A method, system, terminal and storage medium for adaptive enhancement of question and answer data based on a computing power platform

Abstract

The application discloses a kind of based on computing power platform's question and answer data self-adapting enhancement method, system, terminal and storage medium, the method includes: obtaining the original question and answer data of computing power platform operation and maintenance scene, the original question and answer data is cleaned and structured, obtains multiple structured dictionary format, according to all structured dictionary format establishes basic index library;Based on basic index library, the question text in structured dictionary format is extracted to multidimensional feature, obtains routing instruction, according to routing instruction in multiple operator combination Matching target operator combination, dynamically adjust the weight of target operator combination, obtain target weight;Version data and linkage instruction are obtained, according to linkage instruction, version data and target weight question text is carried out question and answer semantic synchronization and question and answer semantic quality control, obtain high-quality answer data.The application is according to the dynamic adaptation of user intention and the automatic balance of sample distribution, and then accurate answer is obtained.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a question-and-answer data adaptive enhancement method, system, terminal, and computer-readable storage medium based on a computing power platform. Background Technology

[0002] With the deep application of cloud computing and intelligent computing centers in scientific research and enterprise scenarios, automated operation and maintenance based on question-answering systems has become an important means to reduce operation and maintenance costs and improve response efficiency. Unlike open question-answering in general fields, the operation and maintenance scenario of computing power platforms has extremely high "instruction rigidity" and "environment sensitivity".

[0003] Existing question-answering data augmentation technologies mainly focus on two aspects: "generative augmentation based on large language models" and "linear augmentation based on basic NLP (Natural Language Processing) rules." However, in the specific scenario of intelligent computing operations and maintenance, both have inherent defects that cannot be solved by simple combinations of technologies. These defects are primarily manifested in the "probabilistic uncertainty" of large language models (LLMs): LLMs are essentially generative models based on probabilistic prediction. Without strong constraints, they inevitably generate incorrect information, such as fabricating non-existent parameters or incorrect combinations of operational logic. This "black box" uncontrollability is unacceptable in the field of operations and maintenance, which seeks determinism, and the high cost of model fine-tuning makes it difficult to adapt to the high-frequency dynamic evolution of platform versions.

[0004] The "linear stacking" and "static rigidity" defects of traditional NLP methods: Existing non-generative augmentation schemes are usually just simple linear concatenations of regular expressions, thesaurus, and similarity calculations. This mechanical combination has obvious shortcomings: Lack of state-aware static execution: Traditional methods use a "one-size-fits-all" pipeline processing for all input data, failing to perceive subtle differences in user intent (e.g., expert vs. novice), leading to a mismatch between augmentation strategies and data features; Lack of open environment with feedback mechanisms: Traditional schemes are "open-loop," unable to perceive the distribution of generated data. As the number of augmentations increases, the sample distribution may skew towards a certain type of sentence structure, and the system cannot automatically correct this; Lack of structured coupling for single-point modifications: Traditional regular expression replacements usually operate in isolation on the "question" or "answer," lacking a structured understanding of the question-answer pair as a whole logical unit, resulting in semantic disconnect where "the question is changed, but the answer is not synchronized."

[0005] Existing question-answering data augmentation methods include generative augmentation methods based on large language models and linear augmentation methods based on basic NLP rules. These traditional methods are limited by the logical uncertainty caused by generation errors in LLM and the "static rigidity" and "open-loop" defects of traditional NLP combination schemes. They cannot effectively support the high-precision and diverse sentence generation in intelligent computing operation and maintenance scenarios. As a result, the use of existing question-answering data augmentation methods cannot solve the risk of term generation errors, and it is difficult to achieve dynamic adaptation based on user intent and automatic balance of sample distribution to obtain accurate answers. Therefore, existing adaptive question-answering data augmentation technologies based on computing power platforms still need to be improved and optimized. Summary of the Invention

[0006] The main objective of this invention is to provide a question-and-answer data adaptive enhancement method, system, terminal, and computer-readable storage medium based on a computing power platform. This invention aims to address the problems that existing question-and-answer data enhancement methods cannot solve the risk of generating incorrect terms and are difficult to achieve dynamic adaptation based on user intent and automatic balancing of sample distribution to obtain accurate answers.

[0007] To achieve the above objectives, the present invention provides a question-answering data adaptive enhancement method based on a computing power platform, the method comprising the following steps: Obtain raw question and answer data from the operation and maintenance scenario of the computing power platform, clean and structure the raw question and answer data to obtain multiple structured dictionary formats, and establish a basic index library based on all the structured dictionary formats; Based on the aforementioned basic index library, multi-dimensional feature extraction is performed on the question text in the structured dictionary format to obtain routing instructions. According to the routing instructions, a target operator combination is matched among multiple operator combinations, and the weights of the target operator combination are dynamically adjusted to obtain the target weights. Obtain version data and linkage instructions, and perform question-and-answer semantic synchronization and question-and-answer semantic quality control on the question text based on the linkage instructions, version data, and target weight to obtain answer data.

