Task processing model training method, task processing method, and query processing method

Abstract

The embodiment of the present specification provides a task processing model training method, a task processing method and a query processing method, wherein the task processing model training method comprises: determining a sample object corresponding to a target object processing task, a label object processing result and a plurality of candidate use cases; inputting the sample object and the plurality of candidate use cases into a task processing model to obtain a predicted object processing result of the sample object, wherein the predicted object processing result is obtained by the task processing model according to a target use case to predict the sample object, the target use case is obtained by the task processing model according to a use case selection strategy to decode a fusion coding feature, and the fusion coding feature is obtained by the task processing model to perform feature fusion processing on the sample object and the plurality of candidate use cases; adjusting parameters of the task processing model according to the predicted object processing result and the label object processing result until a training stop condition is reached to obtain a target task processing model.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of artificial intelligence technology, and in particular to methods for training task processing models, task processing methods, and query processing methods. Background Technology

[0002] In practical applications, with the development of artificial intelligence technology, large-scale models have emerged. These are deep learning models with a large number of parameters, which, after being trained on massive amounts of data, can be widely applied to general domains. However, large-scale models often exhibit poor inference accuracy for downstream tasks in specific domains. In such cases, fine-tuning is usually necessary to adapt the large-scale model to the specific task. However, ensuring the inference performance of large-scale models in specific domains typically requires significant computational resources to achieve in-depth application. Therefore, an effective technical solution is urgently needed to address the problem of reducing the fine-tuning cost of large-scale models while maintaining their inference performance. Summary of the Invention

[0003] In view of the above, embodiments of this specification provide a task processing model training method. One or more embodiments of this specification also relate to a task processing model training apparatus, a task processing method, a task processing device, a query processing method, a query processing device, a computing device, an electronic device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0004] According to a first aspect of the embodiments of this specification, a method for training a task processing model is provided, comprising:

[0005] The sample object, the label object processing result, and multiple candidate use cases corresponding to the target object processing task are determined, wherein the multiple candidate use cases are used to describe the processing flow of the target object processing task;

[0006] The sample object and the multiple candidate use cases are input into the task processing model to obtain the predicted object processing result of the sample object. The predicted object processing result is obtained by the task processing model predicting the sample object according to the target use case. The target use case is obtained by the task processing model decoding the fused encoding features according to the use case selection strategy. The fused encoding features are obtained by the task processing model performing feature fusion processing on the sample object and the multiple candidate use cases.

[0007] Based on the processing results of the predicted object and the processing results of the labeled object, the parameters of the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

[0008] According to a second aspect of the embodiments of this specification, a task processing model training apparatus is provided, comprising:

[0009] The determination module is configured to determine the sample object, the label object processing result, and multiple candidate use cases corresponding to the target object processing task, wherein the multiple candidate use cases are used to describe the processing flow of the target object processing task;

[0010] The input module is configured to input the sample object and the plurality of candidate use cases into the task processing model to obtain the predicted object processing result of the sample object, wherein the predicted object processing result is obtained by the task processing model predicting the sample object according to the target use case, the target use case is obtained by the task processing model decoding the fused encoding features according to the use case selection strategy, and the fused encoding features are obtained by the task processing model performing feature fusion processing on the sample object and the plurality of candidate use cases;

[0011] The adjustment module is configured to adjust the parameters of the task processing model based on the processing results of the predicted object and the processing results of the labeled object, until the training stopping condition is met, thereby obtaining the target task processing model.

[0012] According to a third aspect of the embodiments of this specification, a task processing method is provided, including:

[0013] Identify the objects to be processed corresponding to the target object's processing task;

[0014] The object to be processed is input into the target task processing model to obtain the object processing result of the object to be processed, wherein the target task processing model is trained according to the task processing model training method provided in the first aspect of the embodiments of this specification.

[0015] According to a fourth aspect of the embodiments of this specification, a task processing apparatus is provided, comprising:

[0016] The determination module is configured to determine the objects to be processed corresponding to the target object processing task.

[0017] The input module is configured to input the object to be processed into the target task processing model to obtain the object processing result of the object to be processed, wherein the target task processing model is trained according to the task processing model training method provided in the first aspect of the embodiments of this specification.

[0018] According to a fifth aspect of the embodiments of this specification, a query processing method is provided, including:

[0019] Identify queries to be processed;

[0020] The query to be processed is input into the target task processing model to obtain the query result of the query to be processed, wherein the target task processing model is trained according to the task processing model training method provided in the first aspect of the embodiments of this specification.

[0021] According to a sixth aspect of the embodiments of this specification, a query processing apparatus is provided, comprising:

[0022] The "Determine" module is configured to determine queries to be processed.

[0023] The input module is configured to input the query to be processed into the target task processing model to obtain the query result of the query to be processed, wherein the target task processing model is trained according to the task processing model training method provided in the first aspect of the embodiments of this specification.

[0024] According to a seventh aspect of the embodiments of this specification, a task processing method is provided, including:

[0025] Identify the objects to be processed and multiple candidate use cases corresponding to the target object processing task;

[0026] The object to be processed and the multiple candidate use cases are input into the target task processing model to obtain the object processing result of the object to be processed. The object processing result is obtained by the target task processing model predicting the object to be processed based on the target use cases. The target use cases are obtained by the target task processing model decoding the fused encoding features according to the use case selection strategy. The fused encoding features are obtained by the target task processing model performing feature fusion processing on the object to be processed and the multiple candidate use cases.

[0027] According to an eighth aspect of the embodiments of this specification, a computing device is provided, comprising:

[0028] Memory and processor;

[0029] The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which implement the steps of the above method when executed by the processor.

[0030] According to a ninth aspect of the embodiments of this specification, an electronic device is provided, comprising:

[0031] A memory and a processor, the memory and the processor being connected via a bus;

[0032] The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the above method.

[0033] According to a tenth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0034] According to an eleventh aspect of the embodiments of this specification, a computer program product is provided, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0035] This specification provides an embodiment of a task processing model training method, comprising: determining a sample object, a label object processing result, and multiple candidate use cases corresponding to a target object processing task, wherein the multiple candidate use cases are used to describe the processing flow of the target object processing task; inputting the sample object and the multiple candidate use cases into a task processing model to obtain a predicted object processing result for the sample object, wherein the predicted object processing result is obtained by the task processing model predicting the sample object based on the target use case, the target use case is obtained by the task processing model decoding fused encoding features according to a use case selection strategy, and the fused encoding features are obtained by the task processing model performing feature fusion processing on the sample object and the multiple candidate use cases; adjusting the parameters of the task processing model based on the predicted object processing result and the label object processing result until a training stopping condition is reached to obtain a target task processing model.

[0036] In the above method, to ensure the task processing model meets the processing requirements of the target object processing task, the sample object, the label object processing results, and multiple candidate use cases corresponding to the target object processing task are determined as training data. During training, the task processing model can perform inference according to the processing flow described by the candidate use cases based on the candidate use cases. Furthermore, in the task processing model, feature processing is performed on the sample object and multiple candidate use cases to obtain fused encoded features, realizing the interactive processing between the sample object and the candidate use cases. At the same time, the fused encoded features are decoded according to the use case selection strategy to obtain the target use case, realizing the selection of multiple candidate use cases based on the sample object, and using the selected target use case to predict the sample object. The processing results of the predicted objects are obtained, and the parameters of the task processing model are adjusted based on the processing results of the predicted objects and the processing results of the labeled objects until the target task processing model that meets the training stopping condition is obtained. This allows the target task processing model to select target use cases corresponding to higher quality predicted object processing results through the interaction between sample objects and candidate use cases when selecting target use cases based on the use case selection strategy. This ensures high quality of target use case selection and enables the target task processing model to automatically explore and utilize high-quality data (i.e., target use cases) from a small amount of training data (i.e., sample objects, label object processing results, and multiple candidate use cases). This achieves improved inference performance while reducing the cost of model fine-tuning. Attached Figure Description

[0037] Figure 1 This is a schematic diagram illustrating an application scenario of a task processing model training method provided in one embodiment of this specification.

[0038] Figure 2 This is a flowchart illustrating a task processing model training method provided in one embodiment of this specification;

[0039] Figure 3 This is a flowchart illustrating the processing procedure of a task processing model training method provided in one embodiment of this specification.

