Enterprise evaluation method and evaluation device based on natural language processing

Abstract

The application provides an enterprise evaluation method and device based on natural language processing, which comprises the following steps: obtaining original text data set of an enterprise to be evaluated; preprocessing the original text data set through an ARM instruction set to generate preprocessed enterprise text data set; inputting the preprocessed enterprise text data set into a pre-trained enterprise feature extraction model; performing hierarchical semantic feature extraction through a multi-layer semantic analysis structure in the enterprise feature extraction model to generate an enterprise semantic feature vector set; inputting the enterprise semantic feature vector set into a pre-built enterprise evaluation model; performing dynamic weight distribution processing on the enterprise semantic feature vector set through a multi-dimensional evaluation network in the enterprise evaluation model to generate an enterprise comprehensive evaluation index set; constructing an enterprise evaluation result tree structure according to the correlation between each evaluation index in the enterprise comprehensive evaluation index set; and performing layer-by-layer result screening through the evaluation threshold of each hierarchical node in the enterprise evaluation result tree structure to generate a final enterprise credit risk evaluation report. The application can improve the automation degree and reliability of enterprise credit evaluation.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method and apparatus for enterprise evaluation based on natural language processing. Background Technology

[0002] Corporate credit risk assessment aims to evaluate a company's credit risk level by analyzing relevant textual data. Existing methods typically rely on structured financial data or use basic natural language processing techniques to extract keywords and perform statistical modeling on unstructured text. They then generate assessment results by manually setting weight rules and integrating multi-dimensional indicators. However, existing technologies suffer from low computational efficiency in the cleaning and feature extraction processes of unstructured text. Traditional preprocessing methods struggle to effectively remove redundant noise and fail to guarantee semantic coherence. The lack of contextual analysis in the feature extraction layer leads to the loss of semantic information. Fixed weight rules limit the model's adaptability to different industry characteristics. The assessment results are presented in a single format and lack traceability, resulting in problems such as low assessment efficiency, shallow semantic understanding, weak cross-industry generalization ability, and opaque decision-making basis. Summary of the Invention

[0003] In view of this, embodiments of the present invention provide a method and apparatus for enterprise evaluation based on natural language processing. The technical solution of the embodiments of the present invention is implemented as follows:

[0004] On one hand, embodiments of the present invention provide a method for enterprise assessment based on natural language processing. The method includes: acquiring a set of original text data of the enterprise to be assessed, and preprocessing the original text data set using an ARM instruction set to generate a preprocessed enterprise text data set; inputting the preprocessed enterprise text data set into a pre-trained enterprise feature extraction model, and extracting hierarchical semantic features from the enterprise text data set using a multi-layer semantic analysis structure in the enterprise feature extraction model to generate a set of enterprise semantic feature vectors; inputting the set of enterprise semantic feature vectors into a pre-constructed enterprise assessment model, and dynamically weighting the set of enterprise semantic feature vectors using a multi-dimensional assessment network in the enterprise assessment model to generate a set of comprehensive enterprise assessment indicators; constructing a tree structure of enterprise assessment results based on the correlation between the assessment indicators in the comprehensive enterprise assessment indicator set, and filtering the results layer by layer using the assessment thresholds of each level node in the tree structure to generate a final enterprise credit risk assessment report.

[0005] On the other hand, the present invention provides an enterprise assessment device, comprising: a data acquisition module, used to acquire a set of original text data of the enterprise to be assessed, and preprocess the original text data set using an ARM instruction set to generate a preprocessed enterprise text data set; a feature extraction module, used to input the preprocessed enterprise text data set into a pre-trained enterprise feature extraction model, and extract hierarchical semantic features from the enterprise text data set through a multi-layer semantic analysis structure in the enterprise feature extraction model to generate a set of enterprise semantic feature vectors; an index generation module, used to input the set of enterprise semantic feature vectors into a pre-constructed enterprise assessment model, and dynamically assign weights to the set of enterprise semantic feature vectors through a multi-dimensional assessment network in the enterprise assessment model to generate a set of comprehensive enterprise assessment indicators; and a credit identification module, used to construct a tree structure of enterprise assessment results based on the correlation between the assessment indicators in the set of comprehensive enterprise assessment indicators, and filter the results layer by layer through the assessment thresholds of each level node in the tree structure to generate a final enterprise credit risk assessment report.

[0006] The enterprise credit risk assessment method provided by this invention deeply integrates the parallel processing capabilities of the ARM instruction set with the natural language processing workflow. In the preprocessing stage, SIMD instructions are used to efficiently clean and tag unstructured text, and part-of-speech tagging is achieved. Combined with frequency threshold filtering using part-of-speech distribution histograms and a grammatical coherence detection mechanism, redundant noise is effectively removed while retaining core semantic information. Through a bidirectional attention mechanism and multi-scale convolutional feature fusion in a multi-layer semantic analysis structure, high-order semantic features with contextual relevance are extracted from enterprise text. Adaptive dynamic weight allocation technology of a multi-dimensional evaluation network is used to automatically adjust the contribution of different evaluation dimensions based on semantic features. This overcomes the traditional evaluation model's reliance on manually set rules to construct a tree structure based on the correlation between evaluation indicators and perform layer-by-layer threshold filtering. The complex enterprise credit risk is decomposed into a traceable multi-level decision-making path, improving the accuracy of risk assessment while enhancing the interpretability of results. This enables real-time processing of massive heterogeneous enterprise data, in-depth mining of multi-dimensional semantic features, autonomous optimization of dynamic evaluation strategies, and transparent output of risk assessment conclusions, significantly improving the automation, cross-industry adaptability, and decision support reliability of enterprise credit assessment. Attached Figure Description

[0007] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present invention and, together with the specification, serve to explain the technical solutions of the present invention.

[0008] Figure 1 This is a schematic diagram illustrating the implementation process of an enterprise evaluation method based on natural language processing, provided in an embodiment of the present invention.

[0009] Figure 2 This is a schematic diagram of the composition structure of an enterprise evaluation device provided in an embodiment of the present invention. Detailed Implementation

[0010] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. The described embodiments should not be regarded as limitations on the present invention. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0011] This invention provides a natural language processing-based enterprise evaluation method, which can be executed by a computer system's processor. The computer system can refer to devices with data processing capabilities, such as servers, laptops, tablets, desktop computers, and mobile devices (e.g., mobile phones, portable video players, personal digital assistants).

[0012] Figure 1 This is a schematic diagram illustrating the implementation process of a natural language processing-based enterprise evaluation method provided in an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes the following steps:

[0013] Step S100: Obtain the original text data set of the enterprise to be evaluated, and preprocess the original text data set using the ARM instruction set to generate the preprocessed enterprise text data set.

[0014] In this step, the raw text data set of the company to be evaluated refers to the collection of various textual information related to the company. This textual data may come from various channels such as the company's annual reports, news reports, social media comments, and industry research reports, containing rich information about the company's operations, market performance, and financial status. The ARM instruction set is an instruction set of a Reduced Instruction Set Computer (RISC) architecture, characterized by low power consumption and high performance, and is commonly used in embedded systems and mobile devices. Preprocessing the raw text data set using the ARM instruction set aims to remove noise and redundant information from the data, converting it into a format suitable for subsequent analysis.

[0015] For example, suppose the company being evaluated is a manufacturing enterprise. Its original text data set might include product introduction documents, production reports, customer feedback, etc. The efficient computing power of the ARM instruction set can be used to process this text data. Specifically, the hardware features of the ARM processor, such as parallel computing units, can be leveraged to accelerate the data processing. This preprocessing improves the accuracy and efficiency of subsequent analysis, enabling feature extraction and evaluation to be based on cleaner, more valuable data.

[0016] As one implementation method, step S100, which preprocesses the original text data set using the ARM instruction set to generate a preprocessed enterprise text data set, may specifically include the following steps:

[0017] Step S110: Parallel text cleaning is performed on the original text data set using SIMD instructions in the ARM instruction set to remove unstructured noise data from the original text data set and generate preliminary cleaned text data.

[0018] Here, the SIMD (Single Instruction, Multiple Data) instruction in the ARM instruction set is a type of instruction that can perform the same operation on multiple data sets simultaneously under the control of a single instruction. Unstructured noisy data refers to information in the original text data that does not directly help with enterprise evaluation, such as HTML tags, special characters, and garbled text. By parallelizing text cleaning processing, the parallel computing capabilities of SIMD instructions can be utilized to process multiple blocks of text data simultaneously, improving cleaning efficiency.

[0019] For example, if the original text dataset contains corporate news reports scraped from web crawlers, it may contain unstructured noisy data such as HTML tags and advertising information. Using SIMD instructions from the ARM instruction set, the text data can be divided into multiple small blocks, and these blocks can be cleaned to remove unstructured noise. For instance, preset character matching rules can be set to identify and remove HTML tags and special characters. This parallel cleaning process yields initially cleaned text data, providing a cleaner data foundation for subsequent processing.

[0020] Step S120: Perform instruction-level word segmentation on each text segment in the initially cleaned text data, and use the memory access optimization mechanism of the ARM instruction set to perform part-of-speech tagging on the segmented text segments to generate a set of text data with part-of-speech tags.

[0021] Instruction-level word segmentation refers to the process of dividing each text segment in the initially cleaned text data into individual words according to preset rules, supported by the ARM instruction set. Part-of-speech tagging then assigns the part-of-speech attribute (NOT) to each segmented word, such as noun, verb, or adjective. The memory access optimization mechanism of the ARM instruction set can improve the efficiency of data reading and writing, reduce memory access latency, and thus speed up the part-of-speech tagging process.

[0022] Taking a short product introduction as an example of initial text data cleaning, the text is segmented into individual words such as "high-performance," "electronic products," "possess," "advanced," and "technology" through instruction-level word segmentation. Then, utilizing the memory access optimization mechanism of the ARM instruction set, the segmented words are quickly read and labeled with their parts of speech according to preset part-of-speech tagging rules. For example, "high-performance" is labeled as an adjective, "electronic products" as a noun, and "possess" as a verb. Finally, a text dataset with tagged parts of speech is generated, providing richer semantic information for subsequent text analysis.

[0023] Step S130: Based on the distribution characteristics of each part-of-speech tag in the text data set after part-of-speech tagging, construct a part-of-speech distribution histogram, and filter redundant words in the text data set by the frequency threshold of each part-of-speech tag in the histogram to generate a filtered and concise text data set.

[0024] A part-of-speech tag (POS) histogram is a statistical chart used to display the frequency of each POS tag in a text dataset. By analyzing the distribution characteristics of each POS tag, we can understand the linguistic structure and semantic features of the text. The frequency threshold is a pre-set value used to determine whether the frequency of a particular POS tag is too high or too low. POS tags with excessively high or low frequencies may correspond to redundant words and need to be filtered.

[0025] In the text dataset labeled with parts of speech, the frequency of each part-of-speech tag is counted. Then, according to a preset interval division rule, the counts of the same part-of-speech tag are accumulated into the corresponding histogram interval to generate a part-of-speech distribution histogram. For example, the frequency of noun, verb, and adjective tags can be counted separately and histograms can be drawn. Based on preset frequency thresholds, words corresponding to part-of-speech tags that are higher than a first frequency threshold or lower than a second frequency threshold are filtered out and treated as redundant words and removed. After this filtering process, a simplified text dataset is obtained, removing a large amount of redundant information that is not of great value to enterprise evaluation, making the data more refined.

[0026] As one implementation method, step S130, based on the distribution characteristics of each part-of-speech tag in the annotated text data set, constructs a part-of-speech distribution histogram, and filters redundant words in the text data set using the frequency threshold of each part-of-speech tag in the histogram to generate a filtered and concise text data set, may specifically include the following steps:

[0027] Step S131: Scan each text segment in the text data set after tagging the part of speech word by word, count the number of times each part of speech tag appears in the text data set, and generate a part of speech tag count set.

