Methods and Systems for Data Processing Based on Large Language Models

Abstract

This invention relates to the field of data processing technology, specifically to a method and system for processing data based on a large language model. The method includes the following steps: based on inter-sentence punctuation location and part-of-speech tagging, dividing sentence blocks into functional partitions; extracting themes and relational words to determine the starting points of semantic chains; constructing trigger index sequences; identifying logical jump points and breakpoint locations; mapping the tag structure to the language model output to analyze deviations; and generating a list of tag mapping combinations. In this invention, by analyzing the trend of semantic direction changes, semantic turning points can be captured, and their relationship with verb-noun combinations can be determined. The trigger points for actual information transfer can be extracted. Through the location of keyword starting and ending blocks and the logical judgment of noun group cross-combinations, semantic path breakpoint regions can be identified, giving the path integrity clear breakpoints. A traceable semantic mapping structure path is established within the language model, enhancing the accuracy, coherence, and hierarchical clarity of semantic reconstruction.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to methods and systems for processing data based on large language models. Background Technology

[0002] The field of data processing technology encompasses systematic methods and tools for collecting, transforming, analyzing, modeling, and structuring various types of data. Its core focus is improving the efficiency of data organization and processing to support information extraction, decision-making, and intelligent applications. Data processing broadly covers various aspects, including cleaning structured and unstructured data, feature extraction, data aggregation, classification indexing, and semantic analysis. Common applications include database management, image recognition, and predictive modeling. With the continuous expansion and diversification of data scale, data processing technologies are increasingly incorporating self-learning and reasoning mechanisms to enhance their processing logic and adaptive capabilities, enabling them to play a greater role in intelligent systems.

[0003] The method of processing data using large language models refers to the approach of using large-scale language models based on neural networks to perform semantic modeling, structured reconstruction, and contextual reasoning on textual data. This patent primarily addresses the problems of structuring unstructured textual data, identifying semantic associations, and representing knowledge. It employs a pre-trained language model to generate contextual embedding vectors, then uses vector similarity metrics to determine the dependencies between semantic segments, and combines a rule generation mechanism to parse natural language text into content tags and entity relation mappings that satisfy specific structural definitions, thus completing the data structure representation. This method includes key steps such as large-scale corpus pre-training, semantic embedding generation, entity relation extraction, and structure mapping assembly.

[0004] In existing technologies, when dealing with structured processing tasks involving unstructured textual data, although vector similarity metrics are used to identify semantic dependencies, the overall recognition logic heavily relies on the context vectors output by the language model, lacking an explicit judgment mechanism based on changes in the semantic content itself. Because structural judgments are not made for semantic transitions or logical breaks in the text, label paths are prone to overlap, misalignment, or omissions in actual mapping. Existing methods, when handling nested labels or parallel multi-label paths, use vector density similarity for coarse merging, lacking clear hierarchical boundary definitions. This results in disordered label distribution in the structured mapping output layer, easily leading to label assignment errors, especially in long texts or multi-chain semantic structures. For example, in policy texts or contract documents, the lack of chain start-point identification and path jump breakpoint annotation capabilities can easily lead to mixed output of labels from multiple logical paragraphs, affecting the accuracy of downstream semantic deconstruction and reasoning operations. Summary of the Invention

[0005] To address the shortcomings of existing technologies, such as the high dependence of overall recognition logic on context vectors output by language models and the lack of explicit judgment mechanisms based on semantic content changes, and the failure to perform structural judgments at semantic transitions or logical breaks in the text, label paths are prone to overlap, misalignment, or omission in actual mapping. Existing methods, when handling nested labels or parallel multi-label paths, use vector density similarity for coarse merging, lacking clear hierarchical boundary definitions. This results in disordered label distribution at the structural mapping output layer, easily leading to label assignment errors, especially in long texts or multi-chain semantic structures. For example, in policy texts or contract documents, the lack of chain start-point identification and path jump breakpoint annotation capabilities can easily lead to mixed output of labels from multiple logical paragraphs, affecting the accuracy of downstream semantic deconstruction and reasoning operations. This invention provides a method and system for processing data based on a large language model. The technical solution is as follows:

[0006] On the one hand, a method for processing data based on a large language model is provided, which includes:

[0007] S1: Based on the sentences in the text paragraph, locate the punctuation between sentences, tag the words connecting adjacent sentences with part-of-speech tags, identify verb and subject pairs, divide the sentence block area according to the word matching order, and generate sentence segment functional partitions;

[0008] S2: Call the sentence segment functional partition, extract the topic words and relation words of the text segment, mark the position of the word group in the context of the block, record the semantic trend of the word group according to the direction of position change, identify the direction reversal node in two consecutive blocks, and generate a semantic chain trigger index sequence.

[0009] S3: Use the semantic chain to trigger the index sequence, locate the start and end blocks of keywords in the tag path segment, extract the key noun groups at both ends, cross-combine the two groups of words and determine whether there are three or more groups of non-common words. If so, mark them as logical jump points and generate tag path breakpoint position groups.

[0010] S4: Call the tag path breakpoint position group, extract tag phrases from the corresponding segment, record the position sequence of tag phrases in the sentence, identify whether there are path intersections or position nesting phenomena, analyze the nesting hierarchy relationship, and generate a tag structure nesting hierarchy table.

[0011] As a further aspect of the present invention, the sentence segment functional partition includes semantic connection boundary points, block structure boundary lines, and sentence backbone matching groups; the semantic chain trigger index sequence includes a jump node sequence, semantic direction turning points, and block combination start positions; the tag path breakpoint position group includes tag semantic jump areas, keyword disconnected segments, and path continuity interruption points; and the tag structure nesting hierarchy table includes a tag hierarchy index set, a nested path number column, and a tag interleaving sequence diagram.

[0012] As a further aspect of the present invention, the step of obtaining the segment functional partition is specifically as follows:

[0013] S101: Based on the sentences in the text paragraph, scan the punctuation marks between sentences, analyze the connection relationship between sentences, extract the connecting words after the punctuation at the end of the sentence as the analysis object, identify the part-of-speech tags, determine whether they are conjunctions or adverbs that indicate parallel, contrast or progressive relationship, and then mark the words as connection nodes to generate a connection part-of-speech tag structure set.

