Question and answer model training method, question and answer processing method and task platform

Abstract

The embodiment of the invention provides a question and answer model training method, a question and answer processing method and a task platform, and the question and answer model training method comprises the steps: obtaining first training data which comprises first to-be-analyzed data, a first cognitive question aiming at the first to-be-analyzed data and first answer information corresponding to the first cognitive question, the first reply information is located at a first position in the first to-be-analyzed data; generating a first perception question based on the first cognitive question and the first position; inputting the first to-be-analyzed data and the first perception question into a pre-trained question and answer model, predicting first reference reply information corresponding to the first perception question, and predicting second reference reply information of the first perception question based on the first reference reply information; and based on the first reply information, the first position, the first reference reply information and the second reference reply information, adjusting a pre-trained question and answer model to obtain a trained question and answer model. The question and answer task execution effect of the question and answer model obtained in this way is good.

Description

Technical Field

[0001] The embodiments in this specification relate to the fields of artificial intelligence technology and terminal technology, and in particular to a question-answering model training method, question-answering processing method, and task platform. Background Technology

[0002] With the rapid development of artificial intelligence technology, various models are being widely used to help people perform data analysis and processing, and improve the convenience of people's work and life.

[0003] For example, an AI model can perform question-and-answer tasks, providing answers to questions posed by users. Users can input a document to be analyzed and a question related to that document into the AI ​​model. The AI ​​model can then analyze the document based on the question and output the answer to that question.

[0004] However, currently, for some perceptual and cognitive questions that should correspond to the same document and require the same response, AI models provide different answers. Therefore, the performance of question-answering tasks still needs improvement. Summary of the Invention

[0005] In view of this, embodiments of this specification provide a training method for a question-answering model, which can improve the performance of question-answering tasks. One or more embodiments of this specification also relate to a question-answering processing method, a task platform, a computing device, a computer-readable storage medium, and a computer program product.

[0006] According to a first aspect of the embodiments of this specification, a method for training a question-answering model is provided, comprising:

[0007] Acquire first training data, wherein the first training data includes: first data to be analyzed, a first cognitive question for the first data to be analyzed, a first response information corresponding to the first cognitive question, and a first position of the first response information in the first data to be analyzed;

[0008] Based on the first cognitive question and the first location, a first perceptual question is generated; wherein, the first perceptual question indicates the location of the response information to the first cognitive question in the first data to be analyzed, based on the first location;

[0009] The first data to be analyzed and the first perception question are input into a pre-trained question-answering model to predict the first reference answer information corresponding to the first perception question, and the second reference answer information of the first perception question is predicted based on the first reference answer information.

[0010] Based on the first response information, the first location, the first reference response information, and the second reference response information, the pre-trained question-answering model is adjusted to obtain the trained question-answering model.

[0011] According to a second aspect of the embodiments of this specification, a question-and-answer processing method is provided, including:

[0012] Obtain task data for a question-and-answer task, wherein the task data includes data to be analyzed and question information raised in response to the data to be analyzed;

[0013] The question-answering model is invoked to analyze the task data and obtain the answer information corresponding to the question information; wherein, the question-answering model is the question-answering model trained by the above training method.

[0014] According to a third aspect of the embodiments of this specification, a question-answering processing method is provided, applied to a task platform, including:

[0015] The device receives a model request sent by a terminal device, wherein the model request includes at least one of the following: scene identifier of the target scene, scene input data of the target scene, and target model specification parameters.

[0016] Based on the model request, a corresponding question-answering model is determined from at least one model; wherein the question-answering model is trained based on the training method of the question-answering model described above, and the question-answering model is used to output the answer information corresponding to the question information.

[0017] According to a fourth aspect of the embodiments of this specification, a task platform is provided, including a request interface and a response unit;

[0018] The request interface is used to receive a model request for unit testing sent by a terminal device, wherein the model request includes at least one of the following: a scene identifier of the target scene, scene input data of the target scene, and model specification parameters.

[0019] The response unit is configured to determine a corresponding question-answering model from at least one model based on the model request, wherein the question-answering model is trained using the training method described above.

[0020] According to a fifth aspect of the embodiments of this specification, a computing device is provided, comprising:

[0021] Memory and processor;

[0022] The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions. When the computer programs / instructions are executed by the processor, they implement the steps of the above-mentioned question-answering processing method or question-answering model training method.

[0023] According to a sixth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions, which, when executed by a processor, implement the steps of the above-described question-answering processing method or question-answering model training method.

[0024] According to a seventh 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 question-answering processing method or question-answering model training method.

[0025] In one embodiment of this specification, a first perceptual question can be generated based on a first cognitive question for a first set of data to be analyzed. This first perceptual question indicates the position of the response information for the first cognitive question within the first set of data to be analyzed. The first perceptual question connects the cognitive question and the perceptual question for the first set of data to be analyzed. A pre-trained question-answering model is trained based on the first set of data to be analyzed and the first perceptual question. This involves predicting a first reference response for the first cognitive question and, based on the first reference response, predicting a second reference response for the first perceptual question. The model is then adjusted by combining the first position and the first response information for the first cognitive question. By linking and fusing the cognitive and perceptual tasks for the first set of data to be analyzed to train the question-answering model, the consistency of the cognitive and perceptual tasks in the question-answering model can be improved, the accuracy of the output response information can be enhanced, and the performance of the question-answering task can be improved. Attached Figure Description

[0026] Figure 1 This is a schematic diagram illustrating the performance of cognitive and perceptual tasks on images of a portion of a scanned document;

[0027] Figure 2 This is a schematic diagram illustrating the performance of cognitive and perceptual tasks on real-world images from a report.

[0028] Figure 3 This is an application scenario architecture diagram of a question-and-answer processing method provided in one embodiment of this specification;

[0029] Figure 4 This is a flowchart illustrating a training method for a question-answering model provided in one embodiment of this specification;

[0030] Figure 5 This is a flowchart of a question-and-answer processing method provided in one embodiment of this specification;

[0031] Figure 6 This is a flowchart of another question-and-answer processing method provided in one embodiment of this specification;

[0032] Figure 7This is a flowchart of another question-and-answer processing method provided in one embodiment of this specification;

[0033] Figure 8 This is a schematic diagram of the structure of a task platform provided in one embodiment of this specification;

[0034] Figure 9 This is a schematic diagram of the structure of a training device for a question-answering model provided in one embodiment of this specification;

[0035] Figure 10 This is a schematic diagram of the structure of a question-and-answer processing device provided in one embodiment of this specification;

[0036] Figure 11 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0037] 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.

[0038] 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 limit the scope of one or more embodiments of this specification. The singular forms “a” and “the” 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” used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed procedures. 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 used only 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”, “in response to a determination”, or “when…”.

[0039] 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.

[0040] In one or more embodiments of this specification, the model referred to may be a large model. A large model is 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 a pre-trained model with hundreds of millions of parameters produced by pre-training on a large-scale unlabeled corpus. Such a model can adapt to a wide range of downstream tasks and has good generalization ability. Examples include large-scale language models (LLMs) and multi-modal pre-training models. In practical applications, large models require only a small number of samples to fine-tune the pre-trained model before being 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 NLP tasks such as text-based sentiment classification, text summarization, and machine translation. Major application scenarios for large models include digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design. The models mentioned in the embodiments of this specification include, but are not limited to, the large models described above, and can be any type of machine learning model, such as convolutional neural network (CNN) models, Transformer models, etc.

[0041] With the development of artificial intelligence technology, large models are being used more and more widely, possessing the ability to understand language and generate content. Large models can also process data of multiple modalities (such as text, audio, and images); these are called multimodal large language models (MLLMs). Large models can perform question-answering tasks; users can input questions into the model, and the model can generate answers and provide feedback to the user. However, current large models suffer from knowledge conflict issues. For example, when generating answers to questions with the same semantics but different syntax, the model may exhibit inconsistent behavior. Large models also suffer from object illusion problems, producing inaccurate, erroneous, or fabricated information, leading to poor performance. For instance, when a large model identifies objects in a given image, the output object information may not exist in the image; this phenomenon is called object illusion.

[0042] Large models can perform perceptual and cognitive tasks. Perceptual tasks mainly involve extracting information from input data, recognizing patterns, or performing classification. These tasks typically rely on the model's direct understanding and processing capabilities of the input data. The object illusion phenomenon described above falls under the category of perceptual tasks. Cognitive tasks involve higher-level cognitive processes such as reasoning, understanding, generation, and decision-making. Cognitive tasks require models to possess not only basic comprehension abilities but also logical reasoning, contextual understanding, and knowledge application capabilities.

[0043] Currently, models sometimes experience conflicts when handling cognitive and perceptual tasks. Taking document understanding as an example, cognitive tasks mainly include visual question answering (VQA), which uses the text in the document as the answer; perceptual tasks mainly include optical character recognition (OCR), which identifies the location of text within the document. Ideally, the results of visual question answering and optical character recognition for the same location in the document should be consistent, meaning the model's cognition and perception should be consistent. However, in reality, model cognition and perception often conflict.

[0044] Below are two real-world examples of conflict where the model may give inconsistent responses when the ground truth values ​​for the cognitive and perceptual tasks in a document image involve the same text. In one case, Figure 1This is a schematic diagram illustrating the performance of cognitive and perceptual tasks on a portion of an image from a scanned document. When the input question to the model is "What is the name of the BRMPermit Holder?", the model outputs "Doraf," which is a cognitive task. When the input question is "What is the content within the marked box in the image?", the model outputs "Doral," which is a perceptual task. Therefore, in this example, the model outputs an incorrect response during the cognitive task but a correct response during the perceptual task. In another example… Figure 2 This diagram illustrates the application of cognitive and perceptual tasks to a real-life image of a report. When the input question to the model is "What was the permitted quantity recorded in the 1978 report?", the model outputs "3598," which represents a cognitive task. When the input question is "What is the content within the marked box in the image?", the model outputs "3593," which represents a perceptual task. Therefore, in this case, the model outputs the correct answer during the cognitive task but an incorrect answer during the perceptual task.