[0028] In this step, the text data set after part-of-speech tagging is scanned word by word, which means checking the part-of-speech tag of each word in turn and recording the number of times each part-of-speech tag appears. The part-of-speech tag count set is a set containing each part-of-speech tag and its corresponding number of occurrences.

[0029] For example, for a text dataset labeled with parts of speech, the part of speech of each word is checked. If a noun is encountered, the count for that noun is incremented by 1; if a verb is encountered, the count for that verb is incremented by 1, and so on. Suppose the text dataset contains words such as "enterprise" (noun), "development" (verb), and "rapidly" (adjective). After scanning each word, the frequency of occurrence of nouns, verbs, and adjectives is counted, ultimately generating a set of part-of-speech tag counts. This set reflects the distribution of each part-of-speech tag in the text dataset, providing the foundational data for constructing a histogram of part-of-speech distribution.

[0030] Step S132: Based on the occurrence frequency of each part-of-speech tag in the part-of-speech tag count set, according to the preset histogram interval division rules, the counts of the same part-of-speech tags are accumulated to the corresponding histogram intervals to generate a part-of-speech distribution histogram structure containing interval counts and the correspondence between part-of-speech tags.

[0031] The preset histogram interval division rule is a predefined rule for grouping part-of-speech tags according to a preset method. By accumulating the counts of the same part-of-speech tags into the corresponding histogram intervals, the distribution of each part-of-speech tag can be displayed more intuitively.

[0032] For example, part-of-speech tags can be categorized into several intervals based on their part-of-speech function, such as nouns, verbs, and adjectives. Based on the frequency of each part-of-speech tag in the tag count set, the counts of tags belonging to the noun category are accumulated in the noun interval, the counts of tags belonging to the verb category are accumulated in the verb interval, and so on. This ultimately generates a histogram structure of part-of-speech distribution that includes interval counts and the correspondence between part-of-speech tags. This structure clearly displays the counting situation of different part-of-speech tag intervals, which is helpful for subsequent frequency analysis and redundant word filtering.

[0033] Step S133: Based on the count data of each histogram interval in the part-of-speech distribution histogram structure, calculate the total frequency of each part-of-speech tag, and generate a part-of-speech tag frequency sequence by arranging them in descending order of frequency.

[0034] The total frequency of each part-of-speech tag refers to the proportion of times that part-of-speech tag appears in the text dataset relative to the total number of words. By calculating the total frequency, we can more accurately understand the importance of each part-of-speech tag in the text. Generating a part-of-speech tag frequency sequence by sorting the tags in descending order of frequency allows us to prioritize the most frequent tags, facilitating subsequent filtering.

[0035] Based on the count data of each histogram interval in the part-of-speech tag distribution histogram structure, the total number of words in the text dataset is first calculated. Then, the count of each part-of-speech tag is divided by the total number of words to obtain the total frequency of each part-of-speech tag. For example, if the count of nouns is 100 and the total number of words is 500, then the total frequency of nouns is 100 ÷ 500 = 0.2. After calculating the total frequency of all part-of-speech tags, they are arranged in descending order to generate a part-of-speech tag frequency sequence. This sequence can help quickly identify part-of-speech tags with high and low frequencies, providing a basis for filtering redundant words.

[0036] Step S134: Based on the preset frequency threshold configuration table, generate a set of part-of-speech tags to be filtered for part-of-speech tags in the part-of-speech tag frequency sequence that are higher than the first frequency threshold or lower than the second frequency threshold.

[0037] The preset frequency threshold configuration table is a pre-defined table containing a first frequency threshold and a second frequency threshold. The first frequency threshold is used to filter out part-of-speech tags that appear too frequently, and the second frequency threshold is used to filter out part-of-speech tags that appear too infrequently. The set of part-of-speech tags to be filtered consists of part-of-speech tags that are either above the first frequency threshold or below the second frequency threshold.

[0038] For example, in the preset frequency threshold configuration table, the first frequency threshold is 0.3, and the second frequency threshold is 0.05. In the part-of-speech tag frequency sequence, if the total frequency of a certain part-of-speech tag is higher than 0.3 or lower than 0.05, then that part-of-speech tag is added to the set of part-of-speech tags to be filtered. In this way, the part-of-speech tags that need to be filtered can be accurately identified, providing a clear target for subsequent redundant word filtering operations.

[0039] Step S135: Traverse each word in the text data set after tagging the part of speech, match the part of speech tag corresponding to the word with the set of part of speech tags to be filtered, remove the successfully matched words and their part of speech tags, and generate an intermediate filtered text data set.

[0040] The process involves iterating through each word in the tagged text dataset, checking each word and its corresponding part-of-speech tag sequentially. The part-of-speech tag of each word is then matched against the set of part-of-speech tags to be filtered. If a match is found, the word is considered redundant and needs to be removed. The intermediate filtered text dataset is the remaining text dataset after removing the successfully matched words and their part-of-speech tags.

[0041] For example, in the text data set with词性annotations, there are words such as "的" (auxiliary word, belonging to the set of词性tags to be filtered) and "企业" (noun, not belonging to the set of词性tags to be filtered). Check the词性tags of each word in sequence. When encountering "的", it is found that its词性tag matches successfully with the auxiliary word in the set of词性tags to be filtered, and "的" and its词性tag are removed; when encountering "企业", it is found that its词性tag does not match the set of词性tags to be filtered, and the word and its词性tag are retained. After traversal and matching operations, an intermediate filtered text data set is generated, which removes some redundant words and the data is more concise.

[0042] Step S136: Perform consecutive词性tag coherence detection on the remaining words in the intermediate filtered text data set. When the combination of词性tags of adjacent words conforms to the preset invalid grammar rules, remove the adjacent word combination to generate a grammar-corrected filtered text data set.

[0043] Consecutive词性tag coherence detection refers to checking whether the combination of词性tags of adjacent words in the intermediate filtered text data set conforms to grammar rules. The preset invalid grammar rules are some predefined rules for combinations of词性tags that do not conform to normal grammar structures. The grammar-corrected filtered text data set is the text data set obtained after removing adjacent word combinations that do not conform to grammar rules.

[0044] For example, the preset invalid grammar rules stipulate that an adjective cannot be followed by two adverbs consecutively. In the intermediate filtered text data set, if the situation where "非常" (adverb), "极其" (adverb) follow "美丽" (adjective) consecutively occurs, it is considered that the adjacent word combination does not conform to grammar rules and is removed. By consecutive词性tag coherence detection and removing adjacent word combinations that do not conform to grammar rules, the text data can be further optimized, improving the quality and semantic accuracy of the data, and generating a grammar-corrected filtered text data set.

[0045] Step S137: According to the distribution density of词性tags of each word in the grammar-corrected filtered text data set, perform an intermittent retention process on words with the same词性tag repeated within the same sentence, and only retain the words that conform to the preset interval positions to generate a filtered and refined text data set.

[0046] The distribution density of词性tags refers to the frequency of each词性tag appearing within a set range in the grammar-corrected filtered text data set. The intermittent retention process refers to, for words with the same词性tag repeated within the same sentence, following the preset interval position rules and only retaining some of the words. The preset interval positions are the position rules preset for determining which words to retain.

[0047] For example, in a sentence within a filtered text dataset after grammatical correction, multiple adjectives appear: "red," "blue," and "yellow." Based on a predefined interval rule—such as retaining one adjective every other—it might only retain "red" and "yellow," removing "blue." This interval-based retention process removes excessive repetition of words corresponding to part-of-speech tags within the same sentence, making the text more concise and clear, and generating a filtered, streamlined text dataset.

[0048] Step S140: Perform instruction-level data format conversion processing on the filtered and simplified text data set to uniformly convert text data of different formats into a preset standardized text format, and generate a preprocessed enterprise text data set.

[0049] Instruction-level data format conversion refers to the format conversion operation performed on a filtered and concise text data set with the support of the ARM instruction set. Different text data formats may include different encoding formats, delimiter usage, paragraph structures, etc. A preset standardized text format is a predefined, unified text format that facilitates subsequent analysis and processing.

[0050] For example, the filtered, concise text dataset may contain text in different encoding formats, such as UTF-8 and GBK, as well as different delimiters, such as commas, periods, and semicolons. Through instruction-level data format conversion, utilizing the encoding conversion and delimiter processing capabilities of the ARM instruction set, these text data of different formats are uniformly converted into a preset standardized text format, such as uniformly encoding to UTF-8 and using preset delimiters. The resulting preprocessed enterprise text dataset has a unified format, providing a more convenient data foundation for subsequent feature extraction and evaluation.

[0051] As one implementation method, step S140 involves performing instruction-level data format conversion processing on the filtered, simplified text data set to uniformly convert text data of different formats into a preset standardized text format, generating a preprocessed enterprise text data set. This may specifically include the following steps:

[0052] Step S141: Traverse each text data unit in the filtered and simplified text data set, identify the original format identifier of each text data unit, and generate a format conversion instruction sequence based on the mapping relationship between the original format identifier and the preset standardized text format.

[0053] A text data unit refers to each independent text segment within a filtered and concise collection of text data. An original format identifier is information used to identify the original format of each text data unit, such as encoding format and delimiter type. A mapping relationship is a predefined correspondence between original format identifiers and preset standardized text formats. A format conversion instruction sequence is a set of instructions generated based on the mapping relationship to guide format conversion operations.

[0054] For example, each text data unit in the filtered, simplified text dataset is traversed, and its encoding format (UTF-8 or GBK), and the delimiter (comma or period), etc., is checked to determine its original format identifier. Based on a pre-defined mapping relationship, if the original format is GBK encoding and the preset standardized text format is UTF-8 encoding, an instruction to convert GBK encoding to UTF-8 encoding is generated; if the original delimiter is a comma and the preset standardized delimiter is a semicolon, an instruction to replace the comma with a semicolon is generated. These instructions are arranged in processing order to generate a format conversion instruction sequence.

[0055] Step S142: Perform parallel format parsing operation on the text data unit based on the ARM instruction set according to each instruction in the format conversion instruction sequence, extract the structured text blocks and unstructured text blocks in the text data unit, and generate the segmented text blocks.

[0056] Parallel format parsing based on the ARM instruction set refers to using the parallel computing capabilities of the ARM instruction set to simultaneously parse the format of multiple text data units. Structured text blocks refer to text fragments with a pre-defined structure and rules, such as tabular data and ordered lists; unstructured text blocks refer to text fragments without a clear structure, such as paragraph text and free text.

[0057] For example, text data units are parsed in parallel according to instructions in the format conversion instruction sequence. If the instruction requires the recognition of table data in the text, the parallel processing capability of the ARM instruction set is used to simultaneously recognize the table structure of multiple text data units, extracting the table portion as a structured text block and the remaining portion as an unstructured text block. After the parallel format parsing operation, the text data units are divided into structured and unstructured text blocks, generating segmented text blocks.

[0058] Step S143: Perform instruction-level delimiter insertion processing on each unstructured text block in the segmented text block set, insert dynamic delimiters according to the preset standardized text format delimiter rules, and generate an intermediate text block set marked with delimiters.

[0059] Instruction-level delimiter insertion refers to the insertion of delimiters into unstructured text blocks with the support of the ARM instruction set. Predefined delimiter rules for standardized text formats are predefined rules used to determine the insertion position and type of delimiters. Dynamic delimiters refer to delimiters that are dynamically inserted based on the text content and delimiter rules.