[0014] S102: By connecting the part-of-speech tagging structure set, locate the verbs and subjects in the sentences corresponding to the connecting nodes, extract the subject phrases before and after the position of the verb and determine whether they constitute a subject-predicate structure, identify the subject-predicate pairs with related meanings in continuous sentences, record the start and end positions and semantic attribution of the subject-predicate structure in the original text, and generate a subject-predicate phrase combination mapping sequence.

[0015] S103: Based on the subject-predicate combination mapping sequence, arrange each semantic unit according to the order of subject-predicate combinations in the original text, divide the sentence boundaries and assemble them into semantic blocks, determine whether semantic continuation is formed between semantic groups based on the number of semantic repetitions between the subject and predicate in the semantic blocks, classify the blocks of related subject-predicate combinations into the same interval, and obtain the functional partition of the sentence segment.

[0016] As a further aspect of the present invention, the step of obtaining the semantic chain trigger index sequence specifically comprises:

[0017] S201: Call the statement units in the sentence segment functional partition, extract the noun groups and verb groups with the highest frequency of occurrence in the sentence block, classify the noun groups as topic words, classify the verb groups as relation words, and record the start and end positions in the sentence block as semantic index ranges to obtain the topic relation word position information set;

[0018] S202: Based on the topic relation word position information set, the index range of the word group is compared vertically according to the original arrangement order of the word blocks in the text. The direction of movement of the index position in the previous word block and the next word block are compared. The forward, backward or unchanged states are marked respectively. The original text position is recorded to obtain the semantic turning point index table.

[0019] S203: Call the semantic transition node index table, analyze the sentence content corresponding to the transition node, extract the verb and noun combination segments near the node, determine whether the node falls in the combination segment after the verb and before the noun, if the positional relationship is satisfied, mark it as the semantic chain starting point, and generate the semantic chain trigger index sequence.

[0020] As a further aspect of the present invention, the step of obtaining the tag path breakpoint location group specifically includes:

[0021] S301: Using the semantic chain to trigger the index sequence, locate the starting and ending blocks corresponding to the trigger index in the text, extract the noun groups that appear with priority in the two blocks, and record the position index range and the block identifier in the original text to generate a set of key noun groups for the blocks.

[0022] The extraction of noun phrases from the two language blocks based on frequency is performed using the following formula:

[0023] ;

[0024] in, Score the priority of noun phrases. This represents the frequency of the i-th noun group in the chunk. This represents the average frequency of the noun phrase. represents the relevance score of the i-th noun group to the context, and k represents the number of noun groups;

[0025] S302: Call the noun groups of the starting and ending blocks in the set of key noun groups of the block, perform pairwise cross-combination, identify whether there is word overlap in each combination, count the number of combinations without common words, if the number of combinations without common words reaches three or more, mark the starting and ending block segments as logical jump points, and obtain the label path breakpoint position group.

[0026] As a further aspect of the present invention, the steps for obtaining the nested hierarchy table of the tag structure are specifically as follows:

[0027] S401: Call the text segment corresponding to the jump point in the tag path breakpoint position group, extract the noun phrases that appear more than twice in the text segment as candidate tag phrases, record the first word position and the last word position of the tag phrase in the sentence, and obtain the tag phrase position index table.

[0028] S402: Based on the start and end position indexes of each group of words in the tag phrase position indexing table, arrange the tag phrases in the order of the sentence, compare whether there are overlapping areas or nested relationships between the front and back boundary ranges, and generate a tag path nested and interleaved tag set;

[0029] S403: Call the nested and interleaved tag set of the tag path, establish the tag level index number according to the nesting layer, classify the word groups with the same nesting starting point into the same level path group, and obtain the tag structure nesting level table.

[0030] As a further aspect of the present invention, the comparison of whether there are overlapping areas or nested relationships between the front and rear boundary ranges is performed using the following formula:

[0031] ;

[0032] in, This represents the boundary overlap strength index between the p-th and q-th tag phrases. This represents the starting index of the p-th tag phrase. This represents the index value of the end position of the p-th tag phrase. This represents the starting index of the q-th tag phrase. This represents the index value of the end position of the q-th tag phrase. This represents the starting index of the r-th tag phrase. This represents the average index value of the starting position of the first u group of tag phrases. This represents the index value of the end position of the r-th tag phrase. This represents the average index value of the end position of the preceding u-group of tag phrases. This represents the span value of the r-th tag phrase. This represents the number of tag pairs.

[0033] As a further aspect of the present invention, the method further includes step S5:

[0034] S5: Based on the nested hierarchy table of the label structure, extract the context position of the label segment from the pre-training corpus of the large language model, map the label structure corresponding to the nested hierarchy to the output layer of the language model, compare the deviation of the label word distribution at the corresponding position in the default output of the language model after mapping, identify the combination order of the main components of the label in the traceable path, and generate a list of label mapping combinations for the language model.

[0035] The language model label mapping combination list includes semantic nesting mapping paths, label structure correspondence sets, and word-direction distribution difference marker sequences.

[0036] As a further aspect of the present invention, the step of obtaining the language model tag mapping combination list specifically includes:

[0037] S501: Call the hierarchical index number of each tag path in the nested hierarchy table of the tag structure, retrieve the context segment of the corresponding tag in the original text in the pre-training corpus of the large language model, extract the position identifier range of the tag context segment in the model input, establish the context input index set of the tag segment, and obtain the tag context mapping location set.

[0038] S502: Based on the tag context mapping location set, inject the corresponding tag structure nesting level information into the output mapping sequence of the large language model, extract the tag word direction representation generated at the corresponding position in the output layer before and after injection, compare the distribution trajectory between each group of tag word directions, locate and classify the offset direction, angle and spacing, and obtain the tag word direction distribution deviation information.

[0039] S503: Based on the label combination trajectory with continuous offset and consistent direction in the label word distribution deviation information, extract the word path that can be semantically traced back, determine whether the logical word combination of the first and last words connecting the labels in the word path satisfies the original semantic chain order, filter the label paths that conform to the sequential logic and arrange them into nested combinations, and obtain the list of language model label mapping combinations.