[0040] Figure 4 This is a schematic diagram of the structure of a task processing model training device provided in one embodiment of this specification;

[0041] Figure 5 This is a flowchart illustrating a task processing method provided in one embodiment of this specification;

[0042] Figure 6 This is a structural diagram of a task processing device provided in one embodiment of this specification;

[0043] Figure 7 This is a flowchart illustrating a query processing method provided in one embodiment of this specification;

[0044] Figure 8 This is a structural diagram of a query processing device provided in one embodiment of this specification;

[0045] Figure 9 This is a structural block diagram of a computing device provided in one embodiment of this specification;

[0046] Figure 10 This is a structural block diagram of an electronic device provided in one embodiment of this specification. Detailed Implementation

[0047] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0048] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0049] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0050] Furthermore, it should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, displayed data, etc.) involved in one or more embodiments of this specification are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0051] In one or more embodiments of this specification, a large model refers to a deep learning model with a large number of model parameters, typically containing hundreds of millions, tens of billions, hundreds of billions, trillions, or even tens of trillions of model parameters. A large model can also be called a foundation model. It is pre-trained using large-scale unlabeled corpora to produce a pre-trained model with hundreds of millions of parameters. Such models can adapt to a wide range of downstream tasks and have good generalization ability. Examples include Large Language Models (LLMs) and multi-modal pre-training models.

[0052] In practical applications, large models only require a small number of samples to fine-tune the pre-trained model before they can be applied to different tasks. Large models can be widely used in fields such as Natural Language Processing (NLP) and Computer Vision. Specifically, they can be applied to computer vision tasks such as Visual Question Answering (VQA), Image Captioning (IC), and Image Generation, as well as natural language processing tasks such as text-based sentiment classification, text summarization, and machine translation. The main application scenarios for large models include digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design.

[0053] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0054] In-context Learning (ICL) is a large model hinting technique.

[0055] Demonstration: A use case, which contains a set of target object processing tasks and processing flows.

[0056] Prompt: a prompt word.

[0057] Few-shot prompt: Contextual learning falls under the category of few-shot prompts.

[0058] Modal Embedding: A vector used to identify feature modes. Commonly used in Transformers.

[0059] Positional Embedding: A vector used to identify feature indices. Commonly used in Transformers.

[0060] Latent Code: Latent space coding, which is a fusion coding feature.

[0061] Transformer is a deep learning model architecture based on attention mechanisms. The core idea of ​​Transformer is to capture global dependencies in the input sequence through "self-attention mechanism", thereby overcoming the limitations of traditional recurrent neural networks and convolutional neural networks in processing long sequence data.

[0062] To address the aforementioned technical problems, this specification provides a task processing model training method. This specification also relates to a task processing model training device, a task processing method, a task processing apparatus, a query processing method, a query processing apparatus, a computing device, an electronic device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0063] See Figure 1 , Figure 1 The diagram illustrates an application scenario of a task processing model training method provided according to an embodiment of this specification.

[0064] The training method for this task processing model specifically includes the following steps.

[0065] The sample object, the label object processing result, and multiple candidate use cases corresponding to the target object processing task are determined, wherein the multiple candidate use cases are used to describe the processing flow of the target object processing task.

[0066] The sample object and the multiple candidate use cases are input into the task processing model to obtain the predicted object processing result of the sample object. The predicted object processing result is obtained by the task processing model predicting the sample object based on the target use case. The target use case is obtained by the task processing model decoding the fused encoding features according to the use case selection strategy. The fused encoding features are obtained by the task processing model performing feature fusion processing on the sample object and the multiple candidate use cases.

[0067] Based on the processing results of the predicted object and the processing results of the labeled object, the parameters of the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

[0068] like Figure 1 As shown, Figure 1 It includes an edge device 102 and a model training platform 104.

[0069] In practice, the model trainer can send a model training request to the model training platform 104 via the edge device 102. This request may include training data corresponding to the target object processing task. The model training platform 104 receives the request, determines that the training data for the target object processing task includes sample objects, label object processing results, and multiple candidate use cases, and inputs these data into the task processing model to be trained. Within the model, feature processing is performed on the sample objects and candidate use cases to obtain fused encoded features, enabling interactive processing between the sample objects and candidate use cases. Simultaneously, the fused encoded features are decoded according to a use case selection strategy to obtain the target use case. This allows for the selection of multiple candidate use cases based on the sample objects, and the use of the selected target use case to predict the processing results of the predicted objects. The parameters of the task processing model are then adjusted based on the predicted object processing results and the label object processing results until a target task processing model that meets the training stopping conditions is obtained. The model training platform 104 can send the model information of the target task processing model to the edge device 102, so that the edge device 102 can deploy the target task processing model locally on the edge device 102 according to the received model information. Alternatively, the model training platform 104 can also send the model call interface of the target task processing model to the edge device 102, so that the edge device 102 can directly call the target task processing model according to the model call interface.

[0070] The edge device 102 may include a browser, an app (application), or a web application such as an H5 (Hypertext Markup Language 5) application, a lightweight application (also known as a mini-program), or a cloud application. The edge device can be developed based on a software development kit (SDK) provided by the server, such as a real-time communication (RTC) SDK. The edge device can be deployed in an electronic device and depends on the device's operation or certain apps within the device to run. The electronic device may have a display screen and support information browsing, such as a personal mobile terminal like a mobile phone, tablet, or personal computer. Various other types of applications can also be configured in the electronic device, such as human-computer interaction applications, model training applications, text processing applications, web browser applications, shopping applications, search applications, instant messaging tools, email clients, and social media platform software.

[0071] The model training platform 104 can be understood as a server used for model training.

[0072] See Figure 2 , Figure 2 A flowchart of a task processing model training method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0073] Step 202: Determine the sample object, label object processing results, and multiple candidate use cases corresponding to the target object processing task, wherein the multiple candidate use cases are used to describe the processing flow of the target object processing task.

[0074] Specifically, the task processing model training method provided in the embodiments of this specification can be applied to a model training platform. After receiving a model training request for a task processing model, the model training platform can obtain training data corresponding to the target object processing task. This training data may include multiple sample objects corresponding to the target object processing task, processing results of multiple labeled objects corresponding to the multiple sample objects, and multiple candidate use cases. In practical applications, the training data can be uploaded to the model training platform by model trainers through a client.

[0075] The target object processing task can be understood as the task that the task processing model needs to process. The target object processing task can refer to the downstream task of a specific field, such as including but not limited to text processing tasks, image processing tasks, user query processing tasks, etc. Furthermore, the target object processing task can be medical image processing tasks, question text processing tasks, product consultation question processing tasks, etc. The embodiments in this specification do not limit this.

[0076] The sample object can be understood as the object that needs to be processed in the target object processing task. For example, if the target object processing task is a product inquiry processing task, the sample object can be understood as the product inquiry question, such as "How much does product A cost?" or "What is the model of product A?". The label object processing result can be understood as the accurate object processing result obtained after processing the sample object. When the sample object is a product inquiry question, the label object processing result can be the answer to the product inquiry question, such as "The price of product A is XX yuan" or "The model of product A is XX model?". Understandably, in supervised training of the task processing model, the sample object can be the training sample in the training data, and the label object processing result can be the training label in the training data. The training sample and training label constitute the training data pair used for supervised training of the task processing model.

[0077] Use cases describe the processing flow and / or processing method of a target object's task. Use cases guide the task processing model to reason according to the described flow and / or method. Candidate use cases can be understood as multiple use cases related to the target object's task. Understandably, the processing flow for different sample objects will differ, and therefore the use cases can also differ; one sample object can correspond to multiple candidate use cases. In other words, candidate use cases can be related to the object type of the sample object. For example, if the sample object is a product consultation question, there might be one or more candidate use cases if the product consultation question is a pre-sales consultation question, and another one or more candidate use cases if the product consultation question is a post-sales processing question.

[0078] Based on this, in response to a model training request, the model training platform can determine the target object processing task carried by the model training request, and obtain the sample object, label object processing results, and multiple candidate use cases corresponding to the target object processing task as model training data.

[0079] Taking a query processing task as an example, the sample object can be a sample query, the tag object processing result can be the tag query result, and the candidate use cases can be used to describe the processing flow of the query processing task. These candidate use cases can include examples of queries that have already been processed, such as the query and its corresponding result. Furthermore, if the target object processing task is a question-and-answer task, then the candidate use cases for that task can include examples of questions that have already been processed, such as the question and answer.