[0060] For example, for unstructured text blocks in the segmented text block set, a predefined standard text format delimiter rule is used, such as inserting a semicolon every 10 characters. Using instruction-level operations of the ARM instruction set, each unstructured text block is traversed, and semicolons are inserted at appropriate positions. After this delimiter insertion process, an intermediate text block set containing dynamic delimiters is generated, giving the unstructured text blocks a clearer structure.

[0061] Step S144: Input each text block in the intermediate text block set into the ARM instruction set optimized encoding conversion pipeline. According to the encoding requirements of the format, perform unified character set conversion on the mixed-encoded text characters to generate a text block set with normalized encoding.

[0062] An ARM instruction set-optimized encoding conversion pipeline refers to an encoding conversion process optimized using the characteristics of the ARM instruction set. Mixed-encoding text characters refer to text blocks containing characters with multiple different encoding formats. Unified character set conversion processing converts mixed-encoding text characters into a unified character set encoding.

[0063] For example, a text block in the intermediate text block set might contain characters encoded in both UTF-8 and GBK. This text block is input into an ARM instruction set-optimized encoding conversion pipeline. Based on preset standardized text format encoding requirements, such as uniformly converting to UTF-8 encoding, the pipeline converts the mixed-encoding text characters. This pipelined processing improves encoding conversion efficiency and generates a set of text blocks with normalized encoding, ensuring all text blocks have a unified encoding format.

[0064] Step S145: Perform syntax structure verification processing on the instruction level of the encoded normalized text block, remove redundant character sequences that do not conform to the standardized text format syntax rules, and generate the final text block set after syntax correction.

[0065] Instruction-level syntax verification refers to checking the syntax structure of encoded, normalized text blocks with the support of the ARM instruction set. Standardized text format syntax rules are predefined syntax rules used to standardize text format. Redundant character sequences refer to extra character combinations that do not conform to the syntax rules.

[0066] For example, for text blocks after encoding normalization, the system checks for redundant spaces, non-standard punctuation marks, etc., according to standardized text format syntax rules. Using instruction-level operations of the ARM instruction set, the system performs syntax structure verification on each text block, removing redundant character sequences that do not conform to the rules. After syntax structure verification, a final set of grammatically corrected text blocks is generated, making the syntax structure of the text blocks more standardized.

[0067] Step S146: According to the paragraph splicing rules of the standardized text format, the text blocks in the final text block set are reorganized at the instruction level according to the semantic order of the original text data units to generate the reorganized standardized paragraph set.

[0068] Standardized text format paragraph concatenation rules are predefined rules used to combine text blocks into paragraphs. The semantic order of the original text data units refers to the order and semantic association of text blocks within the original text data set. Instruction-level reassembly refers to the recombination of text blocks according to rules, supported by the ARM instruction set.

[0069] For example, based on standardized text format paragraph concatenation rules, such as each paragraph ending with a period, the text blocks in the final text block set are concatenated according to the semantic order of the original text data units. Using instruction-level operations of the ARM instruction set, related text blocks are sequentially connected to form paragraphs. After instruction-level reorganization, a reorganized standardized paragraph set is generated, giving the text a clearer paragraph structure.

[0070] Step S147: Perform instruction-level binding processing on each paragraph in the standardized paragraph set with the metadata tags of the corresponding original text data units to generate a preprocessed enterprise text data set containing complete format identifiers.

[0071] Metadata tags are additional information used to describe raw text data units, such as source, date, and author. Instruction-level binding processing refers to associating standardized paragraphs with corresponding metadata tags, supported by the ARM instruction set. A complete format identifier is a unified identifier that includes paragraph content and metadata tags.

[0072] For example, for each paragraph in a standardized paragraph set, the metadata tag of its corresponding original text data unit is found. If the paragraph originates from a news report, its metadata tag includes information such as the report's date and author. Using instruction-level operations of the ARM instruction set, the paragraphs are bound to the metadata tags, generating a preprocessed enterprise text data set containing complete format identifiers. In this way, each paragraph has a complete information description, facilitating subsequent management and analysis.

[0073] Step S200: Input the preprocessed enterprise text data set into the pre-trained enterprise feature extraction model, and extract hierarchical semantic features from the enterprise text data set through the multi-layer semantic analysis structure in the enterprise feature extraction model to generate an enterprise semantic feature vector set.

[0074] The preprocessed enterprise text data set is a collection of text data with a unified format and refined content obtained after step S100. The pre-trained enterprise feature extraction model is trained on a large amount of historical data and is used to extract valuable feature information from the text data. A multi-layer semantic analysis structure is one structure within the enterprise feature extraction model; it can perform multi-layer semantic analysis on text data to mine deep semantic information. Hierarchical semantic feature extraction refers to extracting semantic features from text data according to different levels and dimensions. The enterprise semantic feature vector set is a set of vectors composed of the extracted semantic features, representing the semantic features of the enterprise text data.

[0075] For example, a pre-processed dataset of enterprise text data is input into a pre-trained enterprise feature extraction model. This model's multi-layered semantic analysis structure first performs shallow semantic analysis to identify basic concepts and themes within the text; then, it conducts deeper semantic analysis to uncover implicit semantic relationships and sentiment tendencies. Through hierarchical semantic feature extraction, semantic features of the text are extracted from different levels and dimensions, such as the enterprise's market competitiveness, development trends, and product advantages. Finally, a set of enterprise semantic feature vectors is generated, which can be used for subsequent enterprise evaluation and analysis.

[0076] As one implementation method, the training process of the enterprise feature extraction model provided in this embodiment of the invention may specifically include:

[0077] Step S201: Obtain and merge historical enterprise text datasets. Divide the historical enterprise text dataset into time series data to generate training text dataset subsets and validation text dataset subsets.

[0078] Historical corporate text datasets refer to a large amount of textual data related to a company over a past period. This data can originate from the company's historical annual reports, news reports, industry research, etc. Time series partitioning refers to dividing the historical corporate text dataset into different subsets according to chronological order. The training subset is used to train the corporate feature extraction model, while the validation subset is used to validate the model's performance.

[0079] For example, we can obtain a collection of historical text data from the past five years of a company, and divide it chronologically into a subset of data from the first four years for training and a subset of data from the last year for validation. This time-series division allows the model to learn text features from different time periods during training, while the validation subset is used to evaluate the model's generalization ability and accuracy.

[0080] Step S202: Perform batch word segmentation on the training text data subset using the batch processing unit under the ARM instruction set architecture to generate a training word segmentation data set, and perform word vector mapping on the training word segmentation data set to generate a training word vector set.

[0081] In the ARM instruction set architecture, a batch processing unit refers to a hardware or software module based on the ARM instruction set capable of processing multiple data units simultaneously. Batch word segmentation refers to simultaneously performing word segmentation operations on multiple text data sets within a subset of training text data. The training word segmentation dataset is the set of words obtained after word segmentation. Word vector mapping processes map each word to a vector, allowing words to be represented and computed in a vector space. The training word vector set is the collection of word vectors obtained through mapping.

[0082] For example, using a batch processing unit under the ARM instruction set architecture, a large amount of text data in a subset of training text data can be segmented simultaneously. The text data is divided into individual words, generating a training segmentation dataset. Then, a pre-trained word vector model, such as Word2Vec, is used to map each word in the training segmentation dataset into a vector, generating a training word vector set. In this way, words can participate in model training in vector form, facilitating the model's learning of semantic relationships between words.

[0083] Step S203: Construct an initial enterprise feature extraction model, which includes a sequentially connected word embedding layer, a bidirectional semantic analysis layer, and a feature fusion layer.

[0084] The initial enterprise feature extraction model is the model structure built before training begins. The word embedding layer further transforms and represents the input word vectors, enriching the semantic information of the words. The bidirectional semantic analysis layer can simultaneously consider the context of words, performing a more comprehensive semantic analysis. The feature fusion layer is used to fuse features from different levels and dimensions to generate a comprehensive feature representation.

[0085] For example, when building an initial enterprise feature extraction model, a word embedding layer is first defined, which maps the input word vectors to a higher-dimensional space, enhancing the semantic representation of words. Then, a bidirectional semantic analysis layer is connected, which can use structures such as Bidirectional Recurrent Neural Networks (Bi-RNN) or Bidirectional Long Short-Term Memory Networks (Bi-LSTM) to perform contextual semantic association analysis on the word vectors. Finally, a feature fusion layer is connected, which can use structures such as Convolutional Neural Networks (CNN) or fully connected layers to fuse the features output from the bidirectional semantic analysis layer, generating a comprehensive feature representation.

[0086] Step S204: Input the training word vector set into the initial enterprise feature extraction model, and perform contextual semantic association analysis on the training word vector set through the attention mechanism in the bidirectional semantic analysis layer to generate contextual association feature vectors.

[0087] Attention mechanisms are mechanisms that automatically focus on important parts of text. They assign different weights to each word based on its contextual information, thereby better capturing the semantic relationships between words. Contextual semantic association analysis refers to analyzing the semantic relationships and interactions of words within their context. Contextual association feature vectors are feature vectors obtained through attention mechanism analysis that reflect the semantic relationships between words within their context.

[0088] For example, the training set of word vectors is input into the bidirectional semantic analysis layer of the initial enterprise feature extraction model. The attention mechanism in the bidirectional semantic analysis layer first performs positional encoding on each word vector in the training set, adding positional information to each word vector to distinguish the order of words in a sentence. Then, multiple parallel attention heads calculate semantic weights of different dimensions on the word vector sequence, assigning different weights to each word vector. Based on these weights, the word vector sequence undergoes dynamic weighted fusion processing to generate a preliminary contextual feature vector. Finally, the preliminary contextual feature vector is input into a feedforward neural network module for nonlinear transformation and layer normalization to generate a standardized contextual feature vector.

[0089] As one implementation method, step S204, which involves performing contextual semantic association analysis on the training word vector set through the attention mechanism in the bidirectional semantic analysis layer to generate contextual association feature vectors, may specifically include the following steps:

[0090] Step S2041: Perform position encoding on each word vector in the training word vector set to generate a word vector sequence with position information.

[0091] Positional encoding adds positional information to each word vector to distinguish the order of words in a sentence. Because word order is crucial for semantic understanding in natural language, positional encoding allows models to better capture the contextual information of words. A sequence of word vectors with positional information is a sequence of word vectors that have undergone positional encoding.

[0092] For example, for each word vector in the training word vector set, positional encoding is performed using methods such as sine and cosine functions. Assuming a sentence has 10 words, each corresponding to a word vector, positional encoding information is added to each word vector. After positional encoding, a sequence of word vectors with positional information is generated. This sequence allows the model to know the position of each word in the sentence, thereby enabling better contextual semantic association analysis.

[0093] Step S2042: Input the word vector sequence with positional information into the multi-head attention module in the bidirectional semantic analysis layer. Calculate the semantic weights of the word vector sequence in different dimensions through multiple parallel attention heads to generate a multi-head attention weight matrix.

[0094] The multi-head attention module is a crucial component of the bidirectional semantic analysis layer, comprising multiple parallel attention heads. Each attention head can calculate semantic weights on the word vector sequence from different dimensions, thereby capturing richer semantic information. The multi-head attention weight matrix is ​​a matrix composed of weights calculated by multiple attention heads.

[0095] For example, a sequence of word vectors with location information can be input into a multi-head attention module in a bidirectional semantic analysis layer. Assume this module has eight attention heads, each calculating semantic weights independently for the word vector sequence. Each attention head assigns a weight to each word vector by calculating similarity between word vectors, etc. The weights calculated by these eight attention heads are combined to generate a multi-head attention weight matrix. This matrix reflects the importance of word vectors across different dimensions.

[0096] Step S2043: Perform dynamic weighted fusion processing on the word vector sequence based on the multi-head attention weight matrix to generate preliminary context feature vectors.