[0040] On the other hand, the system for processing data based on a large language model is used to execute the above-described method for processing data based on a large language model, and the system includes:

[0041] The sentence block recognition module identifies the position of punctuation marks based on sentences in a text paragraph, detects the part-of-speech category of the connecting words between punctuation marks, determines whether the sentences adjacent to the connecting words contain subject phrases and predicate verb structures respectively, divides the entire text into multiple sentence blocks according to the order of appearance of the phrases in the sentences, and generates sentence block functional partitions.

[0042] The semantic indexing module calls the sentence segment functional partition, extracts the topic words and relation words in the sentence block, records the position of the phrase in the context block, identifies whether the movement direction of the phrase has been reversed, determines whether the reversal point is located between the verb and noun combination segment, and obtains the semantic chain trigger index sequence.

[0043] The logical jump module triggers the index sequence according to the semantic chain, locates the corresponding start and end blocks in the text paragraph, extracts all noun groups in the start and end blocks, and performs full combination and arrangement according to the noun groups at both ends to generate the tag path breakpoint position group.

[0044] The tag extraction module uses the tag path breakpoint location group to extract tag phrases from the sentence block pointed to by the jump node, identifies the intra-sentence position number of the tag phrases in the sentence block, determines whether there are intersecting positions or nested arrangements, and obtains the tag structure nesting hierarchy table.

[0045] The label mapping module uses the nested hierarchical table of the label structure to extract the context of the language model that matches the label phrases, identify the word vector position distribution of the label group in the output layer of the language model, compare the corresponding difference with the default word vector distribution, and generate a list of language model label mapping combinations.

[0046] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0047] By identifying sentence boundaries in text paragraphs and using the structural positioning of connective word part-of-speech tags and verb-subject pairs as the basis for sentence block division, the resulting functional partitioning of sentence segments can logically organize semantics at the beginning of the text, providing a structured input foundation for the accurate positioning of subsequent semantic chain initiation points. By leveraging the semantic direction change trends reflected in the differences in the positional occurrence of phrases before and after a block, semantic turning points can be captured, and their relationship with verb-noun combinations can be determined, extracting trigger points with actual information transfer functions. Through the positioning of keyword start and end blocks and the logical judgment of noun phrase cross-combinations, semantic path interruption areas are identified, giving the path integrity clear break markers. The extraction of nested hierarchical structures within the path is performed by stratifying the relative positions of tags within sentences and defining the logical hierarchy of tags based on path intersection features. After mapping the nested structure to the output layer of the large language model, the rationality of the path is judged by identifying word distribution deviations, realizing the transition from unstructured text to structured expression, and establishing a traceable semantic mapping structure path within the language model, enhancing the accuracy, coherence, and hierarchical clarity of semantic reconstruction. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the workflow of the present invention;

[0049] Figure 2 This is a system flowchart of the present invention. Detailed Implementation

[0050] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0051] In embodiments of the present invention, words such as "exemplarily," "for example," etc., are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" in the present invention should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of the word "exemplary" is intended to present the concept in a concrete manner. Furthermore, in embodiments of the present invention, the meaning expressed by "and / or" can be both, or either one.

[0052] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0053] Please see Figure 1 This invention provides a method for processing data based on a large language model. The processing flow of this method may include the following steps:

[0054] S1: Based on the sentences in the text paragraph, locate the punctuation between sentences, tag the words connecting adjacent sentences with part-of-speech tags, identify verb and subject pairs, divide the sentence block area according to the word matching order, and generate sentence segment functional partitions;

[0055] S2: Call the sentence segment function partition, extract the topic words and relation words of the text segment, mark the position of the phrase in the context before and after the block, record the semantic trend of the phrase according to the direction of position change, identify the direction reversal node in two consecutive blocks, determine whether the node is located between the semantic verb and noun combination segment, if the combination structure is satisfied, define the node as the semantic chain starting point, and generate the semantic chain trigger index sequence.

[0056] S3: Use semantic chain to trigger the index sequence, locate the start and end blocks of keywords in the tag path segment, extract the key noun groups at both ends, cross-combine the two groups of words and determine whether there are three or more groups of non-common words. If so, mark them as logical jump points and generate tag path breakpoint position groups.

[0057] S4: Call the tag path breakpoint position group, extract the tag phrases from the corresponding segment, record the position sequence of the tag phrases in the sentence, identify whether there are path intersections or position nesting phenomena, analyze the nesting hierarchy relationship, and generate a tag structure nesting hierarchy table.

[0058] S5: Based on the nested hierarchy table of the label structure, extract the context position of the label segments from the pre-training corpus of the large language model, map the label structure corresponding to the nested hierarchy to the output layer of the language model, compare the deviation of the label word distribution at the corresponding position in the default output of the language model after mapping, identify the combination order of the main components of the label in the traceable path, and generate a list of label mapping combinations for the language model.

[0059] The sentence segment functional partitioning includes semantic connection boundary points, block structure boundary lines, and sentence core matching groups. The semantic chain trigger index sequence includes jump node sequences, semantic direction turning points, and block combination start positions. The label path breakpoint position group includes label semantic jump areas, keyword disconnected segments, and path coherence interruption points. The label structure nesting hierarchy table includes label hierarchy index set, nested path number column, and label interleaving sequence diagram. The language model label mapping combination list includes semantic nesting mapping paths, label structure correspondence set, and word direction distribution difference marker sequence.

[0060] The specific steps for obtaining the functional partitions of a sentence segment are as follows:

[0061] S101: Based on the sentences in the text paragraph, scan the punctuation marks between sentences, analyze the connection relationship between sentences, extract the connecting words after the punctuation at the end of the sentence as the analysis object, identify the part-of-speech tags, determine whether they are conjunctions or adverbs that indicate parallel, contrast or progressive relationship, and then mark the words as connection nodes to generate a connection part-of-speech tag structure set.

