetermined ratio condition.

[0121] According to an example implementation of the present disclosure, the generation module is further configured to: determine another target type from the multiple types; and generate a portion of the first set of reference queries associated with the other target type using another template associated with the other target type.

[0122] According to an example implementation of the present disclosure, a difference between the number of parts associated with a target type in the first set of reference queries and the number of parts associated with another target type in the first set of reference queries satisfies a predetermined difference condition.

[0123] According to an example implementation of the present disclosure, the determination module is further configured to determine a second knowledge domain in the general knowledge domain; the generation module is further configured to generate a second set of reference queries associated with the second knowledge domain; the acquisition module is further configured to acquire a second set of reference answers matching the second set of reference queries from the repository; and the update module is further configured to update the processing model using the second set of reference queries and the second set of reference answers.

[0124] According to an example implementation of the present disclosure, the number of the first set of reference queries and the number of the second set of reference queries are determined based on the number of user queries belonging to the first knowledge domain and the number of user queries belonging to the second knowledge domain.

[0125] According to an exemplary implementation of the present disclosure, the plurality of types include at least any one of the following: an explanation type, a summary type, a description type, an analysis type, a diagnosis type, a rewrite type, and a generation type.

[0126] According to an example implementation of the present disclosure, the apparatus 1000 further includes: a query module configured to, in response to receiving a user query for the processing model, obtain a query result matching the user query using the updated processing model.

[0127] Fig.11 1 shows a block diagram of a device 1100 capable of implementing multiple implementations of the present disclosure. It should be understood that Fig.11The computing device 1100 shown is merely exemplary and should not constitute any limitation on the functionality and scope of the implementations described herein. Fig.11 The computing device 1100 shown may be used to implement the methods described above.

[0128] like Fig.11 As shown, computing device 1100 is in the form of a general-purpose computing device. Components of computing device 1100 may include, but are not limited to, one or more processors or processing units 1110, memory 1120, storage device 1130, one or more communication units 1140, one or more input devices 1150, and one or more output devices 1160. Processing unit 1110 may be an actual or virtual processor and is capable of performing various processes according to a program stored in memory 1120. In a multi-processor system, multiple processing units execute computer executable instructions in parallel to increase the parallel processing capabilities of computing device 1100.

[0129] The computing device 1100 typically includes a plurality of computer storage media. Such media may be any available media accessible to the computing device 1100, including but not limited to volatile and non-volatile media, removable and non-removable media. The memory 1120 may be a volatile memory (e.g., registers, caches, random access memory (RAM)), a non-volatile memory (e.g., a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), flash memory), or some combination thereof. The storage device 1130 may be a removable or non-removable medium, and may include a machine-readable medium, such as a flash drive, a disk, or any other medium, which may be capable of being used to store information and / or data (e.g., training data for training) and may be accessed within the computing device 1100.

[0130] The computing device 1100 may further include additional removable / non-removable, volatile / non-volatile storage media. Fig.11 As shown in , a disk drive for reading or writing from a removable, non-volatile disk (e.g., a "floppy disk") and an optical drive for reading or writing from a removable, non-volatile optical disk may be provided. In these cases, each drive may be connected to the bus (not shown) by one or more data media interfaces. The memory 1120 may include a computer program product 1125 having one or more program modules that are configured to perform various methods or actions of various implementations of the present disclosure.

[0131] The communication unit 1140 enables communication with other computing devices via a communication medium. Additionally, the functions of the components of the computing device 1100 can be implemented in a single computing cluster or multiple computing machines that can communicate via a communication connection. Therefore, the computing device 1100 can operate in a networked environment using a logical connection with one or more other servers, a network personal computer (PC), or another network node.

[0132] Input device 1150 may be one or more input devices, such as a mouse, keyboard, trackball, etc. Output device 1160 may be one or more output devices, such as a display, speaker, printer, etc. Computing device 1100 may also communicate with one or more external devices (not shown) through communication unit 1140 as needed, such as storage devices, display devices, etc., communicate with one or more devices that allow a user to interact with computing device 1100, or communicate with any device that allows computing device 1100 to communicate with one or more other computing devices (e.g., a network card, a modem, etc.). Such communication may be performed via an input / output (I / O) interface (not shown).

[0133] According to an exemplary implementation of the present disclosure, a computer-readable storage medium is provided, on which computer-executable instructions are stored, wherein the computer-executable instructions are executed by a processor to implement the method described above. According to an exemplary implementation of the present disclosure, a computer program product is also provided, the computer program product is tangibly stored on a non-transitory computer-readable medium and includes computer-executable instructions, and the computer-executable instructions are executed by a processor to implement the method described above. According to an exemplary implementation of the present disclosure, a computer program product is provided, on which a computer program is stored, and when the program is executed by a processor, the method described above is implemented.

[0134] Various aspects of the present disclosure are described herein with reference to the flowcharts and / or block diagrams of the methods, devices, equipment, and computer program products implemented according to the present disclosure. It should be understood that each box in the flowchart and / or block diagram and the combination of each box in the flowchart and / or block diagram can be implemented by computer-readable program instructions.