[0045] By conducting cognitive and perceptual consistency analysis on various datasets, including scanned documents, photographed receipts, charts, and tables, it was found that cognitive and perceptual conflicts exist for different types of data. Even the most advanced multimodal large models currently available exhibit this problem; for example, the consistency of the GPT-4o model is only 62.93%. Therefore, a solution to address this consistency issue is urgently needed.

[0046] This specification provides a training method for a question-answering model, which can achieve high performance in question-answering tasks. One or more embodiments of this specification relate to a question-answering processing method, a question-answering processing apparatus, a question-answering model training apparatus, a task platform, a computing device, a computer-readable storage medium, and a computer program product, which will be described in detail in the following embodiments.

[0047] The question-answering processing method provided in the embodiments of this specification can utilize a large model (i.e., a question-answering model) obtained through the training method of this question-answering model to generate answer information corresponding to the question information. Considering the large number of model parameters in the large model and the limited computing resources of mobile terminals, the question-answering processing method provided in the embodiments of this specification can be applied to, for example, Figure 3 The application scenarios shown are not limited to these. Figure 3 This is an application scenario architecture diagram of a question-and-answer processing method provided in one embodiment of this specification. Figure 3In the application scenario shown, the large model is deployed on server 10. Server 10 can connect to one or more client devices 20 via a local area network (LAN), wide area network (WAN), internet connection, or other types of data network. Client devices 20 may include, but are not limited to, smartphones, tablets, laptops, PDAs, personal computers, smart home devices, and in-vehicle devices. Client devices 20 can interact with users through a graphical user interface to invoke the large model, thereby implementing the question-and-answer processing method provided in the embodiments of this specification. If the client device has sufficient computing resources, the large model can also be directly deployed on the client device, allowing the client device to directly utilize the large model to implement the question-and-answer processing method provided in the embodiments of this specification.

[0048] Figure 4 This is a flowchart illustrating a training method for a question-answering model according to an embodiment of this specification. This method can be applied to a question-answering processing device, such as a question-answering processing device that can be... Figure 3 Server 10 in the middle. Specifically, it includes the following steps 402 and 408.

[0049] Step 402: Obtain the first training data, wherein the first training data includes: the first data to be analyzed, the first cognitive question for the first data to be analyzed, the first response information corresponding to the first cognitive question, and the first position of the first response information in the first data to be analyzed.

[0050] In the embodiments of this specification, the data to be analyzed (such as the first data to be analyzed) can be documents, videos, audio, images, or other modalities of data. The following explanation uses the example where all data to be analyzed is text data. The first cognitive question can be any cognitive question directed at the first data to be analyzed, and the first response information corresponding to the first cognitive question is the correct response information to the first cognitive question. The first response information includes fields or text in the first data to be analyzed, and the first position is the correct position of the fields or text included in the first response information within the first data to be analyzed. The first position in the first training data actually refers to the information at that first position. The first cognitive question is associated with a cognitive task, and the first position information is associated with a perceptual task. The first data to be analyzed and the first cognitive question can be different from the data to be analyzed and the cognitive questions processed by the trained question-answering model in actual application. For example, the first data to be analyzed and the first cognitive question can be data and cognitive questions processed historically by the question-answering model.

[0051] In some implementations, the first training data may include training samples and training labels. The training samples are the data to be input into the model, and the training labels are the data used to compare with the predicted data output by the model. For example, the first data to be analyzed and the first cognitive question can be used as training samples, and the first response information and the first position can be used as training labels. Alternatively, the first position can also be used as a training sample along with the first data to be analyzed and the first cognitive question.

[0052] In the embodiments of this specification, the first training data can be constructed based on the training data used during the pre-training of the question-answering model for cognitive and perceptual tasks, or it can be obtained by labeling new data. The pre-trained question-answering model can be any model currently available for performing question-answering tasks. Optionally, the first response information in the first training data can also be obtained by inputting a first cognitive question into the pre-trained question-answering model, and the first position can be obtained by identifying the position of the first response information in the first data to be analyzed. The model training device can acquire multiple pieces of first training data to train the question-answering model; this embodiment of the specification only uses any one piece of first training data as an example for illustration.

[0053] When a question-answering model is pre-trained, the training samples in the training data for a cognitive task may include: data to be analyzed and a cognitive question for that data. The training label may include the answer information for that cognitive question. Similarly, the training samples in the training data for a perceptual task may include: data to be analyzed and a perceptual question for that data. The training label may include the answer information for that perceptual question. For example, the data to be analyzed in the training samples during the pre-training process can be used as the first data to be analyzed in the first training samples; the cognitive question information can be used as the first cognitive question in the first training samples; based on the answer information for the cognitive question in the training label, position recognition (such as optical character recognition) can be performed on the first data to be analyzed to obtain the information of the first position in the first training data; the answer information for the cognitive question in the training label during the pre-training process can be determined as the first answer information in the first training data.

[0054] This example uses the construction of corresponding perceptual task data (i.e., the information of the first position) based on the training data of the cognitive task used in pre-training. In some implementations, it is also possible to construct corresponding cognitive task data (i.e., the first cognitive question) based on the training data of the perceptual task, so that the answer information of the first cognitive question is the answer information of the perceptual task. In some implementations, if the question-answering model constructs both cognitive task training data and perceptual task training data for the same information in the data to be analyzed during pre-training, then the information of the first position can directly use the position information in the training data of the perceptual task.

[0055] Suppose that the question-answering model uses image x1 as the data to be analyzed during pre-training, and the input information Xp for the perception task performed on image x1 is " <box> 166,567,305,583< / box> The input information Xc (i.e., the cognitive question Q) for the cognitive task is "What is the name of the person in the target area of ​​the picture?", and the output information Yc and Yp (i.e., the training label, the response information A to the cognitive question Q) for both tasks is "Little A". <box> 166,567,305,583< / box> The quotation marks ("") represent the location information of the text within the region (box) in the image. For example, "(166,567x)" represents the coordinates of the top-left vertex of the region, and "(305,583)" represents the coordinates of the bottom-right vertex of the region. The question-and-answer label for image x1 is (Q, A). Based on the training data used in the pre-training process, the first training data can be constructed, including Q, A, and the box. For example, the first training sample Xconn in the first training data includes Q and the box, and the first training label Yconn includes A and the box. The box in the first training label indicates that the box in the first training sample is correct.

[0056] In some implementations, the first training data may further include at least one first auxiliary position in the first data to be analyzed, which is different from the first position. This first auxiliary position in the first training data actually refers to information at that first auxiliary position. The information (such as fields or text) at the first auxiliary position in the first data to be analyzed is different from the first response information corresponding to the first cognitive question. This first auxiliary position is also an error position corresponding to the response information of the first cognitive question. The model training device can randomly determine the first auxiliary position outside the first position. Thus, by combining correct and incorrect position information to train the model, the accuracy of the model's output can be improved.

[0057] Step 404: Generate a first perception question based on the first cognitive question, wherein the first perception question indicates the position of the response information to the first cognitive question in the first data to be analyzed.

[0058] In the embodiments of this specification, the model training device can construct a first perceptual question that links cognitive and perceptual tasks based on the obtained first training data. For example, the first perceptual question instructs the question-answering model to determine the position of the response information to the first cognitive question within the first data to be analyzed. For this first perceptual question, the question-answering model first needs to determine the response information to the first cognitive question; this process belongs to the cognitive task. Then, based on the determined response information, it determines its position; this process belongs to the perceptual task. Thus, the question-answering model can link and integrate cognitive and perceptual tasks through this first perceptual question.

[0059] For example, if the initial cognitive question is "What is the name of the person in the target area of ​​the picture?", the generated initial perception question could be: "Which location in the picture contains the answer to the question 'What is the name of the person in the target area of ​​the picture?'" This initial perception question is information input to the question-answering model during training based on the initial training data. The output information of the question-answering model (i.e., the answer to the initial perception question) could include: the answer to the question is "Little A", and it is located in the area "Little A" in the picture. <box> 166,567,305,583< / box> The result is found in "". <box> 166,567,305,583< / box> ".

[0060] In some implementations, the model training device can generate a first perception question based on the first cognitive question and by incorporating a first position in the first training data. In this case, the first perception question indicates the position of the response information for the first cognitive question in the first data to be analyzed, based on the first position. For example, the first perception question instructs the question-answering model to determine the position of the response information for the first cognitive question in the first data to be analyzed, based on the first cognitive question and the provided response information position (such as the first position).

[0061] Continuing with the first cognitive problem in the above example, and taking the information at the first position as " <box> 166,567,305,583< / box> For example, the first perception question generated could be: "In the image..." <box> 166,567,305,583< / box> Does the region contain the answer to the question, "What is the name of the person in the target area of ​​the image?"? The corresponding response information will be adjusted according to the content of the first perception question. For example, for this first perception question, the output information of the question-answering model may include: the answer to the question is "Little A," and it is located in the region "" in the image. <box> 166,567,305,583< / box> The value was found in the string, so the result is "yes".