[0062] This paper addresses the task of extracting and tagging part-of-speech (POS) words following punctuation marks. It reads input text paragraphs, uses natural language processing tools to segment sentences and tag the POS of each word, identifies sentence-ending punctuation using regular expressions or syntactic dependency parsing models, locates the end markers of each short sentence, and statistically extracts the words immediately following punctuation marks. From the POS tagging results, words belonging to the categories of conjunctions (e.g., CC) or adverbs (e.g., RB) are selected. A custom vocabulary (e.g., "but," "moreover," "therefore," etc.) is used to further determine whether they constitute semantic relationships such as parallelism, contrast, or progression. For example, if the original sentence is "The equipment malfunctioned; therefore, the parameters need to be reconfigured," then words following punctuation marks are identified. Therefore, "therefore" is used as a conjunctive adverb, connecting the two sentences to establish a semantic link. It is marked as a progressive node, with the corresponding annotation structure as: {"Conjunctive word": "therefore", "Part of speech": "RB", "Connection type": "progressive"}. A structure set list is built for each conjunctive word, and its position information and connection type in the text are recorded in combination with contextual syntactic analysis. This information is stored in JSON format, including fields such as "starting position", "conjunctive word", "part of speech tag", and "semantic connection type". This information will be used for subject-predicate structure recognition and semantic sentence block segmentation. After this step is completed, the conjunctive node will be used as the entry point for subsequent semantic analysis to generate a conjunctive part-of-speech tag structure set.

[0063] S102: By connecting the part-of-speech tagging structure set, locate the verbs and subjects in the sentences corresponding to the connecting nodes, extract the subject phrases before and after the position of the verb and determine whether they form a subject-predicate structure, identify the subject-predicate pairs with related meanings in continuous sentences, record the start and end positions and semantic attribution of the subject-predicate structure in the original text, and generate a subject-predicate phrase combination mapping sequence.

[0064] The algorithm parses and connects the part-of-speech tagged structure set, indexes each connection node to the statement it connects to, and extracts the verb root position using an NLP dependency tree model (such as BERT-based semantic dependency analysis) and tracks the subject phrases before and after it. For a structure like "An error occurred; therefore, the module needs to be restarted," "needs" is identified as the predicate verb, and its dependent parent node is the subject "module." Based on part-of-speech combination (NN+VB), it determines whether a subject-predicate structure is formed. By tracing the dependency relationship nodes forward sentence by sentence, it records the word combinations that form the subject-predicate structure. In each structure, it extracts the root, modifiers, and determiners to form a complete subject-predicate phrase, such as "subject": module, "predicate": need to restart. For each structure, a mapping entry is created, recording the starting and ending character positions, the original text position index, and the semantic range to which it belongs, forming a subject-predicate combination mapping table. The table is set as follows: "{Subject: 'Module', Predicate: 'Needs to restart', Starting position: 14, Ending position: 22, Semantic range: 'Restart operation'}". After verifying whether each structure has semantic coherence (such as whether it shares a subject or whether the verb category is consistent), it is written into the mapping sequence for subsequent sentence block construction and functional partitioning. In the example, if there is "Unrecognizable; therefore, the interface needs to be adjusted", the two sentences share a subject and the verb structure matches, so they are merged into a pair of related subject-predicate structures, generating a subject-predicate phrase combination mapping sequence.

[0065] S103: Based on the subject-predicate phrase combination mapping sequence, arrange each semantic unit according to the order of subject-predicate combinations in the original text, divide the sentence boundaries and assemble them into semantic sentence blocks, determine whether semantic continuation is formed between semantic groups based on the number of semantic repetitions between the subject and predicate in the semantic sentence blocks, and classify the sentence blocks with related subject-predicate combinations into the same interval to obtain the sentence segment functional partition.

[0066] The original text is rearranged according to the order of the subject-verb pairs. Each combination is defined as a basic semantic unit. The original text is segmented into semantic blocks by indexing its start and end positions. For example, the original text "The server experienced a delay; therefore, load balancing was performed; and further expansion was carried out" forms three subject-verb structures: "server-experienced," "performed," and "expanded." Based on dependency analysis, "performed load balancing" and "further expanded" are determined to be continuations of the subject's actions, forming a progressive structure. These semantic blocks are grouped into the same interval, and the corresponding sentence segment function is set as "response processing flow," using semantic roots between the subject and verb. The comparison model (such as synonym discrimination and word vector distance comparison) determines whether "load balancing" and "expansion" constitute similar motivations or action groups. If the word vector similarity (cosine similarity) is greater than 0.75, it is identified as a semantic continuation relationship and classified into the same functional partition. The output functional partition block includes: starting subject-predicate combination, semantic attribution, and structural mapping index, such as: "{starting combination: 'perform load balancing', ending combination: 'further expansion', interval semantics: 'response scheduling optimization'}". This partition is used in subsequent tasks for semantic scene merging or information extraction to obtain sentence segment functional partitions.

[0067] The specific steps for obtaining the semantic chain trigger index sequence are as follows:

[0068] S201: Call the statement units in the sentence segment function partition, extract the noun groups and verb groups with the highest frequency of occurrence in the sentence block, classify the noun groups as topic words, classify the verb groups as relation words, and record the start and end positions in the sentence block as semantic index ranges to obtain the topic relation word position information set;

[0069] For each semantic block in the functional partition of the sentence segment, the extraction of noun groups and verb groups is completed. The sentence is segmented, and each word is labeled with its part of speech. Noun groups include general entities, tools, components, names, etc., and verb groups include operation behaviors, state changes, functional reactions, etc. In practical applications, such as processing the text "Restart service module, load driver component, user input login information", the high-frequency nouns and verbs in each sentence block are counted. "Module" and "driver" will be used as high-frequency nouns, while "restart", "load", and "input" are verb groups. According to the frequency of occurrence or the priority of occurrence position, noun groups are selected as subject words and verb groups are used as relation words. The first and last occurrence positions are recorded respectively. The first occurrence is set to the first position in the first sentence and the second occurrence is set to the middle position in the second sentence. The index range of the subject word record extends from 0 to the end of the sentence block. "Load" first appears in the second sentence and also stops at the end. The index range is determined to complete the identification of subject words and relation words in all semantic blocks, establish accurate word group positioning support for subsequent semantic trend judgment, and obtain the subject relation word position information set.

[0070] S202: Based on the topic relation word position information set, the index range of the word group is compared vertically according to the original arrangement order of the word blocks in the text. The direction of the movement of the index position in the preceding and following word blocks is compared, and the states of moving forward, moving backward or no change are marked respectively. The original text position is recorded to obtain the semantic turning point index table.

