Sample data construction method, task platform, computing device, and computer-readable storage medium

Abstract

This specification provides a sample data construction method, task platform, computing device, and computer-readable storage medium. The sample data construction method includes: acquiring a sample document; using a parsing model to split the sample document into multiple sample document fragments; parsing the multiple sample document fragments to obtain multiple sample parsing results; sending the multiple sample parsing results to a front-end for display; and, upon receiving a confirmation instruction from the front-end, using the parsing model to generate sample question-and-answer pairs corresponding to the multiple sample parsing results. In the process of parsing and generating the sample document using the parsing model, the parsing process and the generation process are separated. The sample parsing results are sent to the front-end for display, and sample question-and-answer pairs are only generated upon receiving a confirmation instruction. This overcomes the problem of the black box nature of the model's parsing and generation process, ensuring the quality of the sample question-and-answer pairs.

Description

Technical Field

[0001] This specification relates to the technical field of deep learning, and in particular to a method for constructing sample data, a task platform, a computing device, and a computer-readable storage medium. Background Technology

[0002] With the development of deep learning technology, training question-answering models or building retrieval knowledge bases using sample question-answering pairs can enable professional question-answering in specific fields. For example, in the medical and health field, trained question-answering models or retrieval knowledge bases can provide disease queries and health advice; in legal consultation, trained question-answering models or retrieval knowledge bases can explain regulations and guide users to understand their rights and obligations, providing professional legal advice; and in technical support, trained question-answering models or retrieval knowledge bases can help users solve problems encountered during software or hardware development and use.

[0003] Currently, training question-answering models relies on high-quality sample question-answer pairs, which are typically constructed manually, resulting in inefficiency. Therefore, the parsing capabilities of multimodal models such as natural language models, visual models, and structured data models can be utilized to automatically generate sample question-answer pairs by parsing sample documents. However, since the model's parsing and generation process is a black box, it's impossible to know whether the model fully understands the document content before generating the sample question-answer pairs. This leads to inconsistent quality in the constructed sample question-answer pairs, potentially affecting the training performance of subsequent question-answering models or the quality of the knowledge base built for retrieval. Therefore, an interpretable sample data construction method is urgently needed to ensure the quality of constructed sample question-answer pairs. Summary of the Invention

[0004] In view of this, embodiments of this specification provide a method for constructing sample data. One or more embodiments of this specification also relate to a task platform, a computing device, a computer-readable storage medium, and a computer program product, to address the technical deficiencies existing in the prior art.

[0005] According to a first aspect of the embodiments of this specification, a method for constructing sample data is provided, comprising:

[0006] Obtain sample documents;

[0007] Using a parsing model, the sample document is split into multiple sample document fragments. These fragments are then parsed to obtain multiple parsing results.

[0008] Send the parsing results of multiple samples to the front end for display;

[0009] Upon receiving confirmation instructions from the front end, the parsing model is used to generate sample question-and-answer pairs corresponding to multiple sample parsing results based on the parsing results of multiple samples.

[0010] According to a second aspect of the embodiments of this specification, a task platform is provided, including a sample data construction interface and a response unit;

[0011] The sample data construction interface is used to receive sample documents sent from the front end;

[0012] The response unit is used to split the sample document to obtain multiple sample document fragments, parse the multiple sample document fragments, and obtain multiple sample parsing results;

[0013] The sample data construction interface is also used to send the parsing results of multiple samples to the front end for display and to receive confirmation instructions from the front end.

[0014] The response unit is also used to generate sample question-answer pairs corresponding to multiple sample parsing results based on the parsing results of multiple samples using the parsing model.

[0015] According to a third aspect of the embodiments of this specification, a computing device is provided, comprising:

[0016] Memory and processor;

[0017] The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the above-described sample data construction method.

[0018] According to a fourth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores a computer program / instructions that, when executed by a processor, implement the steps of the sample data construction method described above.

[0019] According to a fifth 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 sample data construction method described above.

[0020] In one embodiment of this specification, a sample document is obtained; the sample document is split into multiple sample document fragments using a parsing model; the multiple sample document fragments are parsed to obtain multiple sample parsing results; the multiple sample parsing results are sent to the front end for display; upon receiving a confirmation instruction from the front end, the parsing model is used to generate sample question-and-answer pairs corresponding to the multiple sample parsing results.

[0021] In the process of parsing and generating sample documents using the parsing model, the parsing process and the generation process are separated. Sample parsing results corresponding to multiple sample document fragments are obtained and sent to the front end for display. Only after receiving confirmation instructions from the front end are sample question-and-answer pairs corresponding to multiple sample parsing results generated. This allows the front end to verify whether the parsing model has fully understood the document content when parsing and generating sample question-and-answer pairs. This overcomes the problem of the black box and lack of visibility in the model's parsing and generation process and ensures the quality of sample question-and-answer pairs. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating a sample data construction method provided in one embodiment of this specification;

[0023] Figure 2 This is a front-end display diagram of the sample parsing results and sample question-and-answer pairs of a sample data construction method provided in one embodiment of this specification;

[0024] Figure 3 This is a schematic diagram of the front-end display of a table in a sample data construction method provided in one embodiment of this specification;

[0025] Figure 4 This is a front-end display diagram illustrating a sample data construction method provided in one embodiment of this specification, showing the adjustment of a table.

[0026] Figure 5 This is a front-end display diagram illustrating a single cell processing method for constructing sample data according to an embodiment of this specification;

[0027] Figure 6 This is a front-end display diagram illustrating the processing of multiple cells in a sample data construction method provided in one embodiment of this specification;

[0028] Figure 7 This is a front-end display diagram of adding columns in a sample data construction method provided in one embodiment of this specification;

[0029] Figure 8 This is a schematic diagram of the front-end display of the add row in a sample data construction method provided in one embodiment of this specification;

[0030] Figure 9 This is a front-end display diagram showing the comparison of table processing before and after a sample data construction method provided in one embodiment of this specification;

[0031] Figure 10 This is a flowchart illustrating the process of constructing sample data for product Q&A, as provided in one embodiment of this specification.

[0032] Figure 11This is a schematic diagram of the structure of a task platform provided in one embodiment of this specification;

[0033] Figure 12 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0034] 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.

[0035] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0036] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0037] 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.

[0038] In one or more embodiments of this specification, a large model refers to a deep learning model with a large number of model parameters, typically containing hundreds of millions, tens of billions, hundreds of billions, trillions, or even tens of trillions of model parameters. A large model can also be called a foundation model. It is pre-trained using large-scale unlabeled corpora to produce a pre-trained model with hundreds of millions of parameters. Such models can adapt to a wide range of downstream tasks and have good generalization ability. Examples include Large Language Models (LLMs) and multi-modal pre-training models.

[0039] In practical applications, large models only require a small number of samples to fine-tune the pre-trained model before they can be applied to different tasks. Large models can be widely used in fields such as Natural Language Processing (NLP) and Computer Vision. Specifically, they can be applied to computer vision tasks such as Visual Question Answering (VQA), Image Captioning (IC), and Image Generation, as well as natural language processing tasks such as text-based sentiment classification, text summarization, and machine translation. The main application scenarios for large models include digital assistants, intelligent robots, search, online education, office software, e-commerce, and intelligent design.

[0040] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0041] Large language models, also known as large models, are natural language models with a large number of parameters, trained using deep learning techniques. These models are typically trained on large amounts of text data and are capable of generating and understanding natural language. They are widely used in tasks such as text generation, machine translation, dialogue systems, and question-answering systems. Examples include GPT, BERT, and T5.

[0042] Deep self-attention model (Transformer model): A deep learning architecture based on attention mechanism for processing sequential data, such as natural language.