[0062] In some implementations, the model training device can generate a first perception question by combining at least one first auxiliary position from the first training data, based on the first cognitive question and the first location. In the embodiments of this specification, a method for generating the first perception question can be pre-set, and then the specific first training data to be constructed can be determined based on this method. In this implementation, the first perception question can indicate the location of the answer information for the first cognitive question among the first location and at least one first auxiliary position. When the question-answering model responds to the first perception question, it can analyze both correct and incorrect location information, making greater use of the model parameters. Based on this analysis result, the model can be trained and adjusted, improving the training effect and ensuring the final accuracy of the question-answering task.

[0063] Using the first cognitive problem and first location from the above example, and taking the information from the first auxiliary location as " <box> 115,295,214,310< / box> The first perceptual question that we get can be:

[0064] Which of the following areas contains the answer to the question, "What is the name of the person in the target area of ​​the image?"

[0065] a: <box> 166,567,305,583< / box> ;

[0066] b: <box> 115,295,214,310< / box> .

[0067] In this example, the question-answering model might output: the answer to the question is "Little A", and it is located in the region "" in the graph. <box> 166,567,305,583< / box> The value was found in the string, so the result is "a".

[0068] In some implementations, the various information required to generate the first perception question can be arranged in a prescribed format to obtain the first perception question. This helps the question-answering model to locate the various information to be analyzed in the first perception question more quickly and accurately, facilitating the understanding of the question-answering model and contributing to more accurate output results. Taking the information required to generate the first perception question as including the first cognitive question, the information of the first position, and the information of the first auxiliary position as an example, the above-mentioned generation of the first perception question based on the first cognitive question, the first position, and at least one first auxiliary position includes: filling the information of the first cognitive question, the information of the first position, and the information of at least one first auxiliary position into the corresponding positions in the query information template to obtain the first perception question. The query information template includes: the stem information indicating the perception question, the filling position information of the first cognitive question, the filling position information of the first position information, and the filling position information of the first auxiliary position information. For first perception questions generated based on other information, the corresponding query information template can also be used for generation.

[0069] The query information template can include the positions of the information required to generate the first perception question, as well as fixed core question information (i.e., the aforementioned question stem information). For example, the query information template can be pre-set by staff and stored in the model training device. After obtaining the information required to generate the first perception question, the model training device can directly arrange the first perception question according to the query information template. This information may include only the first perception question, or it may include the first perception question and the first position, or it may include the first perception question, the first position, and at least one first auxiliary position. The core question information in the corresponding query information template can also differ depending on the information used to generate the first perception question.

[0070] Continuing with the first cognitive question from the previous example, if the first perception question is generated based on the first cognitive question, the core information of the query information template could be "Which area in the image contains the answer to the question 'XXX'?", where XXX represents blank information, corresponding to the filling position of the first cognitive question. Accordingly, the resulting first perception question could be: "Which area in the image contains the answer to the question 'What is the name of the person in the target area of ​​the image?'"

[0071] Continuing with the first cognitive question and first location from the previous example, if the first perception question is generated based on the first cognitive question and first location, the core information of the query information template can be "Does the area 'XX' in the figure contain the answer to the question 'XXX'?", where XX and XXX both represent blank information. XX corresponds to the filling position of the information in the first location, and XXX corresponds to the filling position of the first cognitive question. Accordingly, the resulting first perception question can be: "In the figure…" <box> 166,567,305,583< / box> Does the area contain the answer to the question, "What is the name of the person in the target area of ​​the image?"

[0072] Continuing with the first cognitive question, first location, and first auxiliary location from the previous example, if the first perception question is generated based on the first cognitive question, first location, and first auxiliary location, then the first location information and each of the first auxiliary location information can be constructed as corresponding options, thereby obtaining the corresponding first perception question. For example, the first perception question x... Conn =TempQ Conn (Q,Box,Box ′ TempQ Conn This is a query template for a multiple-choice question. The core information of the query template can be: Which of the following areas contains the answer to the question "XXX"? a: "xx"; b: "xxx". Here, "xx" and "xxx" represent blank information. "xx" corresponds to the fill position for the first position, and "xxx" corresponds to the fill position for the first auxiliary position. Accordingly, the obtained question information can be: "Which of the following areas contains the answer to the question 'What is the name of the person in the target area of ​​the image?'?" a: <box> 166,567,305,583< / box> b: <box> 115,295,214,310< / box> "In the embodiments of this specification, the information of the first position and the information of the first auxiliary position can also be randomly matched with each option. Here, we take setting only one first auxiliary position information as an example. When setting multiple first auxiliary position information, the number of options set in the main problem information can be increased accordingly."

[0073] Step 406: Input the first data to be analyzed and the first perception question into the pre-trained question answering model, predict the first reference answer information corresponding to the first cognitive question, and predict the second reference answer information of the first perception question based on the first reference answer information.

[0074] The pre-trained question-answering model can be any model currently available for performing question-answering tasks. In the embodiments of this specification, the pre-trained question-answering model is further trained based on the first data to be analyzed and the first perceptual question. During this further training process, a training task on cognitive and perceptual connections can be performed to ensure that the question-answering model can link cognitive and perceptual tasks, ensuring consistent alignment between the two tasks.

[0075] In the embodiments of this specification, the first and second reference response information refer to the prediction results output by the question-answering model. The second reference response information indicates the position of the first reference response information in the first data to be analyzed. The model training device can input the first data to be analyzed and a first perceptual question into a pre-trained question-answering model. The first perceptual question can link cognitive and perceptual tasks. Based on the input information, the question-answering model can perform a cognitive-perceptual connection task to analyze the first data to be analyzed and obtain the response information for the first perceptual question. The question-answering model can perform a cognitive task to predict the first reference response information corresponding to the first cognitive question, and then perform a perceptual task based on the first reference response information to predict its position in the first data to be analyzed, thereby obtaining the second reference response information for the first perceptual question. The output of the question-answering model may include the first and second reference response information.

[0076] For different first-perception questions, the content of the prediction results output by the question-answering model (such as the second reference response information) can vary. For example, refer to the aforementioned examples of the question-answering model's output information for different first-perception questions. If the first-perception question is generated solely based on the first cognitive question, the second reference response information output by the question-answering model can be location information. If the first-perception question is generated based on the first cognitive question and the first location, the second reference response information output by the question-answering model can be information indicating whether the first location is correct and whether it is the location of the first reference response information, such as expressed in the form of "yes" or "no". If the first-perception question is generated based on the first cognitive question, the first location, and the first auxiliary location, the second reference response information output by the question-answering model can be the option information corresponding to the location.

[0077] For example, the first perception question is: "In the picture" <box> 166,567,305,583< / box>If the question asks whether the region contains the answer to the question "What is the name of the person in the target area of ​​the image?", then the second reference answer in the predicted output of the question-answering model can be "Yes". The question-answering model first identifies the first cognitive question contained within the first perceptual question, predicts and outputs the corresponding first reference answer for that first cognitive question, and then predicts and outputs the second reference answer based on that first reference answer. For example, the output of the question-answering model could include: the answer to the question is "Little A", and it is located in the region "..." in the image. <box> 166,567,305,583< / box> The answer is "Yes". "Little A" is the first reference response, and "Yes" is the second reference response. In some implementations, the output may also directly include: "Little A" as the first reference response and "Yes" as the second reference response. Here, the output format of the question-answering model for the prediction result is not limited.

[0078] When the first perception question is generated based on at least one first auxiliary position, the question-answering model performs a multiple-choice task by constructing both the first position and the first auxiliary position as alternatives, and then selecting from these options. The selected result represents the second reference response information. If the content of the selected option is the same as the first position, it indicates that the question-answering model has correctly determined the given first position; if the content of the selected option is the same as the first position, it indicates that the question-answering model has incorrectly determined the given first position. For example, for the question "Which of the following areas contains the answer to the question 'What is the name of the person in the target area of ​​the image?'"? a: <box> 166,567,305,583< / box> b: <box> 115,295,214,310< / box> The question-answering model can output the following result: the answer to the question is "Little A", which is located in the region "" in the graph. <box> 166,567,305,583< / box> The option "a" was found in the first reference response, and "a" is the second reference response. In some implementations, the output can also directly include: the first reference response is "A" and the second reference response is option "a". Here, the output format of the prediction result of the question-answering model is not limited.

[0079] In one embodiment of this specification, the first and second reference response information predicted by the question-and-answer model can also be arranged according to a set format before being output, so as to facilitate further processing by the model training device. For example, if the format is indicated by a response information template, step 406 includes: inputting the first data to be analyzed and the first perceptual question into the pre-trained question-and-answer model to obtain the first reference response information corresponding to the first cognitive question and the second reference response information for the first perceptual question, arranged according to the response information template. The second reference response information is predicted based on the first reference response information. The response information template includes: relationship information between the first cognitive question and the first reference response information, filling position information of the first reference response information, filling position information of the second reference response information, and relationship information between the first and second reference response information. Optionally, the response information template may also include relationship information between the second reference response information and the first perceptual question.

[0080] This output information clearly indicates the relationship between each question and the predicted reference responses, reflecting the task execution logic within the question-answering model. This ensures that the model outputs information to the user according to this logic during practical applications. Users can use this logic to determine whether the model's output meets their needs, improving model effectiveness and user experience. Furthermore, the filling position information for each reference response allows the model training device to quickly and accurately locate the first and second reference responses, determining their correctness and enabling rapid model adjustments, thus improving training efficiency.