[0071] The order of word positions within each semantic block is compared vertically, and the trend of change between blocks is scanned block by block. From the previous block to the next block, it is observed whether the noun or verb group moves forward, backward, or remains unchanged. If the position of "module" in a certain block is close to the beginning of the sentence, and the noun appears in the middle or end of the next block, it indicates that it has moved backward. Conversely, if it moves from the end of the sentence to the beginning, it is a forward movement. In real-world scenarios, such as describing "starting the module", "module initialization", and "driver module loading completed", the position of the noun "module" moves from the end of the sentence, the beginning of the sentence, to the middle of the sentence, forming a movement trajectory of forward and backward movement. This process is continuously compared according to the order of the blocks in the original text to determine whether the semantic core has a dynamic change in position as the scene changes. Based on this, the direction change marker of each word group is recorded. The marker can be marked as "forward movement", "no change", or "backward movement". This serves as a key sequence for tracking the evolution of concepts and is used for subsequent reversal point identification and chain trigger point screening to obtain a semantic turning point index table.

[0072] S203: Call the semantic transition node index table, analyze the sentence content corresponding to the transition node, extract the verb and noun combination segments near the node, determine whether the node falls in the combination segment after the verb and before the noun, if the positional relationship is satisfied, mark it as the semantic chain starting point, and generate the semantic chain trigger index sequence.

[0073] Based on the recorded boundary positions, several words are extracted forward and backward to form a context range. Grammatical dependency analysis is performed on verbs and nouns within this range to search for combinations where the turning point occurs after the verb and before the noun. Using the text "Attempted to reconnect the interface module, but connection failed," if the verb "attempt" appears before the turning point and the noun "module" appears after the turning point, then this position meets the semantic chain initiation point condition. In application scenarios such as device control logic, jump points in instruction execution flow accompany this structure. This location is positioned as a new semantic clue start node, and the character position and sentence block number of this node in the text are recorded. This index is used to establish a mapping of relationships between semantic blocks. Through this chain initiation point, all subsequent operation flow statement segments controlled or affected by this node are traced, forming a semantic chain initial path structure diagram. This provides basic data for subsequent logical path analysis and generates a semantic chain trigger index sequence.

[0074] The specific steps for obtaining the tag path breakpoint location group are as follows:

[0075] S301: Using semantic chain trigger index sequence, locate the starting and ending blocks corresponding to the trigger index in the text, extract the noun groups that appear with priority in the two blocks, record the position index range and the block identifier in the original text, and generate a set of key noun groups for the blocks.

[0076] Extract the noun phrases from two text blocks based on their frequency, using the following formula:

[0077] ;

[0078] in, Score the priority of noun phrases. This represents the frequency of the i-th noun group in the chunk. This represents the average frequency of the noun phrase. represents the relevance score of the i-th noun group to the context, and k represents the number of noun groups;

[0079] Meaning of parameters and derivation of formulas:

[0080] Representing the The frequency of noun groups in the text is obtained by scanning a given text block using a text analysis tool. In text block A, noun group 1 appears 5 times and noun group 2 appears 3 times.

[0081] The average frequency of a noun phrase is obtained by calculating the frequency of candidate noun phrases and taking their arithmetic mean.

[0082] It is the first The relevance score of a noun group is quantified by semantic analysis or contextual analysis methods. It is based on the weight of the noun group in the context and its influence on the text content. If a noun group has a high semantic match with the current chunk, it is assigned a high relevance score (e.g., 0.8).

[0083] This represents the number of noun phrases to be analyzed. In this example, three noun phrases are selected for analysis. ;

[0084] The following data is based on results obtained through actual text analysis tools or algorithms:

[0085] Noun phrase frequency ( ): 5 times, 3 times, 2 times;

[0086] Calculate the average frequency of noun phrases:

[0087] ;

[0088] Relevance score ( ): 0.7, 0.8, 0.6;

[0089] Calculate the absolute frequency difference of each noun phrase by multiplying it by its logarithm:

[0090] ;

[0091] ;

[0092] ;

[0093] Summing and calculating the numerator:

[0094] ;

[0095] Calculate the square of the difference between each frequency and the average value, and sum them:

[0096] ;

[0097] ;

[0098] ;

[0099] ;

[0100] Calculate the product of each frequency and its relevance score, and sum them:

[0101] ; ;

[0102] ;

[0103] ;

[0104] Substitute the above results into the formula to calculate:

[0105] ;

[0106] The results show that the noun phrase priority score is 0.97, which indicates that the noun phrase has a high frequency priority in the chunk and a high relevance score in the context. This priority score can be used as the basis for subsequent text analysis to identify some noun phrases as key elements in text semantic analysis.

[0107] S302: Call the noun groups of the starting and ending blocks in the set of key noun groups of the block, perform pairwise cross-combination, identify whether there is word overlap in each combination, count the number of combinations without common words, if the number of combinations without common words reaches three or more, mark the starting and ending block segments as logical jump points, and obtain the label path breakpoint position group.

[0108] The code reads three sets of high-frequency nouns from the pre-annotated start and end blocks of the key noun group set. It then generates cross-pairing groups by combining each pair of nouns: the first noun in the start block is paired with the first to third nouns in the end block, and the same operation is repeated for the second and third nouns, resulting in a total of nine combinations. Each combination is compared for lexical overlap. The start block is set to contain "module," "interface," and "service," and the end block to contain "network" and "driver." After combining, each combination is checked for identical or highly similar word roots. If no identical or similar word roots appear in any combination, the code is rejected. Semantic relation words are counted as not having co-word combinations. The remaining combinations are checked. If the total number reaches 3 or more, it is determined that a semantic break occurs between the start and end blocks. This paragraph is marked as a logical jump point. In the actual corpus, there are no identical entities or operation objects between "load driver module" and "user authorize access permissions", which meets the condition of logical isolation, that is, the tag path breakpoint is triggered. The logical jump point is written into the path breakpoint position group, and its original text character start and end positions are recorded for subsequent semantic chain verification and path structure reconstruction. The tag path breakpoint position group is obtained.

[0109] The specific steps to obtain the nested hierarchical table of the tag structure are as follows:

[0110] S401: Call the text segment corresponding to the jump point in the tag path breakpoint location group, extract the noun phrases that appear more than twice in the text segment as candidate tag phrases, record the first word position and the last word position of the tag phrase in the sentence, and obtain the tag phrase position index table.

