Intelligent analysis method and system for user intent based on large language model

Abstract

The application discloses a user intention intelligent analysis method and system based on a large language model, and relates to the technical field of intelligent information processing. The method comprises the following steps: obtaining original query text input by a user, analyzing a subject qualification type identifier, a behavior qualitative label and an object constraint condition, and generating a deterministic intention anchor point; calculating a weighted coverage density of a word sequence relative to a fuzzy word table to obtain a demand fuzzy entropy, and calculating a semantic compactness based on a co-occurrence distance; dynamically determining a logic expansion level value according to the two, expanding the behavior qualitative label in levels, and generating an expanded behavior set; extracting an associated article index number of each behavior, performing subject qualification screening and text overlap rate reordering, and obtaining a rearranged article index sequence; combining the content of the first article with the demand fuzzy entropy to correct the display size, and generating a search result with an intention analysis feature code. The application solves the problem of inaccurate query intention understanding and matching in regulation search.

Description

Technical Field

[0001] This invention relates to the field of intelligent information processing technology, specifically to a method and system for intelligent analysis of user intent based on a large language model. Background Technology

[0002] With the continuous improvement of the legal and regulatory system, legal retrieval platforms have become an important channel for the public to obtain legal information. Users can retrieve relevant legal and regulatory provisions by entering natural language query text.

[0003] Currently, regulatory searches primarily employ keyword-based matching methods. These methods segment the user-input query text by word, perform literal matching within a regulatory database, and return results based on the match rate. When the user-input query text is clear and the intent is explicit, keyword matching methods can meet search needs to a certain extent.

[0004] However, in practice, user queries on regulatory search platforms often exhibit characteristics such as vague wording, mixed elements, and unclear intent. On one hand, user queries may contain numerous vague expressions, such as ambiguous degrees, scopes, or conditions. Keyword matching methods lack a quantification mechanism for this ambiguity in demand, making it impossible to adaptively adjust search strategies accordingly. On the other hand, the semantic relationships between information elements such as subject qualifications, behavioral characteristics, and object constraints in user queries are difficult to accurately identify through literal matching. This can easily lead to search results either omitting relevant regulations or returning a large amount of irrelevant information, failing to accurately match the user's actual information needs. Summary of the Invention

[0005] This invention provides a user intent intelligent analysis method and system based on a large language model, which solves the technical problem in the prior art that it is difficult to accurately understand and parse the information needs in user query text during legal retrieval, and thus difficult to achieve accurate matching of query intent.

[0006] In a first aspect, the present invention provides a user intent intelligent analysis method based on a large language model, the method comprising: Obtain the original query text input by the user in the legal retrieval platform, and parse the subject qualification type identifier, behavior qualitative label and object constraint conditions in the original query text to generate a deterministic intent anchor point; Calculate the weighted coverage density of the word sequence in the original query text relative to the preset fuzzy vocabulary to obtain the demand fuzzy entropy, and calculate the semantic compactness based on the co-occurrence distance of the subject qualification type identifier, behavior qualitative label and object constraint in the original query text; Based on the required fuzzy entropy and the semantic compactness, the logical expansion level value is dynamically determined, and the behavioral qualitative labels in the deterministic intent anchor are hierarchically expanded according to the logical expansion level value to generate an expanded behavior set; The associated legal provision index number of each behavior is extracted from the extended behavior set, and after being filtered by the subject qualification type identifier and reordered by the text overlap rate of the object constraint conditions and the constituent elements of the legal provision, a rearranged legal provision index sequence is obtained. The content of the legal provision corresponding to the first legal provision index number in the rearranged legal provision index sequence is combined with the required fuzzy entropy to adjust the display length of the legal provision content and generate a search result with intent parsing feature code.

[0007] Secondly, this invention also provides a user intent intelligent analysis system based on a large language model. The system includes: The intent anchor parsing module is used to obtain the original query text input by the user in the legal retrieval platform, and parse it to obtain the subject qualification type identifier, behavior qualitative label and object constraint conditions in the original query text, and generate a deterministic intent anchor. The demand fuzzy quantification module is used to calculate the weighted coverage density of the word sequence in the original query text relative to the preset fuzzy vocabulary, obtain the demand fuzzy entropy, and calculate the semantic compactness based on the co-occurrence distance of the subject qualification type identifier, behavior qualitative label and object constraint conditions in the original query text. The dynamic hierarchical expansion module is used to dynamically determine the logical expansion level value based on the required fuzzy entropy and the semantic compactness, and to expand the behavioral qualitative labels in the deterministic intent anchor point hierarchically according to the logical expansion level value to generate an expanded behavior set. The legal provision matching and sorting module is used to extract the associated legal provision index number of each behavior from the extended behavior set, and re-sort it according to the text overlap rate of the subject qualification type identifier and the object constraint conditions and the constituent elements of the legal provision to obtain a rearranged legal provision index sequence. The retrieval result generation module is used to combine the content of the legal provision corresponding to the first legal provision index number in the rearranged legal provision index sequence with the required fuzzy entropy, correct the display length of the legal provision content, and generate retrieval results with intent parsing feature codes.