[0008] Optionally, the aforementioned question-answering data adaptive enhancement method based on a computing power platform, wherein obtaining the raw question-answering data from the computing power platform's operation and maintenance scenario, cleaning and structuring the raw question-answering data, and establishing a basic index library specifically includes: The raw question and answer data of the computing power platform operation and maintenance scenario is obtained, the raw question and answer data is cleaned to obtain effective data, and the effective data is structured to obtain multiple structured dictionary formats. Input all the structured dictionary formats into the original database to obtain the basic index library.

[0009] Optionally, the aforementioned question-answering data adaptive enhancement method based on a computing power platform, wherein the steps of acquiring raw question-answering data from the computing power platform's operation and maintenance scenario, cleaning and structuring the raw question-answering data, establishing a basic index library, and then further including: Obtain the synonym replacement weight, colloquial expansion weight, and sentence restructuring weight, and configure the synonym replacement weight, colloquial expansion weight, and sentence restructuring weight to obtain a dynamic sampling probability matrix; The words in the original question-and-answer data are mapped according to the dynamic sampling probability matrix to obtain the corrected words.

[0010] Optionally, the question-answering data adaptive enhancement method based on computing power platform includes a combination of operators such as synonym replacement operator, back translation operator, stop word insertion operator, entity recognition operator, knowledge graph reasoning operator, version term mapping operator, sentence inversion operator, and modal word addition operator. The target operator combination includes any one or more of the first operator combination, the second operator combination, and the third operator combination; The first operator combination includes a synonym substitution operator, a back-translation operator, and a stop word insertion operator; The second operator combination includes entity recognition operators, knowledge graph reasoning operators, and version term mapping operators; The third operator combination includes a sentence inversion operator and a modal particle addition operator.

[0011] Optionally, the question-answering data adaptive enhancement method based on the computing power platform, wherein the step of extracting multi-dimensional features from the question text in the structured dictionary format based on the basic index library to obtain routing instructions, matching a target operator combination among multiple operator combinations according to the routing instructions, and dynamically adjusting the weights of the target operator combination to obtain target weights, specifically includes: Based on the aforementioned basic index library, multidimensional feature extraction is performed on the question text in the structured dictionary format to obtain the confidence score; Acquire control signal packets, and perform intent recognition based on the confidence level according to preset policy operators and the control signal packets to obtain routing instructions; According to the routing instructions, match any one or more of the first operator combination, the second operator combination, and the third operator combination in a plurality of operator combinations; If the deviation of any one or more of the first operator combination, the second operator combination, and the third operator combination is greater than a preset value, then the weights of any one or more of the first operator combination, the second operator combination, and the third operator combination are dynamically adjusted through a negative feedback mechanism to obtain the target weight.

[0012] Optionally, the question-answering data adaptive enhancement method based on a computing power platform, wherein the step of obtaining the control signal packet and performing intent recognition on the confidence level according to a preset strategy operator and the control signal packet to obtain a routing instruction specifically includes: Acquire control signal packets, and perform intent recognition on the confidence level based on preset strategy operators and the control signal packets to obtain the intent category; The question text is tagged according to the intent category to obtain routing instructions; The intent categories include novice intents and expert intents.

[0013] Optionally, the aforementioned question-answering data adaptive enhancement method based on a computing power platform, wherein obtaining version data and linkage instructions, and performing question-answering semantic synchronization and question-answering semantic quality control on the question text according to the linkage instructions, the version data, and the target weight to obtain answer data, specifically includes: Obtain version data and linkage instructions; based on semantic anchoring technology, perform question-and-answer semantic synchronization on the question text according to the linkage instructions, version data, and target weight to obtain an initial answer; Obtain the summary extraction rules, and adjust the initial answer according to the summary extraction rules to obtain the target answer; Based on the cosine similarity algorithm, semantic deduplication is performed on the target answer according to the strategies of terminology accuracy, sentence structure rationality, and intent adaptation to obtain the answer data.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a question-answering data adaptive enhancement system based on a computing power platform, wherein the question-answering data adaptive enhancement system based on a computing power platform includes: The question data structuring module is used to obtain raw question and answer data from the operation and maintenance scenario of the computing power platform, clean and structure the raw question and answer data to obtain multiple structured dictionary formats, and establish a basic index library based on all the structured dictionary formats. The question weight adjustment module is used to perform multi-dimensional feature extraction on the question text in the structured dictionary format based on the basic index library to obtain routing instructions, match the target operator combination among multiple operator combinations according to the routing instructions, and dynamically adjust the weight of the target operator combination to obtain the target weight. The answer generation module is used to acquire version data and linkage instructions, and to perform question-and-answer semantic synchronization and question-and-answer semantic quality control on the question text based on the linkage instructions, the version data, and the target weight, so as to obtain answer data.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a question-answering data adaptive enhancement program based on a computing power platform, and when the question-answering data adaptive enhancement program based on the computing power platform is executed by a processor, it implements the steps of the question-answering data adaptive enhancement method based on the computing power platform as described above.