[0081] The response information template may include the appropriate positions for the first and second reference responses, as well as fixed core response information. This core response information may include: the relationship between the first cognitive question and the first reference response, and the relationship between the first and second reference responses. For example, the response information template can be pre-set by staff and stored in the model training device. The model training device can perform cognitive and perceptual connection tasks, and based on the response information template, fill the predicted first and second reference responses into their respective positions to obtain the predicted output.

[0082] For example, the question-answering model predicts the output result y. Conn =TempR Conn (A, Box), where TempR ConnThis represents the response information template. Following the example above, the core information in the response information template could include: the answer to the question is "XXX", which is found in area "xxx" in the diagram, therefore the result is "xxxx". Here, "XXX" corresponds to the filling position of the first reference response information, "xxx" corresponds to the first reference position predicted based on the first reference response information, "xxxx" corresponds to the filling position of the second reference response information, "the answer to the question is" represents the relationship between the first reference response information and the first cognitive question, "it is found in area "xxx" in the diagram" represents the relationship between the first reference response information and the second reference response information, and "therefore the result is" represents the relationship between the second reference response information and the first perceptual question.

[0083] Step 408: Based on the first response information, the first position, the first reference response information, and the second reference response information, adjust the pre-trained question-answering model to obtain the trained question-answering model.

[0084] After the pre-trained question-answering model outputs a result for each first perceived question, the model training device can compare it with the corresponding training label in the first training data and adjust the parameters of the pre-trained question-answering model based on the comparison result. Then, it returns to the step of obtaining the first training data (i.e., step 402), and executes this prediction, comparison, and parameter adjustment process multiple times until the training stopping condition is met, resulting in a fully trained question-answering model. For example, the first reference answer information can be compared with the first response information, and the second reference answer information can be compared with the first position.

[0085] Specifically, based on the comparison results, a loss value can be calculated. Using this loss value, backpropagation can be performed to adjust the model parameters of the pre-trained question-answering model. Then, the process returns to the step of obtaining training data, and the first training data is selected to continue training the pre-trained question-answering model until the training stopping condition is met. Training is then complete, and a usable question-answering model is obtained. This training stopping condition can be that the loss value is less than a loss value threshold, or that the number of iterations reaches a threshold, or a combination of both the loss value threshold and the number of iterations threshold.

[0086] The loss threshold is a critical value for the loss. If the loss value is greater than or equal to the threshold, it indicates that there is still a certain deviation between the prediction results of the pre-trained question-answering model and the actual results, and the model parameters of the pre-trained question-answering model still need to be adjusted. At this point, the training stopping condition has not been met. If the loss value is less than the threshold, it indicates that the prediction results of the pre-trained question-answering model are close enough to the actual results, and training can be stopped. At this point, the training stopping condition can be met. The specific value of the loss threshold is selected based on the actual situation.

[0087] In addition, the number of iterations threshold is set according to the actual situation. When the number of iterations reaches the threshold, it means that the pre-trained question answering model has been trained enough times. At this time, the prediction result of the pre-trained question answering model is close to the actual result, and training can be stopped.

[0088] By applying the training method of the question-answering model provided in the embodiments of this specification, a first perceptual question is constructed based on a first cognitive question. This first perceptual question connects the cognitive question and the perceptual question for the first data to be analyzed. Training the question-answering model using this first perceptual question enables the model to link the cognitive task and the perceptual task, ensuring the consistency and alignment of the finally trained question-answering model in the cognitive and perceptual tasks, and improving the performance of the question-answering task.

[0089] The foregoing content allows for further training of the pre-trained question-answering model on cognitive and perceptual connection tasks. This ensures that the model's output remains consistent across cognitive and perceptual tasks when performing question-answering tasks on the same data. In some cases, for individual cognitive and perceptual tasks, the results obtained by the question-answering model during forward execution may differ from those obtained during backward verification. For example, in a cognitive task, if the question-answering model is asked "Who is sitting on the sofa in this image?", the model outputs "Little A". However, during backward verification, if the question-answering model is asked "Where is Little A sitting in this image?", the model outputs "On the bench".

[0090] In the embodiments of this specification, the pre-trained question-answering model can also be trained for single-task self-consistency, ensuring consistency in the forward execution and backward verification of the trained question-answering model for any task in the cognitive or perceptual tasks, thereby improving the performance of the question-answering task. Furthermore, after completing self-consistency training for the cognitive and perceptual tasks, the consistency between the cognitive and perceptual tasks can also be improved accordingly. The self-consistency training for the cognitive and perceptual tasks is described below. In the embodiments of this specification, performing self-consistency training for the question-answering model on the cognitive task is referred to as enabling the question-answering model to perform a cognitive consistency task; performing self-consistency training for the question-answering model on the perceptual task is referred to as enabling the question-answering model to perform a perceptual consistency task.

[0091] In the embodiments of this specification, for the consistency training of the question-answering model for cognitive tasks, the model training device can perform the following steps s12 to s18.

[0092] Step s12: Obtain the second training data, wherein the second training data includes: the second data to be analyzed, the second cognitive question for the second data to be analyzed, and the second response information corresponding to the second cognitive question.

[0093] Step s12 can refer to the relevant description of step 402 above. The second training data can also be constructed based on the training data used during the pre-training of the question-answering model. Regarding the second data to be analyzed, refer to the relevant description of the first data to be analyzed above. Regarding the second cognitive question, refer to the relevant description of the first cognitive question above. Regarding the second response information, refer to the relevant description of the first response information above. The second response information is the correct answer to the second cognitive question. In this embodiment, the second cognitive question is the same as the first cognitive question. In practical applications, the second cognitive question may be different from the first cognitive question.

[0094] In the embodiments described in this specification, the second response information can be a training label for a pre-annotated second cognitive question. Optionally, the model training device can also input the second cognitive question into a pre-trained question-and-answer model to obtain the second response information for the second cognitive question, and then form second training data based on the second response information.

[0095] In some implementations, the second training data further includes at least one auxiliary response information in the second data to be analyzed that is different from the second response information. This auxiliary response information is the incorrect response information for the second cognitive question, and the auxiliary response information can be compared with the aforementioned first auxiliary position.

[0096] Step s14: Generate a third cognitive question based on the second cognitive question and the second response information, wherein the third cognitive question indicates the response information for the second cognitive question to be determined based on the second response information.

[0097] Step s14 can refer to the aforementioned description of step 404. For example, the third cognitive question can instruct the question-answering model to determine whether the second response is the correct answer to the second cognitive question. Continuing with the above example, the third cognitive question could be: Is "Little A" the answer to the question "What is the name of the person in the target area of ​​the picture?"

[0098] In some implementations, the model training device can generate a third cognitive question by combining auxiliary response information from the second training data, based on the second cognitive question and the second response information. In the embodiments of this specification, a method for generating the third cognitive question can be pre-set, and then the specific second training data to be constructed can be determined based on this method. In this implementation, the third cognitive question can indicate the response information for the third cognitive question from the second response information and at least one auxiliary response information. The question-answering model responds to the third cognitive question, allowing it to analyze both correct and incorrect responses together. This enables greater utilization of the model's parameters, and training and adjusting the model based on the analysis results can better improve the model's training effect and ensure the final accuracy of the question-answering task.

[0099] Using the example above, the third cognitive question could be: "Which name is the answer to the question 'What is the name of the person in the target area of ​​the picture?'? a: Little A; b: Little B."

[0100] In some implementations, for cognitive consistency tasks, the various information required to generate the third cognitive question can be arranged in a prescribed format to obtain the third cognitive question. This helps the question-answering model to locate the various information to be analyzed in the third cognitive question more quickly and accurately, facilitating the understanding of the question-answering model and contributing to more accurate output results. Taking the generation of the third cognitive question as requiring a second cognitive question, second response information, and at least one auxiliary response information as an example, the third cognitive question can be generated based on a query information template for the third cognitive question. For example, this template can also be a query template for multiple-choice questions. Regarding this template, please refer to the relevant introduction to the query information template in step 404 above. The query information template includes fixed format information for the third cognitive question, filling position information for the second cognitive question, filling position information for the second response information, and filling position information for the auxiliary response information.

[0101] For example, taking the setting of auxiliary response information as an example, the input of the third cognitive question x of the question-answering model is... CV =Temp CV (Q,A,A ′ ), where Temp CV This is a query template for a two-choice question, where Q represents the second cognitive question, and A represents the second response information (i.e., the correct response information) to the second cognitive question. ′This represents supplementary response information. In one example, the third cognitive question input into the question-answering model could be, "Which name is the person in the target area of ​​the image? a: Little A; b: Little B." The fixed-format information can include "XX," where a: "xx"; b: "xxx." Here, "XX" corresponds to the position to be filled in for the second cognitive question, "xx" corresponds to the position to be filled in for the second response information, and "xxx" corresponds to the position to be filled in for the supplementary response information.

[0102] Step s16: Input the second data to be analyzed and the third cognitive question into the pre-trained question-answering model to predict the third reference answer information for the second cognitive question.

[0103] Step s16 can be referred to the relevant description of step 406 above, and will not be repeated here. The method for predicting the third reference response information can be analogous to the relevant description of predicting the second reference response information above, and will not be repeated here.

[0104] Step s18: Based on the relationship between the third reference response information and the second response information, adjust the pre-trained question-answering model to obtain the trained question-answering model.

[0105] Step s18 can be referred to the relevant description of step 408 above, and will not be repeated here. The model training device can compare the third reference response information with the second response information to determine whether the essence of the third reference response information and the second response information is consistent, so as to adjust the pre-trained question-answering model if they are inconsistent. In this way, the positive response results (such as the second response information) and the reverse verification results (such as the third reference response information) of the trained question-answering model for each cognitive question can be matched.