[0008] One or more technical solutions provided in this invention have at least the following technical effects or advantages: First, this invention uses a pre-trained large language model to perform word-level tag parsing on the user-input raw query text, extracting subject qualification type identifiers, behavioral qualitative tags, and object constraints, thus transforming unstructured natural language queries into structured deterministic intent anchors. Compared to existing technologies that rely solely on literal keyword matching, this invention can identify deep semantic elements implicit in the query text, such as subject information, behavioral descriptions, and object constraints, solving the problem that traditional methods cannot accurately understand the information needs in user query texts.

[0009] Second, this invention obtains the fuzzy entropy of the query by calculating the weighted coverage density of the word sequence in the original query text relative to a preset fuzzy vocabulary, and calculates semantic compactness based on the co-occurrence distance of the three types of tags in the intent anchor, thus achieving a quantitative description of the fuzziness and compactness of the query request. The logical expansion level value is dynamically determined based on the comparison result of the fuzzy entropy and semantic compactness, and the superordinate word chain is expanded accordingly for the behavioral qualitative tags. The search scope is automatically expanded when the user's expression is vague, and narrowed when the expression is clear, achieving a dynamic balance between the broad coverage of fuzzy queries and the precise matching of clear queries.

[0010] Third, this invention employs subject qualification type identifiers for substring matching and filtering during the legal provision matching stage. It also re-ranks the text based on the text overlap rate between the lexical units of the object constraint conditions and the constituent elements of the legal provision. Furthermore, it incorporates demand fuzzy entropy to correct the display length of the first legal provision. Compared to traditional methods that rely solely on single keyword matching and uniformly display the full text, this invention performs comprehensive filtering from both subject and object matching dimensions, and adaptively adjusts information density based on demand fuzziness, achieving a reasonable balance between information content and accuracy in the search results.

[0011] In summary, this invention solves the technical problems in existing technologies that make it difficult to accurately understand and parse the information needs in user query texts and achieve precise matching of query intents during legal retrieval by using large language model-driven intent parsing, fuzziness-adaptive retrieval strategies, and multi-dimensional legal provision matching and ranking. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the user intent intelligent analysis method based on a large language model provided in an embodiment of the present invention; Figure 2This is a schematic diagram of the user intent intelligent analysis method based on a large language model provided in this embodiment of the invention. Figure 3 This is an example table of the legal provision matching and reordering process of the user intent intelligent analysis method based on a large language model provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of the structure of the user intent intelligent analysis system based on a large language model provided in an embodiment of the present invention; The diagram is labeled as follows: Intent Anchor Parsing Module 11, Demand Fuzzy Measurement Module 12, Dynamic Hierarchical Expansion Module 13, Legal Provision Matching and Sorting Module 14, and Search Result Generation Module 15. Detailed Implementation

[0014] This invention provides a user intent intelligent analysis method and system based on a large language model, which solves the technical problem in the prior art that it is difficult to accurately understand and parse the information needs in user query text during legal retrieval, and thus difficult to achieve accurate matching of query intent.

[0015] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0016] Example 1, as Figure 1 As shown, this invention provides a flowchart illustrating a user intent intelligent analysis method based on a large language model; as... Figure 2 As shown, this invention provides a scheme logic diagram for a user intent intelligent analysis method based on a large language model. The method includes: S100: Obtain the original query text input by the user in the legal retrieval platform, and parse the subject qualification type identifier, behavior qualitative label and object constraint conditions in the original query text to generate a deterministic intent anchor point; Users typically input their legal needs into legal search platforms in natural language, such as "legal liability of shareholders of a limited liability company for failing to pay their capital contributions on time and in full." These texts contain multiple semantic elements, including subject information, behavioral descriptions, and object limitations, but these elements are mixed together in an unstructured form within the same sentence. Traditional keyword-based literal matching search methods cannot distinguish between the subject, behavior, and object in the text, resulting in search results that do not accurately reflect the user's true intent.

[0017] Therefore, this step utilizes a pre-trained large language model to perform word-level semantic parsing on the original query text, extracting subject qualification type identifiers, behavioral qualitative labels, and object constraints. The parsing results are then associated with and stored with various statistical parameters to generate structured deterministic intent anchors, providing structured semantic input for subsequent steps such as intent fuzzy quantification, hierarchical expansion, and legal provision matching.