[0043] Bidirectional Encoder Representations from Transformers (BERT model): A special type of deep self-attention model trained using a bidirectional Transformer encoder and large-scale unlabeled text data.

[0044] This specification provides a method for constructing sample data. It also relates to 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.

[0045] See Figure 1 , Figure 1 A flowchart of a sample data construction method according to an embodiment of this specification is shown, including the following specific steps:

[0046] Step 102: Obtain sample documents.

[0047] This specification applies to applications, websites, or system platforms with sample data construction capabilities. It is used to build sample question-and-answer pairs for specific domains, facilitating subsequent supervised training of question-and-answer models or the construction of retrieval knowledge bases. For example, an online medical service platform can automatically extract information from medical literature and drug instructions to generate high-quality sample question-and-answer pairs. These pairs are used to train question-and-answer models specifically for the medical field or to build retrieval knowledge bases, enabling users to query professional questions related to symptoms, treatment plans, and drug use. Another example is a legal consultation website that parses and generates sample question-and-answer pairs from legal texts. These pairs are used to train question-and-answer models specifically for the legal field or to build retrieval knowledge bases, allowing non-professionals to easily access accurate legal knowledge, understand rights and obligations, and obtain legal guidance. Yet another example is a software development community or internal technical support platform that automatically generates sample question-and-answer pairs related to programming languages, frameworks, libraries, and hardware devices. These pairs are used to train question-and-answer models specifically for the technical field or to build retrieval knowledge bases. This not only improves the speed at which developers solve problems but also provides learning resources for beginners, helping them quickly learn new technologies or solve problems. For example, an online education platform can automatically create sample question-and-answer pairs related to course content from textbooks. These pairs can be used to train question-and-answer models specifically for the education field or to build retrieval knowledge bases, such as explanations of mathematical formulas, analyses of historical events, and explanations of scientific concepts, thereby enhancing the student learning experience and providing an instant feedback mechanism. Another example is the customer service center of an e-commerce or retail company, which can quickly generate sample question-and-answer pairs of frequently asked questions from historical question-and-answer records. These pairs can be used to train question-and-answer models specifically for the customer service field or to build retrieval knowledge bases, improving customer service efficiency and reducing the workload of human customer service representatives.

[0048] Sample documents are the data source documents used to construct sample question-and-answer pairs. Sample documents typically include data from a specific domain. Sample documents can include text modal data and multimodal data. Multimodal data includes, but is not limited to, tabular modal data, image modal data, and structure tree modal data.

[0049] One option for obtaining sample documents is to receive sample documents sent by the front end, and another option is to retrieve sample documents from a database. The database can be an open-source database or a historical database, and there is no limitation on this.

[0050] For example, on a task platform with domain-specific question-answering model training capabilities, users can interactively train a question-answering model for their target domain. User A, representing an online medical service platform, needs to train an online question-answering model for the medical field through this task platform. User A sends multiple sample documents via the front end, one of which is the "Instructions for Use of Inhaled Aerosol Drug A":

[0051] Product Name: A Inhalation Aerosol (A Drug Inhalation Aerosol).

[0052] Ingredients and Specifications: Active Ingredients: Each spray contains ingredients A1 and A2. Specifications: Depending on the packaging, each bottle contains a different dosage of the medication. Please refer to the packaging label for specific information.

[0053] Indications: Drug A inhaler is a combination therapy used for the long-term control of asthma in adults and adolescents aged 12 years and older, and for the maintenance treatment of chronic obstructive pulmonary disease (COPD). It helps improve breathing by reducing airway inflammation and dilating the bronchi.

[0054] Instructions for Use and Dosage: Instructions for Use: This product is for inhalation only and should be used as directed by your doctor. Before first use, shake the bottle well and press to test and ensure the device is functioning correctly. Dosage: The specific frequency and dosage should be determined by your doctor based on the patient's individual condition. Generally, use no more than twice daily, one spray each time, or as directed by your doctor.

[0055] Standard Clauses:

[0056] Section 9.1 Fracture Risk: Patients on long-term inhaled corticosteroid therapy, such as ingredient A1, have an increased risk of fracture. Therefore, physicians should assess each patient's bone health and consider monitoring bone mineral density during treatment.

[0057] Section 9.2 Effects on Bone Mineral Density: Compared with the use of inhaled corticosteroids alone, the use of inhaled aerosol of drug A may result in a decrease in bone mineral density. Caution should be exercised when initiating or continuing use of this product in patients with high-risk factors, and additional precautions may be necessary.

[0058] Section 9.3 does not apply to acute conditions: Inhaled medication A is not indicated for the treatment of acute asthma attacks or acute exacerbations of COPD. If a patient experiences an acute worsening of symptoms, they should seek immediate medical help and adjust their medication regimen as directed by their doctor.

[0059] Precautions: Prescription use only: This product should only be used under the guidance of a physician. Do not purchase or change the dosage on your own. Acute symptom management: This product is not a rapid relief medication. If acute respiratory distress occurs, use emergency medications immediately and consult a physician. Regular follow-up: Patients should have regular follow-up visits with their physician to assess efficacy and safety. Pediatric use: This product is suitable for individuals aged 12 years and older. The safety and effectiveness in children under 12 years of age have not been established. Contraindications: This product is contraindicated in patients with known hypersensitivity to ingredient A1, ingredient A2, or any other ingredient in this product.

[0060] Adverse reactions: Common adverse reactions include, but are not limited to, sore throat, hoarseness, headache, and nausea. Serious side effects are rare, but if symptoms such as chest pain, rapid heartbeat, or severe allergic reactions occur, seek medical attention immediately.

[0061] Storage conditions: Store this product at room temperature, avoiding high temperatures and direct sunlight. Unused medicines should be disposed of properly in accordance with local regulations.

[0062] Packaging: This product is typically packaged with a metered-dose inhaler. Each box contains one inhaler with a counter and a set dose. Please refer to the packaging for detailed packaging information.

[0063] Obtaining sample documents provides the data foundation for subsequent parsing and generating sample question-and-answer pairs.

[0064] Step 104: Using the parsing model, split the sample document to obtain multiple sample document fragments, parse the multiple sample document fragments, and obtain multiple sample parsing results.

[0065] The parsing model is a deep learning model with parsing and question-answer pair generation functions. It can be a pre-trained deep learning model, which is trained based on a pre-trained document, the label parsing results corresponding to multiple pre-trained document fragments in the pre-trained document, and the label question-answer pairs corresponding to multiple label parsing results. For example, a deep self-attention model, or a deep self-attention model with bidirectional encoding representation, or a large language model that achieves the above functions under the guidance of prompt information, without limitation here.

[0066] Sample document fragments are paragraphs or sentences containing specific information in a sample document. Each sample document fragment usually revolves around a central theme or information point and is the basic unit for constructing sample question-and-answer pairs.

[0067] The sample parsing result is the content generated after parsing the sample document fragments. The sample parsing result can be various modalities, including but not limited to: sample summary text, tables, images, and tree structures. Ideally, the sample parsing result more accurately reflects the content of the sample document, assisting the model in clearly identifying which content is suitable as sample questions and which content is suitable as sample answers during the question-answer pair generation process. Compared to sample document fragments, the sample parsing result is more refined and structured, helping to improve the quality and accuracy of the generated sample question-answer pairs.

[0068] Using a parsing model, the sample document is split into multiple sample document fragments. One possible approach is to use the parsing model to split the sample document based on its structural features, such as segmenting based on structural elements like titles, paragraphs, and lists. Another possible approach is to use the parsing model to split the sample document based on its semantic features, and then analyze the semantic coherence of the document content to group semantically similar sentences or paragraphs into a single fragment. This approach is not limited to a single fragment.