[0080] Step 204: Input the sample object and the multiple candidate use cases into the task processing model to obtain the predicted object processing result of the sample object, wherein the predicted object processing result is obtained by the task processing model predicting the sample object according to the target use case, the target use case is obtained by the task processing model decoding the fused encoding features according to the use case selection strategy, and the fused encoding features are obtained by the task processing model performing feature fusion processing on the sample object and the multiple candidate use cases.

[0081] Specifically, after determining the sample object and multiple candidate use cases corresponding to the target object processing task, the sample object and multiple candidate use cases can be input into the task processing model to obtain the predicted object processing result of the sample object output by the task processing model.

[0082] The task processing model can be understood as the task processing model to be trained.

[0083] Based on this, sample objects and multiple candidate use cases can be input into the task processing model to be trained. In this task processing model, feature processing can be performed on the sample objects and multiple candidate use cases to obtain fused encoded features. The fused encoded features are then decoded according to the use case selection strategy to obtain the target use case. Based on the target use case, the sample objects are predicted to obtain and output the predicted object processing result.

[0084] In practical applications, when the sample object is a sample query, the sample query and multiple candidate use cases can be input into the task processing model to be trained. In this task processing model, feature processing can be performed on the sample query and multiple candidate use cases to obtain fused encoded features. The fused encoded features are then decoded according to the use case selection strategy to obtain the target use case. Based on the target use case, the sample query is predicted and the predicted query result is output.

[0085] In specific implementation, the step of inputting the sample object and the multiple candidate use cases into the task processing model to obtain the predicted object processing result of the sample object includes:

[0086] The sample object and the plurality of candidate use cases are input into the task processing model, wherein the task processing model includes a feature encoding unit, a fusion encoding unit, a feature decoding unit and an object processing unit;

[0087] Using the feature encoding unit, feature processing is performed on the sample object and the multiple candidate use cases to obtain multiple encoded features;

[0088] The multiple coding features are fused using the fusion coding unit to obtain fused coding features;

[0089] Using the feature decoding unit, the fused encoded features are decoded according to the use case selection strategy to determine the target use case from the multiple candidate use cases;

[0090] Using the object processing unit, the sample object is predicted based on the target use case to obtain the predicted object processing result.

[0091] Specifically, the sample object and multiple candidate use cases can be input into the feature encoding unit of the task processing model. Feature processing is performed on the sample object and multiple candidate use cases to obtain multiple encoded features. The multiple encoded features are input into the fusion encoding unit to perform fusion processing to obtain fused encoded features. The fused encoded features are input into the feature decoding unit to perform decoding processing according to the use case selection strategy, select the target use case from multiple candidate use cases, input the target use case and the sample object into the object processing unit, and perform prediction processing on the sample object according to the target use case to obtain the prediction object processing result.

[0092] In summary, by processing the sample object and multiple candidate use cases through the task processing model, the interaction between the sample object and the candidate use cases is realized. Combined with the feature decoding unit, the fused encoded features are autoregressively decoded, so that the selection of each target use case takes into account all previously selected target use cases, thereby mitigating the drawbacks of independent selection of target use cases. Furthermore, the target use case is selected from multiple candidate use cases based on the use case selection strategy, which solves the problem of unclear use case selection criteria. This dynamically selects more suitable target use cases for the sample object, avoids always using fixed use cases, and further ensures the inference performance of the subsequent target task processing model.

[0093] Furthermore, the feature encoding unit is used to perform feature processing on the sample object and the multiple candidate use cases to obtain multiple encoded features, including:

[0094] The feature encoding unit is used to encode the multimodal information contained in the sample object and the multiple candidate use cases to obtain multiple multimodal encoded features;

[0095] The multiple multimodal coding features are subjected to feature labeling processing to obtain multiple coding features.

[0096] The sample objects and multiple candidate use cases may contain multimodal information, such as text information, image information, video information, and audio information, etc. This specification does not limit this.

[0097] Based on this, feature encoding units can be used to encode the multimodal information contained in the sample object and multiple candidate use cases to obtain multiple multimodal encoded features, and feature labeling processing can be performed on the multiple multimodal encoded features to obtain multiple encoded features.

[0098] In specific implementation, when the multimodal information includes both text and image information, the feature encoding unit is used to encode the multimodal information contained in the sample object and the multiple candidate use cases to obtain multiple multimodal encoded features, including:

[0099] Using the feature encoding unit, the text and image information contained in the sample object and the multiple candidate use cases are encoded to obtain multiple text encoding features and multiple image encoding features;

[0100] The step of performing feature labeling processing on the multiple multimodal coding features to obtain multiple coding features includes:

[0101] Feature labeling processing is performed on the multiple text encoding features and the multiple image encoding features to obtain multiple encoding features, wherein the multiple encoding features include multiple labeled text encoding features and multiple labeled image encoding features.

[0102] The sample object and multiple candidate use cases can include multiple textual and image information. For example, if the sample object is a sample query, the query could be "What's in this picture?" and an image. In this case, the sample object contains the textual information "What's in this picture?" and the image information "The image". Similarly, candidate use cases can also include textual and image information. Feature labeling can be understood as labeling multiple text-encoded features and multiple image-encoded features separately.

[0103] Specifically, in the feature encoding unit, multiple text information and multiple image information contained in the sample object and multiple candidate use cases can be encoded separately to obtain multiple text encoding features corresponding to multiple text information and multiple image encoding features corresponding to multiple image information. Then, feature labeling processing is performed on the multiple text encoding features and multiple image encoding features to obtain multiple encoding features.

[0104] In practical applications, the feature encoding unit can be a multimodal encoding model. The multimodality can include other modalities such as text and images, such as video modality and audio modality. Whether other modalities are included here can be determined according to the target object processing task. If the target object processing task is an audio and video processing task, then the multimodality here can also include video modality and audio modality. The embodiments in this specification do not limit this, and reserve scalability for input of more other modalities.

[0105] Specifically, a multimodal coding model can encode multiple text information and multiple image information separately to obtain multiple text coding features and multiple image coding features. These multiple text coding features and multiple image coding features contain semantic information of the text information and the image information. Then, feature labeling processing can be performed on the multiple text coding features and multiple image coding features to indicate the differences between the coding features, thus obtaining multiple coding features. In one embodiment of this specification, the multimodal coding model can encode multiple text information and multiple image information into the same feature space; in another embodiment of this specification, the feature coding unit can also encode multiple text information and multiple image information into different feature spaces.

[0106] In summary, by processing multimodal features (i.e., text encoding features and image encoding features), efficient adaptive context learning can be achieved in multimodal scenarios, so that the trained target task processing model is no longer limited to applications in single-modal scenarios.

[0107] Further, the step of performing feature labeling processing on the plurality of text encoding features and the plurality of image encoding features to obtain a plurality of encoding features includes:

[0108] Based on the text information and the image information, a first feature marker and a second feature marker are added to the plurality of text encoding features and the plurality of image encoding features to obtain a plurality of encoding features;

[0109] Wherein, the first feature tag is used to tag the modality of the encoded feature, and the second feature tag is used to tag the association relationship between the encoded features.

[0110] The first feature marker can be used to indicate the modal difference of the encoded feature for subsequent fusion coding units. For example, the first feature marker can be used to mark whether the encoded feature is text modality, image modality, audio modality, or video modality, etc. That is, it can indicate whether the encoded feature is a text encoded feature of text modality or an image encoded feature of image modality for subsequent fusion coding units. For example, the first feature marker T can be used to indicate that the encoded feature is a text encoded feature, and the first feature marker I can be used to indicate that the encoded feature is an image encoded feature. Alternatively, the first feature marker can also be different colors to indicate the modal difference of the encoded feature. For example, text encoded features can be marked in red, and image encoded features can be marked in blue, etc. The embodiments in this specification do not limit this.

[0111] The second feature tag can be used to indicate the correspondence between encoded features in subsequent fusion coding units. Text and image information belonging to the same sample object or candidate use case can be identified as having a relationship. Therefore, the text and image encoded features corresponding to the text and image information belonging to the same sample object or candidate use case can be labeled with the same second feature tag. For example, if a sample object contains bound image and text information, then the image encoded feature corresponding to the image information and the text encoded feature corresponding to the text information can be labeled with the same second feature tag. If sample object 1 contains image information 1 and text information 1, and candidate use case 2 contains image information 2 and text information 2, then the image encoded feature corresponding to image information 1 and the text encoded feature corresponding to text information 1 can be labeled with second feature tag 1, and the image encoded feature corresponding to image information 2 and the text encoded feature corresponding to text information 2 can be labeled with second feature tag 2.