[0018] Step S100 in the method provided in this embodiment of the invention includes: The original query text is input into a pre-trained large language model, and the output is the label type sequence corresponding to each word. Extract consecutive word fragments labeled as subject type from the tag type sequence, including at least natural person name fragments, legal person name fragments, and non-legal person organization name fragments; The total character length, the number of appellations contained in the word fragments of the subject type, and the first occurrence position number in the original query text are counted and used as the subject name length, appellation count, and subject position index, respectively. Concatenate the word fragments labeled as subject type into a text fragment of subject qualification type identifier in the original order, and record the starting character position and ending character position of the subject qualification type identifier; Extract consecutive word segments labeled as behavioral characteristics from the tag type sequence, wherein the behavioral characteristics include at least verb phrase segments and verb-object phrase segments; The occurrence frequency of negative words, total character length, and number of adverbs in the lexical segments used to characterize the behavior are counted and used as the count of negative words, the length of the behavior phrase, and the count of adverb modifications, respectively. The word fragments of the behavior qualitative analysis are concatenated in their original order to form the text fragments of the behavior qualitative labels, and the starting character position and ending character position of the behavior qualitative labels are recorded. Extract consecutive word segments labeled as object constraints from the tag type sequence, wherein the object constraints include at least noun phrase segments and quantity phrase segments; The number of times quantifiers appear, the total length of characters, and the number of adjectives appear in the lexical fragments constrained by the object are counted, and these are used as the quantifier count, the object phrase length, and the adjective modification count, respectively. The word segments marked as object constraints are concatenated in their original order to form the text segment of the object constraint condition, and the starting character position and ending character position of the object constraint condition are recorded. The text fragments of the subject qualification type identifier, the text fragments of the behavior qualitative label, and the text fragments of the object constraint condition are associated and stored with their respective start character positions and end character positions, as well as various statistical parameters, to generate the deterministic intent anchor point.

[0019] The specific implementation method is as follows: First, the original query text input by the user is fed into a pre-trained large language model. The large language model employs a pre-trained network based on the Transformer architecture, containing a shared semantic encoding layer and three parallel output branches, used to predict whether a word belongs to a subject type, behavioral type, or object constraint, respectively. The original query text is segmented into word sequences by character and then input into the model. The shared semantic encoding layer extracts the contextual semantic representation for each word, and the three output branches output the probability distribution of each word belonging to a subject type, behavioral type, or object constraint, respectively. The label type with the highest probability for each word is taken as the labeling result for that word, and the output is the label type sequence corresponding to each word.

[0020] For example, the original query text input by the user is "Shareholders of a limited liability company have failed to pay their capital contributions in full and on time". After inputting the text into the large language model, the output label type sequence is as follows: "Limited Liability Company" is labeled as subject type, "Shareholders" is labeled as subject type, "failed" is labeled as behavior type, "on time" is labeled as behavior type, "in full" is labeled as behavior type, "pay" is labeled as behavior type, and "capital contribution" is labeled as object constraint.

[0021] Further, extract the subject qualification type identifier from the tag type sequence. Filter consecutive word segments labeled as subject type in the tag type sequence; for example, both consecutive word segments "limited liability company" and "shareholder" are labeled as subject type. Count the total character length of the subject type word segment, i.e., the sum of the character counts of all subject type word segments. Count the number of appellations contained in the subject type word segment; appellations refer to words indicating the subject's identity or role, such as "shareholder," "legal representative," "creditor," etc. Record the first occurrence position number of the subject type word segment in the original query text, i.e., the position index of the first word segment labeled as subject type in the entire word sequence, starting from 0. Concatenate the subject type word segments in their original order to form a text segment with the subject qualification type identifier, and record the starting and ending character positions of this text segment in the original query text. For example, the total character length of "limited liability company shareholder" is 8, the appellation "shareholder" is counted as appellation count 1, the first occurrence position number is 0, the starting character position is 0, and the ending character position is 7.

[0022] Further, a behavior qualification tag is extracted from the tag type sequence. Token fragments continuously labeled as behavior qualification in the tag type sequence are screened, for example, four consecutive tokens "wei", "anqi", "zu'e" and "jiaona" are all labeled as behavior qualification. The occurrence count of negative words in the behavior qualification token fragments is counted, and the negative words are, for example, "wei", "bu", "wu", "meiyou", etc. The total character length of the behavior qualification token fragments and the occurrence number of adverbs therein are counted. The behavior qualification token fragments are spliced in the original order to form a text fragment of the behavior qualification tag, and the start character position and the end character position of the text fragment are recorded. Exemplarily, the negative word "wei" in "wei anqi zu'e jiaona" occurs 1 time, the total character length is 6, the adverbs "anqi" and "zu'e" are counted as 2, the start character position is 8, and the end character position is 13.