[0016] In this invention, raw question-and-answer data from a computing platform operation and maintenance scenario is acquired. This raw data is then cleaned and structured to obtain multiple structured dictionary formats. A basic index library is established based on all these formats. Based on this index library, multi-dimensional features are extracted from the question text within the structured dictionary formats to obtain routing instructions. Target operator combinations are matched among multiple operator combinations based on these routing instructions. The weights of the target operator combinations are dynamically adjusted to obtain target weights. Version data and linkage instructions are acquired. Based on these linkage instructions, version data, and target weights, question-and-answer semantic synchronization and quality control are performed on the question text to obtain answer data. This invention, based on the cleaning and structuring of computing platform question-and-answer data, achieves accurate question-and-answer semantic synchronization and quality control through multi-dimensional feature extraction and dynamic weight adjustment, thereby obtaining accurate answers. Attached Figure Description

[0017] Figure 1 This is a flowchart of a preferred embodiment of the question-answering data adaptive enhancement method based on a computing power platform according to the present invention; Figure 2 This is a flowchart illustrating a preferred embodiment of the question-answering data adaptive enhancement method based on a computing power platform according to the present invention. Figure 3 This is a flowchart of feedback adjustment in a preferred embodiment of the question-answering data adaptive enhancement method based on a computing power platform according to the present invention; Figure 4 This is a structural diagram of a preferred embodiment of the question-answering data adaptive enhancement system based on a computing power platform according to the present invention; Figure 5 This is a structural diagram of a preferred embodiment of the terminal of the device of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0019] Traditional enhancement methods, for example, address the following question: Original problem: Using two machines, each configured with only two accelerator cards, running a parallel test program results in test failure. Enhanced problem: In the current intelligent computing node environment, if two computing servers are configured, and each only has two accelerator cards enabled to run a parallel verification program, how to handle the task interruption error? Answer: Current parallel architecture design requirements only support single-machine parallel processing; multi-machine processing is not supported. That is, 2 machines with 4 accelerator cards can be implemented with a single machine with 4 accelerator cards, but not with dual machines with 4 accelerator cards. Solution / workaround: Configure the cards according to the requirements. However, after the question enhancement, the answer cannot be synchronized in a timely manner, resulting in semantic disconnect: the answer is too brief and cannot match the professional context of complex questions.

[0020] Existing question-answering data augmentation methods include generative augmentation methods based on large language models and linear augmentation methods based on basic NLP rules. These traditional methods are limited by the logical uncertainty caused by the generation errors of LLM, and the "static rigidity" and "open-loop" defects of traditional NLP combination schemes. They cannot effectively support the high-precision and diverse sentence generation in intelligent computing operation and maintenance scenarios. Therefore, there is a need for a question-answering data adaptive augmentation method based on computing power platforms. This method should achieve accurate synchronization and quality control of question-answering semantics through multi-dimensional feature extraction and dynamic weight adjustment, avoiding the problem of not being able to dynamically adapt to user intent and automatically balance sample distribution, thus failing to obtain accurate answers.

[0021] The preferred embodiment of the question-answering data adaptive enhancement method based on a computing power platform described in this invention, such as... Figure 1 and Figure 2 As shown, the question-answering data adaptive enhancement method based on the computing power platform includes the following steps: Step S10: Obtain the original question and answer data of the computing power platform operation and maintenance scenario, clean and structure the original question and answer data to obtain multiple structured dictionary formats, and establish a basic index library based on all the structured dictionary formats.

[0022] Step S10 includes: Step S11: Obtain the original question and answer data of the computing power platform operation and maintenance scenario, clean the original question and answer data to obtain effective data, and perform structured processing on the effective data to obtain multiple structured dictionary formats; Step S12: Input all the structured dictionary formats into the original database to obtain the basic index library.

[0023] Specifically, after step S10, the method further includes obtaining synonym replacement weights, colloquial expansion weights, and sentence restructuring weights; configuring the synonym replacement weights, colloquial expansion weights, and sentence restructuring weights to obtain a dynamic sampling probability matrix; mapping the words of the original question-and-answer data according to the dynamic sampling probability matrix to obtain corrected words (the system automatically extracts terminology associations based on the computing power platform technical manual and version update logs to form a dynamic sampling probability matrix (dynamic sampling probability matrix: a two-dimensional matrix structure, rows represent terminology types (core / association), columns represent operation types (retention / synonym replacement / cross-version mapping), and values ​​represent sampling probabilities); performing deterministic discrete sampling based on the probability matrix on the original question, with each operation only performing precise mapping for a specific slot); obtaining the original question-and-answer data of the computing power platform operation and maintenance scenario; cleaning the original question-and-answer data to obtain valid data (valid data at least contains non-empty quadruples:<Query,Answer,Version,Environment> The valid data is processed in a structured manner to obtain multiple structured dictionary formats (the valid data is converted into structured dictionary formats). All the structured dictionary formats are then input into the original database to obtain the basic index library.

