Research and development platform language intelligent analysis method based on multi-system fusion

Abstract

The invention discloses a research and development platform language intelligent analysis method based on multi-system fusion, and relates to the technical field of language intelligent analysis. According to the method, various language data in a multi-system fusion research and development platform are efficiently processed, cleaned and coded to generate a standardized data set by customizing a data adaptation interface, deep lexical and syntactic analysis and semantic role labeling are performed on the standardized data set, related features are extracted, a semantic mapping network is constructed, and semantic feature vectors are generated; cross-modal feature fusion is realized, deep meanings of language data are comprehensively captured, rich feature input is provided for intelligent analysis, a constructed intelligent analysis model can perform accurate intelligent analysis and intention recognition on natural language instructions and technical documents, and the intelligent analysis model can be used for performing intelligent analysis and intention recognition on the technical documents by means of compiling environment simulation, debugging information semantic optimization and the like. The model training effect is improved, and the analysis accuracy and the hardware compatibility are improved through hardware behavior simulation, real-time data flow analysis, user feedback and other modes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of language intelligent analysis, and particularly relates to a language intelligent analysis method for a research and development platform based on multi-system fusion. BACKGROUND

[0002] The language data generated by different systems presents the characteristics of diversification and heterogeneity, including structured code data, semi-structured hardware protocol data, and unstructured natural language description, etc. How to effectively analyze and understand these complex language data becomes a key problem faced by the research and development platform. The traditional language analysis method can only process a single type of data, and is difficult to adapt to the complex requirements in the multi-system fusion environment. For example, for code data, hardware protocol data, and natural language technical documents, there is a lack of unified data input format conversion and deep semantic analysis, and it is impossible to effectively extract semantic features in various language data and construct a semantic mapping network. In the aspect of cross-platform compilation environment adaptation, the feature fusion of different modal data and the construction of the corresponding intelligent analysis model are not involved, and it is difficult to realize the intelligent analysis and intent recognition of natural language instructions and technical documents. Therefore, there is an urgent need for a method that can comprehensively process multi-source heterogeneous language data, realize cross-platform compilation environment adaptation and optimization, and construct an efficient intelligent analysis model, so as to improve the intelligent level of the research and development platform and the research and development efficiency. SUMMARY

[0003] The present application relates to the technical field of language intelligent analysis, and particularly relates to a language intelligent analysis method for a research and development platform based on multi-system fusion.

[0004] To achieve the above-mentioned purpose, the present application provides the following technical scheme:

[0005] The language intelligent analysis method for a research and development platform based on multi-system fusion comprises:

[0006] Preprocessing of multi-source heterogeneous data:

[0007] The structured, semi-structured, and unstructured language data in the multi-system fusion research and development platform are protocol-analyzed through the self-defined data adaptation interface, and the data input format is unified;

[0008] The garbled code, repeated data, and invalid characters in the language data after the unified data input format are identified and removed, the cleaned language data is encoded, and a standardized data set is generated.

[0009] Cross-platform compilation environment adaptation and optimization:

[0010] Lexical analysis, syntactic analysis and semantic role labeling are performed on the standardized dataset to extract lexical, syntactic structure and semantic relationship features.

[0011] A semantic mapping network is constructed to match the extracted lexical, syntactic structure and semantic relationship features with predefined semantic samples, calculate the correlation weight between features, and generate a semantic feature vector.

[0012] Multi-system fusion intelligent analysis model construction:

[0013] The semantic feature vector and each modality data are mapped to a unified dimensional feature space, different modality data weights are assigned, an intelligent analysis model is constructed, and the intelligent analysis model is trained based on the annotated analysis task data. Based on the trained intelligent analysis model, natural language instructions and technical documents are intelligently analyzed and intent recognized.

[0014] Further, multi-source heterogeneous data preprocessing, specifically including:

[0015] Code data analysis: through self-defined adaptive interface, embedded compilation tool chain is connected, C / C++ source code data is scanned and analyzed, code data is converted to abstract syntax tree structure, mapping relationship between code data and target architecture parameters in compilation environment configuration data is established, and comments and invalid syntax units in code data are removed.

[0016] Hardware protocol data standardization: format standardization processing is performed on hardware interface protocol data, and protocol parameters are converted to text semantic units.

[0017] Compilation environment configuration data processing: real-time / semi-real-time compilation environment is analyzed, compilation environment configuration data is parameter extracted, converted to standardized parameter vector and stored in association with code data.

[0018] Multi-source data cleaning: abstract syntax tree structure code data, converted text semantic unit data and standardized parameter vector are uniformly cleaned, hardware description text is identified based on semantic rule library, redundant parameters are removed, electrical parameter format specification is unified, and standardized dataset is generated.

[0019] Further, cross-platform compilation environment adaptation and optimization, specifically including:

[0020] Code semantic feature extraction: code data is selected from the standardized dataset, link intermediate products generated during compilation are symbolically analyzed, deep semantic features in code data are extracted, and a logical correlation graph between code data is constructed.