[0111] The process begins by analyzing the text at the jump point, extracting noun phrases from the text, including common nouns, proper nouns, and phrase combinations of nouns. A parser is used to tag parts of speech, and word frequency statistics are performed on all noun phrases. Phrases appearing more than twice are selected as candidate tag phrases. For example, in a description like "Module loading failed, module interface response was abnormal, module error detected, module restart initialization," the word "module" appears four times and "interface" appears twice. Meeting these criteria, "module" and "interface" are selected as candidate tag phrases. The word order of the first and last words in the sentence is recorded. For instance, "module" first appears at the 2nd word position and last appears at the 20th word position. Each record includes the phrase content, the starting word order, the ending word order, and the corresponding text segment number. This provides a structural index for subsequent tag relationship identification, ensuring the accuracy of relative positions between phrases, and obtaining a tag phrase position index table.

[0112] S402: Based on the start and end position indexes of each group of words in the tag phrase position indexing table, arrange the tag phrases in the order of the sentence, compare whether there are overlapping areas or nested relationships between the front and back boundary ranges, and generate a tag path nested and interleaved tag set;

[0113] To compare whether there are overlapping or nested areas between the preceding and following boundary ranges, use the following formula:

[0114] ;

[0115] in, This represents the boundary overlap strength index between the p-th and q-th tag phrases. This represents the starting index of the p-th tag phrase. This represents the index value of the end position of the p-th tag phrase. This represents the starting index of the q-th tag phrase. This represents the index value of the end position of the q-th tag phrase. This represents the starting index of the r-th tag phrase. This represents the average index value of the starting position of the first u group of tag phrases. This represents the index value of the end position of the r-th tag phrase. This represents the average index value of the end position of the preceding u-group of tag phrases. This represents the span value of the r-th tag phrase. This represents the number of tag pairs.

[0116] Meaning of parameters and derivation of formulas:

[0117] The tag phrase position index table is set to include the following five groups of tag phrases, with their start and end position indexes as follows:

[0118] Group 1: , ;

[0119] Group 2: , ;

[0120] Group 3: , ;

[0121] Group 4: , ;

[0122] Group 5: , ;

[0123] Calculate the average of the starting position index values ​​of the first 5 groups of tag phrases:

[0124] ;

[0125] Calculate the average of the end position index values ​​of the first 5 tag phrases:

[0126] ;

[0127] Calculate the boundary crossover strength index between the tag phrases in group 1 and group 2. :

[0128] First calculation:

[0129] ;

[0130] Second calculation:

[0131] ;

[0132] ;

[0133] Take the square root:

[0134] ;

[0135] Substitute into the formula to calculate :

[0136] ;

[0137] The results show that the boundary overlap strength index between the first and second group of tag phrases is 8.294. The larger the value, the higher the degree of overlap. This index can be used to determine whether there are overlapping areas or nesting relationships between tag phrases and generate a set of nested and overlapping tag paths.

[0138] S403: Call the tag path nesting interleaved tag set, establish tag level index numbers according to the nesting layer, group words with a common nesting starting point into the same level path group, and obtain the tag structure nesting level table;

[0139] The process begins by constructing hierarchical numbers based on nesting depth. The outermost phrase in the nested structure is assigned a first-level path group number. Sub-phrases within the nested structure are then identified and sequentially labeled as second-level, third-level, and so on, according to their nesting order. For example, if "module loading" nests "module loading," and "module loading" nests "module," then the three are classified as first-level, second-level, and third-level path groups, respectively. Each phrase is recorded with a number based on its hierarchical structure, including its content, level, nesting starting point, and position range. If two or more phrases share the same nesting starting point but different ending points, they are classified as parallel path groups at the same level and marked as independent path branches within the same level. For instance, if "module configuration," "module loading," and "module detection" have the same starting position but different ending positions, they are grouped under the same path group number and identified as parallel markers, forming a clear semantic hierarchy map of tags. This provides a complete tag system for text structure classification, summary extraction, or semantic indexing, resulting in a tag structure nesting hierarchy table.

[0140] The specific steps for obtaining the list of language model label mapping combinations are as follows:

[0141] S501: Call the hierarchical index number of each label path in the nested hierarchy table of the label structure, retrieve the context segment of the corresponding label in the original text in the pre-training corpus of the large language model, extract the position identifier range of the label context segment in the model input, establish the context input index set of the label segment, and obtain the label context mapping location set.

[0142] The path structure is extracted from the nested hierarchy table of the tag structure. The level index number corresponding to each tag path is read. Then, for the tag content under each level, a retrieval request is constructed and the context segment is matched and searched in the pre-training corpus of the large language model. The retrieval tag is set as "module loading". The semantic matching interface of the pre-training model is called to filter the context segments containing the tag or its synonyms in the model corpus. The character or token position identifier range of the segment in the original input of the model is obtained, such as segments like "module loading exception" and "module loading process". The corresponding token number range is such as between 115 and 125. This range is recorded as the input interval identifier of the tag in the pre-training data. At the same time, this identifier is bound to the level number of the original tag path. Each set of index records includes fields such as tag name, level, context segment content, and token start and end positions. This set is summarized as the tag context mapping location set, which is used for subsequent model output trajectory monitoring and semantic sequence tracking to obtain the tag context mapping location set.

[0143] S502: Based on the tag context mapping location set, the corresponding tag structure nesting level information is injected into the output mapping sequence of the large language model. The tag word direction representation generated at the corresponding position before and after injection is extracted, and the distribution trajectory between each group of tag word directions is compared. The offset direction, angle and spacing are located and classified to obtain the tag word direction distribution deviation information.

[0144] The system calls the tag context mapping to locate the input position index information of each tag path in the set. Semantic embedding tracking is performed on the input position during the generation task of the language model. The tag word direction representation at the corresponding position in the model output sequence is recorded when no structural information is injected. Then, the nesting level information of the tag structure is injected on the same input. By adding special level prompt tokens or building hierarchical structure templates, the changes in word direction representation at the corresponding tag position are compared with the output sequence before and after injection. The focus is on monitoring the direction of word direction shift (such as concentrating towards the semantic center or diverging towards the remaining tag paths), angle (approximately represented by the angle between high-dimensional vectors), and relative spacing (such as Euclidean distance between vectors). The impact and degree of disturbance of the semantic representation after injecting structural information are determined by the index. This information describes the trajectory of semantic association changes between tags, and a word direction migration pattern is constructed based on this to provide a quantitative reference for semantic path optimization and obtain tag word direction distribution deviation information.