[0106] In the embodiments of this specification, a third cognitive question is used to further train the pre-trained question-answering model. This third cognitive question is constructed based on the second cognitive question and the second response information. This allows the question-answering model to process the second cognitive question in conjunction with the response information, enabling the model to make reverse predictions for the original cognitive question. Based on this, the model can be adjusted, allowing for consistent training of the question-answering model for cognitive tasks. This ensures that the forward execution and reverse verification results of the trained question-answering model for a cognitive task match, improving the performance of the cognitive task.

[0107] In the embodiments of this specification, for the consistency training of the question-answering model for the perception task, the model training device can perform the following steps s22 to s28.

[0108] Step s22: Obtain the third training data, where the third training data includes: the third data to be analyzed, the third response information corresponding to the second perception problem for the third data to be analyzed, and the second position of the third response information in the third data to be analyzed.

[0109] Step s22 can refer to the relevant introductions of the foregoing step 402 and step s12. The third training data can also be constructed based on the training data used by the question-and-answer model during pre-training. Regarding the third data to be analyzed, it can refer to the relevant introductions for the first data to be analyzed above. Regarding the second perception problem, it can refer to the relevant introductions for the first cognitive problem and the second cognitive problem above. Regarding the third response information, it can refer to the relevant introductions for the first response information and the second response information above. Regarding the second position, it can refer to the relevant introduction for the first position above.

[0110] The second perception problem can be any perception problem for the third data to be analyzed. The third response information is the correct response information for the second perception problem, and the second position is the correct position of the third response information in the third data to be analyzed. This second perception problem can indicate determining the content at a specified position in the third data to be analyzed, and the second position can be included in the content of the second perception problem. For example, the second perception problem can be "What is the content in the <box> 166,567,305,583< / box> " area in the figure?" The third response information for this second perception problem is "Xiaoming", and the information for the second position is " <box> 166,567,305,583< / box> ".

[0111] In some embodiments, the third training data further includes: at least one second auxiliary position in the third data to be analyzed that is different from the second position. Regarding this second auxiliary position, it can be analogous to the foregoing first auxiliary position.

[0112] Step s24: Generate a third perception problem based on the third response information, where the third perception problem indicates determining the position of the third response information in the third data to be analyzed.

[0113] Step s26 can refer to the relevant introductions of the foregoing step 406 and step s14. In the embodiments of this specification, the third perception problem can be the reverse problem of the second perception problem. Exemplarily, the third perception problem can indicate determining the position of the third response information in the third data to be analyzed. Continuing with the above example of the second perception problem, the third perception problem can be "Where is the text 'Xiaoming' in the figure?"

[0114] In some implementations, the model training device can generate a third perceptual question based on the third response information and the second location. In this case, the third perceptual question instructs the model to determine the position of the third response information within the third set of data to be analyzed, based on the second location. The third perceptual question can instruct the question-answering model to determine whether the second location is the correct position of the third response information within the third set of data to be analyzed. For example, the third perceptual question could be "region". <box> 166,567,305,583< / box> Is the text "Little A" in the correct position in the image?

[0115] In some implementations, the model training device can generate a third perceptual question by combining at least one second auxiliary position from the third training data, based on the third response information and the second position. In this case, the third perceptual question indicates that the response information for the third perceptual question is determined from the second position and at least one second auxiliary position. The question-answering model responds to this third perceptual question, allowing it to analyze both correct and incorrect positions simultaneously. This enables greater utilization of the model's parameters, and training and adjusting the model based on the analysis results can better improve the model's training performance and ensure the final accuracy of the question-answering task.

[0116] Using the example above, the third perception question could be: "Which area is the correct location of the text 'Little A' in the image?" <box> 166,567,305,583< / box> b: <box> 115,295,214,310< / box> .

[0117] In some implementations, for the perceptual consistency task, the various information required to generate the third perceptual question can be arranged in a prescribed format to obtain the third perceptual question. This helps the question-answering model to locate the various information to be analyzed in the third perceptual question more quickly and accurately, facilitates the understanding of the question-answering model, and helps to output more accurate results. Taking the generation of the third perceptual question as requiring third response information, a second position, and at least one second auxiliary position as an example, the third perceptual question can be generated based on the query information template of the third perceptual question. For example, the template can also be a query template for a multiple-choice question. Regarding the template, please refer to the relevant introduction of the query information template in steps 404 and s14 above. The query information template includes the fixed format information of the third perceptual question, the filling position information of the third response information, the filling position information of the second position, and the filling position information of the second auxiliary position.

[0118] For example, taking setting a second auxiliary position as an example, the third perception question x of the question answering model is input. PV =Temp PV (A, Box, Box) ′ ), where Temp CVThis is a query template for a two-choice question. A represents the third response, Box represents the second position corresponding to the third response (i.e., the correct position), and Box′ represents the second auxiliary position. Box′ can be the position of a randomly sampled bounding box after excluding Box from the image. In one example, the third cognitive question input to the question-answering model could be "Which region contains the text 'Little A'?" <box> 166,567,305,583< / box> b: <box> 115,295,214,310< / box> The fixed format information may include "which area contains the text "XX", a: xx; b: xxx". Here, "XX" corresponds to the third response information, "xx" corresponds to the filling position of the information in the second position, and "xxx" corresponds to the filling position of the information in the second auxiliary position.

[0119] Step s26: Input the third data to be analyzed and the third perception question into the pre-trained question answering model to predict the fourth reference response information for the third perception question.

[0120] Step s26 can be referred to the aforementioned descriptions of steps 406 and s16, and will not be repeated here. The fourth reference response information indicates the position of the third response information within the third set of data to be analyzed. The method for predicting this fourth reference response information is analogous to the aforementioned descriptions of predicting the second and third reference response information, and will not be repeated here.

[0121] Step s28: Based on the relationship between the fourth reference response information and the second position, adjust the pre-trained question-answering model to obtain the trained question-answering model.

[0122] Step s28 can be referred to the relevant descriptions of steps 408 and s18 above, and will not be repeated here. The model training device can compare the position indicated by the fourth reference response information with the second position to determine whether the two positions are essentially consistent, so as to adjust the pre-trained question-answering model if they are inconsistent. In this way, the trained question-answering model can match the positive response results (such as the third response information) and the reverse verification results (such as the position indicated by the fourth reference response information) for each perceived question.

[0123] In the embodiments of this specification, a third perceptual question is used to further train the pre-trained question-answering model. This third perceptual question is constructed in reverse based on the third response information of the second perceptual question. This allows the question-answering model to make reverse predictions of the second perceptual question by combining the response information of the second perceptual question. Based on this, the model is adjusted, enabling consistent training of the question-answering model for perceptual tasks. This ensures that the forward execution and reverse verification results of the trained question-answering model for a perceptual task match, improving the performance of the perceptual task.

[0124] In some implementations, the model training device may first perform self-consistency training on the cognitive task for a pre-trained question-answering model, then perform self-consistency training on the perceptual task for the question-answering model obtained after this training, and finally perform the aforementioned consistency training on the cognitive and perceptual tasks for the question-answering model obtained after this training. The order of these three different training processes can also be arbitrarily adjusted and is not limited here.

[0125] In the embodiments of this specification, the training method provided in the embodiments of this specification is used to train multiple open source MLLMs of different series and parameter sizes. After conducting corresponding experiments, it is determined that the training in the embodiments of this specification significantly improves the consistency between cognition and perception on all models, reduces the conflict between cognitive and perceptual knowledge, and also improves the performance and consistency of cognitive and perceptual tasks.

[0126] After the question-answering model is trained through the above training process, it can be used to perform actual question-answering tasks. The execution device for this question-answering task can be the same as or different from the aforementioned model training device; no limitation is made here. Figure 5 This is a flowchart illustrating a question-and-answer processing method according to an embodiment of this specification. This method can be applied to a question-and-answer processing device, such as a device that can be... Figure 3 Server 10 in the middle. For example... Figure 5 As shown, the question-and-answer processing method includes the following steps 502 and 504.

[0127] Step 502: Obtain the task data for the question-and-answer task, which includes the data to be analyzed and the question information raised for the data to be analyzed.

[0128] In the embodiments of this specification, the data to be analyzed can be documents, videos, audio, images, or other modalities. The following explanation uses the example where all data to be analyzed is text data. The questions raised regarding the data to be analyzed can be questions corresponding to perceptual tasks or cognitive tasks. The data to be analyzed here is actual user-facing data; the first, second, and third data to be analyzed in the aforementioned training process can be data constructed by staff or the computer itself.

[0129] In one implementation, both the data to be analyzed and the questions posed regarding the data can be sent by the user using a client device. For example, the client device can connect to the question-and-answer processing device using an installed application with question-and-answer functionality, or a website or mini-program with question-and-answer functionality, and then send task data for the question-and-answer task to the device, triggering the device to execute the task. In another implementation, the client device can send question information and a description of the data to be analyzed to the question-and-answer processing device. The question-and-answer processing device then retrieves the corresponding data to be analyzed from the internet or storage space based on this description information.

[0130] Step 504: Call the question-answering model to analyze the task data and obtain the answer information corresponding to the question information; wherein, the question-answering model is a question-answering model trained by the aforementioned question-answering model training method.

[0131] Please refer to the relevant introduction in the aforementioned training method for information about this question-answering model. The question-answering processing device can, based on the acquired question-answering task data, call the trained question-answering model obtained after the aforementioned training process, and analyze the data to be analyzed within the task data to find the answer information corresponding to the question information in the task data.