[0112] Based on this, according to the text information and image information, as well as the sample object and candidate use case to which the text information and image information belong, a first feature label is added to the text encoding feature to indicate that the encoding feature is a text modality, a first feature label is added to the image encoding feature to indicate that the encoding feature is an image modality, and a second feature label is added to the text encoding features and image encoding features corresponding to the text information and image information belonging to the same sample object or the same candidate use case, so as to obtain multiple encoding features after labeling.

[0113] In practical applications, the first feature label can be a Modal Embedding, and the second feature label can be a Positional Embedding. The task processing model can overlay different Modal Embeddings on encoded features of different modalities, thereby indicating the modal differences of encoded features for subsequent fusion encoding units. It can also overlay the same Positional Embedding on encoded features of different modalities belonging to the same group of sample objects or candidate use cases, thereby indicating the correspondence between encoded features for subsequent fusion encoding units.

[0114] In summary, by adding the first feature marker and the second feature marker, we can indicate the modal differences of the encoded features and the correlation between the encoded features for the subsequent fusion encoding units, ensuring that when multiple encoded features are fused, the interaction between the multimodal features and between the sample objects and the candidate use cases can be realized.

[0115] Furthermore, after obtaining multiple encoded features, these features can be directly input into the fusion encoding unit without mixing. Interactive encoding is then performed within the fusion encoding unit to achieve interaction between the sample object and candidate use cases, thereby obtaining fused encoded features. In specific implementations, when interactively encoding multiple encoded features in the fusion encoding unit, the multiple encoded features can be directly fused to obtain fused encoded features. Alternatively, self-attention processing can be applied to the multiple encoded features to obtain fused encoded features; this embodiment does not limit the specific approach. This interactive encoding of encoded features from different modalities allows the task processing model to mitigate the drawbacks of independently selecting use cases through the interaction between the sample object and candidate use cases.

[0116] In specific implementation, the step of using the feature decoding unit to decode the fused encoded features according to the use case selection strategy, and determining the target use case from the multiple candidate use cases, includes:

[0117] Using the feature decoding unit, the fused encoded features are decoded according to the use case selection strategy defined by the parameters of the feature decoding unit, and the target use case is determined from the multiple candidate use cases.

[0118] The use case selection strategy can be a strategy defined by the parameters of the feature decoding unit. The use case selection strategy can be understood as a strategy for selecting from multiple candidate use cases. This strategy can include random selection and selection based on the confidence level of candidate use cases, etc., and the embodiments in this specification do not limit this. It is understood that the task processing model can adaptively select different use case selection strategies from multiple candidate use cases to choose the target use case for different input sample objects.

[0119] In practical applications, this decoding process is a self-exploration process for the task processing model. The feature decoding unit can be a Transformer decoder, and the parameters of this Transformer decoder can define the use case selection strategy. Understandably, during subsequent adjustments to the parameters of the task processing model, the parameters of this Transformer decoder can be adjusted, thereby adjusting the use case selection strategy. This allows the trained target task processing model to adaptively select different use case selection strategies based on the input sample objects.

[0120] In summary, by decoding the fused encoded features obtained from processing the sample objects and multiple candidate use cases, the target use case is selected from multiple candidate use cases. The relationship between the sample objects (which may be user queries) and candidate use cases is taken into account, and adaptive context learning use case selection is achieved. It can dynamically select more suitable target use cases for different types of sample objects (which may be different types of user queries), rather than always using fixed target use cases, making context learning more granular and more generalizable.

[0121] Furthermore, there can be multiple target use cases identified from multiple candidate use cases. Understandably, the number of target use cases can be less than the number of candidate use cases. The specific implementation is as follows:

[0122] The step of using the feature decoding unit to decode the fused encoded features according to the use case selection strategy, and determining the target use case from the plurality of candidate use cases, includes:

[0123] Using the feature decoding unit, in response to the query vector, and according to the use case selection strategy, the fused encoded features are subjected to autoregressive decoding processing to determine a first target use case from the plurality of candidate use cases, wherein the query vector is used to instruct the feature decoding unit to start decoding;

[0124] Based on the first target use case, the fused coding features are subjected to autoregressive decoding processing to determine the second target use case from the plurality of candidate use cases;

[0125] Based on the second target use case, the process of performing autoregressive decoding on the fused coding features continues to be executed to obtain multiple target use cases, wherein the multiple target use cases include at least the first target use case and the second target use case.

[0126] The query vector can be understood as a decoding start instruction for the feature decoding unit, which can be used to instruct the feature decoding unit to start autoregressive decoding.

[0127] Specifically, in the autoregressive decoding process, the query vector can be used as the first input of the feature decoding unit to perform autoregressive decoding on the fused encoded features. The first target use case can be understood as one of the multiple candidate use cases output in the first round of autoregressive decoding. This first target use case can be used as the input of the second round of autoregressive decoding. Based on this first target use case, the fused encoded features are further autoregressively decoded to output the second target use case. This second target use case can be understood as one of the multiple candidate use cases output in the second round of autoregressive decoding. Subsequently, this second target use case can be used as the input of the third round of autoregressive decoding to continue autoregressive decoding.

[0128] Furthermore, during the autoregressive decoding process, the process can be terminated based on decoding end markers, maximum length limits, probability thresholds, and heuristic rules to obtain multiple target use cases. During the model training phase, the feature decoding unit can append a decoding end marker to the end of the selected target use cases it outputs. During the decoding phase, if the target use cases output by the feature decoding unit include this decoding end marker, the feature decoding unit can stop the autoregressive decoding process. Alternatively, the autoregressive decoding process will stop when the sequence of target use cases output by the feature decoding unit reaches a preset length. This specification does not limit the specific implementation of these embodiments.

[0129] In practical applications, during the self-exploration process of the task processing model, to model the selection of target use cases as a combinatorial problem, a Transformer decoder can be used to receive fused encoded features output from the fused encoding unit. The decoder's parameters define a use case selection strategy. The decoder takes a query vector as input and selects the target use cases from the fused encoded features according to the defined strategy. In practice, the use case selection strategy can be a probabilistic beam search strategy. The selection of target use cases can be based on probabilistic beam search (Stochastic Beam Search). During the selection process, the autoregressive decoding process ensures that the selection of each target use case is related to both the sample object and the previously selected target use case. This realizes the interaction between use cases and between use cases and objects (which can be the problem) in a combinatorial problem. Probabilistic beam search can avoid the use case selection becoming fixed and falling into suboptimal solutions; even use cases considered poor by the current strategy may be selected, thus achieving stochastic exploration of the strategy. Starting from the current strategy, the task processing model searches and generates multiple sets of use case combinations with different strategies for subsequent use.

[0130] In summary, by treating the selection of use cases in context learning as a composition problem, and through an autoregressive decoding process ensuring that each selected target use case is related to the sample object and the previously selected target use case, it becomes possible to enhance the interactivity between multiple use cases. Different use cases can provide complementary hints to the model, thereby forming a higher-quality combination of use cases (i.e., multiple selected target use cases). Furthermore, if two candidate use cases provide the same information, and both are selected as target use cases, it leads to information duplication. Excessive duplication can mislead model inference. By using autoregressive decoding to ensure that the selected target use case is related to the previously selected target use case, information redundancy is reduced and the accuracy of model inference is improved, thus overcoming the information redundancy problem that affects the quality of use cases.

[0131] Specifically, the step of using the object processing unit to perform prediction processing on the sample object according to the target use case to obtain the prediction object processing result includes:

[0132] Based on the target use case and the sample object, construct the prompt information;

[0133] The object processing unit is used to perform prediction processing on the prompt information to obtain the prediction object processing result.

[0134] The prompts can be used to guide the task processing model in reasoning.

[0135] In practical applications, this process is a self-utilization process of the task processing model. After determining the target use case through the aforementioned self-exploration process, the task processing model can construct a context-learning format of prompt information based on the sample object and the target use case. This prompt information is then input into the object processing unit, which learns and performs analogical reasoning based on the prompt information to provide a predicted object processing result. This object processing unit can be a large model.

[0136] Specifically, in order to achieve multimodal tasks, the constructed prompt information can include text, images referenced in the text, and other modal information.