[0021] Hardware interface semantic mapping: extracting hardware interface protocol data from standardized data sets, establishing a mapping network of hardware interface protocol and semantic concepts; calculating the association weight between different interface protocol characteristics, optimizing the mapping network parameters, and structuring the mapped semantic information for storage, generating a hardware interface semantic knowledge base;

[0022] Cross-modal feature fusion: integrating standardized parameter vectors in standardized data sets, code data and hardware semantic information, constructing a multi-modal semantic feature extraction framework, aligning and fusing feature vectors of different modalities, generating a fusion feature vector containing multi-modal semantics, and finally constructing a generated semantic feature vector.

[0023] Further, the hardware interface protocol data includes serial communication protocol data, bus communication protocol data and storage interface protocol data, and the protocol parameters include transmission rate, transmission mode, verification method and timing configuration parameters.

[0024] Further, the cross-modal feature fusion specifically includes:

[0025] The standardized parameter vector is extracted from the standardized data set, and the code semantic feature matrix and the semantic information of the hardware interface are unified in dimension;

[0026] The feature vectors of different modalities are time-aligned, the corresponding relationship between cross-modal features is established, and the standardized parameter vector is dynamically weighted based on the logical association graph between code data and the hardware interface semantic knowledge base;

[0027] The weighted standardized parameter vector is serially fused with the code semantic feature matrix and the hardware interface semantic vector to construct a cross-modal feature set;

[0028] The fused cross-modal feature set is dimensionally reduced and semantically refined to generate a fusion feature vector containing multi-modal semantics as an input feature of the intelligent analysis model.

[0029] Further, the intelligent analysis model is trained based on the annotated analysis task data, further comprising:

[0030] Compilation environment simulation module: constructs a compilation environment simulation framework, obtains semantic error information generated during the compilation process, and converts error location, type and context into error semantic features;

[0031] Debugging information semantic optimization: when analyzing the loop structure in the code, the variable iteration data obtained by the debugging tool is combined to obtain runtime information, the association mapping between code static semantics and dynamic execution behavior is constructed, and the error semantic feature generation strategy is optimized.

[0032] Further, it also includes the optimization and feedback of the analysis result, specifically including:

[0033] Obtaining the intelligent analysis result of the intelligent analysis model, performing hardware behavior simulation on the intelligent analysis result, collecting key timing characteristics in the simulation process, comparing the standard timing mode of the hardware interface semantic knowledge base, and generating a timing deviation vector;

[0034] Calculating the hardware compatibility score of the analysis result based on the timing deviation vector, and verifying the accuracy of semantic analysis;

[0035] Collecting real-time data streams of hardware runtime, performing timing alignment calculation of semantic similarity between the real-time data streams and the analysis result, and dynamically adjusting the parameter weight of the intelligent analysis model based on the similarity score.

[0036] Further, the optimization and feedback of the analysis result also include:

[0037] Establishing a user feedback channel, collecting manually labeled analysis deviation information, and standardizing the coding of the manually labeled analysis deviation information based on the semantic rule base to generate a standardized feedback vector;

[0038] Mixing the standardized feedback vector and the labeled analysis task data by weighting, adjusting the training priority of data from different sources, and retraining and optimizing the intelligent analysis model.

[0039] Further, when the trained intelligent analysis model is used to intelligently analyze and identify the natural language instructions and technical documents, the target content in the natural language instructions and technical documents that cannot be intelligently analyzed and identified by the trained intelligent analysis model is obtained; wherein the target content at least includes: ambiguous content, ambiguous content;

[0040] Determine the training ability distribution related to the target content from the training ability feature map of the trained intelligent analysis model; wherein the training ability feature map at least includes: a graph that presents the training ability feature of the trained intelligent analysis model in multiple dimensions;

[0041] The fusion allowed value of multiple training abilities in the training ability distribution is calculated by the following formula:

[0042]

[0043] Wherein, Alo is the fusion allowed value, which represents the degree of allowing multiple training abilities in the training ability distribution to be fused, p i is the evaluation value of the multiple training abilities in the training ability distribution under the ith fusion allowed evaluation standard, and the fusion allowed evaluation standard at least includes: the complementary relationship degree between the training abilities, the effect degree of historical fusion, K ia preset weight value corresponding to the ith fusion allowed evaluation criterion, and N is the total number of the fusion allowed evaluation criteria;

[0044] When the fusion allowed value exceeds the threshold value, a knowledge usage situation is constructed with the target content and the multiple training capabilities in the training capability distribution as situation elements;

[0045] The training capability fusion knowledge applicable in the knowledge usage situation is determined from the training capability fusion knowledge base, wherein the training capability fusion knowledge is a preset rule used to indicate the process of training capability fusion of the multiple training capabilities in the training capability distribution;

[0046] Based on the training capability fusion knowledge, the multiple training capabilities in the training capability distribution are instructed to perform training capability fusion, and a fused training capability is obtained;

[0047] The fused training capability is assigned to the trained intelligent parsing model;

[0048] Based on the trained intelligent parsing model to which the fused training capability is assigned, the target content is re-parsed and re-identified.