[0132] The embodiments in this specification utilize the aforementioned question-answering model that enables consistency in the execution of cognitive and perceptual tasks to perform question-answering tasks, thereby improving the accuracy of the response information output for the question-answering task and enhancing the execution effect of the question-answering task.

[0133] Figure 6 This is a flowchart of another question-and-answer processing method provided in one embodiment of this specification. This method can be applied to a question-and-answer processing device, such as a question-and-answer processing device that can be... Figure 3 Server 10 in the middle. Figure 6 The interaction scenario between the question-and-answer processing device and the front-end device is described below. The front-end device can be the aforementioned client device. Figure 6 The provided method can be compared with the previous one. Figure 5 The methods provided are mutually referential. Specifically, this method includes steps 602 and 608.

[0134] Step 602: Receive the Q&A request uploaded by the front-end device, wherein the Q&A request includes the data to be analyzed and the question information raised for the data to be analyzed.

[0135] Step 602 can refer to the relevant introduction of the task data for sending the question-and-answer task by the client device in step 502 above. This task data can be carried in the question-and-answer request and sent to the question-and-answer processing device.

[0136] Step 604: In response to the question-and-answer request, call the question-and-answer model to analyze the task data and obtain the answer information corresponding to the question information; wherein, the question-and-answer model is a question-and-answer model trained by the aforementioned question-and-answer model training method.

[0137] Step 604 can be referred to the aforementioned description of step 504, and will not be repeated here.

[0138] Step 606: Send the response information corresponding to the problem information to the front-end device.

[0139] Step 606 can be referred to the aforementioned description of step 406, and will not be repeated here.

[0140] For example, after generating the response information corresponding to the question information, the question-and-answer processing device can send the response information to the front-end device. The front-end device can have a response information display area for displaying the response information corresponding to the received question information.

[0141] Step 608: Send recommended task information related to the data to be analyzed or the problem information to the front-end device.

[0142] In addition to analyzing the data to be analyzed based on the question information, the question-and-answer processing device can also predict potential task information (i.e., recommended task information) that the user might trigger in response to the data to be analyzed, based on historical question-and-answer patterns, and then send the recommended task information to the front-end device. Upon receiving the recommended task information, the front-end device can display trigger controls for the recommended task information for the user to select.

[0143] For example, continuing with the first cognitive question in the aforementioned model training process, "What is the name of the person in the target area of ​​the image?", the recommended task information sent by the question-answering processing device could be "Analyze the current emotion of the person", "Determine the name of the person next to the person", or "Determine the location of other people in the image", etc.

[0144] Step 610: Upon receiving a trigger instruction from the front-end device for the recommendation task information, call the question-answering model to generate the execution result of the recommendation task information and send the execution result to the front-end device.

[0145] Users can determine whether they are interested in the recommended task information displayed on the front-end device. If interested, they can directly trigger the corresponding recommended task information's trigger control, causing the front-end device to send a trigger command for that recommended task information to the question-and-answer processing device. Accordingly, the question-and-answer processing device can, based on the trigger command, invoke the question-and-answer model to generate the execution result of the recommended task information and send the execution result to the front-end device. If the user is not interested in the recommended task information, they can ignore it or re-enter their new question information.

[0146] Steps 608 and 610 are only one optional implementation. In some cases, steps 608 and 610 may not be performed, which is not limited here.

[0147] Step 612: Upon receiving feedback information from the front-end device regarding the response information, adjust the question-and-answer model based on the feedback information.

[0148] In some implementations, after step 606, the user can determine whether the response to the received question meets their needs and resolves their problem. If the user believes the response does not meet their needs, they can send feedback information to the question-and-answer processing device, which may include a detailed description of the user's opinion. Upon receiving feedback information from the front-end device, the question-and-answer processing device adjusts the question-and-answer model to improve its capabilities, enhance the accuracy and quality of subsequent responses sent to the front-end device, and improve the user's question-and-answer experience. Based on user instructions, the question-and-answer processing device can re-execute the question-and-answer processing method using the adjusted model to regenerate the response and provide it back to the user.

[0149] Step 612 is only one optional implementation. In some cases, step 612 may not be performed, which is not limited here.

[0150] Figure 7 This is a flowchart of another question-and-answer processing method provided in an embodiment of this specification, applied to a task platform, specifically including the following steps 702 and 704.

[0151] Step 702: Receive a model request sent by the terminal device, wherein the model request includes at least one of the following: scene identifier of the target scene, scene input data of the target scene, and target model specification parameters.

[0152] It should be noted that the task platform is an online platform that provides model services, allowing developers, enterprises, or individual users to submit tasks, obtain resources, or implement specific functions through API calls, web interfaces, or other interactive methods. In the embodiments of this specification, the task platform is a cloud service platform capable of providing deep learning models, allowing external applications (such as e-commerce platforms, navigation applications, social applications, film and television production applications, and online education platforms) to access and use it. The task platform ensures efficient and accurate responses to requests from terminal devices by managing the training, updating, optimization, load balancing, scheduling, and inference of the models. The terminal device is the terminal device requesting the acquisition of the target question-answering model.

[0153] In practice, a model request is a request sent by the terminal device to the task platform to obtain or use a specific model service. This model request may include at least one of the following: a scene identifier of the target scene, scene input data of the target scene, and target model specification parameters. In the embodiments described in this specification, this model request may be a request for a question-answering model.

[0154] The target scenario refers to the specific environment or context in which the target question-answering model will be applied, describing the purpose and expected results of the model service. The scenario identifier is the identification information used to identify the target scenario, assisting in quickly locating the model related to the specific scenario, such as "document question answering," "image question answering," "video question answering," "video navigation," "video surveillance," "video quality evaluation," "object localization," "product localization and search," etc. The scenario input data can be the input data used to train the pre-trained question-answering model in the target scenario, adapted to the target scenario. For example, in a video question-answering scenario, the scenario input data can be sample videos with response information. Additionally, the model specification parameters are attribute specifications defining the model architecture, performance, or configuration, such as model model, model type, model size, etc. These model specification parameters are used to determine the model that best suits the user's needs.

[0155] For example, on the task platform client of a terminal device, a user selects the target scene identifier as "Document Q&A," selects the required model size as 32GB, uploads the scene input data of the target scene, and the terminal device generates a model request and sends it to the task platform. The task platform receives the model request sent by the terminal device, which includes the scene identifier of the target scene, the scene input data of the target scene, and the target model specification parameters.

[0156] In the embodiments described in this specification, users can independently select at least one of the following information on the terminal device: the scene identifier of the target scene, the scene input data of the target scene, and the target model specification parameters, to trigger a model request. The task platform can receive the model request for unit testing sent by the terminal device and provide reference information for selecting the target question-answering model based on at least one of the following: the scene identifier of the target scene, the scene input data of the target scene, and the target model specification parameters, thus making it applicable to a wider range of scenarios.

[0157] Step 704: Based on the model request, determine the corresponding question-answering model from at least one model, wherein the question-answering model is trained based on the above-described question-answering model training method, and the question-answering model is used to generate the answer information corresponding to the question information.

[0158] In the embodiments of this specification, a model library may be provided. This module library may store at least one question-answering model adapted to different scenarios, as well as multiple question-answering models with different model specifications and parameters. Based on a model request, the corresponding question-answering model is determined from at least one model. One optional method is to search for the corresponding question-answering model from at least one model included in the model library based on the model request. The model library may store multiple question-answering models with different model specifications and parameters. Another optional method is to train and obtain the question-answering model based on the model request. Yet another optional method is to construct the question-answering model based on the model request, which is not limited here.

[0159] In one implementation of step 704, determining the corresponding question-answering model from at least one model based on the model request may include:

[0160] Based on at least one piece of information in the model request, search for the question-answering model corresponding to that at least one piece of information in the model library; wherein, the scene identifier of the target scene and the scene input data of the target scene both correspond to the model adapted to the target scene, and the target model specification parameters correspond to the model with the same model specification parameters as the target model specification parameters.

[0161] If at least one piece of information includes scene input data of the target scene, the found question-answering model is trained based on the scene input data of the target scene to obtain a trained question-answering model.

[0162] Regarding the implementation of step 704 above, the task platform can determine the required question-answering model by searching for models in the model library, based on both the scene identifier and the target model specifications. For the scene input data of the target scene, the platform determines the required question-answering model by searching for models in the model library and then further training the found models.

[0163] If the model request includes only one piece of information, and if the model request includes a scene identifier for the target scene, then step 704 may include: searching for a question-answering model adapted to the target scene from the model library based on the scene identifier, wherein the model library stores at least one question-answering model adapted to different scenes. If the model request includes scene input data for the target scene, then step 704 may include: determining an initial question-answering model adapted to the target scene from at least one model; training the initial question-answering model based on the scene input data for the target scene to obtain a trained question-answering model. If the model request includes model specification parameters, then step 704 may include: searching for a corresponding question-answering model from the model library based on the model specification parameters, wherein the model library stores multiple question-answering models with different model specification parameters.

[0164] When a model request includes two or three of the following: a scene identifier for the target scene, scene input data for the target scene, and target model specifications, the task platform can search the model library for models that simultaneously satisfy two or three of these requirements to determine the required model. If the model request includes both a scene identifier and target model specifications, the task platform can search the model library for a model that fits the target scene's specifications. If the model request includes both a scene identifier and scene input data for the target scene, the task platform can first search the model library for a model that fits the target scene, and then train that model based on the scene input data to obtain the required model. If the model request includes both target model specifications and scene input data for the target scene, the task platform can first search the model library for a model that matches the target model specifications, and then train that model based on the scene input data to obtain the required model.