[0024] In this embodiment, the multidimensional feature cleaning and structured coupling module obtains the original Q&A data from the intelligent computing center customer service platform. The original Q&A data includes fields such as "question", "solution" and "computing platform version". Invalid data with empty "question" or "solution" fields are removed, and the valid data is converted into a structured dictionary format to establish a basic index library.

[0025] For example, the weight for synonym replacement is 0.6, meaning there is a 60% probability of replacing "view" with "query," "check," or "get," etc.; the weight for colloquial expansion is 0.4, meaning there is a 40% probability of adding colloquial conjunctions such as "please ask," "hello teacher," or "how to do it" at the beginning / end of a sentence; and the weight for sentence restructuring is 0.3, meaning there is a 30% probability of changing a declarative sentence into an inverted sentence or something else.

[0026] Specifically, the dynamic sampling probability matrix is ​​shown in Table 1: Table 1: Dynamic Sampling Probability Matrix

[0027] Step S20: Based on the basic index library, perform multi-dimensional feature extraction on the question text in the structured dictionary format to obtain routing instructions. Match the target operator combination among multiple operator combinations according to the routing instructions, and dynamically adjust the weight of the target operator combination to obtain the target weight.

[0028] Step S20 includes: Step S21: Based on the basic index library, perform multi-dimensional feature extraction on the question text in the structured dictionary format to obtain the confidence score; Step S22: Obtain control signal packets, and perform intent recognition on the confidence level according to the preset strategy operator and the control signal packets to obtain routing instructions; Step S23: Match any one or more of the first operator combination, the second operator combination, and the third operator combination in the multiple operator combinations according to the routing instruction; Step S24: If the deviation of any one or more of the first operator combination, the second operator combination, and the third operator combination is greater than a preset value, then the weights of any one or more of the first operator combination, the second operator combination, and the third operator combination are dynamically adjusted through a negative feedback mechanism to obtain the target weight.

[0029] Specifically, such as Figure 3As shown, based on the basic index library, multi-dimensional feature extraction is performed on the question text in the structured dictionary format to obtain the confidence score (assuming an intent recognition rule library containing multi-dimensional feature slots is used, multi-dimensional feature extraction is performed on the input question text, and its confidence score in dimensions such as "emergency help, routine consultation, and beginner guidance" is calculated. Confidence score: the probability value output by the intent perception module that the input question belongs to a specific intent category (such as emergency help, routine consultation, and beginner question), with a range of [0,1]). Control signal packets are obtained, and intent recognition is performed on the confidence score according to the preset strategy operators and the control signal packets to obtain routing instructions (and the question is tagged as a routing instruction for subsequent enhancement strategies. Routing instructions: control signal packets generated by the system to control which preset strategy operators and their parameters are enabled by the enhancement engine). According to the routing instructions, the first operator combination, the second operator combination, and the third operator combination are matched in multiple operator combinations. If the deviation of any one or more of the first, second, and third operator combinations is greater than a preset value, then the weights of any one or more of the first, second, and third operator combinations are dynamically adjusted through a negative feedback mechanism to obtain the target weight. The operator combinations include synonym substitution operators, back-translation operators, stop word insertion operators, entity recognition operators, knowledge graph reasoning operators, version terminology mapping operators, sentence inversion operators, and modifier addition operators. The target operator combination includes any one or more of the first, second, and third operator combinations. The first operator combination includes synonym substitution operators, back-translation operators, and stop word insertion operators. The second operator combination includes entity recognition operators, knowledge graph reasoning operators, and version terminology mapping operators. The third operator combination includes sentence inversion operators and modifier addition operators.

[0030] In this embodiment, the confidence level of the urgent request for help, "The server is down, restarting is ineffective, urgent!", is 0.92 (triggering the emergency strategy), the confidence level of a novice's question is 0.65, and the confidence level of a regular consultation is 0.10. The adaptive enhancement engine matches the corresponding operator combination according to the routing instructions. At the same time, the distribution monitor records the strategy distribution of the current generated data in real time. When the output ratio of a certain strategy, such as "sentence rewriting" (the weight of any one or more of the first operator combination, the second operator combination, and the third operator combination) deviates from the preset threshold (e.g., deviation > 5%), the negative feedback adjustment mechanism is triggered to automatically reduce the probability of the strategy being selected, ensuring that the distribution entropy of the generated corpus is maintained in a balanced state. Here, the balanced state means that the statistical distribution characteristics of various intentions, sentence lengths, or versions of data in the generated corpus meet the preset mathematical standards (usually measured by information entropy or variance).