[0049] Further, when the fusion allowed value is lower than the threshold value, the target content is converted to be closest to a conversion target, the trained intelligent parsing model is used to parse and identify the converted target content, and new parsing and identification results are output as results in comparison with the target content before and after conversion; wherein the comparison information includes a multi-modal comparison result between the target content before and after conversion.

[0050] The conversion target includes:

[0051] The training capabilities with the top W training capability degree values in the multiple training capabilities in the training capability distribution can each enable the trained intelligent parsing model to individually parse and identify part of the content in the target content; wherein the training capability degree value represents the capability degree of the training capability.

[0052] The ratio of W to the total number of the multiple training capabilities in the training capability distribution is closest to a preset ratio threshold value, so that W has a unique value.

[0053] Compared with the prior art, the present application has the following advantages:

[0054] By customizing the data adaptation interface, the various language data in the multi-system fusion research and development platform are efficiently processed, the data format is unified, the standardized data set is generated by cleaning and encoding, and a solid foundation is laid for subsequent analysis; the standardized data set is subjected to in-depth morphological, syntactic analysis and semantic role labeling, relevant features are extracted and a semantic mapping network is constructed, a semantic feature vector is generated, cross-modal feature fusion is realized, the semantic information of multiple data is integrated, the deep meaning of language data is comprehensively captured, rich feature input is provided for intelligent analysis, and the intelligent analysis model can accurately analyze and identify the natural language instructions and technical documents.

[0055] By analyzing the ambiguity or ambiguity of the target content and combining the fusion of training capabilities, the intelligent analysis model can more accurately understand complex natural language instructions or technical documents. This fusion training capability mechanism improves the flexibility and accuracy of the model, enabling it to handle a wider range of language scenarios, thereby improving its analysis and intent recognition performance.

[0056] When the fusion capability is insufficient, the content is automatically converted, and the intelligent analysis model is used for analysis and intent recognition to ensure that even if multiple training capabilities cannot be fused, the independent advantages of each training capability can be effectively utilized to improve the accuracy and intelligence of the analysis results. At the same time, by setting a preset ratio threshold and comparing information analysis, the system can maintain efficient and stable language intelligent analysis performance under varying input data conditions, ensuring that the output results have high analysis value and practical application significance. BRIEF DESCRIPTION OF DRAWINGS

[0057] Figure 1 The flowchart of the language intelligent analysis method based on the multi-system fusion research and development platform of the present application. DETAILED DESCRIPTION

[0058] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0059] To solve the technical problems of low processing efficiency of multi-source heterogeneous language data, insufficient semantic analysis depth, weak cross-modal feature fusion capability, and poor accuracy and hardware compatibility of intelligent analysis models, please refer to Figure 1 The technical solutions of the present embodiment are as follows:

[0060] The language intelligent analysis method based on multi-system fusion research and development platform includes:

[0061] Multi-source heterogeneous data preprocessing:

[0062] Through the self-defined data adaptation interface, the structured, semi-structured and unstructured language data in the multi-system fusion research and development platform are protocol analyzed, and the data input format is unified;

[0063] The regular expression and semantic rule library are used for data cleaning, and the language data after unified data input format is identified and removed. The language data after cleaning is encoded according to the unified data input format to generate a standardized data set, which lays a foundation for subsequent feature extraction;

[0064] Cross-platform compilation environment adaptation and optimization:

[0065] The standardized data set is extracted by using the statistical and rule combination feature extraction algorithm, and the lexical analysis, syntax analysis and semantic role labeling are carried out to extract the vocabulary, syntax structure and semantic relationship features;

[0066] The extracted vocabulary, syntax structure and semantic relationship features are matched with the pre-defined semantic samples, the correlation weight between features is calculated by using the graph neural network, the precise mapping of language features to semantic concepts is realized, and the semantic feature vector is generated to provide the semantic basis for intelligent analysis;

[0067] Multi-system fusion intelligent analysis model construction:

[0068] The semantic feature vector, code logic, system configuration and other modal data are mapped to a unified dimensional feature space, the attention mechanism is used to dynamically allocate the weight of different modal data, the intelligent analysis model is constructed, the intelligent analysis model is trained based on the annotated analysis task data, and the natural language instruction and technical document are intelligently analyzed and intention recognized based on the trained intelligent analysis model.

[0069] In this embodiment, by customizing the adaptation interface and the semantic rule library, the protocol analysis and unified cleaning of structured, semi-structured and unstructured data are realized, the standardized data set is generated, the consistency data foundation is laid for subsequent feature extraction, the data processing efficiency and quality are significantly improved, the language features are extracted by combining statistical and rule algorithms, the semantic mapping network is constructed by using the graph neural network, the representation limitation of traditional semantic analysis is broken through, the accurate mapping from language features to semantic concepts is realized, the feature vector containing deep semantic relationship is generated, the powerful semantic foundation is provided for intelligent analysis, the multi-modal data is mapped to a unified feature space, the efficient intelligent analysis model is constructed by dynamically allocating weights through the attention mechanism, the collaborative processing of multi-modal semantic information is realized, the analysis accuracy and intention recognition ability of natural language instructions and technical documents are significantly improved, and the cross-platform adaptability and intelligence of the model are enhanced.