[0137] Understandably, in practical applications, the above process can be repeated in order to adjust the parameters of the task processing model. In the self-utilization process, the sample objects of each repeated execution and the target use cases selected each time can be constructed into context learning format prompt information. The target use cases selected each time can be different. Thus, multiple sets of different target use cases can be obtained during the training process, and each set of target use cases includes multiple target use cases selected each time.

[0138] In summary, the self-exploration and self-utilization processes of the task processing model described above solve the problem of unclear use case selection criteria. It explores better use case combinations through iterative exploration and probabilistic beam search, and adaptively optimizes the sampling policy network (i.e., feature decoding unit) in the utilization phase using reinforcement learning.

[0139] Step 206: Adjust the parameters of the task processing model according to the processing results of the predicted object and the processing results of the labeled object until the training stopping condition is met, and obtain the target task processing model.

[0140] Specifically, after obtaining the predicted object processing result output by the task processing model, the parameters of the task processing model can be adjusted based on this predicted object processing result and the label object processing result until the training stopping condition is met, thus obtaining the target task processing model. It can be understood that one sample object can correspond to a set of target use cases (including multiple selected target use cases), and correspond to one predicted object processing result and one label object processing result. When multiple sample objects are input into the task processing model, the above processing procedure is performed for each sample object, selecting the corresponding target use case and obtaining the corresponding predicted object processing result.

[0141] The training stopping condition can be understood as the model loss value reaching a preset loss value threshold and / or the number of training rounds reaching a preset round threshold. The target task processing model can be the task processing model after the parameters have been adjusted.

[0142] In one embodiment of this specification, adjusting the parameters of the task processing model based on the processing results of the predicted object and the processing results of the labeled object until the training stopping condition is met to obtain the target task processing model includes:

[0143] Calculate the model loss value based on the processing results of the predicted object and the processing results of the labeled object;

[0144] Based on the model loss value, the parameters of the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

[0145] Specifically, the model loss value can be calculated based on the processing results of the predicted object and the processing results of the labeled object, and the parameters of the task processing model can be adjusted according to the model loss value until the training stopping condition is reached to obtain the target task processing model.

[0146] In another embodiment of this specification, the model loss value can be a scoring result. Then, the step of adjusting the parameters of the task processing model based on the processing results of the predicted object and the processing results of the labeled object until the training stopping condition is met to obtain the target task processing model includes:

[0147] The processing results of the predicted object and the processing results of the label object are compared to obtain the comparison results;

[0148] Based on the comparison results, determine the scoring result corresponding to the processing result of the predicted object;

[0149] Based on the scoring results, the parameters of the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

[0150] The scoring result can be understood as the score of the predicted object processing result. For example, if the similarity between the predicted object processing result and the label object processing result reaches a preset similarity threshold based on the comparison results, a higher score can be assigned to the predicted object processing result; if the similarity between the predicted object processing result and the label object processing result does not reach the preset similarity threshold, a lower score can be assigned to the predicted object processing result. Alternatively, a correspondence can be pre-configured between the comparison results and the scoring results; for example, if the comparison results show a high similarity, the corresponding scoring result will also be high. This specification does not impose any limitations on this. Comparing the predicted object processing result and the label object processing result can involve comparing the semantic information of the predicted object processing result and the label object processing result.

[0151] Specifically, the processing results of the predicted object and the processing results of the labeled object can be compared to obtain the comparison results. The processing results of the predicted object can be scored based on the comparison results to obtain the score of the processing results of the predicted object. The parameters of the task processing model can be adjusted based on the score until the training stopping condition is reached to obtain the target task processing model.

[0152] In practical applications, taking the sample object as the sample question as an example, the result of the prediction object processing can be the predicted answer, and the result of the label object processing can be the label answer of the sample question. Then, the predicted answer and the label answer can be compared to obtain the score of the predicted answer.

[0153] In summary, by calculating the score corresponding to the processing result of the predicted object and the processing result of the labeled object, and adjusting the parameters of the task processing model according to the score, reinforcement learning of the task processing model is achieved, thus realizing high-quality contextual learning.

[0154] Further, adjusting the parameters of the task processing model based on the scoring results until the training stopping condition is met to obtain the target task processing model includes:

[0155] The scoring results are standardized to obtain standardized scoring results;

[0156] Based on the standardized scoring results, the parameters of the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

[0157] Specifically, when adjusting parameters based on the scoring results, the scoring results can be standardized to obtain standardized scoring results. Based on the standardized scoring results, the parameters of the task processing model can be adjusted until the target task processing model that meets the training stopping condition is obtained.

[0158] In summary, since the standardized scoring results reflect the superiority or inferiority relationship between different use case selection strategies, the task processing model uses this superiority or inferiority relationship as a signal to achieve supervised adjustment of the parameters of the task processing model, thereby further ensuring the inference performance of the target task processing model.

[0159] Furthermore, the step of adjusting the parameters of the task processing model based on the processing results of the predicted object and the processing results of the labeled object until the training stopping condition is met to obtain the target task processing model also includes:

[0160] Based on the processing results of the predicted object and the processing results of the labeled object, the parameters of the fusion encoding unit and the feature decoding unit in the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

[0161] Specifically, when adjusting the parameters of the task processing model, the parameters of the fusion encoding unit and the feature decoding unit can be adjusted, thereby optimizing the use case selection strategy for the parameter definition of the feature decoding unit, and also optimizing the feature fusion processing of the fusion encoding unit.

[0162] In practical applications, the superiority or inferiority relationship between different use case selection strategies reflected by the standardized scoring results can be used as a signal to supervise the updating of parameters of the fusion coding unit and the feature decoding unit.

[0163] Understandably, the above process can be repeated continuously, thereby using only a small amount of training data (i.e., sample objects, label object processing results, and multiple candidate use cases) to continuously optimize the selection strategy of context learning use cases and continuously enhance the reasoning ability of the task processing model.

[0164] In practical applications, the trained target task processing model was tested on multiple different multimodal large models (i.e., object processing units) and multiple datasets. The test results all showed that the reasoning ability of the target task processing model was significantly improved.

[0165] In summary, the above method, to ensure the task processing model meets the processing requirements of the target object processing task, determines the sample object, label object processing results, and multiple candidate use cases corresponding to the target object processing task as training data. During training, the task processing model can infer according to the processing flow described by the candidate use cases based on the candidate use cases. Furthermore, in the task processing model, feature processing is performed on the sample object and multiple candidate use cases to obtain fused encoded features, realizing the interactive processing between the sample object and candidate use cases. Simultaneously, the fused encoded features are decoded according to the use case selection strategy to obtain the target use case, realizing the selection of multiple candidate use cases based on the sample object, and utilizing the selected target use case. The use cases predict the processing results of the sample objects, and adjust the parameters of the task processing model based on the processing results of the predicted objects and the processing results of the labeled objects until the target task processing model that meets the training stopping condition is obtained. This allows the target task processing model to select target use cases with higher quality corresponding to the processing results of the predicted objects through the interaction between the sample objects and the candidate use cases when selecting target use cases based on the use case selection strategy. This ensures the high quality of target use case selection and enables the target task processing model to automatically explore and utilize high-quality data (i.e., target use cases) from a small amount of training data (i.e., candidate use cases). This achieves the improvement of inference performance while ensuring the model fine-tuning cost.

[0166] The following is in conjunction with the appendix Figure 3 Taking the application of the task processing model training method provided in this specification in a question-answering scenario as an example, the task processing model training method will be further explained. Among other things, Figure 3 The flowchart of a task processing model training method provided in one embodiment of this specification is shown, which specifically includes the following steps.

[0167] Step 302: Determine the sample questions, tagged answers, and multiple candidate use cases corresponding to the question-answering task.

[0168] Step 304: Input the sample problem and multiple candidate use cases into the feature encoding unit, and encode the text information and image information contained in the sample problem and the multiple candidate use cases to obtain multiple text encoding features and multiple image encoding features.

[0169] Specifically, the feature encoding unit is a multimodal encoding model, which includes at least a text encoder and an image encoder. Text information can be encoded based on the text encoder, and image information can be encoded based on the image encoder to obtain multiple text-encoded features and multiple image-encoded features.

[0170] Step 306: Add a first feature tag and a second feature tag to multiple text encoding features and multiple image encoding features to obtain multiple encoding features.

[0171] Step 308: Input multiple coding features into the fusion coding unit, perform fusion processing on the multiple coding features, and obtain fused coding features.