[0165] When a model request includes the scene identifier of the target scene, the scene input data of the target scene, and the specification parameters of the target model, the task platform can first search for at least one pre-trained question-answering model in the model library based on the scene identifier of the target scene. Then, based on the specification parameters of the target model, it can select a question-answering model of the corresponding size from the at least one question-answering model. Finally, based on the scene input data of the target scene, it can train the question-answering model of the corresponding size to obtain a question-answering model suitable for the user's needs.

[0166] The question-answering model obtained by the task platform is based on Figure 4 The question-answering model described above was trained using the training method described above. For details on the training method, please refer to the content on model training mentioned above, which will not be repeated here.

[0167] In the embodiments of this specification, based on scenario requirements, the question-answering model suitable for the scenario is accurately found through scenario identification, and then these models are used to generate the corresponding answer information for the question information. This can ensure that the answer information corresponding to the question information is more accurate and fits the scenario, thereby improving the user experience and the quality of question-answering.

[0168] Optionally, the response information corresponding to the questions required for unit testing of code in different scenarios may vary. As an example, the task platform can provide question-and-answer models for various scenarios, and can provide corresponding question-and-answer models and code modification models based on the model requests for unit testing sent by the terminal device, so as to generate accurate and effective response information corresponding to the question information in the logical units of the program in the corresponding domain.

[0169] In this embodiment of the specification, after step 704, the task platform can also deploy a question-and-answer model and build a question-and-answer interface based on the question-and-answer model, so that the terminal device can schedule the question-and-answer model through the question-and-answer interface to generate the answer information corresponding to the question information.

[0170] Corresponding to the above method embodiments, this specification also provides task platform embodiments. Figure 8 This is a schematic diagram of the structure of a task platform provided in one embodiment of this specification. Figure 8 As shown, the task platform 800 includes: a request interface 802 and a response unit 804;

[0171] Request interface 802 is used to receive a model request sent by a terminal device, wherein the model request includes at least one of the following: scene identifier of the target scene, scene input data of the target scene, and model specification parameters.

[0172] The response unit 804 is used to determine the corresponding question-answering model from at least one model based on the model request, wherein the question-answering model is trained based on the training method of the above-mentioned question-answering model.

[0173] Optionally, the model library stores at least one question-answering model adapted to different scenarios, as well as multiple question-answering models with different model specifications and parameters; the response unit 804 is used for:

[0174] Based on at least one piece of information in the model request, search the model library for the question-answering model corresponding to at least one piece of information; wherein, the scene identifier of the target scene and the scene input data of the target scene both correspond to the model adapted to the target scene, and the target model specification parameters correspond to the model with the same model specification parameters as the target model specification parameters.

[0175] If at least one piece of information includes scene input data of the target scene, the found question-answering model is trained based on the scene input data of the target scene to obtain a trained question-answering model.

[0176] Optionally, the task platform also includes a question-and-answer interface, which is constructed based on a question-and-answer model; the question-and-answer interface is used for scheduling by the terminal device to provide answer information for the question information.

[0177] The above is an illustrative scheme of a task platform according to this embodiment. It should be noted that the technical solution of this task platform and the technical solution of the question-and-answer processing method described above belong to the same concept. For details not described in detail in the technical solution of the task platform, please refer to the description of the technical solution of the question-and-answer processing method described above.

[0178] Corresponding to the above training method embodiments, this specification also provides embodiments of a training device for a question-answering model. Figure 9 This is a schematic diagram of the structure of a training device for a question-answering model provided in one embodiment of this specification. Figure 9 As shown, the device includes:

[0179] The first acquisition module 902 is used to acquire first training data, wherein the first training data includes: first data to be analyzed, a first cognitive question for the first data to be analyzed, a first response information corresponding to the first cognitive question, and a first position of the first response information in the first data to be analyzed;

[0180] The first generation module 904 is used to generate a first perception question based on a first cognitive question, wherein the first perception question indicates the position of the response information to the first cognitive question in the first data to be analyzed;

[0181] The first prediction module 906 is used to input the first data to be analyzed and the first perception question into a pre-trained question answering model, predict the first reference answer information corresponding to the first sample perception question, and predict the second reference answer information of the first perception question based on the first reference answer information.

[0182] The first adjustment module 908 is used to adjust the pre-trained question-answering model based on the first response information, the first position, the first reference response information, and the second reference response information to obtain the trained question-answering model.

[0183] Optionally, the first training data further includes: at least one first auxiliary position in the first data to be analyzed that is different from the first position; the first generation module 904 is used to: generate a first perception question based on the first cognitive question, the first position and at least one first auxiliary position; wherein the first perception question indicates the location of the response information of the first cognitive question in the first position and at least one first auxiliary position.

[0184] Optionally, the first generation module 904 is used to: fill the first cognitive question, the information of the first position, and the information of at least one first auxiliary position into the corresponding position in the query information template to obtain the first perception question; wherein, the query information template includes: the question stem information indicating the perception question, the filling position information of the first cognitive question, the filling position information of the information of the first position, and the filling position information of the information of the first auxiliary position.

[0185] Optionally, the first prediction module 906 is used to: input the first data to be analyzed and the first perception question into a pre-trained question-answering model to obtain the first reference response information corresponding to the first cognitive question and the second reference response information of the first perception question arranged according to the response information template; wherein, the second reference response information is predicted based on the first reference response information, and the response information template includes: the relationship information between the first cognitive question and the first reference response information, the filling position information of the first reference response information, the filling position information of the second reference response information, and the relationship information between the first reference response information and the second reference response information.

[0186] Optionally, the training apparatus for the question-answering model also includes:

[0187] The second acquisition module is used to acquire the second training data, wherein the second training data includes: the second data to be analyzed, the second cognitive question for the second data to be analyzed, and the second response information corresponding to the second cognitive question;

[0188] The second generation module is used to generate a third cognitive question based on the second cognitive question and the second response information, wherein the third cognitive question indicates the response information of the second cognitive question to be determined based on the second response information;

[0189] The second prediction module is used to input the second data to be analyzed and the third cognitive question into the pre-trained question-answering model to predict the third reference answer information for the second cognitive question.

[0190] The second adjustment module is used to adjust the pre-trained question-answering model based on the relationship between the third reference response information and the second response information, so as to obtain the trained question-answering model.

[0191] Optionally, the second training data further includes: at least one auxiliary response information in the second data to be analyzed that is different from the second response information; the second generation module is used to: generate a third cognitive question based on the second cognitive question, the second response information and at least one auxiliary response information; wherein the third cognitive question indicates that the response information for the second cognitive question is determined from the second response information and at least one auxiliary response information.

[0192] Optionally, the training apparatus for the question-answering model also includes:

[0193] The third acquisition module is used to acquire the third training data, wherein the third training data includes: the third data to be analyzed, the third response information corresponding to the second perception question of the third data to be analyzed, and the second position of the third response information in the third data to be analyzed;

[0194] The third generation module is used to generate a third perception question based on the third response information, wherein the third perception question indicates the position of the third response information in the third data to be analyzed;

[0195] The third prediction module is used to input the third data to be analyzed and the third perception question into the pre-trained question answering model to predict the fourth reference response information of the third perception question.

[0196] The third adjustment module is used to adjust the pre-trained question-answering model based on the relationship between the fourth reference response information and the second position, so as to obtain the trained question-answering model.

[0197] Optionally, the third training data further includes: at least one second auxiliary position in the third data to be analyzed that is different from the second position; the third generation module is used to: generate a third perception question based on the third response information, the second position and at least one second auxiliary position; wherein the third perception question indicates the response information of the third perception question to be determined in the second position and at least one second auxiliary position.

[0198] The technical solutions of the training device for the question-answering model and the training method for the question-answering model mentioned above belong to the same concept, and the technical solutions of the training device can be found in the description of the technical solutions of the training method mentioned above.

[0199] Corresponding to the above-described question-and-answer processing method embodiments, this specification also provides embodiments of question-and-answer processing devices. Figure 10 This is a schematic diagram of the structure of a question-and-answer processing device provided in one embodiment of this specification. Figure 10 As shown, the question-and-answer processing device includes:

[0200] The acquisition module 1002 is used to acquire task data for the question-and-answer task, wherein the task data includes data to be analyzed and question information raised for the data to be analyzed;

[0201] The analysis module 1004 is used to call the question-answering model to analyze the task data and obtain the answer information corresponding to the question information; wherein, the question-answering model is the question-answering model trained by the aforementioned training method.

[0202] Optionally, the question-and-answer processing apparatus further includes:

[0203] The receiving module is used to receive question-and-answer requests uploaded by the front-end device, wherein the question-and-answer request includes the task data of the question-and-answer task;

[0204] The first sending module is used to call the question-answering model to analyze the task data, obtain the answer information corresponding to the question information, and then send the answer information corresponding to the question information to the front-end device.

[0205] Optionally, the question-and-answer processing apparatus further includes:

[0206] The adjustment module is used to adjust the question-answering model based on the question feedback information after sending the response information corresponding to the question information to the front-end device.

[0207] Optionally, the question-and-answer processing apparatus further includes:

[0208] The second sending module is used to send the response information corresponding to the problem information to the front-end device, and then send the recommended task information related to the data to be analyzed or the problem information to the front-end device.

[0209] The processing module is used to call the question-answering model to generate the execution result of the recommendation task information when it receives the trigger instruction from the front-end device for the recommendation task information, and then send the execution result to the front-end device.