[0070] In this embodiment, the multi-source heterogeneous data preprocessing specifically includes:

[0071] Code data analysis: by customizing the adaptation interface to connect embedded compilation tool chains such as assembly tools and C compilation tools, the C / C++ source code data is subjected to lexical scanning and syntax analysis, the code data is converted into an abstract syntax tree (AST) structure, a mapping relationship with target architecture parameters in the compilation environment configuration data is established, and comments and invalid syntax units in the code data are removed; for example, by using the scanning and analysis function of the compilation tool chain, the variable definition, function call and other syntax units in the code are identified, and noise data such as comments and invalid blank lines are removed to meet the requirements of unified data input format;

[0072] Hardware protocol data standardization: for hardware interface protocol data such as UART, SPI and CAN, such as communication component output, format standardization processing is performed through regular expressions and protocol rule library, binary or semi-structured protocol parameters are converted into textual semantic units such as "UART baud rate configuration 115200" and "SPI transmission mode 1", and protocol analysis and format unification of hardware protocol data are realized;

[0073] Compilation environment configuration data processing: the real-time / semi-real-time compilation environment is analyzed, the target architecture, optimization level and other parameters of the compilation environment configuration data are extracted, the standardized parameter vector is converted and stored in association with the code data, and the protocol analysis and format specification of the compilation environment configuration data are completed;

[0074] Multi-source data cleaning: the code data in the abstract syntax tree structure, the converted textual semantic unit data and the standardized parameter vector are subjected to unified cleaning, the hardware description text such as power integrity and signal line circuit is identified based on the semantic rule library, redundant parameters are removed, the format specification of electrical parameters such as voltage and impedance is unified, and the standardized data set is generated.

[0075] In this embodiment, the compilation tool chain is deeply integrated with code parsing through self-defined adaptation interface, the mapping relationship between code AST and compilation environment target architecture parameters is established, the accuracy and cross-platform adaptation capability of code parsing are significantly improved, a unified hardware protocol semantic representation is formed, standardized input is provided for hardware interface semantic mapping, the processing efficiency and consistency of hardware related data are greatly improved, the compilation environment parameters are extracted and stored in association with code data, a mapping bridge between compilation parameters and code logic is built, dynamic compilation parameters are converted into structured vectors, the adaptability of the model to different compilation configurations is enhanced, multi-modal data is integrated for collaborative cleaning, a high-quality standardized data set is generated, laying a data foundation for subsequent feature extraction and semantic mapping, and the reliability of the overall parsing process is improved.

[0076] In this embodiment, the cross-platform compilation environment adaptation and optimization specifically includes:

[0077] Code semantic feature extraction: code data is selected from the standardized data set, and the symbol parsing capability of the linking tool is combined to perform symbol parsing on the linking intermediate product generated in the compilation process, and deep semantic features such as function dependency relationship and variable scope in the code data are extracted; for example, by analyzing the map file generated in the linking process, a function call graph is constructed, logical associations between code data are identified, and a logical association graph between code data is constructed;

[0078] Hardware interface semantic mapping: hardware interface protocol data is extracted from the standardized data set, and a mapping network of hardware interface protocol and semantic concept is established based on the hardware component function definition specification; for example, “SPI transmission mode 1” is mapped to the semantic description of “SCK low level when idle, rising edge sampling”, the association weight between different interface protocol features is calculated by constructing a hardware semantic knowledge network, the mapping network parameters are optimized, and the mapped semantic information is stored in a structured manner to generate a hardware interface semantic knowledge base;

[0079] Cross-modal feature fusion: the standardized parameter vectors in the standardized data set, the code data and the hardware semantic information are integrated to construct a multi-modal semantic feature extraction framework, the understanding ability of the model to the code execution context is enhanced, the feature vectors of different modalities are aligned and fused, a fusion feature vector containing multi-modal semantics is generated, and finally a semantic feature vector is generated.

[0080] In this embodiment, the hardware interface protocol data includes serial communication protocol data, bus communication protocol data and storage interface protocol data, and the protocol parameters include transmission rate, transmission mode, verification mode and timing configuration parameters.

[0081] In this embodiment, by extracting deep semantic features, a logical correlation graph is constructed to provide structured logical representation for code semantic understanding, accurately capture the logical correlation between code modules, improve the depth of complex code semantic analysis, based on the function specification of hardware components, calculate the protocol feature correlation weight through the graph neural network, construct the mapping network of hardware interface protocol and semantic concept and generate a structured knowledge base, solve the fragmentation problem of hardware protocol semantic representation in the prior art, provide standardized semantic reference for hardware related code parsing, improve the accuracy and consistency of hardware operation instruction parsing, integrate multi-modal data such as code, hardware, and compilation parameters, construct a semantic feature vector containing complete execution context through feature alignment and fusion, break through the limitations of single-modal analysis in the prior art, realize the collaborative representation of multi-dimensional semantic information, provide more comprehensive input features for intelligent analysis model, and significantly enhance the semantic understanding ability of complex research and development scenarios.