[0145] S503: Based on the label combination trajectory with continuous offset and consistent direction in the label word distribution deviation information, extract the word path that can be semantically backtracked, determine whether the logical word combination of the first and last words connecting the labels in the word path satisfies the original semantic chain order, filter the label paths that conform to the sequential logic and arrange them into nested combinations, and obtain the list of language model label mapping combinations.

[0146] By analyzing the tag combination trajectories with continuous offsets and consistent directions, word-direction paths with stable semantic evolution directions are selected. These paths continuously progress along a certain direction in space, exhibiting an arrangement pattern where the angle between several tag word-direction vectors is small and the spacing gradually decreases. The word-direction representations of the three sets of tag paths, namely "module detection → module initialization → module loading," show a distribution direction that continuously shifts to the upper right. Based on these trajectories, semantic backtracking is performed on the word-direction paths to determine whether there is a logical connection between the tags at both ends of the path in the original text. For example, if "detection" is used as a pre-operation and "loading" is used as the final state, and the two also show a sequential relationship in the original semantic chain, then the semantic order requirement is met, and the path is retained as a valid tag sequence. The tag paths that meet the conditions are reorganized and arranged according to the nested structure for multi-level tag tracking, cross-layer semantic integration, and semantic diffraction analysis of nested structures, forming a complete list of language model tag mapping combinations.

[0147] Please see Figure 2 A system for processing data based on a large language model, the system includes:

[0148] The sentence block recognition module identifies the position of punctuation marks based on sentences in a text paragraph, detects the part-of-speech category of the connecting words between punctuation marks, determines whether the sentences adjacent to the connecting words contain subject phrases and predicate verb structures respectively, divides the entire text into multiple sentence blocks according to the order of appearance of the phrases in the sentences, and generates sentence block functional partitions.

[0149] The semantic indexing module calls the sentence segment functional partition, extracts the topic words and relation words in the sentence block, records the position of the phrase in the context block, identifies whether the movement direction of the phrase has been reversed, determines whether the reversal point is located between the verb and noun combination segment, and obtains the semantic chain to trigger the index sequence.

[0150] The logical jump module triggers the index sequence based on the semantic chain, locates the corresponding start and end blocks in the text paragraph, extracts all noun groups in the start and end blocks, and performs full combination and arrangement according to the noun groups at both ends to generate the tag path breakpoint position group.

[0151] The tag extraction module uses the tag path breakpoint location group to extract tag phrases from the sentence block pointed to by the jump node, identifies the position number of the tag phrase in the sentence block, determines whether there are intersecting positions or nested arrangements, and obtains the tag structure nesting hierarchy table.

[0152] The label mapping module utilizes a nested hierarchical table of label structures to extract the context of the language model that matches the label phrases, identifies the word vector position distribution of the label group in the language model output layer, compares the corresponding difference with the default word vector distribution, and generates a list of language model label mapping combinations.

[0153] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for processing data based on a large language model, characterized in that, Includes the following steps: S1: Based on the sentences in the text paragraph, locate the punctuation between sentences, tag the words connecting adjacent sentences with part-of-speech tags, identify verb and subject pairs, divide the sentence block area according to the word matching order, and generate sentence segment functional partitions; S2: Call the sentence segment functional partition, extract the topic words and relation words of the text segment, mark the position of the word group in the context of the block, record the semantic trend of the word group according to the direction of position change, identify the direction reversal node in two consecutive blocks, and generate a semantic chain trigger index sequence. S3: Use the semantic chain to trigger the index sequence, locate the start and end blocks of keywords in the tag path segment, extract the key noun groups at both ends, cross-combine the two groups of words and determine whether there are three or more groups of non-common words. If so, mark them as logical jump points and generate tag path breakpoint position groups. S4: Call the tag path breakpoint position group, extract tag phrases from the corresponding segment, record the position sequence of tag phrases in the sentence, identify whether there are path intersections or position nesting phenomena, analyze the nesting hierarchy relationship, and generate a tag structure nesting hierarchy table. S5: Based on the nested hierarchy table of the label structure, extract the context position of the label segment from the pre-training corpus of the large language model, map the label structure corresponding to the nested hierarchy to the output layer of the language model, compare the deviation of the label word distribution at the corresponding position in the default output of the language model after mapping, identify the combination order of the main components of the label in the traceable path, and generate a list of label mapping combinations for the language model. The language model label mapping combination list includes semantic nesting mapping paths, label structure correspondence sets, and word-direction distribution difference marker sequences.

2. The method for processing data based on a large language model according to claim 1, characterized in that, The sentence segment functional partition includes semantic connection boundary points, block structure boundary lines, and sentence backbone matching groups. The semantic chain trigger index sequence includes jump node sequence, semantic direction turning point, and block combination start position. The tag path breakpoint position group includes tag semantic jump area, keyword disconnected segment, and path continuity interruption point. The tag structure nesting hierarchy table includes tag hierarchy index set, nested path number column, and tag interlacing sequence diagram.

3. The method for processing data based on a large language model according to claim 1, characterized in that, The specific steps for obtaining the segment functional partitions are as follows: S101: Based on the sentences in the text paragraph, scan the punctuation marks between sentences, analyze the connection relationship between sentences, extract the connecting words after the punctuation at the end of the sentence as the analysis object, identify the part-of-speech tags, determine whether they are conjunctions or adverbs that indicate parallel, contrast or progressive relationship, and then mark the words as connection nodes to generate a connection part-of-speech tag structure set. S102: By connecting the part-of-speech tagging structure set, locate the verbs and subjects in the sentences corresponding to the connecting nodes, extract the subject phrases before and after the position of the verb and determine whether they constitute a subject-predicate structure, identify the subject-predicate pairs with related meanings in continuous sentences, record the start and end positions and semantic attribution of the subject-predicate structure in the original text, and generate a subject-predicate phrase combination mapping sequence. S103: Based on the subject-predicate combination mapping sequence, arrange each semantic unit according to the order of subject-predicate combinations in the original text, divide the sentence boundaries and assemble them into semantic blocks, determine whether semantic continuation is formed between semantic groups based on the number of semantic repetitions between the subject and predicate in the semantic blocks, classify the blocks of related subject-predicate combinations into the same interval, and obtain the functional partition of the sentence segment.