[0069] Using a parsing model, document parsing is performed on the sample document to obtain sample parsing results corresponding to multiple sample document fragments in the sample document. One optional approach is to use the parsing model to perform semantic parsing on the sample document to obtain sample parsing results corresponding to multiple sample document fragments in the sample document. The semantic parsing method includes at least one of the following: semantic segmentation, semantic generalization, and semantic classification. Another optional approach is to use the parsing model to perform multimodal data extraction on the sample document to obtain sample parsing results corresponding to multiple sample document fragments in the sample document. No limitation is imposed here.

[0070] It should be noted that step 104 can be implemented directly using a pre-trained parsing model, or it can be implemented using a large language model with the guidance of prompts, such as "You are an expert in the medical field. Please parse the input sample document and obtain the sample parsing results corresponding to multiple sample document fragments". No limitation is made here.

[0071] For example, the parsing model is a pre-trained deep attention model, which is obtained through two stages of supervised training based on sample data (including pre-trained documents, label parsing results corresponding to multiple pre-trained document fragments in the pre-trained documents, and label question-answer pairs corresponding to multiple label parsing results).

[0072] Using this analytical model, semantic parsing was performed on the aforementioned sample document, "Instructions for Use of Drug A Inhaler," yielding parsing results for eight document fragments. The parsing result for the document fragment containing the "Standard Clauses" is as follows:

[0073] Section 9.1: Patients on long-term inhaled corticosteroid therapy are at risk of fractures.

[0074] Section 9.2: Compared with inhaled corticosteroids, the use of this product may result in decreased bone mineral density.

[0075] Article 9.3: This product is not suitable for the treatment of acute asthma attacks or acute exacerbations of COPD.

[0076] In the process of parsing and generating sample documents using the parsing model, the parsing process and the generation process are separated, resulting in sample parsing results corresponding to multiple sample document fragments and sample question-and-answer pairs corresponding to multiple sample parsing results, providing comprehensive data support for subsequent direct display on the front end.

[0077] Step 106: Send the parsing results of multiple samples to the front end for display.

[0078] The front-end is the user interface of an application, website, or system platform that has sample data construction capabilities. It is used for direct interaction with the front-end user and serves as the display interface for the sample data construction results, sample parsing results, and sample question-and-answer pairs. The front-end can not only display the constructed sample data but also receive feedback to interactively adjust the constructed sample data to improve its quality. The front-end can communicate with the back-end services of the application, website, or system platform through mechanisms such as Application Programming Interfaces (APIs) or message queues to complete the above interaction process.

[0079] For example, the front-end page displays the above sample analysis results. Users can click on different analysis results to view detailed content and make adjustments or confirmations as needed. The front-end can also provide interactive tools, allowing users to manually modify the analysis results or add additional annotations to further improve the accuracy of the analysis results.

[0080] Sending the parsing results of multiple samples to the front end for display provides an interactive basis for subsequent confirmation and generation of sample question-and-answer pairs.

[0081] Step 108: Upon receiving the confirmation instruction from the front end, use the parsing model to generate sample question-answer pairs corresponding to multiple sample parsing results based on the parsing results of multiple samples.

[0082] The confirmation command is triggered when the user confirms the parsing results on the front-end interface. It indicates that the user believes the sample parsing results output by the parsing model are reasonable and correct, and that sample question-and-answer pairs can be generated based on these results. The confirmation command can be triggered by clicking the "Confirm" button, or through messages, code commands, or voice commands.

[0083] The sample question-and-answer pairs corresponding to the sample parsing results are sample questions and sample answers generated based on the sample parsing results. These sample question-and-answer pairs constitute the supervision data for training the question-and-answer model, or the knowledge base data for constructing the retrieval knowledge base. Sample question-and-answer pairs can be directly extracted from the sample parsing results, or they can be rewritten based on the sample parsing results; this is not limited here. Compared to sample documents, sample question-and-answer pairs more directly reflect the questions users might ask and their corresponding correct answers, enabling the question-and-answer model to better serve the question-and-answer needs of specific domains.

[0084] Using a parsing model, based on the parsing results of multiple samples, sample question-answer pairs corresponding to the parsing results of multiple samples are generated. One possible approach is to use the parsing model to perform Augmentation Generation (AG) based on the parsing results of multiple samples to generate sample question-answer pairs corresponding to the parsing results of multiple samples. Another possible approach is to use the parsing model to perform template matching based on the parsing results of multiple samples to generate sample question-answer pairs corresponding to the parsing results of multiple samples according to a predefined question-answer template. No particular limitation is made here.

[0085] Optionally, upon receiving a confirmation instruction from the front end, after generating sample question-and-answer pairs corresponding to multiple sample parsing results using the parsing model, the following specific steps are also included:

[0086] Send sample question-and-answer pairs to the front end for display.

[0087] For example, upon receiving a confirmation instruction from the front end, the parsing model is used to enhance the generation based on the above three sample parsing results, generating sample question-answer pairs corresponding to the three sample parsing results:

[0088] Sample Question 1: What are the risks for patients who use inhaled corticosteroids long-term?

[0089] Sample response 1: Patients who use inhaled corticosteroids for a long period of time are at risk of fractures.

[0090] Sample Question 2: What are the adverse reactions of inhaled corticosteroids?

[0091] Sample response 2: Compared with inhaled corticosteroids, the use of this product may result in decreased bone density.

[0092] Sample Question 3: What should I do if I experience an acute asthma attack while using this product?

[0093] Sample response 3: If this occurs, please discontinue use of this product and consider other treatment options.

[0094] Sample Question 4: Can this product be used to treat acute asthma attacks?

[0095] Sample response 4: This product is not suitable for treating acute asthma attacks or acute exacerbations of COPD.

[0096] Sample Question 5: Can it be used by patients with asthma who are in remission?

[0097] Sample response 5: This product can be used by patients with asthma who are experiencing relief.

[0098] Send the parsing results of multiple samples, along with the corresponding question-and-answer pairs, to the front end for display. Figure 2 This diagram illustrates the front-end display of sample parsing results and sample question-answer pairs from an embodiment of the sample data construction method provided in this specification. Figure 2 As shown:

[0099] The original document, "Instructions for Use of Drug A Inhaler," sent from the front end, is displayed on the left side of the front-end page, as shown in the example above. The parsing results are displayed in the middle of the front-end page. These results include summary text, a structure tree, tables, and images. The summary text displays summary text for multiple document fragments from the original document.

[0100] A Inhalation Aerosol is a combination medication used to treat asthma and chronic obstructive pulmonary disease (COPD). Each spray contains active ingredients A1 and A2. Dosage varies depending on the packaging; please refer to the packaging label for details.

[0101] This inhaled aerosol is suitable for long-term asthma control in adults and adolescents aged 12 years and older, and for maintenance treatment in patients with COPD, by reducing airway inflammation and dilating the bronchi to improve respiratory status.

[0102] When using this product, it should be inhaled at the prescribed time as directed by the doctor. Shake well and test the device before first use. The specific frequency and dosage will be determined by the doctor based on the patient's condition, usually no more than twice a day, one spray each time.

[0103] Section 9.1: Patients on long-term inhaled corticosteroid therapy are at risk of fracture. Section 9.2: Compared to inhaled corticosteroids, this product may cause decreased bone mineral density. Section 9.3: This product is not indicated for the treatment of acute asthma attacks or acute exacerbations of COPD.

[0104] This product is for prescription use only and should not be purchased or have its dosage altered without medical advice. It is not a rapid-response medication; if acute respiratory distress occurs, use emergency medications immediately and consult a doctor. Regular follow-up appointments are recommended to assess efficacy and safety. This product is suitable for individuals aged 12 years and older; its safety and efficacy in children under 12 years of age have not been established. This product is contraindicated in individuals with known hypersensitivity to ingredients A1, A2, or any other ingredient.