[0082] In this embodiment, the cross-modal feature fusion specifically includes:

[0083] The standardized parameter vector is extracted from the standardized data set, and at the same time, the code semantic feature matrix and the semantic information of the hardware interface are subjected to dimension unification processing to ensure that the features of each modality are represented in the same semantic space;

[0084] The feature vectors of different modalities are time-aligned to establish a corresponding relationship between the cross-modal features, eliminate semantic drift caused by differences in data acquisition frequency or processing flow, and dynamically allocate weights to the standardized parameter vector based on the logical correlation graph between code data and the hardware interface semantic knowledge base;

[0085] The weighted standardized parameter vector is concatenated and fused with the code semantic feature matrix and the hardware interface semantic vector to construct a cross-modal feature set and enhance the semantic understanding of the code execution environment;

[0086] The fused cross-modal feature set is subjected to dimension reduction processing and semantic extraction to generate a fusion feature vector containing multi-modal semantics as the input feature of the intelligent analysis model.

[0087] In this embodiment, the intelligent analysis model is trained based on the annotated analysis task data, and further includes:

[0088] The compilation environment simulation module: based on the simulator function of the Docker container environment, a compilation environment simulation framework is constructed to support the simulation of different systems such as iOS and Android compilation processes, obtain semantic error information generated during the compilation process such as type mismatch and interface call exception, and convert the error position, type and context into error semantic features, for example:

[0089] Type error feature: [error type identifier, conflict type pair, error code line number]

[0090] Interface exception features: [interface name, parameter type list, expected return value, actual error code];

[0091] Debugging information semantic optimization: use the breakpoint setting and parameter monitoring functions of the debugging tool to dynamically inject runtime semantic information during model reasoning; for example, when parsing the loop structure in the code, the variable iteration data obtained by the debugging tool is combined to obtain runtime information, the association mapping between code static semantics and dynamic execution behavior is constructed, the parsing accuracy of the loop semantics is optimized, and the error semantic feature generation strategy is optimized.

[0092] In this embodiment, the weight distribution based on the logical association graph and the semantic knowledge base is realized through the multi-modal fusion framework, which improves the logicality and accuracy of feature fusion, enhances the understanding of the code execution environment, provides more comprehensive semantic feature input for intelligent parsing, constructs an extensible compilation simulation framework based on container technology, converts semantic errors into structured features, fills the gap in the existing technology of compiling error feature application, provides real compilation error supervision signals for model training, improves the model's ability to recognize code semantic correctness, supports multi-system compilation process simulation to enhance cross-platform adaptability, dynamically injects runtime data during the reasoning stage, establishes an association mechanism between static semantics and dynamic execution behavior, breaks through the limitations of existing technologies that rely solely on static analysis, optimizes the parsing accuracy of complex structures such as loops through runtime information such as variable iteration, and at the same time, it also benefits the error feature generation strategy, forming a virtuous cycle of "parsing-debugging-optimization".

[0093] In this embodiment, it also includes parsing result optimization and feedback, specifically including:

[0094] Obtain the intelligent parsing result of the intelligent parsing model, use the simulator of the real-time / quasi-real-time environment to simulate the hardware behavior of the intelligent parsing result, collect the key timing features in the simulation process, and compare them with the standard timing mode of the hardware interface semantic knowledge base to generate a timing deviation vector;

[0095] Calculate the hardware compatibility score of the parsing result based on the timing deviation vector to verify the accuracy of semantic parsing, such as after parsing the SPI communication code, verify whether the data transmission timing meets the protocol requirements through the simulator;

[0096] Collect real-time data streams of hardware runtime, including power ripple, signal strength, and other electrical parameters, and calculate the semantic similarity between the real-time data streams and the parsing result through timing alignment, and dynamically adjust the parameter weight of the intelligent parsing model based on the similarity score;

[0097] Establish a user feedback channel, collect manually annotated parsing deviation information, and based on the semantic rule base, standardize the coding of the manually annotated parsing deviation information to generate a standardized feedback vector;

[0098] The standardized feedback vector is mixed with the labeled analysis task data, the training priority of different source data is adjusted, the intelligent analysis model is retrained and optimized, a closed-loop iteration process of "analysis-feedback-optimization" is formed, and the accuracy and applicability of language intelligent analysis are continuously improved.

[0099] In the embodiment, the timing deviation vector is generated by the simulator and the compatibility score is calculated, the quantitative verification of the analysis result and the actual response of the hardware is realized, the "analysis-simulation-verification" mechanism is established, the protocol compliance of the hardware operation instruction timing is guaranteed, the research and development efficiency is improved, the real-time data stream of the hardware operation is collected and aligned with the timing of the analysis result, the dynamic environmental parameters are introduced into the model optimization, the limitations of the prior art relying on static data are broken, the model parameters are dynamically adjusted, the environmental adaptability of semantic analysis in complex scenarios is enhanced, the standardized artificial annotation feedback and the weighted mixed training data are constructed, the "artificial knowledge-machine learning" collaborative optimization mechanism is constructed, the analysis accuracy is continuously improved through closed-loop iteration, and the dynamic changes of the research and development requirements are adapted.