4. The method for processing data based on a large language model according to claim 3, characterized in that, The steps for obtaining the semantic chain trigger index sequence are as follows: S201: Call the statement units in the sentence segment functional partition, extract the noun groups and verb groups with the highest frequency of occurrence in the sentence block, classify the noun groups as topic words, classify the verb groups as relation words, and record the start and end positions in the sentence block as semantic index ranges to obtain the topic relation word position information set; S202: Based on the topic relation word position information set, the index range of the word group is compared vertically according to the original arrangement order of the word blocks in the text. The direction of movement of the index position in the previous word block and the next word block are compared. The forward, backward or unchanged states are marked respectively. The original text position is recorded to obtain the semantic turning point index table. S203: Call the semantic transition node index table, analyze the sentence content corresponding to the transition node, extract the verb and noun combination segments near the node, determine whether the node falls in the combination segment after the verb and before the noun, if the positional relationship is satisfied, mark it as the semantic chain starting point, and generate the semantic chain trigger index sequence.

5. The method for processing data based on a large language model according to claim 4, characterized in that, The specific steps for obtaining the tag path breakpoint location group are as follows: S301: Using the semantic chain to trigger the index sequence, locate the starting and ending blocks corresponding to the trigger index in the text, extract the noun groups that appear with priority in the two blocks, and record the position index range and the block identifier in the original text to generate a set of key noun groups for the blocks. The extraction of noun phrases from the two language blocks based on frequency is performed using the following formula: ； in, Score the priority of noun phrases. This represents the frequency of the i-th noun group in the chunk. This represents the average frequency of the noun phrase. represents the relevance score of the i-th noun group to the context, and k represents the number of noun groups; S302: Call the noun groups of the starting and ending blocks in the set of key noun groups of the block, perform pairwise cross-combination, identify whether there is word overlap in each combination, count the number of combinations without common words, if the number of combinations without common words reaches three or more, mark the starting and ending block segments as logical jump points, and obtain the label path breakpoint position group.

6. The method for processing data based on a large language model according to claim 5, characterized in that, The specific steps for obtaining the nested hierarchy table of the tag structure are as follows: S401: Call the text segment corresponding to the jump point in the tag path breakpoint position group, extract the noun phrases that appear more than twice in the text segment as candidate tag phrases, record the first word position and the last word position of the tag phrase in the sentence, and obtain the tag phrase position index table. S402: Based on the start and end position indexes of each group of words in the tag phrase position indexing table, arrange the tag phrases in the order of the sentence, compare whether there are overlapping areas or nested relationships between the front and back boundary ranges, and generate a tag path nested and interleaved tag set; S403: Call the nested and interleaved tag set of the tag path, establish the tag level index number according to the nesting layer, classify the word groups with the same nesting starting point into the same level path group, and obtain the tag structure nesting level table.

7. The method for processing data based on a large language model according to claim 6, characterized in that, The comparison of whether there are overlapping areas or nested relationships between the front and rear boundary ranges is performed using the following formula: ； in, This represents the boundary overlap strength index between the p-th and q-th tag phrases. This represents the starting index of the p-th tag phrase. This represents the index value of the end position of the p-th tag phrase. This represents the starting index of the q-th tag phrase. This represents the index value of the end position of the q-th tag phrase. This represents the starting index of the r-th tag phrase. This represents the average index value of the starting position of the first u group of tag phrases. This represents the index value of the end position of the r-th tag phrase. This represents the average index value of the end position of the preceding u-group of tag phrases. This represents the span value of the r-th tag phrase. This represents the number of tag pairs.

8. The method for processing data based on a large language model according to claim 1, characterized in that, The specific steps for obtaining the language model label mapping combination list are as follows: S501: Call the hierarchical index number of each tag path in the nested hierarchy table of the tag structure, retrieve the context segment of the corresponding tag in the original text in the pre-training corpus of the large language model, extract the position identifier range of the tag context segment in the model input, establish the context input index set of the tag segment, and obtain the tag context mapping location set. S502: Based on the tag context mapping location set, inject the corresponding tag structure nesting level information into the output mapping sequence of the large language model, extract the tag word direction representation generated at the corresponding position in the output layer before and after injection, compare the distribution trajectory between each group of tag word directions, locate and classify the offset direction, angle and spacing, and obtain the tag word direction distribution deviation information. S503: Based on the label combination trajectory with continuous offset and consistent direction in the label word distribution deviation information, extract the word path that can be semantically traced back, determine whether the logical word combination of the first and last words connecting the labels in the word path satisfies the original semantic chain order, filter the label paths that conform to the sequential logic and arrange them into nested combinations, and obtain the list of language model label mapping combinations.

9. A system for processing data based on a large language model, characterized in that, The system is used to implement the method for processing data based on a large language model as described in any one of claims 1-8, the system comprising: The sentence block recognition module identifies the position of punctuation marks based on sentences in a text paragraph, detects the part-of-speech category of the connecting words between punctuation marks, determines whether the sentences adjacent to the connecting words contain subject phrases and predicate verb structures respectively, divides the entire text into multiple sentence blocks according to the order of appearance of the phrases in the sentences, and generates sentence block functional partitions. The semantic indexing module calls the sentence segment functional partition, extracts the topic words and relation words in the sentence block, records the position of the phrase in the context block, identifies whether the movement direction of the phrase has been reversed, determines whether the reversal point is located between the verb and noun combination segment, and obtains the semantic chain trigger index sequence. The logical jump module triggers the index sequence according to the semantic chain, locates the corresponding start and end blocks in the text paragraph, extracts all noun groups in the start and end blocks, and performs full combination and arrangement according to the noun groups at both ends to generate the tag path breakpoint position group. The tag extraction module uses the tag path breakpoint location group to extract tag phrases from the sentence block pointed to by the jump node, identifies the intra-sentence position number of the tag phrases in the sentence block, determines whether there are intersecting positions or nested arrangements, and obtains the tag structure nesting hierarchy table. The tag mapping module uses the nested hierarchical table of the tag structure to extract the context of the segment that matches the tag phrase in the large language model, identify the word vector position distribution of the tag group in the output layer of the language model, compare the corresponding difference with the default word vector distribution, and generate a list of language model tag mapping combinations.

Images