[0097] Dynamic weighted fusion processing refers to the weighted summation of each word vector in a word vector sequence based on the weights in the multi-head attention weight matrix, thereby fusing semantic information from different dimensions. The initial context feature vector is the feature vector obtained after dynamic weighted fusion processing, which reflects the contextual semantic relationships of words.

[0098] For example, based on the weights in the multi-head attention weight matrix, a weighted sum is performed on each word vector in the word vector sequence with positional information. Assuming there are 10 word vectors in the sequence, each corresponding to a weight, each word vector is multiplied by its corresponding weight, and then these weighted word vectors are summed to obtain a preliminary contextual feature vector. Through this dynamic weighted fusion process, semantic information from different dimensions can be integrated to generate a more representative preliminary contextual feature vector.

[0099] Step S2044: Input the initial context feature vector into the feedforward neural network module in the bidirectional semantic analysis layer for nonlinear transformation processing to generate the enhanced context feature vector.

[0100] Feedforward neural network modules are a common neural network architecture that can perform nonlinear transformations on input feature vectors, increasing the model's expressive power. Nonlinear transformation processing refers to using nonlinear activation functions to transform the input feature vectors, enabling the model to learn more complex semantic relationships. The enhanced context feature vector, obtained after processing by the feedforward neural network module, possesses stronger semantic representation capabilities.

[0101] For example, the initial context feature vector is input into the feedforward neural network module in the bidirectional semantic analysis layer. This module contains multiple fully connected layers and non-linear activation functions, such as the ReLU function. The feedforward neural network module performs a series of linear transformations and non-linear activation operations on the initial context feature vector, mapping the input feature vector to a higher-dimensional space, thereby increasing the model's expressive power. After processing by the feedforward neural network module, an enhanced context feature vector is generated, which better reflects the contextual semantic relationships of words.

[0102] Step S2045: Perform layer normalization on the enhanced context feature vector to generate a standardized context-related feature vector.

[0103] Layer normalization is a method for standardizing feature vectors, which can make the mean of the feature vectors zero and the variance one, thereby speeding up model training and improving model stability. The standardized context-related feature vectors, obtained after layer normalization, have better numerical stability and comparability.

[0104] For example, layer normalization can be applied to the enhanced context feature vector. The mean and variance of the enhanced context feature vector are calculated, and then the mean is subtracted from each element and divided by the variance to obtain the standardized context-related feature vector. Layer normalization can make the model more stable during training, avoiding problems such as vanishing or exploding gradients, and improving the model's training performance.

[0105] Step S205: Perform feature fusion processing on the context-related feature vectors through multi-scale convolutional kernels in the feature fusion layer to generate fused training feature vectors. Then, perform iterative optimization of model parameters based on the difference between the fused training feature vectors and the validation text data subset to generate the trained enterprise feature extraction model.

[0106] Multi-scale convolutional kernels are a type of convolutional kernel used in feature fusion layers. They possess different scales and receptive fields, enabling them to capture feature information of varying sizes. Feature fusion processing involves convolving context-related feature vectors using multi-scale convolutional kernels to fuse feature information from different scales. The fused training feature vector, obtained after feature fusion, contains richer feature information. The difference between the fused training feature vector and the feature representation of the validation text data subset can be measured using metrics such as mean squared error. Iterative optimization of model parameters involves continuously adjusting the model's parameters based on the difference, thereby continuously improving the model's performance. The successfully trained enterprise feature extraction model is a high-performance model obtained after multiple iterations of optimization.

[0107] For example, context-dependent feature vectors are input into a feature fusion layer, and multi-scale convolutional kernels are used to perform convolution operations on them. Convolutional kernels of different scales can capture feature information of different sizes, such as local and global features. Through convolution, these feature information of different scales are fused to generate a fused training feature vector. Then, the difference between the fused training feature vector and a subset of validation text data is calculated, for example, using mean squared error as the difference metric. Based on the difference, the model parameters are iteratively optimized using methods such as backpropagation. This process is repeated until the difference reaches a small value, generating a trained enterprise feature extraction model.

[0108] Step S300: Input the set of enterprise semantic feature vectors into the pre-built enterprise evaluation model, and use the multi-dimensional evaluation network in the enterprise evaluation model to dynamically assign weights to the set of enterprise semantic feature vectors to generate a set of comprehensive enterprise evaluation indicators.

[0109] The enterprise semantic feature vector set is a set of vectors generated in step S200 that represents the semantic features of the enterprise's textual data. The pre-built enterprise evaluation model is designed for evaluating enterprises and includes a multi-dimensional evaluation network. This network analyzes and evaluates the enterprise semantic feature vector set from multiple dimensions. Dynamic weight allocation processing involves assigning different weights to each feature based on its characteristics and importance, and these weights can be dynamically adjusted according to actual conditions. The enterprise comprehensive evaluation index set is a set of multiple evaluation indicators obtained after dynamic weight allocation processing, comprehensively reflecting the enterprise's overall situation. For example, the enterprise semantic feature vector set is input into the pre-built enterprise evaluation model's multi-dimensional evaluation network. The multi-dimensional evaluation network first extracts multiple semantic feature subsets from the enterprise semantic feature vector set, such as market competitiveness, financial status, and innovation capabilities. Then, it constructs a multi-dimensional feature space based on the feature types of each semantic feature subset. In the multi-dimensional feature space, an adaptive weight allocator performs real-time weight calculations on each semantic feature subset, assigning different weights to each subset. Based on these weights, the corresponding semantic feature subsets are weighted and fused to generate a weighted comprehensive feature vector. Finally, the weighted comprehensive feature vector is input into the evaluation index generator, where it undergoes index mapping processing through a fully connected layer to generate a preliminary evaluation index set. The values ​​of each index in the preliminary evaluation index set are then normalized and mapped to a preset evaluation interval to generate the enterprise's comprehensive evaluation index set.

[0110] It is understandable that the enterprise evaluation model is used to integrate the semantic feature vectors of enterprises and output comprehensive evaluation indicators, while the multi-dimensional evaluation network is the key part of the model, responsible for dynamic weight allocation to more accurately reflect the importance of various characteristics of the enterprise.

[0111] In a feasible architecture, the enterprise evaluation model can adopt a deep learning-based architecture, such as combining a multilayer perceptron (MLP) and an attention mechanism. In the enterprise evaluation model, the input layer receives a set of semantic feature vectors of the enterprise, the hidden layer performs nonlinear transformations and feature combinations on the input features, and the output layer outputs preliminary evaluation metrics.

[0112] For example, the number of neurons in the input layer is the same as the dimension of the enterprise semantic feature vector to ensure complete reception of the input feature information. Multiple hidden layers can be set, and the number of neurons in each hidden layer can be adjusted according to the specific task and data characteristics. In the hidden layers, neurons perform a non-linear transformation on the input through activation functions (such as the ReLU function, i.e., the corrected linear unit, whose mathematical expression is f(x)=max(0, x)), thereby learning more complex feature relationships. The number of neurons in the output layer is determined by the number of metrics to be evaluated. The introduction of an attention mechanism allows the model to pay more attention to features that have a greater impact on the evaluation results. Specifically, the attention mechanism module can be set between the hidden layers of a multilayer perceptron. First, the output of the previous hidden layer is used as input and mapped to a new feature space through a linear transformation. Then, the attention weight of each feature vector is calculated. The weight calculation can be implemented through a fully connected layer and a softmax function. The softmax function transforms the input feature vector into a probability distribution, i.e., the attention weight. Finally, the attention weight and the feature vector are weighted and summed to obtain the feature vector processed by the attention mechanism, which is then input into the next hidden layer for further processing.

[0113] The multi-dimensional evaluation network is responsible for dynamically assigning weights to the set of semantic feature vectors of the enterprise. The multi-dimensional evaluation network can employ an adaptive weight allocation strategy, dynamically adjusting the weights of each feature based on the input features. One implementation method is to use a learnable parameter matrix for weight allocation.

[0114] Specifically, the multi-dimensional evaluation network can consist of a feature weight allocation module and an indicator generation module. The feature weight allocation module contains a learnable parameter matrix, the elements of which can be adjusted during training based on the input features. The input set of enterprise semantic feature vectors is multiplied by the learnable parameter matrix to obtain the weight of each feature. The indicator generation module then generates specific evaluation indicators based on the weighted features. The indicator generation module can employ non-linear transformation layers, such as fully connected layers, and activation functions (such as the sigmoid function), to perform non-linear transformations on the weighted features, mapping them to the evaluation indicator space. For example, the weighted feature vectors can be input into a fully connected layer, and the output of this fully connected layer can be transformed by the sigmoid function, restricting the output value to the range [0, 1], serving as a preliminary evaluation indicator.

[0115] As one implementation method, step S300, which involves dynamically assigning weights to the set of semantic feature vectors of an enterprise through a multi-dimensional evaluation network in the enterprise evaluation model to generate a set of comprehensive enterprise evaluation indicators, may specifically include the following steps:

[0116] Step S310: Extract semantic feature subsets of multiple dimensions from the set of enterprise semantic feature vectors, and construct a multi-dimensional feature space according to the feature type of each semantic feature subset.

[0117] A semantic feature subset refers to a set of feature vectors with the same or similar feature types extracted from the set of semantic feature vectors of an enterprise. A multidimensional feature space is a space composed of multiple semantic feature subsets, which can be used to analyze and evaluate different dimensions of an enterprise.

[0118] For example, semantic feature subsets across multiple dimensions, such as market competitiveness, financial status, and innovation capability, can be extracted from a set of corporate semantic feature vectors. The semantic feature subset for market competitiveness might include feature vectors such as market share and brand awareness; the semantic feature subset for financial status might include feature vectors such as operating revenue and profit; and the semantic feature subset for innovation capability might include feature vectors such as R&D investment and number of patents. Based on the feature types of these semantic feature subsets, a multi-dimensional feature space can be constructed, with each dimension representing a corresponding feature type. This allows for a more intuitive analysis of a company's performance across different dimensions.

[0119] Step S320: In the multidimensional feature space, the adaptive weight allocator in the multidimensional evaluation network performs real-time weight calculation on each semantic feature subset to generate a dynamic feature weight set.

[0120] The adaptive weight allocator is a component in a multi-dimensional evaluation network that automatically adjusts the weights of each semantic feature subset based on the actual situation of the enterprise semantic feature vector set. Real-time weight calculation refers to dynamically calculating the weights of each semantic feature subset while processing the enterprise semantic feature vector set. The dynamic feature weight set is a collection composed of the weights of each semantic feature subset calculated in real time.

[0121] For example, in a multidimensional feature space, the adaptive weight allocator calculates the weights of each semantic feature subset in real time based on factors such as the feature values ​​and the correlation between features. Assuming the semantic feature subset representing market competitiveness is more important in the current enterprise evaluation, the adaptive weight allocator will assign a higher weight to this subset; while the semantic feature subset representing financial status is relatively less important in the current evaluation, and will be assigned a correspondingly lower weight. Through real-time weight calculation, a dynamic feature weight set is generated, which reflects the importance of each semantic feature subset under different circumstances.

[0122] Step S330: Perform weighted fusion processing on the corresponding semantic feature subsets according to the weight values ​​in the dynamic feature weight set to generate a weighted comprehensive feature vector.

[0123] Weighted fusion processing involves multiplying the feature vectors of each semantic feature subset by their corresponding weight values, and then summing these weighted feature vectors to obtain a comprehensive feature vector. The weighted comprehensive feature vector is the vector obtained after weighted fusion processing; it contains information from each semantic feature subset and is synthesized according to the weight values.