[0172] Step 310: Input the fused encoded features into the feature decoding unit, and decode the fused encoded features according to the use case selection strategy defined by the parameters of the feature decoding unit to determine the target use case from multiple candidate use cases.

[0173] Specifically, the query vector BOS can be input into the feature decoding unit as the start of the autoregressive decoding process. According to the use case selection strategy defined by the parameters of the feature decoding unit, the fused encoded features are subjected to autoregressive decoding processing, and the first target use case d1 is output. The first target use case d1 is used as the input for the next round of autoregressive decoding processing, and the second target use case d2 is output. The second target use case d2 is used as the input for the next round of autoregressive decoding processing, and the third target use case d3 is output. The first target use case d1, the second target use case d2, and the third target use case d3 are all selected as target use cases from multiple candidate use cases.

[0174] Step 312: Based on the sample question and the target use case, construct the prompt information, input the prompt information into the object processing unit, and obtain the predicted answer.

[0175] Specifically, for multiple input sample questions, prompts can be constructed for each sample question based on its corresponding target use case. These prompts are then input into the object processing unit (i.e., the multimodal large model) to obtain predicted answers for the multiple sample questions. Taking three sample questions as an example: for sample question 1, target use case 1 and predicted answer 1 are obtained; for sample question 2, target use case 2 and predicted answer 2 are obtained; and for sample question 3, target use case 3 and predicted answer 3 are obtained.

[0176] Step 314: Determine the score corresponding to the predicted answer based on the predicted answer and the labeled answer.

[0177] Specifically, the score for each predicted answer can be determined based on each predicted answer and the corresponding labeled answer for each sample question.

[0178] Step 316: Adjust the parameters of the fusion coding unit and the feature decoding unit according to the scores.

[0179] Specifically, the score corresponding to each predicted answer can be standardized to obtain a standardized score, and the parameters of the fusion coding unit and the feature decoding unit can be adjusted based on the standardized score.

[0180] Steps 302-316 above can be repeated to fine-tune the task processing model.

[0181] In summary, this paper treats demonstration selection in in-context learning as a combinatorial problem, innovatively considering the interaction between questions and demonstrations, and between demonstrations themselves. It also examines the challenge of lacking clear guidance and judgment criteria for demonstration selection in in-context learning. Finally, combining reinforcement learning approaches, a multimodal large-scale model reasoning capability enhancement scheme based on in-context learning technology is constructed, effectively supporting the wide-ranging application of large-scale intelligent agents. This framework (i.e., the task processing model) uses the provided training data as the basis for judging the correctness of the model's output answer for any pre-trained multimodal large-scale model, and then explores the selection strategy for context learning samples on the large model. Starting with a random selection strategy, the framework explores multiple new demonstration selection strategies, generating multiple sets of different multimodal demonstration sets for user questions. These different demonstration sets serve as the learning basis for the large model to learn and for analogical reasoning to fine-tune the data's answer results. Then, the framework judges the correctness of the answer results, selects the strategy corresponding to the higher-quality answer as the optimization direction, and updates the current random selection strategy. This process is continuously repeated to continuously optimize strategy performance. This framework automates the exploration of multimodal context learning use case selection strategies applicable to diverse scenarios and questions, effectively supporting reasoning enhancement for large model agents in multimodal scenarios, and achieving a better balance between performance and cost.

[0182] Corresponding to the above method embodiments, this specification also provides embodiments of a task processing model training device. Figure 4 A schematic diagram of a task processing model training device according to one embodiment of this specification is shown. Figure 4 As shown, the device includes:

[0183] The determination module 402 is configured to determine the sample object, the label object processing result, and multiple candidate use cases corresponding to the target object processing task, wherein the multiple candidate use cases are used to describe the processing flow of the target object processing task;

[0184] Input module 404 is configured to input the sample object and the plurality of candidate use cases into a task processing model to obtain a predicted object processing result of the sample object, wherein the predicted object processing result is obtained by the task processing model predicting the sample object based on a target use case, the target use case is obtained by the task processing model decoding the fused encoding features according to a use case selection strategy, and the fused encoding features are obtained by the task processing model performing feature fusion processing on the sample object and the plurality of candidate use cases;

[0185] The adjustment module 406 is configured to adjust the parameters of the task processing model based on the processing results of the predicted object and the processing results of the labeled object until the training stopping condition is met, thereby obtaining the target task processing model.

[0186] In an optional embodiment, the input module 404 is further configured to:

[0187] The sample object and the plurality of candidate use cases are input into the task processing model, wherein the task processing model includes a feature encoding unit, a fusion encoding unit, a feature decoding unit and an object processing unit;

[0188] Using the feature encoding unit, feature processing is performed on the sample object and the multiple candidate use cases to obtain multiple encoded features;

[0189] The multiple coding features are fused using the fusion coding unit to obtain fused coding features;

[0190] Using the feature decoding unit, the fused encoded features are decoded according to the use case selection strategy to determine the target use case from the multiple candidate use cases;

[0191] Using the object processing unit, the sample object is predicted based on the target use case to obtain the predicted object processing result.

[0192] In an optional embodiment, the input module 404 is further configured to:

[0193] The feature encoding unit is used to encode the multimodal information contained in the sample object and the multiple candidate use cases to obtain multiple multimodal encoded features;

[0194] The multiple multimodal coding features are subjected to feature labeling processing to obtain multiple coding features.

[0195] In an optional embodiment, the input module 404 is further configured to:

[0196] Using the feature encoding unit, the text and image information contained in the sample object and the multiple candidate use cases are encoded to obtain multiple text encoding features and multiple image encoding features;

[0197] Feature labeling processing is performed on the multiple text encoding features and the multiple image encoding features to obtain multiple encoding features, wherein the multiple encoding features include multiple labeled text encoding features and multiple labeled image encoding features.

[0198] In an optional embodiment, the input module 404 is further configured to:

[0199] Based on the text information and the image information, a first feature marker and a second feature marker are added to the plurality of text encoding features and the plurality of image encoding features to obtain a plurality of encoding features;

[0200] Wherein, the first feature tag is used to tag the modality of the encoded feature, and the second feature tag is used to tag the association relationship between the encoded features.

[0201] In an optional embodiment, the input module 404 is further configured to:

[0202] Using the feature decoding unit, the fused encoded features are decoded according to the use case selection strategy defined by the parameters of the feature decoding unit, and the target use case is determined from the multiple candidate use cases.

[0203] In an optional embodiment, the input module 404 is further configured to:

[0204] Using the feature decoding unit, in response to the query vector, and according to the use case selection strategy, the fused encoded features are subjected to autoregressive decoding processing to determine the first target use case from the plurality of candidate use cases;

[0205] Based on the first target use case, the fused coding features are subjected to autoregressive decoding processing to determine the second target use case from the plurality of candidate use cases;

[0206] Based on the second target use case, the process of performing autoregressive decoding on the fused coding features continues to be executed to obtain multiple target use cases, wherein the multiple target use cases include at least the first target use case and the second target use case.

[0207] In an optional embodiment, the input module 404 is further configured to:

[0208] Based on the target use case and the sample object, construct the prompt information;

[0209] The object processing unit is used to perform prediction processing on the prompt information to obtain the prediction object processing result.

[0210] In an optional embodiment, the adjustment module 406 is further configured to:

[0211] Calculate the model loss value based on the processing results of the predicted object and the processing results of the labeled object;

[0212] Based on the model loss value, the parameters of the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

[0213] In an optional embodiment, the adjustment module 406 is further configured to:

[0214] The processing results of the predicted object and the processing results of the label object are compared to obtain the comparison results;

[0215] Based on the comparison results, determine the scoring result corresponding to the processing result of the predicted object;

[0216] Based on the scoring results, the parameters of the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

[0217] In an optional embodiment, the adjustment module 406 is further configured to:

[0218] The scoring results are standardized to obtain standardized scoring results;

[0219] Based on the standardized scoring results, the parameters of the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

[0220] In an optional embodiment, the adjustment module 406 is further configured to:

[0221] Based on the processing results of the predicted object and the processing results of the labeled object, the parameters of the fusion encoding unit and the feature decoding unit in the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

[0222] In the aforementioned apparatus, to enable the task processing model to meet the processing requirements of the target object processing task, the sample object, the processing results of the labeled object, and multiple candidate use cases corresponding to the target object processing task are determined as training data. During training, the task processing model is able to infer according to the processing flow described by the candidate use cases based on the candidate use cases. Furthermore, in the task processing model, feature processing is performed on the sample object and multiple candidate use cases to obtain fused encoded features, realizing the interactive processing between the sample object and the candidate use cases. Simultaneously, the fused encoded features are decoded according to the use case selection strategy to obtain the target use case, realizing the selection of multiple candidate use cases based on the sample object, and using the selected target use case to predict the sample object. The processing results of the predicted objects are obtained, and the parameters of the task processing model are adjusted based on the processing results of the predicted objects and the processing results of the labeled objects until the target task processing model that meets the training stopping condition is obtained. This allows the target task processing model to select target use cases corresponding to higher quality predicted object processing results through the interaction between sample objects and candidate use cases when selecting target use cases based on the use case selection strategy. This ensures high quality of target use case selection, and enables the target task processing model to automatically explore and utilize high-quality data (i.e., target use cases) from a small amount of training data (i.e., sample objects, label object processing results, and multiple candidate use cases). This achieves improved inference performance while ensuring the model fine-tuning cost.