[0105] Common adverse reactions include sore throat, hoarseness, headache, and nausea. Serious side effects are rare, but if chest pain, rapid heartbeat, or severe allergic reactions occur, seek medical attention immediately.

[0106] This product should be stored at room temperature, avoiding high temperatures and direct sunlight. Unused medicines should be disposed of properly in accordance with local regulations.

[0107] Inhaled medication A typically comes with a metered inhaler. Each box contains one inhaler with a counter and a specific dose. Please refer to the instructions on the packaging for details.

[0108] After selecting the parsing result for "Article 9.1...", the corresponding question-and-answer pair will be displayed on the right, as shown in the example above.

[0109] In the embodiments of this specification, during the process of parsing and generating sample documents using the parsing model, the parsing process and the generation process are separated. Sample parsing results corresponding to multiple sample document fragments are obtained and sent to the front end for display. Only after receiving the confirmation instruction from the front end is the sample question-and-answer pair corresponding to the multiple sample parsing results generated. This allows the front end to verify whether the parsing model has fully understood the document content when parsing and generating the sample question-and-answer pair. This overcomes the problem of the black box and lack of visibility in the model parsing and generation process and ensures the quality of the sample question-and-answer pair.

[0110] In one optional embodiment of this specification, after step 106, the following specific steps are further included:

[0111] Upon receiving the splitting and adjustment instructions from the front end, the parsing model is used to split the sample document according to the splitting and adjustment instructions, obtaining multiple adjusted sample document fragments of the sample document. The multiple adjusted sample document fragments are then parsed to obtain multiple adjusted sample parsing results.

[0112] The analysis results of multiple adjusted samples are sent to the front end for display.

[0113] The splitting and adjustment command is triggered by the user's adjustments to the sample document fragments or sample parsing results generated by the parsing model on the front-end interface. It indicates that the user believes the sample document fragments or sample parsing results generated by the parsing model need further optimization or correction. The splitting and adjustment command can be triggered by clicking the "Adjust" button, dragging the document fragment boundaries, inputting adjustment suggestions, or through voice commands. The content of the splitting and adjustment command can include re-dividing the sample document fragments, modifying the sample parsing results, or re-extracting multimodal data.

[0114] The adjusted sample document fragment is a fragment obtained by re-splitting the sample document according to the splitting and adjustment instructions. The adjusted sample document fragment may better meet the user's expectations than the original sample document fragment, for example, it may more accurately revolve around a central theme or more clearly express a specific information point. The generation of the adjusted sample document fragment can be based on paragraphs, sentences, or semantic units manually divided by the user, or it can be automatically completed by the parsing model based on the adjustment suggestions provided by the user.

[0115] The adjusted sample parsing results are the content generated after re-parseing the sample document fragments according to the splitting and adjustment instructions. The adjusted sample parsing results may be more accurate, more complete, or more in line with the expression habits of a specific domain than the original sample parsing results. For example, users may revise the semantic summaries generated by the parsing model to make them closer to the meaning of the original document; or users may adjust the extracted multimodal data to better meet the needs of actual application scenarios.

[0116] For example, in Figure 2 Based on this, the expert clicked on the sample analysis result "Item 9.1..." and adjusted the corresponding sample document fragment. The task platform received the splitting and adjustment instruction from the expert via the front end: "Further refine the analysis result of Item 9.1, clarifying the specific causes and preventive measures of fracture risk." Using this analysis model and based on this splitting and adjustment instruction, the sample document was re-splitted to obtain the adjusted sample document fragment. The adjusted sample document fragment was then analyzed to obtain the adjusted sample analysis result.

[0117] Adjust sample analysis results:

[0118] Section 9.1: Patients on long-term inhaled corticosteroid therapy, such as ingredient A1, may have an increased risk of fracture. Therefore, physicians should assess each patient's bone health and consider monitoring bone mineral density during treatment to prevent potential skeletal problems.

[0119] In the embodiments described in this specification, a front-end feedback mechanism is introduced to enable real-time adjustments to the splitting process, thereby achieving real-time adjustment and optimization of the sample parsing results. The interactive approach, allowing direct participation in the sample data construction process, further improves the quality of the sample question-and-answer pairs.

[0120] In one optional embodiment of this specification, after step 106, the following specific steps are further included:

[0121] Upon receiving the parsing and adjustment instructions from the front end, the parsing model is used to parse multiple sample document fragments according to the parsing and adjustment instructions, and multiple adjusted sample parsing results are obtained.

[0122] The analysis results of multiple adjusted samples are sent to the front end for display.

[0123] The parsing adjustment command is triggered by the user's adjustments to the sample parsing results generated by the parsing model on the front-end interface, indicating that the user believes the sample parsing results generated by the parsing model need further optimization or correction. The parsing adjustment command can be triggered by clicking the "Adjust" button, entering modification suggestions, or through voice commands. The content of the parsing adjustment command can include restating the sample parsing results, re-extracting multimodal data, or further refining specific information points.

[0124] The adjusted sample parsing results are the content generated after re-parseing the sample document fragments according to the parsing adjustment instructions. The adjusted sample parsing results may be more accurate, more complete, or more in line with the expression habits of a specific domain than the original sample parsing results. For example, users may revise the semantic summaries generated by the parsing model to make them closer to the meaning of the original document; or users may adjust the extracted multimodal data to better meet the needs of actual application scenarios.

[0125] For example, in Figure 2 Based on this, the expert clicked on "Item 9.1..." in the sample analysis result and adjusted it. The task platform received the analysis adjustment instruction from the expert via the front end: "Please further refine the specific causes and preventive measures of fracture risk." Using the analysis model and based on this adjustment instruction, the sample document fragment was re-analyzed to obtain the adjusted sample analysis result:

[0126] Adjust sample analysis results:

[0127] Section 9.1: Patients on long-term inhaled corticosteroid therapy, such as ingredient A1, may have an increased risk of fracture. This is because these drugs may inhibit bone formation and increase bone resorption. Therefore, physicians should assess each patient's bone health and consider monitoring bone mineral density during treatment, while also advising patients to increase their calcium and vitamin D intake and engage in appropriate exercise to improve bone health.

[0128] In the embodiments described in this specification, a front-end feedback mechanism is introduced to enable real-time adjustments to the parsing process, allowing for real-time adjustment and optimization of the sample parsing results. The interactive approach, allowing direct participation in the sample data construction process, further improves the quality of the sample question-and-answer pairs.

[0129] In one optional embodiment of this specification, after step 108, the following specific steps are further included:

[0130] Send sample question-and-answer pairs to the front end for display;

[0131] Upon receiving the generation and adjustment instructions from the front end, the parsing model is used to generate adjustment sample question-answer pairs corresponding to multiple sample parsing results based on the generation and adjustment instructions. The question-answering model is then trained using the adjustment sample question-answer pairs.

[0132] and / or;

[0133] Send sample question-and-answer pairs to the front end for display;

[0134] Upon receiving adjusted question-and-answer pairs corresponding to multiple sample parsing results from the front end, the question-and-answer model is trained using these adjusted question-and-answer pairs.

[0135] The "Generate Adjustment Command" is a command triggered by the user's adjustments to the sample question-and-answer pairs generated by the parsing model on the front-end interface. It indicates that the user believes the sample question-and-answer pairs generated by the parsing model need further optimization or correction. The "Generate Adjustment Command" can be triggered by clicking the "Adjust" button, entering modification suggestions, or through voice commands. The "Generate Adjustment Command" can include the following: Modifying the question: making the question clearer, more accurate, or more relevant to the actual application scenario; Optimizing the answer: polishing, supplementing, or simplifying the answer to make it more readable and professional; Adding keywords: adding relevant keywords to the question or answer to improve the model's understanding and matching capabilities; Adjusting the format: adjusting the format of the question or answer to better meet the requirements of a specific domain.