[0124] For example, based on the weight values ​​in the dynamic feature weight set, a weighted fusion process is performed on semantic feature subsets for dimensions such as market competitiveness, financial condition, and innovation capability. Assume the weight of the semantic feature subset for the market competitiveness dimension is 0.4, the weight for the semantic feature subset for the financial condition dimension is 0.3, and the weight for the semantic feature subset for the innovation capability dimension is 0.3. The feature vectors of each dimension's semantic feature subset are multiplied by their corresponding weight values, and then these weighted feature vectors are summed to obtain a weighted comprehensive feature vector. This vector comprehensively reflects the company's performance across different dimensions.

[0125] Step S340: Input the weighted comprehensive feature vector into the evaluation index generator in the multi-dimensional evaluation network, and perform index mapping processing through the fully connected layer in the evaluation index generator to generate a preliminary evaluation index set.

[0126] The evaluation metric generator is a component in a multi-dimensional evaluation network, used to transform a weighted composite feature vector into specific evaluation metrics. A fully connected layer is a common type of neural network layer that maps the input feature vector to an output space, generating corresponding evaluation metrics. Metric mapping refers to the process of converting the weighted composite feature vector into a preliminary set of evaluation metrics using a fully connected layer.

[0127] For example, the weighted composite feature vector is input into the evaluation metric generator in a multi-dimensional evaluation network. The fully connected layer in the evaluation metric generator contains multiple neurons, each corresponding to an evaluation metric. The fully connected layer performs linear transformations and non-linear activation operations on the weighted composite feature vector, mapping the input feature vector to an output space and generating a preliminary set of evaluation metrics. This set includes multiple evaluation metrics, such as the company's overall score and risk level.

[0128] Step S350: Normalize the values ​​of each indicator in the preliminary evaluation indicator set, and map the normalized indicator values ​​to a preset evaluation interval to generate a comprehensive enterprise evaluation indicator set.

[0129] Normalization refers to standardizing the values ​​of each indicator in the initial evaluation indicator set, making them have the same scale and range. The preset evaluation interval is a predefined range used to represent the evaluation results, such as [0, 100]. The comprehensive enterprise evaluation indicator set is a set of multiple evaluation indicators obtained after normalization and mapping. These indicator values ​​are within the preset evaluation interval, facilitating comparison and analysis.

[0130] For example, the values ​​of each indicator in the initial evaluation indicator set are normalized, such as using a min-max normalization method, mapping each indicator value to the range [0, 1]. Then, the normalized indicator values ​​are multiplied by the length of a preset evaluation interval and the lower limit of the evaluation interval is added to map the indicator values ​​into the preset evaluation interval. Assuming the preset evaluation interval is [0, 100], multiplying the normalized indicator values ​​by 100 yields indicator values ​​within the [0, 100] interval. After this processing, a comprehensive enterprise evaluation indicator set is generated, and the indicator values ​​in this set can intuitively reflect the overall situation of the enterprise.

[0131] As one implementation method, the training process of the multi-dimensional evaluation network provided in this embodiment of the invention may include:

[0132] Step S301: Obtain historical enterprise evaluation datasets and perform feature annotation processing on the historical enterprise evaluation datasets to generate an labeled evaluation training dataset.

[0133] Historical enterprise assessment datasets refer to the collection of data obtained from past enterprise assessments, containing various enterprise characteristics and assessment results. Feature annotation processing involves labeling each data sample in the historical enterprise assessment dataset with its corresponding actual assessment result, such as the enterprise's risk level and overall score. The labeled assessment training dataset is the dataset obtained after feature annotation processing, and it can be used to train multi-dimensional assessment networks.

[0134] For example, acquire a set of enterprise assessment data from the past 10 years, including financial statements, market data, news reports, etc. Perform feature annotation on each data sample, labeling its risk level as high, medium, or low, or its overall score as 80, 90, etc., based on the enterprise's actual situation. After feature annotation, a labeled assessment training dataset is generated, which can be used as training data for a multi-dimensional assessment network.

[0135] Step S302: Construct an initial multidimensional evaluation network, which includes a feature weight allocation module and an index generation module.

[0136] The initial multi-dimensional evaluation network is the network structure built before training begins. The feature weight allocation module is used to assign weights to the input features, determining the importance of each feature in the evaluation. The metric generation module is used to generate specific evaluation metrics based on the weighted features.

[0137] For example, when constructing the initial multidimensional evaluation network, a feature weight allocation module is first defined, which can adaptively allocate weights to input features using a learnable parameter matrix. Then, an index generation module is defined, which can use a nonlinear transformation layer to process the dynamic weight allocation results and generate a set of predicted evaluation indices. These two modules are connected to construct the initial multidimensional evaluation network.

[0138] Step S303: Input the labeled evaluation training data set into the initial multi-dimensional evaluation network, and perform adaptive weight allocation on the input features through the learnable parameter matrix in the feature weight allocation module to generate dynamic weight allocation results.

[0139] The learnable parameter matrix is ​​a matrix within the feature weight allocation module, and its parameters can be continuously learned and adjusted during training. Adaptive weight allocation refers to automatically assigning different weights to each feature based on the characteristics of the input features and the values ​​of the learnable parameter matrix. The result of dynamic weight allocation is the set of weights for each feature obtained after adaptive weight allocation.

[0140] For example, labeled evaluation training data is input into the feature weight allocation module of the initial multi-dimensional evaluation network. The learnable parameter matrix in the feature weight allocation module adaptively assigns weights to each feature based on the numerical value of the input feature and its own parameters. Assuming that the input features include market share, operating revenue, R&D investment, etc., the learnable parameter matrix will assign different weights to each feature according to the importance and relevance of these features, generating a dynamic weight allocation result.

[0141] Step S304: Perform feature transformation processing on the dynamic weight allocation results through the nonlinear transformation layer in the index generation module to generate a set of prediction and evaluation indicators.

[0142] The nonlinear transformation layer is a component in the index generation module. It performs nonlinear transformations on the input features, increasing the model's expressive power. Feature transformation processing refers to the process of converting the dynamic weight allocation results into a set of prediction and evaluation indicators through the nonlinear transformation layer.

[0143] For example, the dynamic weight allocation results are input into the nonlinear transformation layer in the indicator generation module. The nonlinear transformation layer can use nonlinear activation functions such as ReLU and Sigmoid to transform the input features, mapping them to a higher-dimensional space and increasing the model's expressive power. After processing by the nonlinear transformation layer, a set of predictive evaluation indicators is generated, containing the multi-dimensional evaluation network's predictive evaluation results for the enterprise.

[0144] Step S305: Calculate the difference between the predicted evaluation index set and the labeled results in the labeled evaluation training data set, and update the values ​​of the learnable parameter matrix according to the difference through the backpropagation algorithm.

[0145] The dissimilarity refers to the difference between the predicted evaluation metric set and the labeled results in the labeled evaluation training data set, and can be measured using metrics such as mean squared error and cross-entropy loss. Backpropagation is an algorithm used to train neural networks; it calculates the gradient of each parameter based on the dissimilarity and updates the values ​​of the learnable parameter matrix based on the gradients.

[0146] For example, the mean squared error between the predicted evaluation metric set and the labeled results in the labeled evaluation training dataset is calculated as the dissimilarity. Using the backpropagation algorithm, the gradient of each parameter in the learnable parameter matrix is ​​calculated based on the mean squared error. Based on the gradients, an optimization algorithm, such as stochastic gradient descent, is used to update the values ​​of the learnable parameter matrix. By continuously updating the values ​​of the learnable parameter matrix, the dissimilarity between the predicted evaluation metric set and the labeled results gradually decreases.

[0147] Step S306: Repeat the dynamic weight allocation and parameter update process until the difference reaches the preset convergence threshold, and generate the trained multi-dimensional evaluation network.

[0148] The preset convergence threshold is a pre-defined value used to determine whether the model has converged. When the difference reaches the preset convergence threshold, it indicates that the model's performance has reached a good level, and training can stop. The trained multi-dimensional evaluation network is a network with good performance obtained after multiple dynamic weight allocation and parameter update processes.

[0149] For example, steps S303 to S305 are repeated, continuously performing dynamic weight allocation and parameter updates. After each update, the difference between the predicted evaluation index set and the labeled results is calculated. When the difference is less than a preset convergence threshold, it indicates that the model has converged, training stops, and a multi-dimensional evaluation network that has been trained is generated. This network can be used to accurately evaluate enterprises.

[0150] Step S400: Based on the relationships between the various evaluation indicators in the enterprise comprehensive evaluation indicator set, construct a tree structure of enterprise evaluation results, and filter the results layer by layer through the evaluation thresholds of each node in the tree structure to generate the final enterprise credit risk assessment report.

[0151] The relationships between the various evaluation indicators in the comprehensive enterprise evaluation indicator set refer to the mutual influence and dependence among different evaluation indicators. The enterprise evaluation result tree structure is a tree-like structure that visually displays the hierarchical and interrelationships between the evaluation indicators. Evaluation thresholds are pre-set values ​​used to determine whether a node meets the requirements. Layer-by-layer result filtering refers to starting from the root node of the tree structure and filtering each node sequentially according to the hierarchy, retaining only nodes that meet the evaluation thresholds. The final enterprise credit risk assessment report is a report containing the enterprise credit risk assessment results generated after layer-by-layer result filtering.

[0152] For example, a tree structure of enterprise evaluation results can be constructed based on the relationships between the various evaluation indicators in the comprehensive enterprise evaluation indicator set. Assume the comprehensive enterprise evaluation indicator set includes indicators such as market competitiveness, financial condition, and innovation capability, and these indicators have pre-defined relationships. Using market competitiveness as the root node, a tree structure is constructed with sub-indicators related to market competitiveness as child nodes. A corresponding evaluation threshold is set for each node in the tree structure; for example, the evaluation threshold for the market competitiveness node is 80 points. Starting from the root node, the node matching degree is calculated based on the actual values ​​of each indicator in the comprehensive enterprise evaluation indicator set, and child node paths with matching degrees higher than the pre-defined threshold are selected. The evaluation results corresponding to the finally selected child node paths are summarized and integrated to generate a final enterprise credit risk assessment report containing multi-dimensional evaluation information.

[0153] As one implementation method, step S400, constructing a tree structure of enterprise assessment results, and filtering the results layer by layer through the assessment thresholds of each node in the tree structure to generate the final enterprise credit risk assessment report, may specifically include the following steps:

[0154] Step S410: Construct an initial evaluation node based on the evaluation values ​​of each indicator in the enterprise comprehensive evaluation indicator set, and use the initial evaluation node as the root node of the tree structure.

[0155] The initial evaluation node is constructed based on the evaluation values ​​of each indicator in the enterprise's comprehensive evaluation indicator set, and it represents the overall evaluation status of the enterprise. The root node of the tree structure is the starting node of the tree structure, and all child nodes are associated with the root node.

[0156] For example, an initial evaluation node can be constructed based on the evaluation values ​​of indicators such as market competitiveness, financial condition, and innovation capability from the enterprise comprehensive evaluation indicator set. Assuming the evaluation value for market competitiveness is 85 points, the evaluation value for financial condition is 80 points, and the evaluation value for innovation capability is 75 points, these evaluation values ​​are considered comprehensively to construct an initial evaluation node representing the overall evaluation of the enterprise. This initial evaluation node serves as the root node of a tree structure, and subsequent child nodes will revolve around this root node.

[0157] Step S420: Starting from the root node, recursively divide each evaluation index into levels according to the preset hierarchical division rules to generate a tree-like hierarchical structure containing multiple child nodes.

[0158] The preset hierarchical partitioning rules are predefined rules used to hierarchically divide evaluation indicators. Recursive hierarchical partitioning refers to starting from the root node and continuously dividing the evaluation indicator into sub-indicators until the preset partitioning depth is reached. The tree-like hierarchical structure is a tree structure composed of a root node and multiple child nodes, which can clearly show the hierarchical relationship between various evaluation indicators.