[0023] Further, an object constraint condition is extracted from the tag type sequence. Token fragments continuously labeled as object constraint in the tag type sequence are screened. The occurrence count of quantifiers in the object constraint token fragments is counted, the total character length thereof, and the occurrence number of adjectives therein are counted. The object constraint token fragments are spliced in the original order to form a text fragment of the object constraint condition, and the start character position and the end character position of the text fragment are recorded. Exemplarily, the quantifier count of "chuzi" is 0, the total character length is 2, and the adjective modification count is 0.

[0024] Finally, the text fragments, start and end positions and statistical parameters of the three types of tags described above are stored in association, so as to generate a deterministic intention anchor point.

[0025] Step S100 in the method provided by the embodiment of the present invention further comprises: the pre-training step of the large language model comprises: Collecting original query text samples in historical query logs of a regulation retrieval platform, and manually labeling a tag type for each token in each sample, wherein the tag type at least comprises a subject type, a behavior qualification, an object constraint and an irrelevant type, so as to form a labeled sample set; Constructing a pre-trained network based on a Transformer architecture, wherein the pre-trained network comprises a shared semantic encoding layer and three parallel output branches; Inputting the original query text samples in the labeled sample set into the shared semantic encoding layer to obtain a context semantic representation of each token, inputting the context semantic representation into the three parallel output branches respectively, and each output branch outputs a probability distribution of each token belonging to the corresponding tag type through full connection mapping and a normalized exponential function; For each output branch, a cross-entropy loss function is used to calculate a branch loss value between the predicted probability distribution and the manually labeled tag type, the three branch loss values are added to serve as a total loss value, and internal parameters of the pre-trained network are updated iteratively through back propagation; When the decrease in the total loss value on the independent validation sample set is less than the convergence threshold for multiple consecutive iterations, training is stopped, and a pre-trained large language model is obtained.

[0026] The specific implementation method is as follows: First, raw query text samples were collected from the historical query logs of the legal retrieval platform. These logs record a large number of query texts actually entered by users on the platform, from which representative samples can be selected to construct a training dataset. Each word in each sample was manually labeled. Labelers assigned labels based on the semantic role of the word in the query text, classifying it as subject type, behavior type, object constraint, or irrelevant type. Subject type labeling indicates that the word describes the legal subject involved in the query, such as the name of a natural person, legal person, or unincorporated organization; behavior type labeling indicates that the word describes an action performed or not performed by the subject; object constraint labeling indicates that the word describes the object, condition, or result of the action; irrelevant type labeling indicates that the word is a function word, punctuation mark, or other non-core semantic component. All labeled samples were then aggregated to form a labeled sample set.

[0027] Furthermore, a pre-trained network based on the Transformer architecture is constructed. The pre-trained network consists of a shared semantic encoding layer and three parallel output branches. The shared semantic encoding layer contains multiple Transformer encoders, each consisting of a multi-head self-attention sublayer and a feedforward fully connected sublayer, followed by residual connections and layer normalization operations. The shared semantic encoding layer receives the token embedding sequence of the original query text as input and, through layer-by-layer abstraction by the multi-layer Transformer encoder, outputs a vector representation containing contextual semantic information for each token.

[0028] The three parallel output branches are the first, second, and third output branches. The first output branch predicts whether a word belongs to the subject type, the second output branch predicts whether a word belongs to the behavioral type, and the third output branch predicts whether a word belongs to the object type. Each output branch contains a fully connected mapping layer and a normalized exponential function. The fully connected mapping layer maps the context semantic representation vector output by the shared semantic encoding layer into a score vector with the same dimension as the number of label categories. The normalized exponential function performs an exponential operation on each element in the score vector and divides it by the sum of all exponent values, transforming the score vector into a predicted probability distribution for each label category.

[0029] During training, the original query text samples from the labeled sample set are input into the shared semantic encoding layer to obtain the contextual semantic representation of each word. This contextual semantic representation is then input into three output branches, each outputting the probability distribution of each word belonging to its corresponding label type. For each output branch, the cross-entropy loss function is used to calculate the branch loss value between the predicted probability distribution and the manually labeled label type. The cross-entropy loss function takes the negative logarithm of the predicted probability of the true label's corresponding category as the loss value for that word, and the average of all word loss values ​​within a batch is the branch loss value for that branch. The three branch loss values ​​are summed to obtain the total loss value. The gradient of the total loss value with respect to the network's layer parameters is calculated using the backpropagation algorithm and the chain rule. A gradient descent optimizer is then used to iteratively update the network parameters along the gradient's inverse direction. During training, the labeled sample set is divided into a training set and an independent validation set. After each iteration, the parameters are updated on the training set, and the total loss value is calculated on the independent validation set. When the total loss value decreases by less than a preset convergence threshold for multiple consecutive iterations on the independent validation sample set, for example, when the decrease is less than one-thousandth for five consecutive iterations, training stops and a pre-trained large language model is obtained.