[0031] As an example, the intention distribution is balanced: the deviation of emergency help (20%), routine consultation (50%), and novice questions (30%) does not exceed 5%. The length distribution is balanced: short texts (<10 characters) account for 30%, and medium and long texts account for 70%. The strategy coverage is balanced: there is no situation where a single strategy (such as only performing synonym replacement) accounts for more than 80% (i.e., the strategy distribution entropy > threshold).

[0032] Step S22 includes: Step S221: Obtain control signal packets, and perform intent recognition on the confidence level according to the preset strategy operator and the control signal packets to obtain the intent category; Step S222: Tag the question text according to the intent category to obtain routing instructions.

[0033] Specifically, a control signal packet is acquired, and the confidence level is subjected to intent recognition based on the preset strategy operator and the control signal packet to obtain the intent category. The question text is then tagged according to the intent category to obtain routing instructions (the question text is tagged as routing instructions for subsequent enhancement strategies). The intent categories include novice intent and expert intent (Instruction A (for novice intent): Enable colloquial expansion and concept simplification operators, use a friendly tone, and generate long text; Instruction B (for expert / emergency intent): Enable terminology accuracy verification and syntactic compression operators, use a professional tone, and generate short text).

[0034] Furthermore, a first operator combination is obtained by performing generalization enhancement operations through the synonym replacement operator, the back-translation operator, and the stop word insertion operator; a second operator combination is obtained by performing professional enhancement operations through the entity recognition operator, the knowledge graph reasoning operator, and the version term mapping operator; and a third operator combination is obtained by performing style transfer operations through the sentence inversion operator and the modal particle addition operator.

[0035] Step S30: Obtain version data and linkage instructions; perform question-and-answer semantic synchronization and question-and-answer semantic quality control on the question text based on the linkage instructions, the version data, and the target weight to obtain answer data.

[0036] Step S30 includes: Step S31: Obtain version data and linkage instructions. Based on semantic anchoring technology, perform question-and-answer semantic synchronization on the question text according to the linkage instructions, the version data, and the target weight to obtain the initial answer. Step S32: Obtain the summary extraction rules, and adjust the initial answer according to the summary extraction rules to obtain the target answer; Step S33: Based on the cosine similarity algorithm, semantic deduplication is performed on the target answer according to the terminology accuracy strategy, sentence structure rationality strategy, and intent adaptation strategy to obtain the answer data.

[0037] Specifically, version data (version-specific data: applicable only to a specific version, containing question and answer data with features unique to that version or obsolete features, not transferable to other versions; the cross-version adaptation module provides a dual-path parallel generation pipeline based on a version difference table: path A generates "version-specific data" based on the current version mapping table; path B generates "cross-version migration data" based on the cross-version evolution relationship, and automatically adds the "adapted version" metadata tag) and linkage instructions are obtained. Based on semantic anchoring technology, according to the linkage instructions (linkage instruction: target_answer_update: `true` (indicating that the answer end must synchronously find and replace relevant descriptions) and the version data and target weight are used to perform question-and-answer semantic synchronization on the question text to obtain the initial answer (when processing the "solution", the corresponding terms are deterministically updated based on the anchor information. At the same time, based on the length gradient after the question is rewritten, the summary extraction rules are adaptively called to adjust the level of detail of the solution, where anchor information: metadata recorded when the "question" is modified, used to lock the logical association between the question and the answer), obtain the summary extraction rules (summary extraction rules: the specific logic for dynamically adjusting the level of detail of the answer based on the change in question length), and adjust the initial answer according to the summary extraction rules to obtain the target answer. The system identifies high-quality answer data (question-answer pairs whose total score exceeds a preset threshold (e.g., 0.85) by using a cosine similarity algorithm and employing strategies for terminology accuracy, sentence structure rationality, and intent fit. Semantic deduplication involves removing duplicates between newly generated augmented data and existing original / historical data, as well as within the newly generated data itself. For example, if 100 variations of "how to find an IP address" are generated, but 50 of them are identical (e.g., "how to find an IP address"), the system calculates the vector cosine similarity. If the similarity is greater than 0.98, it is considered duplicate data, and only the highest-scoring answer is retained.

[0038] As an example, the original question was: Using two machines, each configured with only two cards, to run a parallel test program, the test failed. The original answer was: The current parallel architecture design requires that parallel requirements that can be run on a single machine are not supported by multiple machines; that is, 2 machines with 4 cards can be implemented with a single machine with 4 cards, but it is not supported to run with two machines with 4 cards. The solution / workaround was: Configure the cards according to the requirements. The enhanced new question is: In the current intelligent computing node environment, if two computing servers are configured and each only has two acceleration cards running a parallel verification program, a task interruption error occurs. How to handle this? The enhanced new answer is: Due to the limitations of the current intelligent computing architecture design, any parallel requirements that can be met by single-machine resources do not support cross-machine scheduling. The "2 machines with 4 cards" mode you mentioned should be optimized to "single machine with 4 cards" to ensure topology efficiency. The workaround is: Please reset the card quota to 4 cards per node in the resource scheduling backend. The calculation formula is as follows: ; in, Indicates an answer. Represents a function, To indicate a problem, Indicates length and term weight.