[0136] The adjusted sample question-and-answer pairs are regenerated based on the adjustment instructions. The adjusted sample question-and-answer pairs may be more accurate, more complete, or more in line with the expression habits of a specific domain than the original sample question-and-answer pairs.

[0137] Using adjusted sample question-answer pairs to train a question-answering model, one possible approach is to construct a reward function based on the adjusted sample question-answer pairs and use the reward function to reinforce the training of the question-answering model. Another possible approach is to train the question-answering model by comparing the sample question-answer pairs before and after adjustment. Yet another possible approach is to supervise the training of the question-answering model using the adjusted sample question-answer pairs and the sample question-answer pairs before adjustment. No further restrictions are imposed here.

[0138] For example, in Figure 2 Based on this, the expert clicked on the third sample question-and-answer pair on the right and adjusted it. The adjustment instruction was then sent to the task platform via front-end feedback: "Please refine the answer." Using the analytical model, and following the adjustment instruction, a new adjusted sample question-and-answer pair was generated based on the sample analysis results.

[0139] Sample Question 3: If an acute asthma attack occurs while using inhaled drug A, what measures should be taken?

[0140] Sample Response 3: If an acute asthma attack occurs while using Medication A inhaler, discontinue use immediately and seek emergency medical help. Simultaneously, adjust the medication regimen or seek other suitable treatments as directed by your doctor. Please note that Medication A inhaler is not intended for rapid relief of acute symptoms.

[0141] For example, in Figure 2 Based on this, the expert clicked on the third sample question-and-answer pair on the right and adjusted it. The task platform then received the adjusted sample question-and-answer pair from the expert via the front end.

[0142] Adjusting sample question-and-answer pairs:

[0143] Sample Question 3: If an acute asthma attack occurs while using inhaled drug A, what measures should be taken?

[0144] Sample Response 3: If an acute asthma attack occurs while using Medication A inhaler, discontinue use immediately and seek emergency medical help. Simultaneously, adjust the medication regimen or seek other suitable treatments as directed by your doctor. Please note that Medication A inhaler is not intended for rapid relief of acute symptoms.

[0145] In the embodiments described in this specification, a front-end feedback mechanism is introduced to achieve real-time adjustment of the generation process, enabling real-time adjustment and optimization of the sample question-answer pairs. Using this high-quality sample data, such as the adjusted question-answer pairs, for question-answering model training significantly improves the accuracy and professionalism of the question-answering model in specific domains, enhances the model's performance, and increases training efficiency.

[0146] In one optional embodiment of this specification, the sample parsing result includes sample summary text;

[0147] Step 104 uses a parsing model to split the sample document into multiple sample document fragments, and then parses these fragments to obtain multiple sample parsing results. The specific steps include:

[0148] Using a parsing model, the sample documents are semantically segmented to obtain multiple sample document fragments. Then, the multiple sample document fragments are semantically summarized to obtain multiple sample summary texts.

[0149] The sample summary text corresponding to the sample document fragment is the text content generated by summarizing and generalizing the sample document fragment based on its semantic features. The sample summary text retains the main information of the sample document fragment but removes redundant information, making the text expression more concise and clear, and more clearly indicating which content is suitable as a question and which content is suitable as an answer.

[0150] Semantic decomposition is a natural language parsing method used to divide a document into paragraphs or sentences based on its semantic features. Each segmented paragraph or sentence typically revolves around a main piece of information, which is then used for subsequent natural language processing such as information extraction and question-answer pair generation.

[0151] Semantic summarization is a natural language parsing method used to extract key information from document fragments based on their semantic features and summarize them into a concise text. It preserves the main information of the text content while removing redundant information, making the text more concise and clear.

[0152] For example, using this analytical model, the above-mentioned sample document "Instructions for Use of Drug A Inhaler" was semantically segmented, resulting in 8 sample document fragments:

[0153] Sample document fragment 1: Product name: ..., Ingredients and specifications: ...

[0154] Sample document fragment 2: Indications: ...

[0155] Sample document fragment 3: Usage and dosage: ...

[0156] Sample document fragment 4: Specification clauses: ...

[0157] Sample document fragment 5: Notes: ...

[0158] Sample document fragment 6: Adverse reactions: ...

[0159] Sample document fragment 7: Storage conditions: ...

[0160] Sample document fragment 8: Packaging: ...

[0161] Using this analytical model, semantic summarization was performed on 8 sample document fragments to obtain sample summary text corresponding to the 8 sample document fragments in the sample documents:

[0162] Sample document fragment 1 corresponds to the sample summary text: A Inhalation Aerosol is a combination medication used to treat asthma and chronic obstructive pulmonary disease (COPD). Each spray contains active ingredients A1 and A2. The dosage varies depending on the packaging; please refer to the packaging label for details.

[0163] Sample summary text corresponding to sample document fragment 2: This inhaled aerosol is suitable for long-term asthma control and maintenance treatment of COPD patients in adults and adolescents aged 12 years and older, by reducing airway inflammation and dilating the bronchi to improve respiratory status.

[0164] Sample summary text corresponding to sample document fragment 3: When using, inhale the medication at the prescribed time as directed by your doctor. Shake well and test the device before first use. The specific frequency and dosage will be determined by your doctor based on the patient's condition, usually no more than twice daily, one spray each time.

[0165] Sample document fragment 4 corresponds to the sample summary text: Section 9.1: Patients using inhaled corticosteroids long-term are at risk of fracture. Section 9.2: Compared with inhaled corticosteroids, the use of this product may result in decreased bone mineral density. Section 9.3: This product is not suitable for the treatment of acute asthma attacks or acute exacerbations of COPD.

[0166] Sample summary text corresponding to sample document fragment 5: This product is for prescription use only and should not be purchased or have its dosage altered without medical advice. It is not a rapid-relief medication; if acute respiratory distress occurs, use emergency medications immediately and consult a doctor. Regular follow-up visits are recommended to assess efficacy and safety. This product is suitable for individuals aged 12 years and older; its safety and efficacy in children under 12 years of age have not been established. Contraindicated in individuals with known hypersensitivity to ingredients A1, A2, or any other ingredient.

[0167] Sample summary text corresponding to sample document fragment 6: Common adverse reactions include sore throat, hoarseness, headache, nausea, etc. Serious side effects are rare, but if chest pain, rapid heartbeat, severe allergic reactions, etc. occur, seek medical attention immediately.

[0168] Sample summary text corresponding to sample document fragment 7: This product should be stored at room temperature, avoiding high temperature and direct sunlight. Unused medicines should be properly disposed of in accordance with local regulations.

[0169] Sample document fragment 8 corresponds to the sample summary text: A. Inhaled drug aerosols are usually equipped with a metered inhaler. Each box contains one inhaler with a counter and contains a certain number of doses. Please refer to the instructions on the packaging for details.

[0170] In this embodiment of the specification, semantic segmentation and summarization techniques are used to perform fine-grained analysis of sample documents, generating concise and clear sample summary text. This effectively improves the quality and accuracy of sample question-and-answer pairs, ensuring the complete preservation of the main information of the document and removing redundant content. This makes the sample questions and answers more focused on the main information, further improving the quality of sample question-and-answer pairs.