[0030] Step S100 in the method provided in this embodiment of the invention further includes: storing the text fragments of the subject qualification type identifier, the behavior qualitative label, and the object constraint condition, along with their respective corresponding start and end character positions and statistical parameters, to generate the deterministic intent anchor point, including: The text fragment of the subject qualification type identifier, the start and end character positions of the subject qualification type identifier, the subject name length, the title count, and the subject position index are encapsulated into a first feature area; The text fragments of the behavioral qualitative labels, the starting and ending character positions of the behavioral qualitative labels, the count of negative words, the length of behavioral phrases, and the count of adverb modifications are encapsulated into a second feature region; The text fragment of the object constraint, the starting and ending character positions of the object constraint, the quantifier count, the object phrase length, and the adjective modification count are encapsulated into a third feature region; Separators are inserted between adjacent feature areas in the order of the first feature area, the second feature area, and the third feature area to form an anchor content sequence. The length of the subject name, the number of titles, the number of negative words, and the number of quantifiers are added together to obtain a verification cumulative value. The verification cumulative value is converted into a fixed-length string and appended to the end of the anchor content sequence to generate the deterministic intent anchor.

[0031] The specific implementation method is as follows: The text fragments of the subject qualification type identifiers obtained in the above steps, along with their start and end character positions, together with the subject name length, title count, and subject position index, are encapsulated into the first feature region. The three statistical parameters—subject name length, title count, and subject position index—will serve as inputs in the semantic compactness calculation and the first compactness threshold calculation in subsequent steps.

[0032] The text fragments containing the behavioral qualitative labels obtained in the above steps, along with their start and end character positions, together with the negation word count, behavioral phrase length, and adverb modification count, are encapsulated into a second feature region. The negation word count will be used as a criterion in determining the logical expansion level value and generating the expanded behavior set in subsequent steps, while the behavioral phrase length will be used as a parameter in the calculation of the first entropy threshold.

[0033] The text fragments of the object constraints obtained in the above steps, along with their start and end character positions, quantifier counts, object phrase lengths, and adjective modifier counts, are encapsulated into a third feature region. The text fragments of the object constraints and the quantifier counts will be used as the basis for matching in the text overlap rate calculation in subsequent steps.

[0034] Following the order of the first feature area, the second feature area, and the third feature area, a preset separator, such as the special character "|", is inserted between adjacent feature areas to connect the three feature areas into a whole, forming an anchor content sequence.

[0035] Further, the verification cumulative value is calculated. This value is used in subsequent steps to generate intent parsing feature codes, facilitating the verification and traceability of search results. The verification cumulative value is obtained by adding the four statistical parameters: subject name length, title count, negative word count, and quantifier count. This value is then converted into a fixed-length string and appended to the end of the anchor content sequence. The conversion method is as follows: the verification cumulative value is treated as a decimal integer, its absolute value is taken, and zero-padding is performed according to a fixed number of digits. If the number of digits is insufficient, leading zeros are added; if it exceeds the fixed number of digits, the last digit is truncated. For example, if the fixed number of digits is set to 8, and the verification cumulative value is 123, the converted value is "00000123".

[0036] For example, the original query text "Shareholders of a limited liability company failed to pay their capital contributions on time and in full" has a subject name length of 8, a title count of 1, a negative word count of 1, and a quantifier count of 0. The cumulative check value is 8+1+1+0=10, which is converted to an 8-character fixed-length string "00000010". The three feature areas are as follows: The first feature area contains the subject qualification type identifier text fragment "shareholders of a limited liability company", starting character position 0, ending character position 7, subject name length 8, title count 1, and subject position index 0; the second feature area contains the behavior qualitative label text fragment "failed to pay on time and in full", starting character position 8, ending character position 13, negative word count 1, behavior phrase length 6, and adverb modification count 2; the third feature area contains the object constraint condition text fragment "capital contribution", starting character position 14, ending character position 15, quantifier count 0, object phrase length 2, and adjective modification count 0. Insert a separator “|” between adjacent feature regions to connect them, and append the checksum value “00000010” to the end of the anchor content sequence to generate a deterministic intent anchor.

[0037] The following technical effects were achieved through this step: First, by using a pre-trained large language model to perform word-level semantic parsing on the original query text input by the user, the system extracts the subject qualification type identifier, behavior qualitative label and object constraint conditions, and transforms the unstructured natural language query into a structured intent anchor containing text fragments, location information and statistical parameters, so that subsequent retrieval steps can accurately identify the core legal elements in the user query.