[0039] In this embodiment, the summary extraction rules are as follows: First rule (minimalist mode): If the question length is <8 characters (e.g., "check memory"), the answer only retains the code block or core instruction line, and removes all explanatory text. Second rule (summary mode): If the question is identified as an intent to "quickly preview", the first sentence of the answer and the sentence containing numbers / parameters are extracted, and the remaining paragraphs are folded. Third rule (full mode): If the question length is >20 characters and contains keywords such as "detailed steps", the full text of the answer is retained.

[0040] As an example, terminology accuracy (1.0): key parameters (such as cuda_11.0) have no character errors (this is a hard constraint), sentence structure rationality (>0.8): conforms to Chinese grammatical habits, with no obvious traces of machine translation, and intent fit (>0.9): when the question is in a "beginner" style, the answer cannot be entirely obscure and difficult-to-understand professional descriptions; when the question is in an "expert" style, the answer cannot be entirely plain language / casual talk.

[0041] Furthermore, such as Figure 4 As shown, based on the above-described adaptive enhancement method for question-answering data based on a computing power platform, the present invention also provides a corresponding adaptive enhancement system for question-answering data based on a computing power platform, wherein the adaptive enhancement system for question-answering data based on a computing power platform includes: The question data structuring module 51 is used to obtain the original question and answer data of the computing power platform operation and maintenance scenario, clean and structure the original question and answer data to obtain multiple structured dictionary formats, and establish a basic index library based on all the structured dictionary formats. The question weight adjustment module 52 is used to perform multi-dimensional feature extraction on the question text in the structured dictionary format based on the basic index library to obtain routing instructions, match the target operator combination in multiple operator combinations according to the routing instructions, and dynamically adjust the weight of the target operator combination to obtain the target weight. The answer generation module 53 is used to obtain version data and linkage instructions, and to perform question-and-answer semantic synchronization and question-and-answer semantic quality control on the question text according to the linkage instructions, the version data and the target weight, so as to obtain answer data.

[0042] Furthermore, such as Figure 5 As shown, based on the above-mentioned question-answering data adaptive enhancement method and system based on computing power platform, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 5 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0043] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as the program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a computing platform-based question-and-answer data adaptive enhancement program 40, which can be executed by the processor 10 to implement the computing platform-based question-and-answer data adaptive enhancement method of this application.

[0044] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the question-and-answer data adaptive enhancement method based on the computing power platform.

[0045] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The terminals communicate with each other via a system bus.

[0046] In one embodiment, when the processor 10 executes the question-answering data adaptive enhancement program 40 based on the computing platform in the memory 20, the following steps are performed: Obtain raw question and answer data from the operation and maintenance scenario of the computing power platform, clean and structure the raw question and answer data to obtain multiple structured dictionary formats, and establish a basic index library based on all the structured dictionary formats; Based on the aforementioned basic index library, multi-dimensional feature extraction is performed on the question text in the structured dictionary format to obtain routing instructions. According to the routing instructions, a target operator combination is matched among multiple operator combinations, and the weights of the target operator combination are dynamically adjusted to obtain the target weights. Obtain version data and linkage instructions, and perform question-and-answer semantic synchronization and question-and-answer semantic quality control on the question text based on the linkage instructions, version data, and target weight to obtain answer data.

[0047] The process of acquiring raw question-and-answer data from the operation and maintenance scenario of the computing power platform, cleaning and structuring the raw question-and-answer data, and establishing a basic index library specifically includes: The raw question and answer data of the computing power platform operation and maintenance scenario is obtained, the raw question and answer data is cleaned to obtain effective data, and the effective data is structured to obtain multiple structured dictionary formats. Input all the structured dictionary formats into the original database to obtain the basic index library.

[0048] The process of acquiring raw question-and-answer data from the computing power platform's operation and maintenance scenario, cleaning and structuring the raw question-and-answer data, establishing a basic index library, and then further includes: Obtain the synonym replacement weight, colloquial expansion weight, and sentence restructuring weight, and configure the synonym replacement weight, colloquial expansion weight, and sentence restructuring weight to obtain a dynamic sampling probability matrix; The words in the original question-and-answer data are mapped according to the dynamic sampling probability matrix to obtain the corrected words.