[0223] The above is an illustrative scheme of a task processing model training device according to this embodiment. It should be noted that the technical solution of this task processing model training device and the technical solution of the task processing model training method described above belong to the same concept. For details not described in detail in the technical solution of the task processing model training device, please refer to the description of the technical solution of the task processing model training method described above.

[0224] Corresponding to the above method embodiments, this specification provides a task processing method, see [link to relevant documentation]. Figure 5 , Figure 5 A flowchart of a task processing method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0225] Step 502: Determine the object to be processed corresponding to the target object processing task;

[0226] Step 504: Input the object to be processed into the target task processing model to obtain the object processing result of the object to be processed, wherein the target task processing model is trained according to the task processing model training method provided in the first aspect of the embodiments of this specification.

[0227] Specifically, after training the target task processing model according to the above task processing model training method, the object to be processed corresponding to the target object processing task can be determined, and the object to be processed can be input into the target task processing model to obtain the object processing result of the object to be processed.

[0228] For example, if the target object processing task is a text processing task, it can receive the text to be processed sent by the user, input the text to be processed into the target task processing model, and obtain the text processing result.

[0229] The above is an illustrative scheme of a task processing method according to this embodiment. It should be noted that the technical solution of this task processing method belongs to the same concept as the technical solution of the task processing model training method described above. For details not described in detail in the technical solution of the task processing method, please refer to the description of the technical solution of the task processing model training method described above.

[0230] Corresponding to the above method embodiments, this specification provides a task processing apparatus, see [link to relevant documentation]. Figure 6 , Figure 6 A structural diagram of a task processing apparatus according to an embodiment of this specification is shown, the apparatus comprising:

[0231] The determination module 602 is configured to determine the object to be processed corresponding to the target object processing task;

[0232] The input module 604 is configured to input the object to be processed into the target task processing model to obtain the object processing result of the object to be processed, wherein the target task processing model is trained according to the task processing model training method provided in the first aspect of the embodiments of this specification.

[0233] The above is an illustrative scheme of a task processing device according to this embodiment. It should be noted that the technical solution of this task processing device and the technical solution of the task processing model training method described above belong to the same concept. For details not described in detail in the technical solution of the task processing device, please refer to the description of the technical solution of the task processing model training method described above.

[0234] Corresponding to the above method embodiments, this specification provides a query processing method, see [link to relevant documentation]. Figure 7 , Figure 7 A flowchart of a query processing method according to an embodiment of this specification is shown, which specifically includes the following steps.

[0235] Step 702: Determine the query to be processed;

[0236] Step 704: Input the query to be processed into the target task processing model to obtain the query result of the query to be processed, wherein the target task processing model is trained according to the task processing model training method provided in the first aspect of the embodiments of this specification.

[0237] The target task processing model can be trained using the above-mentioned task processing model training method when the target object processing task is a query processing task.

[0238] Specifically, it can receive pending queries sent by users through the client, input the pending query into the target task model, and obtain the query results of the pending query.

[0239] The above is an illustrative scheme of a query processing method according to this embodiment. It should be noted that the technical solution of this query processing method and the technical solution of the task processing model training method described above belong to the same concept. For details not described in detail in the technical solution of the query processing method, please refer to the description of the technical solution of the task processing model training method described above.

[0240] Corresponding to the above method embodiments, this specification provides a query processing device, see [link to documentation]. Figure 8 , Figure 8 A structural diagram of a query processing apparatus according to an embodiment of this specification is shown. The apparatus includes:

[0241] Module 802 is configured to determine queries to be processed.

[0242] The input module 804 is configured to input the query to be processed into the target task processing model to obtain the query result of the query to be processed, wherein the target task processing model is trained according to the task processing model training method provided in the first aspect of the embodiments of this specification.

[0243] The above is an illustrative scheme of a query processing device according to this embodiment. It should be noted that the technical solution of this query processing device and the technical solution of the task processing model training method described above belong to the same concept. For details not described in detail in the technical solution of the query processing device, please refer to the description of the technical solution of the task processing model training method described above.

[0244] Corresponding to the above method embodiments, this specification also provides a task processing method, which specifically includes the following steps.

[0245] Identify the objects to be processed and multiple candidate use cases corresponding to the target object processing task;

[0246] The object to be processed and the multiple candidate use cases are input into the target task processing model to obtain the object processing result of the object to be processed. The object processing result is obtained by the target task processing model predicting the object to be processed based on the target use cases. The target use cases are obtained by the target task processing model decoding the fused encoding features according to the use case selection strategy. The fused encoding features are obtained by the target task processing model performing feature fusion processing on the object to be processed and the multiple candidate use cases.

[0247] The target task processing model can be trained using the task processing model training method described above.

[0248] Specifically, it can receive the object to be processed corresponding to the target object processing task input by the user, determine multiple candidate use cases corresponding to the target object processing task, and input the object to be processed and multiple candidate use cases into the target task processing model. In the target task processing model, the object to be processed and multiple candidate use cases are fused to obtain fused encoded features. The fused encoded features are decoded according to the use case selection strategy to obtain the target use case. The object to be processed is predicted according to the target use case to obtain the object processing result.

[0249] The above is an illustrative scheme of a task processing method according to this embodiment. It should be noted that the technical solution of this task processing method belongs to the same concept as the technical solution of the task processing model training method described above. For details not described in detail in the technical solution of the task processing method, please refer to the description of the technical solution of the task processing model training method described above.

[0250] Figure 9 A structural block diagram of a computing device 900 provided according to one embodiment of this specification is shown.

[0251] The computing device 900 includes:

[0252] Memory 910 and processor 920;

[0253] The memory 910 is used to store computer programs / instructions, and the processor 920 is used to execute the computer programs / instructions, which, when executed by the processor 920, implement the steps of the above method.

[0254] In one or more embodiments of this specification, the computing device can be understood as an integrated smart terminal, including but not limited to a server, desktop computer, PC (Personal Computer), all-in-one model machine, mobile phone, tablet computer or other portable smart terminal, etc., and the computing device may have the model described in the above embodiments of this application pre-installed.

[0255] Specifically, this computing device can pre-install various types of models, including but not limited to models in natural language processing, visual processing, speech processing, code processing, and multimodal task processing, thus providing diverse model selection. In different product forms, this computing device can support one or more model usage methods, including but not limited to model training, model invocation, model fine-tuning, model deployment, model inference, and application. In some product forms, this computing device also supports model management, including but not limited to multi-type model management (supporting the management of discriminative, generative, and other types of models), model version control (supporting the control of different model versions), and model evaluation (evaluating model performance and effectiveness based on model evaluation tools). In other product forms, this computing device can also create applications based on models, providing API (Application Programming Interface) calling capabilities. Users can call models into created applications through the API interface, and application management tools are also provided to manage and monitor the applications.

[0256] Furthermore, the computing device may also include data management (supporting the creation and management of model tuning datasets), a training center (providing abundant training resources to help users learn and master AI technology), and basic control capabilities (providing enterprise-level basic control capabilities to ensure the security and efficient operation of the system). Through the above functions, it provides a comprehensive and integrated device for AI development, training, deployment, and application.

[0257] Figure 10 A structural block diagram of an electronic device 1000 provided according to one embodiment of this specification is shown.

[0258] The memory 1010 and the processor 1020 are connected via a bus 1030;

[0259] The memory 1010 is used to store computer programs / instructions, and the processor 1020 is used to execute the computer programs / instructions, which, when executed by the processor 1020, implement the steps of the method.