[0038] Second, the statistical parameters included in the intent anchor, such as the length of the subject name, the number of titles, the number of negative words, the length of the action phrase, and the number of quantifiers, provide directly usable quantitative features for subsequent steps such as demand fuzzy entropy calculation, semantic compactness calculation, determination of logical expansion level value, and legal provision matching and filtering, thus avoiding the computational overhead of repeatedly parsing the original text in subsequent steps.

[0039] S200: Calculate the weighted coverage density of the word sequence in the original query text relative to the preset fuzzy vocabulary, obtain the demand fuzzy entropy, and calculate the semantic compactness based on the co-occurrence distance of the subject qualification type identifier, behavior qualitative label and object constraint in the original query text; The above steps generated deterministic intent anchors through large language model parsing, extracting subject qualification type identifiers, behavioral qualitative labels, and object constraints. However, the original query texts input by users vary in their expression. Some query texts contain a large number of vague terms, such as "approximately," "various types," and "if," while the distribution of subject, behavior, and object elements in some query texts is relatively scattered. Too many vague terms will make the search intent unclear, and a scattered distribution of elements means a loose semantic structure. Both situations affect the selection of subsequent search strategies and the accuracy of legal provision matching.

[0040] Step S200 in the method provided in this embodiment of the invention includes: Obtain a pre-constructed fuzzy word list, wherein the fuzzy word list contains three categories: degree fuzzy words, range fuzzy words, and condition fuzzy words, and each category of fuzzy words corresponds to a fuzzy contribution coefficient; The original query text is segmented by character to obtain a word sequence; Traverse each word in the word sequence, and when the word matches any fuzzy word in the fuzzy word list, extract the fuzzy contribution coefficient of the category to which the fuzzy word belongs; The unit contribution value of a lexical is obtained by dividing the fuzzy contribution coefficient by the absolute distance from each lexical to the starting character position of the behavior qualitative label. The unit contribution values ​​corresponding to all lexical units that hit the fuzzy word in the lexical sequence are summed, and combined with the total character length of the lexical sequence, the required fuzzy entropy is obtained. The starting character positions of the subject qualification type identifier, the starting character positions of the behavior qualitative label, and the starting character positions of the object constraint conditions are extracted from the deterministic intent anchor point and denoted as the first position, the second position, and the third position, respectively. Calculate the first absolute difference between the second position and the first position, calculate the second absolute difference between the third position and the second position, and combine the total character length of the original query text to calculate the normalized co-occurrence span; Subtracting the normalized co-occurrence span from the value 1 yields the semantic compactness.

[0041] The specific implementation method is as follows: First, obtain the pre-constructed fuzzy term list. The fuzzy term list contains three categories of fuzzy terms: degree fuzzy terms, scope fuzzy terms, and conditional fuzzy terms. Degree fuzzy terms describe uncertainty in degree, such as "approximately," "almost," "roughly," "around," and "up and down." These terms indicate that the user cannot provide a precise description of the query results, and their corresponding fuzzy contribution coefficient is set to 0.8. Scope fuzzy terms describe unclear scope or boundaries, such as "various types," "parts," "some," "multiple," and "several," and their corresponding fuzzy contribution coefficient is set to 0.6. Conditional fuzzy terms introduce assumptions or conditions, such as "if," "unless," "assuming," and "suppose," and their corresponding fuzzy contribution coefficient is set to 0.4. Degree fuzzy terms have the highest fuzzy contribution coefficient because uncertainty in degree directly leads to fuzziness in the search intent; the influence of conditional fuzzy terms is relatively weakest because conditional limitations can narrow the search scope.

[0042] Furthermore, the original query text is segmented by character to obtain a word segment sequence. Word segmentation can be performed using a dictionary-based word segmentation algorithm, which scans the original query text from left to right and segments it into words according to a preset dictionary. Each word segment corresponds to a continuous character fragment in the original query text.

[0043] Further, each lexical in the lexical sequence is traversed, and it is checked whether the lexical matches any fuzzy word in the fuzzy word list. If it matches, the fuzzy contribution coefficient of the category to which the fuzzy word belongs is extracted. Simultaneously, the starting character position of the behavioral qualitative label is extracted from the deterministic intent anchor point generated in step S100, denoted as the second position. This position is the index of the first character of the behavioral qualitative label in the original query text, counted from 0. The absolute distance between the position number of the currently matched lexical and the second position is calculated. The position number is the index of the current lexical in the entire lexical sequence, also counted from 0. The absolute distance is the absolute value of the difference between the two. The fuzzy contribution coefficient is divided by the sum of the absolute distance plus 1, and the quotient is the unit contribution value of the lexical. The denominator uses the absolute distance plus 1 to avoid the denominator being zero, and also to ensure that fuzzy words closer to the behavioral qualitative label contribute more to the unit contribution value.