[0159] For example, starting from the root node, the evaluation indicators are recursively hierarchically divided according to a preset hierarchical division rule. Assume the preset rule divides market competitiveness indicators into sub-indicators such as market share and brand awareness, and financial status indicators into sub-indicators such as operating revenue and profit. Starting with the market competitiveness indicator at the root node, it is divided into market share and brand awareness sub-indicators. A corresponding child node is created for each sub-indicator, and the child nodes are connected to the root node. The indicators corresponding to each child node are clustered and ranked by importance, generating an indicator importance sequence. The child nodes are then prioritized according to the indicator importance sequence, generating an ordered tree-like hierarchical structure. This clustering analysis and child node creation process is repeated until a preset tree structure depth threshold is reached.

[0160] As one implementation method, step S420, which involves recursively dividing each evaluation index according to a preset hierarchical division rule to generate a tree-like hierarchical structure containing multiple child nodes, may specifically include the following steps:

[0161] Step S421: Starting from the current level node, perform cluster analysis on the indicators based on the correlation coefficients of each indicator in the enterprise comprehensive evaluation indicator set to generate multiple indicator cluster groups.

[0162] The correlation coefficient is a numerical value used to measure the correlation between two evaluation indicators, reflecting the degree of mutual influence between them. Indicator cluster analysis refers to grouping evaluation indicators according to their correlation coefficients, ensuring that indicators within the same group have a high degree of correlation. The resulting indicator cluster group is the set of indicator groups obtained after indicator cluster analysis.

[0163] For example, starting from the current level node, calculate the correlation coefficients of each indicator in the enterprise's comprehensive evaluation indicator set. Assume the current level node is a market competitiveness indicator, which includes sub-indicators such as market share, brand awareness, and customer satisfaction. Calculate the correlation coefficients between these sub-indicators; for example, the correlation coefficient between market share and brand awareness is 0.8, and the correlation coefficient between market share and customer satisfaction is 0.6. Perform cluster analysis based on the correlation coefficients, grouping highly correlated indicators into one group, generating multiple indicator clusters. For instance, market share and brand awareness might be grouped together, and customer satisfaction might be grouped into another.

[0164] Step S422: Create corresponding child nodes for each indicator cluster group, and establish connection relationships between the child nodes and the current level nodes.

[0165] Child nodes are nodes in a tree-like hierarchical structure, representing a clustering or grouping of indicators. Establishing connections between child nodes and nodes at the current level clarifies the hierarchical relationships between indicators.

[0166] For example, create corresponding child nodes for each indicator cluster. For an indicator cluster consisting of market share and brand awareness, create a child node to represent that cluster; for an indicator cluster consisting of customer satisfaction, create another child node to represent that cluster. Connect these child nodes to the current level node (market competitiveness indicator node) to form part of a tree-like hierarchical structure.

[0167] Step S423: Perform importance ranking on the clustered groups of indicators corresponding to each child node to generate an indicator importance sequence.

[0168] Importance ranking refers to sorting indicators within a cluster based on their importance. The indicator importance sequence is the ordered list of indicators obtained after importance ranking.

[0169] For example, the importance of clustered indicators for each sub-node can be ranked. The importance of an indicator can be determined based on factors such as its impact on enterprise evaluation and data reliability. For an indicator cluster composed of market share and brand awareness, assuming market share has a greater impact on market competitiveness, market share is ranked first, followed by brand awareness, generating an indicator importance sequence.

[0170] Step S424: Sort the child nodes according to the importance sequence of the indicators to generate an ordered tree-like hierarchical structure.

[0171] Priority sorting refers to ranking child nodes according to their importance sequence, placing more important child nodes first. An ordered tree-like hierarchical structure is the tree structure obtained after priority sorting, which can more clearly show the importance and hierarchical relationship of each indicator.

[0172] For example, child nodes can be prioritized based on the importance sequence of indicators. For child nodes under the market competitiveness indicator node, based on the importance sequence of market share and brand awareness, child nodes representing market share and brand awareness are ranked first, and child nodes representing customer satisfaction are ranked last, generating an ordered tree-like hierarchical structure.

[0173] Step S425: Repeat the cluster analysis and child node creation process until the preset tree structure depth threshold is reached.

[0174] The preset tree structure depth threshold is a pre-defined value used to control the depth of the tree structure. When the depth of the tree structure reaches the preset threshold, the recursive hierarchical division stops.

[0175] For example, steps S421 to S424 are repeated to continuously perform index clustering analysis and child node creation. After each partition, it is checked whether the depth of the tree structure has reached the preset tree structure depth threshold. Assuming the preset tree structure depth threshold is 3 levels, when the tree structure reaches 3 levels, the recursive hierarchical partitioning stops, and the final tree hierarchy containing multiple child nodes is generated.

[0176] Step S430: Set a corresponding evaluation threshold range for each node in the tree-like hierarchical structure, and calculate the node matching degree based on the actual values ​​of each indicator in the enterprise comprehensive evaluation indicator set.

[0177] The evaluation threshold range is a pre-defined numerical range used to determine whether a node meets the requirements. Node matching degree calculation refers to calculating the degree of matching between the actual value of each node and the evaluation threshold range, based on the actual values ​​of each indicator in the enterprise's comprehensive evaluation indicator set.

[0178] For example, set a corresponding evaluation threshold range for each node in the tree-like hierarchical structure. For the market competitiveness node, set the evaluation threshold range to [70, 100]; for the market share sub-node, set the evaluation threshold range to [30%, 50%]. Based on the actual values ​​of each indicator in the enterprise comprehensive evaluation indicator set, calculate the degree of matching between the actual value of each node and the evaluation threshold range. Assuming the actual value of the market competitiveness indicator is 85 points and the actual value of the market share indicator is 40%, calculate their matching degree with the corresponding evaluation threshold range.

[0179] Step S440: Based on the node matching degree calculation result, start from the root node of the tree structure and traverse the nodes layer by layer to filter out the child node paths with a matching degree higher than the preset threshold.

[0180] The preset threshold is a pre-defined value used to filter nodes. Layer-by-layer node traversal refers to starting from the root node of the tree structure and traversing each node sequentially according to its level. The child node path refers to the path from the root node to a specific child node; it represents a set of relevant evaluation metrics.

[0181] For example, based on the node matching degree calculation results, the nodes are traversed layer by layer starting from the root node of the tree structure. For each node, its matching degree is checked to see if it is higher than a preset threshold. Assuming the preset threshold is 0.8, when the matching degree of the market competitiveness node is 0.9, which is higher than the preset threshold, the traversal of its child nodes continues; when the matching degree of the market share child node is 0.85, which is also higher than the preset threshold, the path of that child node is retained. By traversing the nodes layer by layer, child node paths with matching degrees higher than the preset threshold are filtered out.

[0182] Step S450: Summarize and integrate the evaluation results corresponding to the finally selected sub-node paths to generate a final corporate credit risk assessment report containing multi-dimensional evaluation information.

[0183] The final selected sub-node paths correspond to the evaluation results of the evaluation indicators represented by the sub-node paths that were retained after layer-by-layer screening. The aggregation and integration process involves organizing and merging these evaluation results to form a complete report. The final corporate credit risk assessment report is a report containing multi-dimensional evaluation information, which can comprehensively reflect the company's credit risk situation.

[0184] For example, the evaluation results corresponding to the finally selected sub-node paths are summarized and integrated. Assuming the selected sub-node paths include evaluation results for indicators such as market competitiveness, market share, and operating revenue, these results are organized and merged to analyze the company's performance across different dimensions. Based on the analysis results, a final corporate credit risk assessment report containing multi-dimensional evaluation information is generated. The report may include the company's overall score, risk level, strengths and weaknesses analysis, etc., providing a comprehensive and detailed reference for the company's credit risk assessment.

[0185] As one implementation method, the method provided in this embodiment of the invention further includes a continuous optimization process for the enterprise evaluation model, specifically comprising the following steps:

[0186] Step S307: Collect enterprise evaluation result data in real time and build an incremental training dataset.

[0187] Enterprise evaluation results data refers to various data generated during the enterprise evaluation process, including the values ​​of each indicator in the comprehensive enterprise evaluation indicator set and intermediate results during the evaluation process. Real-time collection means continuously collecting this evaluation results data during the enterprise's operation to reflect the latest status of the enterprise. The incremental training dataset is a dataset built from the real-time collected enterprise evaluation results data, used to continuously optimize the enterprise evaluation model.

[0188] For example, in the daily operations of a company, various data collection systems are used to collect real-time market sales data, financial statement data, customer feedback data, etc., and this data is processed and transformed into enterprise evaluation results data. These data collected in real-time over a period of time are then integrated to construct an incremental training dataset. This dataset contains various evaluation information of the company in the latest time period, reflecting the company's dynamic changes.

[0189] Step S308: Perform real-time feature extraction processing on the incremental training dataset through the data pipeline under the ARM instruction set architecture to generate an incremental feature vector set.

[0190] The ARM instruction set architecture-based data pipeline is a high-efficiency data processing architecture that breaks down the data processing process into multiple stages, each handled by a dedicated hardware or software module, thereby achieving parallel data processing and efficient data transmission. Real-time feature extraction processing refers to immediately extracting features from the received incremental training dataset to obtain key information. The incremental feature vector set is the set of feature vectors obtained after real-time feature extraction processing, containing important feature information from the incremental training dataset.

[0191] As one implementation method, step S308, which involves real-time feature extraction processing of the incremental training dataset through a data pipeline under the ARM instruction set architecture to generate an incremental feature vector set, may specifically include the following steps:

[0192] Step S3081: Load the incremental training dataset into the memory buffer using the parallel loading instruction in the ARM instruction set.

[0193] The parallel load instructions in the ARM instruction set can load multiple data units simultaneously, improving data loading efficiency. A memory buffer is an area used for temporary data storage; loading the incremental training dataset into the memory buffer facilitates subsequent data processing operations.

[0194] For example, by utilizing the parallel load instruction in the ARM instruction set, multiple data blocks from the incremental training dataset can be loaded into the memory buffer simultaneously. Suppose the incremental training dataset contains multiple parts such as a company's sales data, financial data, and market data; through the parallel load instruction, these data parts can be loaded into the memory buffer at the same time, reducing the time overhead of data loading.

[0195] Step S3082: Use the vectorization processing unit in the ARM NEON instruction set to perform batch feature calculation processing on the data in the memory buffer to generate an intermediate feature data set.

[0196] The ARM NEON instruction set is a SIMD (Single Instruction, Multiple Data) instruction set under the ARM architecture. It provides vectorized processing units that can perform the same operation on multiple data elements simultaneously. Batch feature computation processing refers to using vectorized processing units to perform feature computation on multiple data elements in a memory buffer simultaneously, improving the efficiency of feature computation. The intermediate feature data set is the intermediate result data set obtained after batch feature computation processing; it contains partial feature information from the data.

[0197] For example, the vectorization processing unit in the ARM NEON instruction set can be used to perform batch feature calculations on enterprise sales data in a memory buffer. It can simultaneously calculate the mean, standard deviation, and other feature values ​​of multiple sales data elements, generating an intermediate feature data set. Vectorization processing significantly improves the speed of feature calculation and accelerates the data processing process.

[0198] Step S3083: Perform instruction-level data alignment processing on the intermediate feature data set to ensure that the storage address of the feature data meets the requirements of vectorization operation.