[0100] In the embodiment, when the natural language instruction and the technical document are intelligently analyzed and the intention is recognized based on the trained intelligent analysis model, target content in the natural language instruction and the technical document that cannot be intelligently analyzed and the intention recognized by the trained intelligent analysis model is acquired; wherein the target content at least includes: ambiguous content, ambiguous content;

[0101] From the training ability feature map of the trained intelligent analysis model, the training ability distribution related to the target content is determined; wherein the training ability feature map at least includes: a graph that presents the training ability feature of the trained intelligent analysis model in multiple dimensions;

[0102] The fusion allowed value of the multiple training abilities in the training ability distribution is calculated by the following formula:

[0103]

[0104] Wherein, Alo is the fusion allowed value, the fusion allowed value represents the degree of allowing fusion of the multiple training abilities in the training ability distribution, p i is the evaluation value of the multiple training abilities in the training ability distribution under the i-th fusion allowed evaluation standard, the fusion allowed evaluation standard at least includes: the complementary relationship degree between the training abilities, the effect degree of historical fusion, K i is the preset weight value corresponding to the i-th fusion allowed evaluation standard, and N is the total number of fusion allowed evaluation standards;

[0105] When the fusion allowed value exceeds the threshold value, the target content and the multiple training abilities in the training ability distribution are taken as situation elements to construct a knowledge usage situation.

[0106] determine training capability fusion knowledge applicable in the knowledge use case from the training capability fusion knowledge base; wherein the training capability fusion knowledge is a preset rule used to indicate a process of training capability fusion of multiple training capabilities in the training capability distribution;

[0107] based on the training capability fusion knowledge, instruct the multiple training capabilities in the training capability distribution to perform training capability fusion, and obtain a fused training capability;

[0108] assign the fused training capability to the trained intelligent parsing model;

[0109] based on the trained intelligent parsing model to which the fused training capability is assigned, re-perform intelligent parsing and intent recognition on the target content.

[0110] The working principle and beneficial effects of the above technical solutions are as follows:

[0111] When the trained intelligent parsing model processes language, it may encounter some content that cannot be fully understood. Generally, some content is ambiguous or ambiguous in language, and the trained intelligent parsing model may not accurately understand the intent of such content. Therefore, the target content needs to be obtained first.

[0112] The training capability feature map shows the training capability of the model in different dimensions, such as understanding of certain contexts, parsing ability of vocabulary, etc. From it, the training capability distribution related to the target content is determined, that is, multiple training capabilities that may have the ability to understand the context and parse the vocabulary in the target content are determined to form a distribution.

[0113] The fusion allowed evaluation criteria are set in advance, wherein the complementary relationship degree between the training capabilities refers to the complementary degree of the complementary relationship between the training capabilities (training capability complementarity refers to the respective complement of the other's capability defects), and the effect degree of historical fusion refers to the effect degree of using the corresponding intelligent parsing model to process language after the historical fusion of the training capabilities. The evaluation values of the multiple training capabilities in the training capability distribution are obtained by using different fusion allowed evaluation criteria, and the weights of the influence degrees of the different fusion allowed evaluation criteria on the final calculation of the fusion allowed value are set based on the pre-set different fusion allowed evaluation criteria. According to the evaluation values under different fusion allowed evaluation criteria, weighted calculation is performed, and finally the fusion allowed value representing the degree of allowing fusion of the multiple training capabilities in the training capability distribution is obtained.

[0114] The threshold is a pre-set value representing a large fusion permission value, and if the fusion permission value exceeds the threshold, the target content and multiple training capabilities in the training capability distribution are used as situation elements to construct a knowledge usage situation, determine the training capability fusion knowledge applicable in the knowledge usage situation from the training capability fusion knowledge base, and train the multiple training capabilities in the training capability distribution based on the indication of the training capability fusion knowledge to obtain the fusion training capability. With the fusion training capability, the trained intelligent parsing model can re-parses and identifies the intent of the target content.

[0115] The embodiment analyzes the ambiguity or ambiguity of the target content and combines the fusion of training capabilities to enable the intelligent parsing model to more accurately understand complex natural language instructions or technical documents. This fusion training capability mechanism improves the flexibility and accuracy of the model and can handle a wider range of language scenarios, thereby improving its parsing and intent recognition effect.

[0116] In the embodiment, when the fusion permission value is lower than the threshold, the target content is converted to the closest conversion target, the trained intelligent parsing model is used to intelligently parse and identify the intent of the converted target content, and the new intelligent parsing and intent recognition result and the comparison information before and after the target content conversion are output as the result. The comparison information includes the multi-modal comparison result between the target content before and after conversion.

[0117] The conversion target includes:

[0118] The top W training capabilities in the training capability distribution can each enable the trained intelligent parsing model to individually intelligently parse and identify the intent of part of the content in the target content. The training capability degree value represents the degree of the training capability.

[0119] The ratio of W to the total number of multiple training capabilities in the training capability distribution is closest to the pre-set ratio threshold, so that W has a unique value.