[0049] The operator combination includes synonym substitution operator, back translation operator, stop word insertion operator, entity recognition operator, knowledge graph reasoning operator, version term mapping operator, sentence inversion operator, and modal word addition operator; The target operator combination includes any one or more of the first operator combination, the second operator combination, and the third operator combination; The first operator combination includes a synonym substitution operator, a back-translation operator, and a stop word insertion operator; The second operator combination includes entity recognition operators, knowledge graph reasoning operators, and version term mapping operators; The third operator combination includes a sentence inversion operator and a modal particle addition operator.

[0050] Specifically, the process of extracting multi-dimensional features from the question text in the structured dictionary format based on the basic index library to obtain routing instructions, matching target operator combinations among multiple operator combinations according to the routing instructions, and dynamically adjusting the weights of the target operator combinations to obtain target weights includes: Based on the aforementioned basic index library, multidimensional feature extraction is performed on the question text in the structured dictionary format to obtain the confidence score; Acquire control signal packets, and perform intent recognition based on the confidence level according to preset policy operators and the control signal packets to obtain routing instructions; According to the routing instructions, match any one or more of the first operator combination, the second operator combination, and the third operator combination in a plurality of operator combinations; If the deviation of any one or more of the first operator combination, the second operator combination, and the third operator combination is greater than a preset value, then the weights of any one or more of the first operator combination, the second operator combination, and the third operator combination are dynamically adjusted through a negative feedback mechanism to obtain the target weight.

[0051] The step of obtaining the control signal packet, which involves performing intent recognition on the confidence level based on a preset policy operator and the control signal packet to obtain routing instructions, specifically includes: Acquire control signal packets, and perform intent recognition on the confidence level based on preset strategy operators and the control signal packets to obtain the intent category; The question text is tagged according to the intent category to obtain routing instructions; The intent categories include novice intents and expert intents.

[0052] The step of obtaining version data and linkage instructions, and performing question-and-answer semantic synchronization and question-and-answer semantic quality control on the question text based on the linkage instructions, the version data, and the target weight to obtain answer data, specifically includes: Obtain version data and linkage instructions; based on semantic anchoring technology, perform question-and-answer semantic synchronization on the question text according to the linkage instructions, version data, and target weight to obtain an initial answer; Obtain the summary extraction rules, and adjust the initial answer according to the summary extraction rules to obtain the target answer; Based on the cosine similarity algorithm, semantic deduplication is performed on the target answer according to the strategies of terminology accuracy, sentence structure rationality, and intent adaptation to obtain the answer data.

[0053] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a question-answering data adaptive enhancement program based on a computing power platform, and the question-answering data adaptive enhancement program based on the computing power platform, when executed by a processor, implements the steps of the question-answering data adaptive enhancement method based on a computing power platform as described above.

[0054] In summary, this invention provides a method, system, terminal, and storage medium for adaptive enhancement of question-and-answer data based on a computing power platform. The method includes: acquiring raw question-and-answer data from a computing power platform operation and maintenance scenario; cleaning and structuring the raw question-and-answer data to obtain multiple structured dictionary formats; establishing a basic index library based on all the structured dictionary formats; extracting multi-dimensional features from the question text in the structured dictionary formats based on the basic index library to obtain routing instructions; matching a target operator combination among multiple operator combinations according to the routing instructions; dynamically adjusting the weights of the target operator combination to obtain a target weight; acquiring version data and linkage instructions; and performing question-and-answer semantic synchronization and quality control on the question text based on the linkage instructions, the version data, and the target weight to obtain answer data. This invention, based on the cleaning and structuring of question-and-answer data from a computing power platform, achieves accurate synchronization and quality control of question-and-answer semantics through multi-dimensional feature extraction and dynamic weight adjustment, thereby obtaining accurate answers.

[0055] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal system that includes that element.

[0056] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0057] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A question-answering data adaptive enhancement method based on a computing power platform, characterized in that, The question-answering data adaptive enhancement method based on the computing power platform includes: Obtain raw question and answer data from the operation and maintenance scenario of the computing power platform, clean and structure the raw question and answer data to obtain multiple structured dictionary formats, and establish a basic index library based on all the structured dictionary formats; Based on the aforementioned basic index library, multi-dimensional feature extraction is performed on the question text in the structured dictionary format to obtain routing instructions. According to the routing instructions, a target operator combination is matched among multiple operator combinations, and the weights of the target operator combination are dynamically adjusted to obtain the target weights. Obtain version data and linkage instructions, and perform question-and-answer semantic synchronization and question-and-answer semantic quality control on the question text based on the linkage instructions, version data, and target weight to obtain answer data.

2. The question-answering data adaptive enhancement method based on a computing power platform according to claim 1, characterized in that, The process of acquiring raw question-and-answer data from the computing power platform's operation and maintenance scenario, cleaning and structuring the raw question-and-answer data, and establishing a basic index library specifically includes: The raw question and answer data of the computing power platform operation and maintenance scenario is obtained, the raw question and answer data is cleaned to obtain effective data, and the effective data is structured to obtain multiple structured dictionary formats. Input all the structured dictionary formats into the original database to obtain the basic index library.