[0135] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing device, thereby producing a machine, so that when these instructions are executed by the processing unit of the computer or other programmable data processing device, a device that implements the functions / actions specified in one or more boxes in the flowchart and / or block diagram is generated. These computer-readable program instructions can also be stored in a computer-readable storage medium, and these instructions cause the computer, programmable data processing device, and / or other equipment to work in a specific manner, so that the computer-readable medium storing the instructions includes a manufactured product, which includes instructions for implementing various aspects of the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0136] Computer-readable program instructions can be loaded onto a computer, other programmable data processing apparatus, or other device so that a series of operational steps are performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, so that the instructions executed on the computer, other programmable data processing apparatus, or other device implement the functions / actions specified in one or more boxes in the flowchart and / or block diagram.

[0137] The flow chart and block diagram in the accompanying drawings show the possible architecture, function and operation of the system, method and computer program product according to multiple implementations of the present disclosure. In this regard, each square box in the flow chart or block diagram can represent a part of a module, program segment or instruction, and a part of a module, program segment or instruction includes one or more executable instructions for realizing the logical function of the specification. In some implementations as replacements, the function marked in the square box can also occur in a sequence different from that marked in the accompanying drawings. For example, two continuous square boxes can actually be executed substantially in parallel, and they can sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each square box in the block diagram and / or flow chart, and the combination of the square boxes in the block diagram and / or flow chart can be realized by a special hardware-based system that performs the function or action of the specification, or can be realized by a combination of special hardware and computer instructions.

[0138] The above descriptions of various implementations of the present disclosure are exemplary, non-exhaustive, and not limited to the disclosed implementations. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described implementations. The selection of terms used herein is intended to best explain the principles of the implementations, practical applications, or improvements to the technology in the market, or to enable other persons of ordinary skill in the art to understand the various implementations disclosed herein.

Claims

1. A model processing method, comprising: determining a first knowledge domain among the general knowledge domains involved in the processing model; generating a first set of reference queries associated with the first knowledge domain; obtaining a first set of reference answers matching the first set of reference queries from a repository associated with the processing model, the repository at least relating to the first knowledge domain; as well as The processing model is updated using the first set of reference queries and the first set of reference answers.

2. The method of claim 1, wherein generating the first set of reference queries comprises: In the first knowledge domain, determining a target type among a plurality of types for a query input to the process model; as well as The first set of reference queries is generated using a template associated with the target type. The method according to claim 2 , wherein the template indicates at least any one of the following: the first knowledge domain, and the target type.

4. The method according to claim 1, further comprising: In response to determining that the similarity between a first reference query and a second reference query in the first set of reference queries satisfies a predetermined condition, any one of the first reference query and the second reference query is removed from the first set of reference queries.

5. The method of claim 1 , wherein updating the process model comprises: For a target reference query in the first group of reference queries, creating a target reference sample for updating the processing model by using the target reference query and a target answer in the first group of reference answers that matches the target reference query; as well as The process model is updated using the target reference samples.

6. The method of claim 5, wherein updating the process model comprises: Acquire at least one general reference sample for updating the processing model, the general reference sample comprising a reference query and a reference answer in the general knowledge domain; as well as The process model is updated using the target reference sample and the at least one common reference sample. 7 . The method according to claim 6 , wherein a ratio between the number of the target reference samples and the number of the at least one common reference sample satisfies a predetermined ratio condition.

8. The method of claim 1, wherein generating the first set of reference queries further comprises: determining another target type from among the plurality of types; as well as A portion of the first set of reference queries associated with the other target type is generated using another template associated with the other target type. 9 . The method of claim 8 , wherein a difference between the number of portions associated with the target type in the first set of reference queries and the number of portions associated with the other target type in the first set of reference queries satisfies a predetermined difference condition.

10. The method according to claim 1, further comprising: determining a second knowledge area within the general knowledge area; generating a second set of reference queries associated with the second knowledge domain; Retrieving from the repository a second set of reference answers matching the second set of reference queries; as well as The processing model is updated using the second set of reference queries and the second set of reference answers. 11 . The method of claim 10 , wherein the number of the first set of reference queries and the number of the second set of reference queries are determined based on the number of user queries belonging to the first knowledge domain and the number of user queries belonging to the second knowledge domain.

12. The method according to claim 2, wherein the plurality of types comprises at least any one of the following: an explanation type, a summary type, a description type, an analysis type, a diagnosis type, a rewrite type, and a generation type.

13. The method according to claim 1, further comprising: In response to receiving a user query for the processing model, the updated processing model is used to obtain query results matching the user query.

14. A model processing device, comprising: A determination module, configured to determine a first knowledge domain in the general knowledge domains involved in the processing model; a generating module configured to generate a first set of reference queries associated with the first knowledge domain; an acquisition module configured to acquire a first set of reference answers matching the first set of reference queries from a repository associated with the processing model, the repository at least relating to the first knowledge domain; as well as An updating module is configured to update the processing model using the first set of reference queries and the first set of reference answers.

15. An electronic device, comprising: at least one processing unit; as well as At least one memory, the at least one memory being coupled to the at least one processing unit and storing instructions for execution by the at least one processing unit, the instructions causing the electronic device to perform the method according to any one of claims 1 to 13 when executed by the at least one processing unit.

16. A computer-readable storage medium having a computer program stored thereon, the computer program causing a processor to implement the method according to any one of claims 1 to 13 when the computer program is executed by the processor.

Images