[0260] Specifically, the components of the electronic device 1000 include, but are not limited to, a memory 1010 and a processor 1020. The processor 1020 is connected to the memory 1010 via a bus 1030, and the database 1050 is used to store data.

[0261] Electronic device 1000 also includes access device 1040, which enables electronic device 1000 to communicate via one or more networks 1060. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. Access device 1040 may include one or more of any type of wired or wireless network interface (e.g., network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, a Near Field Communication (NFC) interface, and so on.

[0262] In one embodiment of this specification, the above-described components of the electronic device 1000 and Figure 10 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 10 The block diagram of the electronic device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0263] Electronic device 1000 can be any type of stationary or mobile electronic device, including mobile computers or mobile electronic devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable electronic devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary electronic devices such as desktop computers or personal computers (PCs). Electronic device 1000 can also be a mobile or stationary server.

[0264] The above is an illustrative scheme of an electronic device according to this embodiment. It should be noted that the technical solution of this electronic device and the technical solution of the above method belong to the same concept, and all details not described in detail in the technical solution of the electronic device can be referred to the description of the technical solution of the above method.

[0265] An embodiment of this specification also provides a computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0266] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium and the technical solution of the method described above belong to the same concept, and all details not described in detail in the technical solution of the storage medium can be referred to the description of the technical solution of the method described above.

[0267] An embodiment of this specification also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described method.

[0268] The above is an illustrative scheme of a computer program product according to this embodiment. It should be noted that the technical solution of this computer program product and the technical solution of the above method belong to the same concept, and all details not described in detail in the technical solution of the computer program product can be referred to the description of the technical solution of the above method.

[0269] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0270] The computer program / instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0271] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0272] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0273] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A method for training a task processing model, comprising: The sample object, label object processing results, and multiple candidate use cases corresponding to the target object processing task are determined, wherein the multiple candidate use cases are used to describe the processing flow of the target object processing task; The sample object and the multiple candidate use cases are input into the task processing model to obtain the predicted object processing result of the sample object. The predicted object processing result is obtained by the task processing model predicting the sample object according to the target use case. The target use case is obtained by the task processing model decoding the fused encoding features according to the use case selection strategy. The fused encoding features are obtained by the task processing model performing feature fusion processing on the sample object and the multiple candidate use cases. Based on the processing results of the predicted object and the processing results of the labeled object, the parameters of the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

2. The method according to claim 1, wherein inputting the sample object and the plurality of candidate use cases into the task processing model to obtain the predicted object processing result of the sample object includes: The sample object and the plurality of candidate use cases are input into the task processing model, wherein the task processing model includes a feature encoding unit, a fusion encoding unit, a feature decoding unit and an object processing unit; Using the feature encoding unit, feature processing is performed on the sample object and the multiple candidate use cases to obtain multiple encoded features; The multiple coding features are fused using the fusion coding unit to obtain fused coding features; Using the feature decoding unit, the fused encoded features are decoded according to the use case selection strategy to determine the target use case from the multiple candidate use cases; Using the object processing unit, the sample object is predicted based on the target use case to obtain the predicted object processing result.

3. The method according to claim 2, wherein the step of using the feature encoding unit to perform feature processing on the sample object and the plurality of candidate use cases to obtain a plurality of encoded features includes: The feature encoding unit is used to encode the multimodal information contained in the sample object and the multiple candidate use cases to obtain multiple multimodal encoded features; The multiple multimodal coding features are subjected to feature labeling processing to obtain multiple coding features.

4. The method according to claim 3, wherein the step of using the feature encoding unit to encode the multimodal information contained in the sample object and the plurality of candidate use cases to obtain a plurality of multimodal encoded features includes: Using the feature encoding unit, the text and image information contained in the sample object and the multiple candidate use cases are encoded to obtain multiple text encoding features and multiple image encoding features; The step of performing feature labeling processing on the multiple multimodal coding features to obtain multiple coding features includes: Feature labeling processing is performed on the multiple text encoding features and the multiple image encoding features to obtain multiple encoding features, wherein the multiple encoding features include multiple labeled text encoding features and multiple labeled image encoding features.

5. The method according to claim 4, wherein performing feature labeling processing on the plurality of text encoding features and the plurality of image encoding features to obtain a plurality of encoding features includes: Based on the text information and the image information, a first feature marker and a second feature marker are added to the plurality of text encoding features and the plurality of image encoding features to obtain a plurality of encoding features; Wherein, the first feature tag is used to tag the modality of the encoded feature, and the second feature tag is used to tag the association relationship between the encoded features.

6. The method according to claim 2, wherein the step of using the feature decoding unit to decode the fused encoded features according to the use case selection strategy to determine the target use case from the plurality of candidate use cases includes: Using the feature decoding unit, the fused encoded features are decoded according to the use case selection strategy defined by the parameters of the feature decoding unit, and the target use case is determined from the multiple candidate use cases.

7. The method according to claim 2, wherein the step of using the feature decoding unit to decode the fused encoded features according to the use case selection strategy to determine the target use case from the plurality of candidate use cases includes: Using the feature decoding unit, in response to the query vector, and according to the use case selection strategy, the fused encoded features are subjected to autoregressive decoding processing to determine a first target use case from the plurality of candidate use cases, wherein the query vector is used to instruct the feature decoding unit to start decoding; Based on the first target use case, the fused coding features are subjected to autoregressive decoding processing to determine the second target use case from the plurality of candidate use cases; Based on the second target use case, the process of performing autoregressive decoding on the fused coding features continues to be executed to obtain multiple target use cases, wherein the multiple target use cases include at least the first target use case and the second target use case.

8. The method according to claim 2, wherein the step of using the object processing unit to perform prediction processing on the sample object according to the target use case to obtain the prediction object processing result includes: Based on the target use case and the sample object, construct the prompt information; The object processing unit is used to perform prediction processing on the prompt information to obtain the prediction object processing result.

9. The method according to any one of claims 1-8, wherein adjusting the parameters of the task processing model based on the processing result of the predicted object and the processing result of the labeled object until the training stopping condition is met to obtain the target task processing model includes: Calculate the model loss value based on the processing results of the predicted object and the processing results of the labeled object; Based on the model loss value, the parameters of the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

10. The method according to any one of claims 1-8, wherein adjusting the parameters of the task processing model based on the processing result of the predicted object and the processing result of the labeled object until the training stopping condition is met to obtain the target task processing model includes: The processing results of the predicted object and the processing results of the label object are compared to obtain the comparison results; Based on the comparison results, determine the scoring result corresponding to the processing result of the predicted object; Based on the scoring results, the parameters of the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

11. The method according to claim 10, wherein adjusting the parameters of the task processing model based on the scoring result until the training stopping condition is met to obtain the target task processing model includes: The scoring results are standardized to obtain standardized scoring results; Based on the standardized scoring results, the parameters of the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

12. The method according to any one of claims 1-8, wherein adjusting the parameters of the task processing model based on the processing result of the predicted object and the processing result of the labeled object until the training stopping condition is met to obtain the target task processing model includes: Based on the processing results of the predicted object and the processing results of the labeled object, the parameters of the fusion encoding unit and the feature decoding unit in the task processing model are adjusted until the training stopping condition is met, thereby obtaining the target task processing model.

13. A task processing method, comprising: Identify the objects to be processed corresponding to the target object's processing task; The object to be processed is input into the target task processing model to obtain the object processing result of the object to be processed, wherein the target task processing model is trained according to any one of claims 1-12.

14. A query processing method, comprising: Identify queries to be processed; The query to be processed is input into the target task processing model to obtain the query result of the query to be processed, wherein the target task processing model is trained by the method according to any one of claims 1-12.

15. A task processing method, comprising: Identify the objects to be processed and multiple candidate use cases corresponding to the target object processing task; The object to be processed and the multiple candidate use cases are input into the target task processing model to obtain the object processing result of the object to be processed. The object processing result is obtained by the target task processing model predicting the object to be processed based on the target use cases. The target use cases are obtained by the target task processing model decoding the fused encoding features according to the use case selection strategy. The fused encoding features are obtained by the target task processing model performing feature fusion processing on the object to be processed and the multiple candidate use cases.

16. A computing device, comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 15.

17. An electronic device comprising: A memory and a processor, the memory and the processor being connected via a bus; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 15.

18. A computer-readable storage medium storing a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 15.

19. A computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1 to 15.

Images