[0044] Further, the fuzzy entropy of the demand is calculated. The unit contribution values ​​corresponding to all units hitting fuzzy words in the lexical sequence are summed to obtain the total unit contribution value. This total unit contribution value is then divided by the total length of the characters in the lexical sequence; the quotient is the fuzzy entropy of the demand. The value of the fuzzy entropy of the demand ranges from 0 to 1. A larger value indicates that the original query text contains more fuzzy expressions and is closer to the behavior center, indicating a less clear user demand; a smaller value indicates that the query text is clearer, contains fewer fuzzy words, or is far from the behavior center, indicating a relatively clear user demand.

[0045] Finally, semantic compactness is calculated. The starting character positions of the subject qualification type identifier, the behavioral qualitative label, and the object constraint condition are extracted from the deterministic intent anchor points generated in step S100, and denoted as the first position, second position, and third position, respectively. The first absolute difference between the second position and the first position is calculated, i.e., |second position - first position|; the second absolute difference between the third position and the second position is calculated, i.e., |third position - second position|. The first and second absolute differences are added together and divided by the total character length of the original query text to obtain the normalized co-occurrence span. The normalized co-occurrence span reflects the spatial distribution of the three types of elements—subject, behavior, and object—in the original query text. A smaller span value indicates that the three types of elements are more concentrated in the text, closer to the situation where the user directly expresses their complete needs in a short sentence; a larger span value indicates that the three types of elements are more dispersed, and the semantic structure is more loose. The difference between the value 1 and the normalized co-occurrence span is the semantic compactness. The semantic compactness value ranges from 0 to 1. The closer the value is to 1, the more compact the semantic relationship between the three types of elements and the clearer the user's intent; the closer the value is to 0, the more dispersed the elements and the looser the semantic structure.

[0046] For example, the starting character position of the subject qualification type identifier is 0, the starting character position of the behavior qualitative label is 8, the starting character position of the object constraint condition is 14, and the total length of the original query text is 16 characters. The first absolute difference is |8-0|=8, the second absolute difference is |14-8|=6, the normalized co-occurrence span is (8+6)÷16=14÷16=0.875, and the semantic compactness is 1-0.875=0.125, indicating that the distribution of the three types of elements in the query text is relatively dispersed.

[0047] The following technical effects were achieved through this step: First, by using three types of fuzzy words in the fuzzy word list and their differentiated fuzzy contribution coefficients, the fuzzy words in the user query text are quantitatively evaluated. The contribution value of the unit is calculated by combining the distance weight between the fuzzy words and the behavioral qualitative labels, so that the demand fuzzy entropy not only reflects the number of fuzzy words, but also reflects the correlation strength between fuzzy words and the core query behavior.

[0048] Second, by calculating the normalized co-occurrence span of the three types of elements—subject, behavior, and object—semantic compactness is obtained, which quantifies the spatial aggregation degree of core elements in the query text. This provides another dimension of evaluation index that complements the demand fuzzy entropy for the dynamic determination of logical expansion level values ​​in subsequent steps.

[0049] S300: Based on the required fuzzy entropy and the semantic compactness, dynamically determine the logical expansion level value, and expand the behavioral qualitative labels in the deterministic intent anchor point hierarchically according to the logical expansion level value to generate an expanded behavior set; The two quantitative indicators, fuzzy entropy and semantic compactness, reflect the degree of coverage of vague terms and the semantic compactness of core elements in user query texts, respectively. In legal retrieval, if the user query text is clear and the intent is explicit, a precise search should be conducted based on the user's directly expressed behavioral description. If the query text contains many vague terms, or if the three elements—subject, behavior, and object—are scattered throughout the text, the behavioral description should be appropriately expanded to cover other relevant legal situations that the user may have implicitly implied. If the same search scope is applied to all queries, fuzzy queries may miss relevant regulations, while explicit queries may return too many irrelevant results.

[0050] Step S300 in the method provided in this embodiment of the invention includes: dynamically determining the logical expansion level value based on the required fuzzy entropy and the semantic compactness, including: The demand fuzzy entropy is compared with a first entropy threshold to determine whether the demand fuzzy entropy exceeds the first entropy threshold, and a first judgment result is obtained, wherein the first entropy threshold is determined based on the count of negative words and the length of the behavior phrase; The semantic compactness is compared with a first compactness threshold to determine whether the semantic compactness is lower than the first compactness threshold, and a second determination result is obtained, wherein the first compactness threshold is determined based on the length of the subject name; When the first judgment result is yes and the second judgment result is yes, the logical expansion level value is set to the maximum value; When the first judgment result is negative and the second judgment result is negative, the logical expansion level value is set to the minimum value; When the first judgment result is inconsistent with the second judgment result, obtain the count of negative words contained in the behavior qualitative label. If the count of negative words is greater than zero, set the logical expansion level value to a value that is relatively large in the middle; otherwise, set it to a value that is relatively small in the middle. The logical expansion level value is an integer between 1 and 4.