[0199] Instruction-level data alignment refers to the rearrangement and storage of data in intermediate feature data sets, supported by the ARM instruction set, to ensure that their storage addresses conform to the requirements of vectorized operations. Vectorized operations typically require data to be stored at contiguous memory addresses, and the starting address of the data must satisfy predefined alignment rules. Data alignment improves the efficiency of vectorized operations and avoids performance degradation caused by incompatible data storage addresses.

[0200] For example, examining the data in an intermediate feature dataset reveals that the storage addresses of some data do not meet the requirements for vectorization operations. Using relevant instructions from the ARM instruction set, these data are rearranged and stored to ensure their storage addresses conform to alignment rules. This allows for more efficient data manipulation in subsequent vectorization processing, improving data processing performance.

[0201] Step S3084: The aligned feature data is processed by pipelined feature extraction using the ARM hardware prefetching mechanism to generate an incremental feature vector set.

[0202] ARM's hardware prefetching mechanism is a hardware-level optimization technique that prefetches data from memory into the cache before the processor needs it, reducing data access latency. Pipeline feature extraction processing refers to breaking down the feature extraction process into multiple stages, each handled by a dedicated hardware module, enabling continuous data processing. The incremental feature vector set is the final set of feature vectors obtained after pipelined feature extraction processing; it contains all the important feature information from the incremental training dataset.

[0203] For example, by utilizing ARM's hardware prefetching mechanism, data can be pre-fetched from memory into a cache when the processor needs aligned feature data for feature extraction. Then, through pipelined feature extraction processing, the data sequentially passes through multiple feature extraction stages, such as feature selection and feature transformation, ultimately generating an incremental feature vector set. This pipelined processing approach can improve the efficiency of feature extraction and make data processing smoother.

[0204] Step S309: Input the incremental feature vector set into the current enterprise evaluation model for prediction processing to generate an incremental prediction result set.

[0205] The current enterprise evaluation model refers to the enterprise evaluation model in the continuous optimization process; it is obtained through previous training and optimization. When the incremental feature vector set is input into the current enterprise evaluation model, the model calculates and analyzes the input feature vectors based on its own parameters and structure, generating prediction results. The incremental prediction result set is the set of results obtained by the current enterprise evaluation model from the incremental feature vector set; it contains the evaluation prediction of the enterprise's latest state.

[0206] For example, the generated set of incremental feature vectors is input into the current enterprise evaluation model. The model's multi-dimensional evaluation network dynamically assigns weights to the input feature vectors, using an adaptive weight allocator to assign different weights to each feature vector, and then performs weighted fusion to generate a weighted comprehensive feature vector. Next, the weighted comprehensive feature vector is input into the evaluation index generator, where it undergoes index mapping processing through a fully connected layer to generate a set of incremental prediction results. The prediction results in this set reflect the enterprise's comprehensive evaluation status in the latest time period.

[0207] Step S3010: Calculate the degree of difference between the incremental prediction result set and the actual evaluation result, and generate the model optimization signal based on the degree of difference.

[0208] Actual assessment results refer to the true assessment results of enterprises obtained through actual assessment methods or data collection. They can serve as a reference standard for measuring the accuracy of model predictions. The degree of discrepancy refers to the difference between the incremental prediction result set and the actual assessment results, which can be measured using various indicators such as mean squared error and absolute error. The model optimization signal is a signal generated based on the degree of discrepancy to guide the adjustment of model parameters; it contains information on the direction and magnitude of the model adjustments needed.

[0209] For example, the mean squared error between each predicted indicator in the incremental prediction result set and its corresponding indicator in the actual evaluation result is calculated as the degree of discrepancy. If the degree of discrepancy is large, it indicates that the model's prediction results deviate significantly from the actual situation, and the model needs to be adjusted. Based on the magnitude and direction of the degree of discrepancy, model optimization signals are generated, such as instructing the model to increase the weight of certain features or decrease the values ​​of certain parameters.

[0210] Step S3011: Fine-tune the parameters in the current enterprise evaluation model online based on the model optimization signal to generate an optimized enterprise evaluation model.

[0211] Online fine-tuning refers to adjusting some parameters of the current enterprise evaluation model based on the model's optimization signals without retraining the entire model. This fine-tuning can be performed in real-time or near real-time, enabling the model to quickly adapt to the dynamic changes of the enterprise. The optimized enterprise evaluation model, obtained after online fine-tuning, has better predictive performance and adaptability.

[0212] For example, based on model optimization signals, the learnable parameter matrix in the current enterprise assessment model is adjusted. If the model optimization signal indicates that the weights of certain features need to be increased, the parameter values ​​at the corresponding positions in the learnable parameter matrix are increased accordingly; if the values ​​of certain parameters need to be decreased, adjustments are made accordingly. Through this online fine-tuning process, the model can better fit the latest data of the enterprise, generating an optimized enterprise assessment model. This optimized model can be used for subsequent enterprise assessments, improving the accuracy and reliability of the assessment.

[0213] In summary, the enterprise assessment method based on natural language processing provided in this invention, through a series of steps including preprocessing of the enterprise's original text data, feature extraction, assessment model construction, and continuous optimization, can comprehensively and accurately assess enterprises and generate a final enterprise credit risk assessment report. Simultaneously, by collecting data in real time and continuously optimizing the model, the assessment model can adapt to the dynamic changes of enterprises, improving the timeliness and accuracy of the assessment. This method utilizes the high-efficiency computing power and parallel processing characteristics of the ARM instruction set, improving data processing efficiency and performance, and providing an effective solution for enterprise assessment.

[0214] In practical applications, this method can be applied to various fields, including credit assessment by financial institutions, investment decisions by investors, and self-assessment by enterprises. For example, when approving loans, financial institutions can use this method to assess a company's credit risk, thereby more accurately determining its repayment ability and creditworthiness and reducing credit risk. Investors can refer to the enterprise assessment report generated by this method when making investment decisions to understand a company's overall strength and development potential, making more informed investment choices. Enterprises themselves can also use this method for regular self-assessment, identifying their strengths and weaknesses, formulating corresponding development strategies, and improving their competitiveness.

[0215] Furthermore, this method can be combined with other technologies to further enhance the effectiveness of enterprise assessments. For example, it can be combined with big data technology to collect more enterprise-related data, enriching the data sources for assessment; and it can be combined with artificial intelligence technologies, such as deep learning and reinforcement learning, to further optimize the performance of the assessment model and improve the accuracy and intelligence of the assessment.

[0216] In the future, with the continuous development of natural language processing and computer technologies, this method can be further improved and perfected. For example, more effective text preprocessing methods can be explored to improve the quality and usability of text data; more advanced feature extraction and evaluation models can be researched to better capture the characteristics of enterprises and assess their risks; and the interpretability of the models can be strengthened to make the evaluation results easier to understand and apply. In summary, the enterprise evaluation method provided by the embodiments of this invention has broad application prospects and development potential, and is expected to bring new breakthroughs and developments to the field of enterprise evaluation.

[0217] Based on the foregoing embodiments, this invention provides an enterprise evaluation device. The units and modules included in the device can be implemented by a processor in a computer device; of course, they can also be implemented by specific logic circuits. In the implementation process, the processor can be a central processing unit (CPU), a microprocessor unit (MPU), a digital signal processor (DSP), or a field programmable gate array (FPGA), etc.

[0218] Figure 2 This is a schematic diagram of the composition structure of an enterprise evaluation device provided in an embodiment of the present invention, as shown below. Figure 2 As shown, the enterprise evaluation device 200 includes:

[0219] Data acquisition module 210 is used to acquire the original text data set of the enterprise to be evaluated, and preprocess the original text data set through ARM instruction set to generate a preprocessed enterprise text data set;

[0220] Feature extraction module 220 is used to input the preprocessed enterprise text data set into a pre-trained enterprise feature extraction model, and to extract hierarchical semantic features from the enterprise text data set through the multi-layer semantic analysis structure in the enterprise feature extraction model to generate an enterprise semantic feature vector set.

[0221] The indicator generation module 230 is used to input the set of enterprise semantic feature vectors into a pre-built enterprise evaluation model, and to perform dynamic weight allocation processing on the set of enterprise semantic feature vectors through the multi-dimensional evaluation network in the enterprise evaluation model to generate a set of comprehensive enterprise evaluation indicators.

[0222] The credit identification module 240 is used to construct a tree structure of enterprise evaluation results based on the correlation between the evaluation indicators in the enterprise comprehensive evaluation indicator set, and to filter the results layer by layer through the evaluation threshold of each node in the tree structure to generate the final enterprise credit risk assessment report.

[0223] The descriptions of the above device embodiments are similar to those of the above method embodiments, and have similar beneficial effects. In some embodiments, the functions or modules included in the device provided by the present invention can be used to execute the methods described in the above method embodiments. For technical details not disclosed in the device embodiments of the present invention, please refer to the descriptions of the method embodiments of the present invention for understanding. It should be noted that, in the embodiments of the present invention, if the above-mentioned enterprise evaluation method based on natural language processing is implemented in the form of software functional modules and sold or used as an independent product, it can also be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the embodiments of the present invention, or the part that contributes to the related technology, can be embodied in the form of a software product. This software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media that can store program code, such as USB flash drives, mobile hard drives, read-only memory (ROM), magnetic disks, or optical disks. Thus, the embodiments of the present invention are not limited to any specific hardware, software, or firmware, or any combination of hardware, software, and firmware.

[0224] The above description is merely an embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A business evaluation method based on natural language processing, characterized in that, The method includes: The process involves acquiring the original text data set of the enterprise to be evaluated, and preprocessing the original text data set using the ARM instruction set to generate a preprocessed enterprise text data set. Specifically, this includes: performing parallel text cleaning on the original text data set using SIMD instructions in the ARM instruction set to remove unstructured noise data and generate preliminary cleaned text data; performing instruction-level word segmentation on each text segment in the preliminary cleaned text data; and using the memory access optimization mechanism of the ARM instruction set to perform part-of-speech tagging on the segmented text segments to generate a part-of-speech tagged text data set. Finally, the process is repeated based on the number of part-of-speech tagged text segments. Based on the distribution characteristics of each part-of-speech tag in the set, a part-of-speech distribution histogram is constructed. Redundant words are filtered from the text data set using the frequency thresholds of each part-of-speech tag in the histogram, generating a filtered, concise text data set. The filtered, concise text data set undergoes instruction-level data format conversion processing, uniformly converting text data of different formats into a preset standardized text format, generating the preprocessed enterprise text data set. Specifically, the process of constructing a part-of-speech distribution histogram based on the distribution characteristics of each part-of-speech tag in the annotated text data set, and filtering redundant words from the text data set using the frequency thresholds of each part-of-speech tag in the histogram, is described. Generating a filtered, simplified text dataset includes: scanning each text segment in the tagged text dataset word by word, counting the occurrences of each part-of-speech tag in the text dataset, and generating a part-of-speech tag count set; based on the occurrences of each part-of-speech tag in the part-of-speech tag count set, according to a preset histogram interval division rule, accumulating the counts of the same part-of-speech tag into the corresponding histogram interval, generating a part-of-speech distribution histogram structure containing interval counts and the correspondence between part-of-speech tags; based on the count data of each histogram interval in the part-of-speech distribution histogram structure, calculating the total occurrence frequency of each part-of-speech tag, and generating part-of-speech tag frequency distribution histograms in descending order of frequency. The process involves: 1) traversing a frequency sequence; 2) matching the part-of-speech tags corresponding to the words with frequencies higher than a first frequency threshold or lower than a second frequency threshold in the part-of-speech tag frequency sequence; 3) performing continuous part-of-speech tag coherence detection on the remaining words in the intermediate filtered text data set; 4) removing adjacent word combinations that conform to preset invalid grammar rules when the combination of adjacent word combinations conforms to a preset invalid grammar rule; and 5) generating a grammar-corrected filtered text data set.Based on the part-of-speech tag distribution density of each word in the filtered text data set after grammatical correction, words with the same part-of-speech tag that appear repeatedly in the same sentence are retained intermittently, and only words that meet the preset interval position are retained, thus generating the filtered simplified text data set. The preprocessed enterprise text data set is input into a pre-trained enterprise feature extraction model. The enterprise text data set is then subjected to hierarchical semantic feature extraction through the multi-layer semantic analysis structure in the enterprise feature extraction model, generating an enterprise semantic feature vector set. The set of enterprise semantic feature vectors is input into a pre-built enterprise evaluation model. The set of enterprise semantic feature vectors is dynamically weighted through the multi-dimensional evaluation network in the enterprise evaluation model to generate a set of comprehensive enterprise evaluation indicators. Based on the relationships between the various evaluation indicators in the comprehensive enterprise evaluation indicator set, a tree structure of enterprise evaluation results is constructed. The results are then filtered layer by layer using the evaluation thresholds of each node in the tree structure to generate the final enterprise credit risk assessment report.