[0171] In one optional embodiment of this specification, the sample parsing results include multimodal data, wherein the multimodal data includes at least one of tables, images, and structure trees;

[0172] Step 104 involves parsing multiple sample document fragments to obtain multiple sample parsing results, including the following specific steps:

[0173] Multimodal data was extracted from multiple sample document fragments.

[0174] Multimodal data refers to data that contains multiple modal representations, including at least one of tables, images, and structure trees. In the process of generating sample question-and-answer pairs, utilizing multimodal data allows the parsing model to more comprehensively understand the content of the sample document, thereby generating more accurate and richer sample question-and-answer pairs. For example, medical literature may contain information in the form of charts, structured expressions, flowcharts, etc., in addition to textual descriptions; these are all important components for constructing high-quality question-and-answer pairs.

[0175] Tables represent data modally using rows and columns. In natural language processing tasks, tables are often used to display structured numerical or categorical information. Images are graphics composed of pixels and can take the form of photographs, illustrations, charts, etc. In natural language processing tasks, images are often used as supplementary information to help interpret or supplement text content. Structure trees are hierarchical data representation methods that use parent-child relationships between nodes to express the composition and structure of information. In fields such as chemistry and biology, structure trees are often used to depict complex relational networks such as molecular structures and biological classification systems.

[0176] Multimodal data can be extracted from multiple sample document fragments. One possible approach is to extract tables corresponding to multiple sample document fragments from the sample documents in a structured manner. Another possible approach is to extract images corresponding to multiple sample document fragments from the sample documents visually. Yet another possible approach is to extract a structure tree corresponding to multiple sample document fragments from the sample documents in a structured manner. No specific approach is required here.

[0177] For example, using an analytical model, the table of "drug efficacy comparison" corresponding to multiple sample document fragments is extracted from the above-mentioned "Instructions for Use of Drug A Inhaler", the image of "Steps for Using Inhaler" corresponding to multiple sample document fragments is extracted visually, and the tree structure of "Chemical Structure of Components in Drug" corresponding to multiple sample document fragments is extracted structurally.

[0178] In the embodiments described in this specification, by extracting and utilizing multimodal data, the analytical model can more comprehensively understand the content of the sample document and generate more accurate and richer sample question-and-answer pairs.

[0179] In one optional embodiment of this specification, the sample parsing result further includes text modal data; step 108 includes the following specific steps:

[0180] By using a parsing model and multimodal data as augmented knowledge, semantic understanding of textual modal data is performed, generating sample question-answer pairs corresponding to multiple sample parsing results.

[0181] Augmented knowledge is additional information or data introduced during the parsing and generation of sample question-and-answer pairs. It helps the parsing model better understand the text modal data to generate corresponding sample question-and-answer pairs, which helps improve the quality and accuracy of the generated question-and-answer pairs and ensures that they are closer to the needs of real-world application scenarios.

[0182] Semantic understanding is a method of natural language generation used for deep analysis of natural language text, identifying and interpreting its meaning, intent, and contextual relationships to generate more accurate question-answer pairs. Semantic understanding involves multiple levels of understanding, including vocabulary, grammar, and syntax, and can capture deep information from document fragments.

[0183] For example, using this parsing model, the multimodal data corresponding to the eight sample document fragments (a table of “drug efficacy comparison”, an image of “steps for using inhaled aerosol”, and a tree structure of “chemical structure of drug components”) are used as augmented knowledge to perform semantic understanding on the text modal data of the eight sample document fragments, and generate sample question-answer pairs corresponding to the parsing results of the eight samples.

[0184] In the embodiments of this specification, by introducing multimodal data as enhanced knowledge and combining it with deep semantic understanding technology, the parsing model can more accurately grasp the core information and contextual relationships of text modal data, generate high-quality sample question-and-answer pairs that are close to the needs of actual applications, and improve the accuracy and practicality of the question-and-answer pairs.

[0185] In one optional embodiment of this specification, the multimodal data is a table;

[0186] Following step 106, the following specific steps are also included:

[0187] Receive the revised adjustment table from the front end, wherein the adjustment method of the adjustment table includes at least one of the following:

[0188] Merge at least two cells in a table;

[0189] Split at least one cell in the table;

[0190] Add at least one column to the table;

[0191] Add at least one row to the table;

[0192] Delete at least one column from the table;

[0193] Delete at least one row from the table.

[0194] Adjusting a table involves structurally modifying the original table to achieve clearer data presentation, better data comparison, or to meet specific project logic. Adjustments can be simple, such as merging or splitting cells, or complex, such as adding new calculated columns or rows, or even deleting unnecessary data columns or rows. The goal of adjusting a table is to improve data readability, usability, and accuracy.

[0195] For example, a table of "drug efficacy comparison" corresponding to multiple sample document fragments was extracted from the above-mentioned "Instructions for Use of Drug A Inhaler", as shown in Table 1:

[0196]

[0197]

[0198] Table 1

[0199] Send the form to the front end for display, such as Figure 3 As shown, Figure 3 This diagram illustrates the front-end display of a table in a sample data construction method according to an embodiment of this specification:

[0200] The original document, "Instructions for Use of Drug A Inhaler," sent from the front end, is displayed on the left side of the front-end page.

[0201] In addition, the acute exacerbation rates of moderate or severe COPD are shown in the table below. The severity of acute exacerbations of COPD is defined as follows: moderate exacerbations requiring systemic corticosteroids and / or antibiotics for 3 days or longer are defined as acute exacerbations; severe exacerbations leading to hospitalization or death are defined as acute exacerbations.

[0202] Table 1 above.

[0203] The prescribed standard dosage for COPD is given (C drug 160 μg / D drug 4.5 μg, two tablets each time, twice daily).

[0204] The period during and 7 days after an acute exacerbation are not included in the exposure time.

[0205] If a second acute exacerbation occurs more than 7 days after the previous acute exacerbation has subsided, it should be considered as two independent acute exacerbation events.

[0206] Total number of acute exacerbations / total number of years of exposure for all patients receiving the treatment. The period during an acute exacerbation and within 7 days following an acute exacerbation are not included in the exposure time.

[0207] The negative binomial regression model includes the following covariates: percentage of predicted values ​​after inhalation of drug E, cell count, history of COPD exacerbations (0, 1, 2 or more), and whether drug F was used, with exposure time as the offset variable.

[0208] Click on the table in the parsing results to complete the table parsing. The original table is displayed at the top right, and the adjusted table is displayed at the bottom. You can adjust the table by clicking the edit controls, or you can click the original table comparison control to compare the table before and after the adjustment.

[0209] By clicking the edit control, a table is displayed on the front end to adjust the page. Figure 4 This diagram illustrates a front-end display illustration of a table adjustment method for constructing sample data according to an embodiment of this specification:

[0210] The top of the front-end page displays table numbers indicating the following: 1.1, No manual adjustment - Normal; 1.2, No manual adjustment - Hover and select; 1.3, Manually adjusted - Normal; 1.4, Manually adjusted - Hover and select.

[0211] Multiple editing controls, including: 1.1 Merge cells, 1.2 Split cells, 1.3 Add column, 1.4 Delete column, 1.5 Add row, 1.6 Delete row.

[0212] This includes splitting a cell in a table, adding a column or row to the table, and deleting a column or row from the table, such as... Figure 5 As shown, Figure 5 This diagram illustrates a front-end display of a single cell processing method for a sample data construction method provided in one embodiment of this specification.

[0213] After hovering over and selecting the "Product - Total Exposure Time" cell, you can select at least one of the following: 1.2, Split Cell, 1.3, Add Column, 1.4, Delete Column, 1.5, Add Row, 1.6, Delete Row to complete the corresponding editing operation for the cell, column or row.