[0210] Figure 11A structural block diagram of a computing device according to an embodiment of this specification is shown. Components of the computing device 1200 include, but are not limited to, a memory 1210 and a processor 1220. The processor 1220 is connected to the memory 1210 via a bus 1230, and a database 1250 is used to store data. The computing device 1200 also includes an access device 1240, which enables the computing device 1200 to communicate via one or more networks 1260. Examples of these networks include a Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the Internet. Access device 1240 may include one or more of any type of wired or wireless network interface (e.g., network interface controller (NIC)), such as IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, Wi-MAX (Worldwide Interoperability for Microwave Access) interface, Ethernet interface, Universal Serial Bus (USB) interface, cellular network interface, Bluetooth interface, and Near Field Communication (NFC).

[0211] In one embodiment of this specification, the aforementioned components of the computing device 1200 and Figure 11 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 11 The block diagram of the computing 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.

[0212] The computing device 1200 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 1200 can also be a mobile or stationary server. The memory 1210 is used to store computer programs / instructions, and the processor 1220 is used to execute the computer programs / instructions stored in the memory 1210, which, when executed by the processor, implement the steps of the aforementioned question-answering processing method or question-answering model training method.

[0213] The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device belongs to the same concept as the technical solution of the question-answering processing method or the question-answering model training method described above. For details not described in detail in the technical solution of the computing device, please refer to the description of the technical solution of the question-answering processing method or the question-answering model training method described above.

[0214] 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 question-answering processing method or question-answering model training method. 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 belongs to the same concept as the technical solution of the above-described question-answering processing method or question-answering model training method. Details not described in detail in the technical solution of the storage medium can be found in the description of the technical solution of the above-described question-answering processing method or question-answering model training method.

[0215] This specification also provides a computer program product in one embodiment, including a computer program / instructions that, when executed by a processor, implement the steps of the above-described question-answering processing method or question-answering model training method. 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 belongs to the same concept as the technical solution of the above-described question-answering processing method or question-answering model training method. Details not described in detail in the technical solution of the computer program product can be found in the description of the technical solution of the above-described question-answering processing method or question-answering model training method.

[0216] 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.

[0217] Computer programs / instructions include computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. Computer-readable media can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, 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 computer-readable media can be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.

[0218] 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. Secondly, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification. In the above embodiments, the descriptions of each embodiment have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0219] 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 question-answering model, comprising: Acquire first training data, wherein the first training data includes: first data to be analyzed, a first cognitive question for the first data to be analyzed, a first response information corresponding to the first cognitive question, and a first position of the first response information in the first data to be analyzed; A first perception question is generated based on the first cognitive question, wherein the first perception question indicates the position of the response information to the first cognitive question in the first data to be analyzed; The first data to be analyzed and the first perception question are input into a pre-trained question-answering model to predict the first reference answer information corresponding to the first perception question, and the second reference answer information of the first perception question is predicted based on the first reference answer information. Based on the first response information, the first location, the first reference response information, and the second reference response information, the pre-trained question-answering model is adjusted to obtain the trained question-answering model.

2. The method according to claim 1, wherein the first training data further comprises: At least one first auxiliary position in the first data to be analyzed that is different from the first position; The generation of the first perception question based on the first cognitive question includes: Based on the first cognitive question, the first location, and the at least one first auxiliary location, a first perceptual question is generated; wherein, the first perceptual question indicates the location of the response information for the first cognitive question among the first location and the at least one first auxiliary location.

3. The method according to claim 2, wherein generating the first perceptual question based on the first cognitive question, the first location, and the at least one first auxiliary location comprises: The first cognitive question, the information of the first location, and the information of the at least one first auxiliary location are filled into the corresponding positions in the query information template to obtain the first perception question; wherein, the query information template includes: the question stem information indicating the perception question, the filling position information of the first cognitive question, the filling position information of the information of the first location, and the filling position information of the information of the first auxiliary location.

4. The method according to claim 1, wherein inputting the first data to be analyzed and the first perceptual question into a pre-trained question-answering model, predicting the first reference response information corresponding to the first cognitive question, and predicting the second reference response information of the first perceptual question based on the first reference response information, comprises: The first data to be analyzed and the first perception question are input into a pre-trained question-answering model to obtain the first reference answer information corresponding to the first cognitive question and the second reference answer information of the first perception question, which are arranged according to the response information template. The second reference response information is predicted based on the first reference response information. The response information template includes: the relationship information between the first cognitive question and the first reference response information, the filling position information of the first reference response information, the filling position information of the second reference response information, and the relationship information between the first reference response information and the second reference response information.

5. The method according to any one of claims 1 to 4, further comprising: Acquire second training data, wherein the second training data includes: second data to be analyzed, a second cognitive question for the second cognitive question, and second response information corresponding to the second cognitive question; A third cognitive question is generated based on the second cognitive question and the second response information, wherein the third cognitive question indicates the response information for determining the second cognitive question based on the second response information; The second data to be analyzed and the third cognitive question are input into the pre-trained question-answering model to predict the third reference answer information for the second cognitive question; Based on the relationship between the third reference response information and the second response information, the pre-trained question-answering model is adjusted to obtain the trained question-answering model.

6. The method according to claim 5, wherein the second training data further comprises: At least one auxiliary response information that differs from the second response information in the second data to be analyzed; The generation of the third cognitive question based on the second cognitive question and the second response information includes: A third cognitive question is generated based on the second cognitive question, the second response information, and the at least one auxiliary response information; wherein the third cognitive question indicates that the response information for the second cognitive question is determined from the second response information and the at least one auxiliary response information.

7. The method according to any one of claims 1 to 4, further comprising: Acquire third training data, wherein the third training data includes: third data to be analyzed, third response information corresponding to a second perception question for the third data to be analyzed, and a second position of the third response information in the third data to be analyzed; A third perception question is generated based on the third response information, wherein the third perception question indicates the location of the third response information in the third data to be analyzed; The third data to be analyzed and the third perception question are input into the pre-trained question-answering model to predict the fourth reference response information for the third perception question; Based on the relationship between the fourth reference response information and the second position, the pre-trained question-answering model is adjusted to obtain the trained question-answering model.

8. The method according to claim 7, wherein the third training data further comprises: The third set of data to be analyzed contains at least one second auxiliary position that is different from the second position; The generation of the third perception question based on the third response information includes: Based on the third response information, the second location, and the at least one second auxiliary location, a third perception question is generated; wherein the third perception question indicates that the response information for the third perception question is determined in the second location and the at least one second auxiliary location.

9. A question-and-answer processing method, comprising: Obtain task data for a question-and-answer task, wherein the task data includes data to be analyzed and question information raised in response to the data to be analyzed; The question-answering model is invoked to analyze the task data to obtain the response information corresponding to the question information; wherein, the question-answering model is a question-answering model trained by any one of the training methods described in claims 1 to 8.

10. The method according to claim 9, wherein obtaining the task data for the question-answering task includes: Receive a question-and-answer request uploaded by a front-end device, wherein the question-and-answer request includes the task data of the question-and-answer task; After the step of calling the question-answering model to analyze the task data and obtain the answer information corresponding to the question information, it also includes: Send the response information corresponding to the problem information to the front-end device.

11. The method according to claim 10, after sending the response information corresponding to the problem information to the front-end device, further comprising: Upon receiving feedback information from the front-end device regarding the response information, the question-and-answer model is adjusted based on the feedback information.

12. The method according to claim 10 or 11, after sending the response information corresponding to the problem information to the front-end device, further comprising: Send recommended task information related to the data to be analyzed or the problem information to the front-end device; Upon receiving a trigger instruction from the front-end device regarding the recommended task information, the system invokes the question-and-answer model to generate the execution result of the recommended task information and sends the execution result to the front-end device.

13. A question-answering processing method, applied to a task platform, comprising: The device receives a model request sent by a terminal device, wherein the model request includes at least one of the following: scene identifier of the target scene, scene input data of the target scene, and target model specification parameters. Based on the model request, a corresponding question-answering model is determined from at least one model; wherein the question-answering model is trained based on the training method described in any one of claims 1 to 8, and the question-answering model is used to output the answer information corresponding to the question information.

14. The method according to claim 13, wherein the model library stores at least one question-answering model adapted to different scenarios, and multiple question-answering models with different model specification parameters; the step of determining the corresponding question-answering model from at least one model based on the model request includes: Based on at least one piece of information in the model request, a question-answering model corresponding to the at least one piece of information is searched in the model library; wherein, the scene identifier of the target scene and the scene input data of the target scene both correspond to a model adapted to the target scene, and the target model specification parameters correspond to a model with the same model specification parameters as the target model specification parameters; If at least one piece of information includes scene input data of the target scene, the found question-answering model is trained based on the scene input data of the target scene to obtain a trained question-answering model.

15. The method according to claim 13 or 14, further comprising, after determining the corresponding question-answering model from at least one model based on the model request: Deploy the question-and-answer model and build a question-and-answer interface based on the question-and-answer model, so that the terminal device can schedule the question-and-answer model through the question-and-answer interface to generate answer information corresponding to the question information.

16. A task platform, comprising a request interface and a response unit; The request interface is used to receive model requests sent by the terminal device, wherein... The model request includes at least one of the following: the scene identifier of the target scene, the scene input data of the target scene, and the model specification parameters; The response unit is configured to determine a corresponding question-answering model from at least one model based on the model request, wherein the question-answering model is trained based on the training method described in any one of claims 1 to 8.

17. The task platform according to claim 16 further includes a question-and-answer interface, wherein the question-and-answer interface is constructed based on the question-and-answer model; the question-and-answer interface is used for scheduling by the terminal device to provide answer information for question information.

18. 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, wherein when the computer programs / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 15.

19. 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.

20. 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