2. The method according to claim 1, characterized in that, The training process of the enterprise feature extraction model includes: The historical enterprise text dataset is acquired and merged. The historical enterprise text dataset is then divided into time series data to generate a training text dataset subset and a validation text dataset subset. The training text data subset is processed by batch word segmentation using a batch processing unit under the ARM instruction set architecture to generate a training word segmentation data set, and then word vector mapping is performed on the training word segmentation data set to generate a training word vector set. An initial enterprise feature extraction model is constructed, which includes a sequentially connected word embedding layer, a bidirectional semantic analysis layer, and a feature fusion layer. The training word vector set is input into the initial enterprise feature extraction model, and the training word vector set is subjected to contextual semantic association analysis through the attention mechanism in the bidirectional semantic analysis layer to generate contextual association feature vectors; The context-related feature vectors are fused using multi-scale convolutional kernels in the feature fusion layer to generate fused training feature vectors. The model parameters are then iteratively optimized based on the difference between the fused training feature vectors and the validation text data subset to generate a trained enterprise feature extraction model.

3. The method according to claim 2, characterized in that, The step of performing contextual semantic association analysis on the training word vector set through the attention mechanism in the bidirectional semantic analysis layer to generate contextual association feature vectors includes: Each word vector in the training word vector set is subjected to position encoding to generate a word vector sequence with position information; The word vector sequence with location information is input into the multi-head attention module in the bidirectional semantic analysis layer. The word vector sequence is subjected to semantic weights of different dimensions by multiple parallel attention heads to generate a multi-head attention weight matrix. The word vector sequence is dynamically weighted and fused according to the multi-head attention weight matrix to generate a preliminary context feature vector. The initial context feature vector is input into the feedforward neural network module in the bidirectional semantic analysis layer for nonlinear transformation processing to generate an enhanced context feature vector. The enhanced context feature vector is subjected to layer normalization to generate a standardized context association feature vector.

4. The method according to claim 1, characterized in that, The step of dynamically weighting the set of semantic feature vectors of the enterprise through the multi-dimensional evaluation network in the enterprise evaluation model to generate a set of comprehensive enterprise evaluation indicators includes: Extract semantic feature subsets of multiple dimensions from the set of enterprise semantic feature vectors, and construct a multi-dimensional feature space based on the feature type of each semantic feature subset; In the multidimensional feature space, the adaptive weight allocator in the multidimensional evaluation network performs real-time weight calculation on each semantic feature subset to generate a dynamic feature weight set. Based on the weight values ​​in the dynamic feature weight set, the corresponding semantic feature subsets are weighted and fused to generate a weighted comprehensive feature vector. The weighted comprehensive feature vector is input into the evaluation index generator in the multi-dimensional evaluation network. The index mapping process is performed through the fully connected layer in the evaluation index generator to generate a preliminary evaluation index set. The values ​​of each indicator in the preliminary evaluation indicator set are normalized, and the normalized indicator values ​​are mapped to a preset evaluation interval to generate the enterprise comprehensive evaluation indicator set.

5. The method according to claim 4, characterized in that, The training process of the multi-dimensional evaluation network includes: The historical enterprise evaluation dataset is acquired and merged. Feature annotation processing is performed on the historical enterprise evaluation dataset to generate an labeled evaluation training dataset. An initial multidimensional evaluation network is constructed, which includes a feature weight allocation module and an index generation module; The labeled evaluation training dataset is input into the initial multi-dimensional evaluation network, and the input features are adaptively weighted using the learnable parameter matrix in the feature weight allocation module to generate dynamic weight allocation results. The dynamic weight allocation result is processed by the nonlinear transformation layer in the index generation module to generate a set of prediction and evaluation indicators. Calculate the difference between the predicted evaluation index set and the labeled results in the labeled evaluation training data set, and update the value of the learnable parameter matrix according to the difference using the backpropagation algorithm; Repeat the dynamic weight allocation and parameter update process until the difference reaches a preset convergence threshold, and generate a trained multi-dimensional evaluation network.

6. The method according to claim 1, characterized in that, The process involves constructing a tree structure of enterprise assessment results, and then filtering the results layer by layer using assessment thresholds for each level of the tree structure to generate a final enterprise credit risk assessment report, including: An initial evaluation node is constructed based on the evaluation values ​​of each indicator in the enterprise comprehensive evaluation indicator set, and the initial evaluation node is used as the root node of the tree structure. Starting from the root node, each evaluation index is recursively divided into hierarchical levels according to the preset hierarchical division rules, generating a tree-like hierarchical structure containing multiple child nodes. Set a corresponding evaluation threshold range for each node in the tree-like hierarchical structure, and calculate the node matching degree based on the actual values ​​of each indicator in the enterprise comprehensive evaluation indicator set; Based on the node matching degree calculation result, the nodes are traversed layer by layer starting from the root node of the tree structure to filter out the child node paths with a matching degree higher than a preset threshold. The evaluation results corresponding to the finally selected sub-node paths are summarized and integrated to generate the final enterprise credit risk assessment report containing multi-dimensional evaluation information.

7. The method according to claim 6, characterized in that, The process of recursively dividing each evaluation index according to a preset hierarchical division rule to generate a tree-like hierarchical structure containing multiple child nodes includes: Starting from the current level node, cluster analysis is performed on each indicator in the enterprise comprehensive evaluation indicator set to generate multiple indicator cluster groups; Create corresponding child nodes for each index cluster group, and establish connection relationships between the child nodes and the current level nodes; For each child node, the indicators are clustered and grouped according to their importance, and an indicator importance sequence is generated. The child nodes are prioritized according to the importance sequence of the indicators to generate an ordered tree-like hierarchical structure; Repeat the clustering analysis and child node creation process until the preset tree structure depth threshold is reached.

8. The method according to claim 1, characterized in that, The method also includes a continuous optimization process for the enterprise evaluation model, including: Real-time collection of enterprise evaluation results data and construction of incremental training datasets; The incremental training dataset is processed in real time using a data pipeline under the ARM instruction set architecture to generate an incremental feature vector set. The incremental feature vector set is input into the current enterprise evaluation model for prediction processing to generate an incremental prediction result set. Calculate the degree of difference between the incremental prediction result set and the actual evaluation result, and generate a model optimization signal based on the degree of difference; Based on the model optimization signal, the parameters in the current enterprise evaluation model are fine-tuned online to generate an optimized enterprise evaluation model. The step of performing real-time feature extraction processing on the incremental training dataset through a data pipeline under the ARM instruction set architecture to generate an incremental feature vector set includes: The incremental training dataset is loaded into the memory buffer using parallel loading instructions from the ARM instruction set; The data in the memory buffer is processed in batches using the vectorization processing unit in the ARM NEON instruction set to generate an intermediate feature data set. The intermediate feature data set is subjected to instruction-level data alignment processing to ensure that the storage address of the feature data meets the requirements of vectorization operation. The aligned feature data is processed using a pipelined feature extraction mechanism based on ARM's hardware prefetching mechanism to generate the incremental feature vector set.

9. A business evaluation device, characterized in that, include: The data acquisition module is used to acquire the original text data set of the enterprise to be evaluated, and preprocess the original text data set using the ARM instruction set to generate a preprocessed enterprise text data set. Specifically, this includes: performing parallel text cleaning processing on the original text data set using SIMD instructions in the ARM instruction set to remove unstructured noise data and generate preliminary cleaned text data; performing instruction-level word segmentation processing on each text segment in the preliminary cleaned text data; using the memory access optimization mechanism of the ARM instruction set to perform part-of-speech tagging on the segmented text segments to generate a part-of-speech tagged text data set; and then performing part-of-speech tagging based on the tagged words. The distribution characteristics of each part-of-speech tag in the annotated text dataset are used to construct a part-of-speech distribution histogram. Redundant words are then filtered from the text dataset using frequency thresholds for each part-of-speech tag in the histogram, generating a filtered, concise text dataset. The filtered, concise text dataset undergoes instruction-level data format conversion to uniformly convert text data of different formats into a preset standardized text format, generating the preprocessed enterprise text dataset. Specifically, the process of constructing a part-of-speech distribution histogram based on the distribution characteristics of each part-of-speech tag in the annotated text dataset and filtering redundant words from the text dataset using frequency thresholds for each part-of-speech tag in the histogram is described. Word filtering, generating a filtered and concise text data set, includes: scanning each text segment in the part-of-speech tagged text data set word by word, counting the occurrence frequency of each part-of-speech tag in the text data set, and generating a part-of-speech tag count set; based on the occurrence frequency of each part-of-speech tag in the part-of-speech tag count set, according to a preset histogram interval division rule, accumulating the counts of the same part-of-speech tag into the corresponding histogram interval, generating a part-of-speech distribution histogram structure containing interval counts and the correspondence between part-of-speech tags; based on the count data of each histogram interval in the part-of-speech distribution histogram structure, calculating the total occurrence frequency of each part-of-speech tag, and generating part-of-speech tags in descending order of frequency. The process involves: 1) Marking frequency sequences; 2) Generating a set of part-of-speech tags to be filtered for part-of-speech tags that are higher than a first frequency threshold or lower than a second frequency threshold in the part-of-speech tag frequency sequence, 3) Traversing each word in the text data set after part-of-speech tagging, matching the part-of-speech tag corresponding to the word with the set of part-of-speech tags to be filtered, removing successfully matched words and their part-of-speech tags, and generating an intermediate filtered text data set; 4) Performing continuous part-of-speech tag coherence detection on the remaining words in the intermediate filtered text data set, and removing adjacent word combinations that conform to preset invalid grammar rules, thus generating a grammatically corrected filtered text data set.Based on the part-of-speech tag distribution density of each word in the filtered text data set after grammatical correction, words with the same part-of-speech tag that appear repeatedly in the same sentence are retained intermittently, and only words that meet the preset interval position are retained, thus generating the filtered simplified text data set. The feature extraction module is used to input the preprocessed enterprise text data set into a pre-trained enterprise feature extraction model, and to extract hierarchical semantic features from the enterprise text data set through the multi-layer semantic analysis structure in the enterprise feature extraction model, thereby generating an enterprise semantic feature vector set. The indicator generation module is used to input the set of enterprise semantic feature vectors into a pre-built enterprise evaluation model, and to perform dynamic weight allocation processing on the set of enterprise semantic feature vectors through the multi-dimensional evaluation network in the enterprise evaluation model to generate a set of comprehensive enterprise evaluation indicators. The credit identification module is used to construct a tree structure of enterprise assessment results based on the correlation between the assessment indicators in the enterprise comprehensive assessment indicator set, and to filter the results layer by layer through the assessment threshold of each node in the tree structure to generate the final enterprise credit risk assessment report.

Images