[0120] The working principle and beneficial effects of the above technical solution are:

[0121] When the fusion permission value is lower than the threshold, it means that the multiple training capabilities in the training capability distribution are not suitable for fusion. To compensate for this defect, the target content is passively converted to the closest conversion target, the trained intelligent parsing model is used to intelligently parse and identify the intent of the converted target content, and the new intelligent parsing and intent recognition result and the comparison information before and after the target content conversion are output as the result. When outputting, the new intelligent parsing and intent recognition result can be analyzed selectively based on the comparison information.

[0122] When the conversion target is set, the training capabilities ranked in the top W in the training capability distribution can each enable the trained intelligent analysis model to independently analyze and identify the intent of part of the target content. The value of W is constrained by the constraint that the ratio of W divided by the total number of training capabilities in the training capability distribution is closest to the preset ratio threshold. Ultimately, a sufficient proportion of training capabilities with greater training capability degrees in the training capability distribution can each enable the trained intelligent analysis model to independently analyze and identify the intent of part of the target content. This enables the new intelligent analysis and intent recognition result to have sufficient language analysis value. The preset ratio threshold can be set in advance according to actual needs.

[0123] The embodiment automatically performs content conversion when the fusion capability is insufficient, analyzes and identifies the intent in combination with the intelligent analysis model, ensures that even if multiple training capabilities cannot be fused, the independent advantages of each training capability can be effectively utilized, thereby improving the accuracy and intelligence level of the analysis result. At the same time, by setting the preset ratio threshold and analyzing the comparison information, the system can maintain efficient and stable language intelligent analysis performance under changing input data conditions, and ensure that the output result has high analysis value and practical application significance.

[0124] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can make equivalent replacements or changes to the technical solution and inventive concept of the present application within the scope of the disclosed technology, and such should be covered within the protection scope of the present application.

Claims

1. A language intelligent parsing method based on a multi-system fusion R&D platform, characterized by: include: Multi-source heterogeneous data preprocessing: Through the custom data adaptation interface, protocol analysis is performed on structured, semi-structured and unstructured language data in the multi-system integration R&D platform to unify the data input format; Identify and remove garbled characters, duplicate data, and invalid characters from the language data after unified data input format, encode the cleaned language data, and generate a standardized data set; Cross-platform compilation environment adaptation and optimization: Perform lexical analysis, syntactic analysis, and semantic role labeling on standardized datasets to extract vocabulary, grammatical structure, and semantic relationship features; Construct a semantic mapping network to match the extracted vocabulary, grammatical structure and semantic relationship features with predefined semantic samples, calculate the correlation weights between features, and generate a semantic feature vector; Construction of multi-system fusion intelligent analysis model: Map the semantic feature vectors and each modal data to a feature space of unified dimension, assign weights to different modal data, build an intelligent parsing model, and train the intelligent parsing model based on the labeled parsing task data. Based on the trained intelligent parsing model, perform intelligent parsing and intent recognition on natural language instructions and technical documents.

2. The language intelligent parsing method based on multi-system integration of the R&D platform according to claim 1 is characterized in that: Multi-source heterogeneous data preprocessing, including: Code data parsing: Connect to the embedded compilation tool chain through a custom adapter interface, perform lexical scanning and syntactic analysis on C / C++ source code data, convert the code data into an abstract syntax tree structure, establish a mapping relationship with the target architecture parameters in the compilation environment configuration data, and remove comments and invalid syntax units in the code data; Hardware protocol data standardization: Standardize the format of hardware interface protocol data and convert protocol parameters into textual semantic units; Compilation environment configuration data processing: parses the real-time / quasi-real-time compilation environment, extracts parameters from the compilation environment configuration data, converts them into standardized parameter vectors, and stores them in association with the code data; Multi-source data cleaning: The code data in the abstract syntax tree structure, the converted textual semantic unit data, and the standardized parameter vector are uniformly cleaned. The hardware description text is identified based on the semantic rule base, redundant parameters are removed, the electrical parameter format specifications are unified, and a standardized data set is generated.

3. The language intelligent parsing method based on multi-system integration of the R&D platform according to claim 2 is characterized in that: Cross-platform compilation environment adaptation and optimization, including: Code semantic feature extraction: Filter code data from standardized datasets, perform symbolic analysis on the link intermediate products generated during the compilation process, extract deep semantic features from the code data, and construct a logical association map between the code data; Hardware interface semantic mapping: Extract hardware interface protocol data from standardized data sets and establish a mapping network between hardware interface protocols and semantic concepts. Calculate the association weights between different interface protocol features, optimize the mapping network parameters, and structure the mapped semantic information to generate a hardware interface semantic knowledge base. Cross-modal feature fusion: Integrate the standardized parameter vectors in the standardized data set with code data and hardware semantic information to build a multimodal semantic feature extraction framework. Align and fuse the feature vectors of different modalities to generate a fused feature vector containing multimodal semantics, and finally construct a semantic feature vector.

4. The language intelligent parsing method based on multi-system integration of the R&D platform according to claim 3 is characterized in that: The hardware interface protocol data includes serial communication protocol data, bus communication protocol data and storage interface protocol data, and the protocol parameters include transmission rate, transmission mode, verification method and timing configuration parameters.