[0214] This involves merging at least two cells in a table, adding at least two columns or rows to the table, and deleting at least two columns or rows from the table, such as... Figure 6 As shown, Figure 6 This diagram illustrates a front-end display of multiple cell processing in a sample data construction method according to an embodiment of this specification.

[0215] After hovering over and selecting the "Total Exposure Time" cell and the "Number of Patients with Acute Exacerbations" cell, you can select at least one of the following: 1.1 Merge Cells, 1.3 Add Column, 1.4 Delete Column, 1.5 Add Row, 1.6 Delete Row to complete the corresponding editing operation on these two cells, columns, or rows.

[0216] Among them, add columns such as Figure 7 As shown, Figure 7 This diagram illustrates the front-end display of an added column in a sample data construction method according to an embodiment of this specification.

[0217] Add a blank column to the left of "This Product" when hovering over it and selecting it.

[0218] Among them, adding rows is as follows Figure 8 As shown, Figure 8 This diagram illustrates the front-end display of an add row in a sample data construction method according to an embodiment of this specification;

[0219] Add a blank line above the selected "Total Exposure Time" field.

[0220] After editing, you can click Figure 3 The original table comparison control shown in Figure 9 completes the original table comparison. Figure 9 This document illustrates a front-end display diagram showing a comparison of table processing before and after processing in a sample data construction method according to an embodiment of this specification:

[0221] The original table is displayed on the left, and the adjusted chart is displayed on the right.

[0222] In this embodiment of the specification, an intuitive front-end editing method is provided to adjust the sample analysis results of the table. This allows for convenient structural modification of the table data, such as merging or splitting cells, adding or deleting rows and columns, thereby optimizing the data display logic. This not only enhances the user experience but also significantly improves the efficiency and accuracy of data analysis and interpretation, increases the flexibility and response speed of data processing, and ensures the integrity and consistency of the data.

[0223] Currently, the parsing capabilities of multimodal models, such as natural language models, visual models, and structured data models, are used to parse sample documents and automatically generate sample question-answer pairs. This approach involves directly initiating the parsing process after the user uploads the document, and then using the parsed document for model training. However, this method has the following drawbacks:

[0224] 1. The waiting time is unpredictable; users do not know how long it will take to complete the parsing.

[0225] 2. The parsing process is completely black box; users are unaware of the parsing logic and can only see the final result.

[0226] 3. If the training effect is not good using the parsed document, it is impossible to manually intervene in the parsing process to make corrections.

[0227] The following is in conjunction with the appendix Figure 10 Taking the sample data construction method provided in this manual as an example in the application of product Q&A, this paper further explains the sample data construction method in response to the above-mentioned problems. Figure 10 The present specification illustrates a process flowchart of a sample data construction method for product question answering, according to an embodiment of this specification, including the following specific steps:

[0228] Step 1002: Receive the sample product manual sent by the front end.

[0229] Step 1004: Using the parsing model, semantically decompose the sample product manual to obtain multiple sample product manual fragments, and semantically summarize the multiple sample product manual fragments to obtain the sample summary text corresponding to the multiple sample product manual fragments in the sample product manual. Extract the multimodal data corresponding to the multiple sample product manual fragments from the sample product manual, wherein the multimodal data includes at least one of tables, images and structure trees.

[0230] Step 1006: Send the sample summary text and multimodal data to the front end for display.

[0231] Step 1008: Receive the parsing adjustment instructions and the adjusted table from the front end. Using the parsing model, parse multiple sample document fragments according to the parsing adjustment instructions to obtain the adjusted sample summary text. The adjustment method of the table includes at least one of the following: merging at least two cells of the table; splitting at least one cell of the table; adding at least one column to the table; adding at least one row to the table; deleting at least one column to the table; deleting at least one row to the table.

[0232] Step 1010: Send the adjusted sample summary text and multimodal data to the front end for display.

[0233] Step 1012: Receive confirmation instructions from the front end; using multimodal data corresponding to multiple sample product manual fragments as augmented knowledge, perform semantic understanding on the sample summary text corresponding to the multiple sample product manual fragments, and generate sample question-answer pairs corresponding to the multiple sample parsing results.

[0234] Step 1014: Send the sample summary text, multimodal data, and sample question-and-answer pairs corresponding to multiple sample parsing results to the front end for display.

[0235] Step 1016: Receive the adjusted sample question-and-answer pairs from the front end.

[0236] Step 1018: Train an online question-answering model for providing product technical support by adjusting sample question-answer pairs.

[0237] This embodiment of the specification achieves transparency and user control over the document parsing process. It receives product manuals uploaded by users, performs semantic segmentation and summarization using a parsing model, and extracts multimodal data as augmented knowledge to generate sample question-and-answer pairs. This not only accelerates the parsing speed but also makes the parsing logic clearly visible, overcoming the problems of uncontrollable parsing time and black-box operation in traditional solutions. Furthermore, it allows users to directly provide feedback on adjusted parsing results and question-and-answer pairs on the front end, supporting flexible editing of tables and other data, ensuring parsing quality. Finally, by using the adjusted samples to train an online question-and-answer model, the model's accuracy and applicability are improved, while also enhancing user engagement and trust, ensuring a significant improvement in model training effectiveness. The entire process is visualized and interactive, greatly improving work efficiency and user experience.

[0238] Corresponding to the above method embodiments, this specification also provides task platform embodiments. Figure 11 A schematic diagram of the structure of a task platform provided in one embodiment of this specification is shown. Figure 11 As shown, the task platform 1100 includes a sample data construction interface 1102 and a response unit 1104;

[0239] The sample data construction interface 1102 is used to receive sample documents sent from the front end;

[0240] The response unit 1104 is used to split the sample document to obtain multiple sample document fragments, parse the multiple sample document fragments, and obtain multiple sample parsing results;

[0241] The sample data construction interface 1102 is also used to send the parsing results of multiple samples to the front end for display and to receive confirmation instructions from the front end.

[0242] The response unit 1104 is also used to generate sample question-answer pairs corresponding to multiple sample parsing results based on the parsing results of multiple samples using the parsing model.

[0243] Optionally, the sample data construction interface 1102 is also used to receive splitting and adjustment instructions from the front end;

[0244] The response unit 1104 is also used to use the parsing model to split the sample document according to the splitting and adjustment instructions, obtain multiple adjusted sample document fragments of the sample document, parse the multiple adjusted sample document fragments, and obtain multiple adjusted sample parsing results;

[0245] The sample data construction interface 1102 is also used to send the parsing results of multiple adjusted samples to the front end for display.

[0246] Optionally, the sample data construction interface 1102 is also used to receive parsing and adjustment instructions from the front end;

[0247] The response unit 1104 is also used to use the parsing model to parse multiple sample document fragments according to the parsing adjustment instructions, and obtain multiple adjusted sample parsing results;

[0248] The sample data construction interface 1102 is also used to send the parsing results of multiple adjusted samples to the front end for display.

[0249] Optionally, the sample data construction interface 1102 is also used to receive generation and adjustment instructions from the front end;

[0250] The response unit 1104 is also used to generate, according to the generation adjustment instruction, multiple sample parsing results corresponding to the multiple sample parsing results using the parsing model, and to train the question answering model using the adjusted sample question answering pairs.

[0251] And / or,

[0252] The sample data construction interface 1102 is also used to receive the adjusted sample question-answer pairs corresponding to multiple sample parsing results fed back from the front end;

[0253] The response unit 1104 is also used to train the question-answering model by adjusting the sample question-answer pairs.

[0254] Optionally, the sample analysis results include multimodal data, which is presented in tabular form.

[0255] The response unit 1104 is specifically used to extract tables corresponding to multiple sample document fragments from the sample document using a parsing model.