3. The question-answering data adaptive enhancement method based on a computing power platform according to claim 1, characterized in that, The process of acquiring raw question-and-answer data from the computing power platform operation and maintenance scenario, cleaning and structuring the raw question-and-answer data, establishing a basic index library, and then further includes: Obtain the synonym replacement weight, colloquial expansion weight, and sentence restructuring weight, and configure the synonym replacement weight, colloquial expansion weight, and sentence restructuring weight to obtain a dynamic sampling probability matrix; The words in the original question-and-answer data are mapped according to the dynamic sampling probability matrix to obtain the corrected words.

4. The question-answering data adaptive enhancement method based on a computing power platform according to claim 1, characterized in that, The operator combination includes synonym replacement operator, back translation operator, stop word insertion operator, entity recognition operator, knowledge graph reasoning operator, version term mapping operator, sentence inversion operator, and modal word addition operator; The target operator combination includes any one or more of the first operator combination, the second operator combination, and the third operator combination; The first operator combination includes a synonym substitution operator, a back-translation operator, and a stop word insertion operator; The second operator combination includes entity recognition operators, knowledge graph reasoning operators, and version term mapping operators; The third operator combination includes a sentence inversion operator and a modal particle addition operator.

5. The question-answering data adaptive enhancement method based on a computing power platform according to claim 4, characterized in that, Based on the basic index library, multi-dimensional feature extraction is performed on the question text in the structured dictionary format to obtain routing instructions. Based on these routing instructions, a target operator combination is matched among multiple operator combinations. The weights of the target operator combination are dynamically adjusted to obtain the target weights. Specifically, this includes: Based on the aforementioned basic index library, multidimensional feature extraction is performed on the question text in the structured dictionary format to obtain the confidence score; Acquire control signal packets, and perform intent recognition based on the confidence level according to preset policy operators and the control signal packets to obtain routing instructions; According to the routing instructions, match any one or more of the first operator combination, the second operator combination, and the third operator combination in a plurality of operator combinations; If the deviation of any one or more of the first operator combination, the second operator combination, and the third operator combination is greater than a preset value, then the weights of any one or more of the first operator combination, the second operator combination, and the third operator combination are dynamically adjusted through a negative feedback mechanism to obtain the target weight.

6. The question-answering data adaptive enhancement method based on a computing power platform according to claim 5, characterized in that, The step of acquiring the control signal packet, and performing intent recognition based on the confidence level according to the preset policy operator and the control signal packet to obtain routing instructions, specifically includes: Acquire control signal packets, and perform intent recognition on the confidence level based on preset strategy operators and the control signal packets to obtain the intent category; The question text is tagged according to the intent category to obtain routing instructions; The intent categories include novice intents and expert intents.

7. The question-answering data adaptive enhancement method based on a computing power platform according to claim 1, characterized in that, The process of acquiring version data and linkage instructions, and performing question-and-answer semantic synchronization and question-and-answer semantic quality control on the question text based on the linkage instructions, version data, and target weight to obtain answer data, specifically includes: Obtain version data and linkage instructions; based on semantic anchoring technology, perform question-and-answer semantic synchronization on the question text according to the linkage instructions, version data, and target weight to obtain an initial answer; Obtain the summary extraction rules, and adjust the initial answer according to the summary extraction rules to obtain the target answer; Based on the cosine similarity algorithm, semantic deduplication is performed on the target answer according to the strategies of terminology accuracy, sentence structure rationality, and intent adaptation to obtain the answer data.

8. A question-answering data adaptive enhancement system based on a computing power platform, characterized in that, The question-answering data adaptive enhancement system based on the computing power platform includes: The question data structuring module is used to obtain raw question and answer data from the operation and maintenance scenario of the computing power platform, clean and structure the raw question and answer data to obtain multiple structured dictionary formats, and establish a basic index library based on all the structured dictionary formats. The question weight adjustment module is used to perform multi-dimensional feature extraction on the question text in the structured dictionary format based on the basic index library to obtain routing instructions, match the target operator combination among multiple operator combinations according to the routing instructions, and dynamically adjust the weight of the target operator combination to obtain the target weight. The answer generation module is used to acquire version data and linkage instructions, and to perform question-and-answer semantic synchronization and question-and-answer semantic quality control on the question text based on the linkage instructions, the version data, and the target weight, so as to obtain answer data.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a computing platform-based question-answering data adaptive enhancement program stored in the memory and executable on the processor. When the computing platform-based question-answering data adaptive enhancement program is executed by the processor, it implements the steps of the computing platform-based question-answering data adaptive enhancement method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a question-and-answer data adaptive enhancement program based on a computing power platform, which, when executed by a processor, implements the steps of the question-and-answer data adaptive enhancement method based on a computing power platform as described in any one of claims 1-7.

Images