5. The language intelligent parsing method based on multi-system integration of the R&D platform according to claim 4 is characterized in that: Cross-modal feature fusion, specifically including: Extract standardized parameter vectors from the standardized dataset, and unify the code semantic feature matrix and the hardware interface semantic information into unified dimensions. Perform temporal alignment on feature vectors of different modalities, establish correspondence between cross-modal features, and dynamically assign weights to standardized parameter vectors based on the logical association graph between code data and the hardware interface semantic knowledge base. The weighted normalized parameter vector is serially fused with the code semantic feature matrix and the hardware interface semantic vector to construct a cross-modal feature set. The fused cross-modal feature set is subjected to dimensionality reduction and semantic extraction to generate a fused feature vector containing multimodal semantics, which serves as the input feature of the intelligent parsing model.

6. The language intelligent parsing method based on multi-system integration of the R&D platform according to claim 5 is characterized in that: Training the intelligent parsing model based on the annotated parsing task data also includes: Compilation environment simulation module: Builds a compilation environment simulation framework, obtains semantic error information generated during the compilation process, and converts error location, type, and context into error semantic features; Debug information semantic optimization: When parsing loop structures in the code, the variable iteration data obtained by the debugging tool is combined with runtime information to build an association mapping between the static semantics of the code and the dynamic execution behavior, and optimize the error semantic feature generation strategy.

7. The language intelligent parsing method based on multi-system integration of the R&D platform according to claim 6 is characterized in that: It also includes analysis result optimization and feedback, including: Obtain the intelligent analysis results of the intelligent analysis model, perform hardware behavior simulation on the intelligent analysis results, collect key timing features in the simulation process, compare them with the standard timing patterns of the hardware interface semantic knowledge base, and generate a timing deviation vector; Calculate the hardware compatibility score of the parsing result based on the timing deviation vector to verify the accuracy of the semantic parsing; Collect real-time data streams during hardware runtime, align the real-time data streams with the parsing results to calculate semantic similarity, and dynamically adjust the parameter weights of the intelligent parsing model based on the similarity score.

8. The language intelligent parsing method based on multi-system integration of the R&D platform according to claim 7 is characterized in that: Analysis result optimization and feedback also include: Establish a user feedback channel to collect manually annotated parsing deviation information, and standardize the manually annotated parsing deviation information based on the semantic rule base to generate a standardized feedback vector; The standardized feedback vector is weightedly mixed with the labeled parsing task data, the training priority of data from different sources is adjusted, and the intelligent parsing model is retrained and optimized.

9. The language intelligent parsing method based on multi-system integration of the R&D platform according to claim 1, characterized in that: When performing intelligent parsing and intent recognition on natural language instructions and technical documents based on a trained intelligent parsing model, obtaining target content in the natural language instructions and technical documents that the trained intelligent parsing model cannot intelligently parse and identify; wherein the target content includes at least: fuzzy content and ambiguous content; Determining a distribution of training capabilities related to the target content from a training capability characteristic graph of the trained intelligent parsing model; wherein the training capability characteristic graph includes at least: a graph presenting a multi-dimensional representation of the training capability characteristics of the trained intelligent parsing model; The fusion allowable value of multiple training capabilities in the training capability distribution is calculated using the following formula: Among them, Alo is the fusion allowable value, which represents the degree to which multiple training capabilities in the training capability distribution are allowed to be fused, and p i K is the evaluation value of multiple training capabilities in the training capability distribution under the i-th fusion allowed evaluation criteria. The fusion allowed evaluation criteria at least include: the degree of complementary relationship between the training capabilities and the degree of historical fusion effect. i is the preset weight value corresponding to the i-th fusion allowed evaluation criteria, and N is the total number of fusion allowed evaluation criteria; When the fusion allowable value exceeds the threshold, the target content and multiple training capabilities in the training capability distribution are used as situation elements to construct the knowledge usage situation; Determining, from a training capability fusion knowledge base, training capability fusion knowledge applicable in a knowledge usage scenario; wherein the training capability fusion knowledge is a preset rule for instructing a process of training capability fusion of multiple training capabilities in a training capability distribution; Based on the training capability fusion knowledge, multiple training capabilities in the training capability distribution are instructed to perform training capability fusion to obtain fused training capability; Empower trained intelligent parsing models with fusion training capabilities; Based on the trained intelligent parsing model that has been endowed with fusion training capabilities, the target content is re-intelligently parsed and intent recognized.

10. The language intelligent parsing method based on multi-system integration of the R&D platform according to claim 9, characterized in that: When the fusion allowable value is lower than the threshold, the target content is transformed to the closest conversion target. The trained intelligent parsing model is used to perform intelligent parsing and intent recognition on the transformed target content. The new intelligent parsing and intent recognition results are compared with the target content before and after the transformation as the output result. The comparison information includes: a multimodal comparison result between the target content before and after the transformation; The conversion goals include: The training capabilities ranked top W in the training capability distribution can each enable the trained intelligent parsing model to independently perform intelligent parsing and intent recognition on a portion of the target content; wherein the training capability degree value represents the degree of the training capability; The ratio of W divided by the total number of the multiple training capabilities in the training capability distribution is closest to the preset ratio threshold, so that the value of W is unique.

Images