[0256] The sample data construction interface 1102 is also used to receive the adjusted table from the front end. The adjustment method of the table includes at least one of the following: merging at least two cells of the table; splitting at least one cell of the table; adding at least one column to the table; adding at least one row to the table; deleting at least one column to the table; deleting at least one row to the table.

[0257] In the embodiments of this specification, the response unit separates the parsing process and the generation process when using the parsing model to parse and generate sample documents, obtaining sample parsing results corresponding to multiple sample document fragments. These results are sent to the front end for display through the sample data construction interface. Only after receiving a confirmation instruction from the front end does the response unit generate sample question-and-answer pairs corresponding to the multiple sample parsing results. This allows the front end to verify whether the parsing model parses and generates sample question-and-answer pairs with a full understanding of the document content. This overcomes the problem of the black box nature of the model parsing and generation process and ensures the quality of the sample question-and-answer pairs.

[0258] 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 sample data construction 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 sample data construction method described above.

[0259] Figure 12 A structural block diagram of a computing device according to one embodiment of this specification is shown. The 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.

[0260] 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 Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or a combination of communication networks such as the Internet. The 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 an IEEE 802.11 Wireless Local Area Network (WLAN) interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or Near Field Communication (NFC).

[0261] In one embodiment of this specification, the aforementioned components of the computing device 1200 and Figure 12 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 12 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.

[0262] 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.

[0263] The processor 1220 is used to execute the following computer program / instructions, which, when executed by the processor, implement the steps of the above-described sample data construction method.

[0264] 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 and the technical solution of the sample data construction method described above belong to the same concept. 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 sample data construction method described above.

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

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

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

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

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

[0270] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash 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 the computer-readable medium may be appropriately added or removed according to the requirements of patent practice. For example, in some regions, according to patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

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

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

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

Claims

1. A method for constructing sample data, comprising: Obtain sample documents; Using a parsing model, the sample document is split into multiple sample document fragments, and these fragments are parsed to obtain multiple sample parsing results. The parsing results of the multiple samples are sent to the front end for display. Upon receiving the confirmation instruction from the front end, the parsing model is used to generate sample question-answer pairs corresponding to the multiple sample parsing results based on the parsing results of the multiple samples.

2. The method according to claim 1, further comprising, after sending the plurality of sample parsing results to the front end for display: Upon receiving the splitting and adjustment instruction from the front end, the parsing model is used to split the sample document according to the splitting and adjustment instruction to obtain multiple adjusted sample document fragments of the sample document. The multiple adjusted sample document fragments are then parsed to obtain multiple adjusted sample parsing results. The analysis results of the multiple adjusted samples are sent to the front end for display.

3. The method according to claim 1, further comprising, after sending the plurality of sample parsing results to the front end for display: Upon receiving the parsing adjustment instruction from the front end, the parsing model is used to parse the multiple sample document fragments according to the parsing adjustment instruction to obtain multiple adjusted sample parsing results; The analysis results of the multiple adjusted samples are sent to the front end for display.

4. The method according to claim 1, after generating sample question-answer pairs corresponding to the multiple sample parsing results based on the multiple sample parsing results using the parsing model upon receiving the confirmation instruction from the front end, further includes: The sample question-and-answer pairs are sent to the front end for display. Upon receiving the generation and adjustment instruction from the front end, the parsing model is used to generate adjusted sample question-answer pairs corresponding to the multiple sample parsing results based on the multiple sample parsing results, according to the generation and adjustment instruction, and the question-answering model is trained using the adjusted sample question-answer pairs; and / or; The sample question-and-answer pairs are sent to the front end for display. Upon receiving the adjusted sample question-answer pairs corresponding to the multiple sample parsing results fed back from the front end, the question-answering model is trained using the adjusted sample question-answer pairs.

5. The method according to claim 1, wherein the sample parsing result includes sample summary text; The process involves using a parsing model to split the sample document into multiple sample document fragments, parsing these fragments to obtain multiple sample parsing results, including: Using a parsing model, the sample documents are semantically segmented to obtain multiple sample document fragments, and the multiple sample document fragments are semantically summarized to obtain multiple sample summary texts.

6. The method according to any one of claims 1-5, wherein the sample parsing result includes multimodal data, wherein, The multimodal data includes at least one of the following: tables, images, and structure trees; The process of parsing the multiple sample document fragments to obtain multiple sample parsing results includes: Multimodal data is extracted from the multiple sample document fragments.

7. The method according to claim 6, wherein the sample parsing result further includes text modal data; The step of using an analytical model to generate sample question-answer pairs corresponding to the multiple sample analysis results includes: Using the aforementioned parsing model and the multimodal data as augmented knowledge, semantic understanding is performed on the text modal data to generate sample question-answer pairs corresponding to the multiple sample parsing results.

8. The method according to claim 6, wherein the multimodal data is a table; After sending the parsing results of the multiple samples to the front end for display, the method further includes: The system receives the adjusted table from the front-end feedback, wherein the adjustment method of the table includes at least one of the following: Merge at least two cells in the table; Split at least one cell of the table; Add at least one column to the table; Add at least one row to the table; Delete at least one column from the table; Delete at least one row from the table.

9. A task platform, comprising a sample data construction interface and a response unit; The sample data construction interface is used to receive sample documents sent from the front end; The response unit is used to split the sample document to obtain multiple sample document fragments, and to parse the multiple sample document fragments to obtain multiple sample parsing results; The sample data construction interface is also used to send the parsing results of the multiple samples to the front end for display and to receive the confirmation instruction from the front end. The response unit is further configured to use the parsing model to generate sample question-answer pairs corresponding to the multiple sample parsing results based on the multiple sample parsing results.

10. The task platform according to claim 9, wherein the sample data construction interface is further configured to receive the splitting and adjustment instructions fed back by the front end; The response unit is further configured to use the parsing model to split the sample document according to the splitting and adjustment instructions, obtain multiple adjusted sample document fragments of the sample document, parse the multiple adjusted sample document fragments, and obtain multiple adjusted sample parsing results; The sample data construction interface is also used to send the parsing results of the multiple adjusted samples to the front end for display.

11. The task platform according to claim 9, wherein the sample data construction interface is further configured to receive the parsing and adjustment instructions fed back by the front end; The response unit is further configured to use the parsing model to parse the plurality of sample document fragments according to the parsing adjustment instructions, and obtain the parsing results of the plurality of adjusted samples; The sample data construction interface is also used to send the parsing results of the multiple adjusted samples to the front end for display.

12. The task platform according to claim 9, wherein the sample data construction interface is further configured to receive the generation and adjustment instructions fed back by the front end; The response unit is further configured to use the parsing model, according to the generation adjustment instruction, to generate adjusted sample question-answer pairs corresponding to the multiple sample parsing results based on the multiple sample parsing results, and to train the question-answering model using the adjusted sample question-answer pairs; And / or, The sample data construction interface is also used to receive the adjusted sample question-and-answer pairs corresponding to the multiple sample parsing results fed back by the front end; The response unit is also used to train a question-answering model using the adjusted sample question-answer pairs.

13. The task platform according to claim 9, wherein the sample parsing results include multimodal data, and the multimodal data is a table; The response unit is specifically used to extract a table corresponding to multiple sample document fragments from the sample document using a parsing model. The sample data construction interface is also used to receive the adjusted table from the front end, wherein... The adjustment methods for the adjustment table include at least one of the following: Merge at least two cells in the table; Split at least one cell of the table; Add at least one column to the table; Add at least one row to the table; Delete at least one column from the table; Delete at least one row from the table.

14. A computing device, comprising: Memory and processor; The memory is used to store computer programs / instructions, and the processor is used to execute the computer programs / instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1 to 8.

15. 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 8.

16. 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 8.

Images