[0051] The specific implementation method is as follows: First, the first entropy threshold is calculated. The first entropy threshold is determined based on the negative word count and behavior phrase length obtained in step S100. It is calculated by multiplying the negative word count by the behavior phrase length and then dividing by 100. The first entropy threshold reflects a reasonable benchmark for comparing the fuzzy entropy of the demand, considering whether negative expressions and behavior description lengths exist in the current query. For example, if the negative word count is 1 and the behavior phrase length is 6 obtained in step S100, then the first entropy threshold is (1 × 6) ÷ 100 = 0.06.

[0052] Further, a first compactness threshold is calculated. The first compactness threshold is determined based on the subject name length obtained in step S100. It is calculated by dividing the subject name length by the sum of the subject name length and 10. A longer subject name indicates more complete subject information in the query text, allowing for a more relaxed compactness requirement; therefore, the first compactness threshold increases accordingly. For example, if the subject name length obtained in step S100 is 8, then the first compactness threshold is 8 ÷ (8 + 10) = 8 ÷ 18 ≈ 0.444.

[0053] Further, a first judgment and a second judgment are performed. The demand fuzzy entropy obtained in step S200 is compared with a first entropy threshold to determine whether the demand fuzzy entropy exceeds the first entropy threshold, thus obtaining a first judgment result. The semantic compactness obtained in step S200 is compared with a first compactness threshold to determine whether the semantic compactness is lower than the first compactness threshold, thus obtaining a second judgment result. For example, if the demand fuzzy entropy is 0.046 and the first entropy threshold is 0.06, the demand fuzzy entropy does not exceed the first entropy threshold, and the first judgment result is negative. If the semantic compactness is 0.125 and the first compactness threshold is 0.444, the semantic compactness is lower than the first compactness threshold, and the second judgment result is positive. The two judgment results are inconsistent.

[0054] Furthermore, the logical expansion level value is determined based on the combination of the two judgment results.

[0055] When both the first and second judgment results are true, it indicates that the query text has high coverage of fuzzy words and a loose semantic structure. The requirement's fuzzy entropy exceeding the first entropy threshold indicates a large number of fuzzy terms, while the semantic compactness falling below the first compactness threshold indicates a scattered distribution of the three types of elements. In this case, setting the logical expansion level value to the maximum value of 4 indicates that the highest level of superordinate term expansion is needed to cover various legal situations that the user may be involved in.

[0056] When both the first and second judgment results are negative, it indicates that the query text has low coverage of fuzzy words and a compact semantic structure. The fact that the demand fuzzy entropy does not exceed the first entropy threshold suggests that there are few fuzzy words, and that the semantic compactness is not lower than the first compactness threshold indicates that the three types of elements are concentrated. In this case, setting the logical expansion level value to the minimum value of 1 indicates that the user's intent is clear, and there is no need for higher-level word expansion; the search can be performed directly using the user's original behavior description.

[0057] When the first and second judgment results are inconsistent, further judgment is needed by combining the negative word count. Obtain the negative word count contained in the behavioral qualitative label from the deterministic intent anchor point in step S100. If the negative word count is greater than zero, it indicates that the user query contains negative expressions, such as "not on time" or "not in full." The presence of negative words means that the user may be describing a legal issue involving a negative scenario. In this case, the logical expansion level value is set to a relatively high value of 3. If the negative word count is equal to zero, it indicates that the user query does not contain negative expressions, and the behavioral description is in an affirmative form. In this case, the logical expansion level value is set to a relatively low value of 2.

[0058] For example, if the first judgment result is negative and the second judgment result is positive, the two are inconsistent, the negative word count is 1 and greater than zero, so the logic expansion level value is set to a relatively large value of 3.

[0059] Step S300 in the method provided in this embodiment of the invention further includes: hierarchically expanding the behavioral qualitative labels in the deterministic intent anchor point according to the logical expansion hierarchy value to generate an expanded behavior set, including: Obtain a pre-constructed behavior hypernym mapping relationship, wherein each behavior verb phrase in the behavior hypernym mapping relationship is associated with a hypernym chain arranged in ascending order of abstraction level; Extract the hypernym chain that matches the text fragment of the behavior qualitative label from the behavior hypernym mapping relationship. If the match fails, the text fragment of the behavior qualitative label itself is taken as the unique hypernym. The expansion quantity is determined based on the logical expansion level value; Select the first extended number of hypernyms starting from the starting position in the hypernym chain, merge them with the text fragments of the behavior qualitative labels, and remove duplicate terms to form a basic behavior set; Obtain the count of negative words contained in the behavioral qualitative label. If the count of negative words is greater than zero, add a negative prefix to each behavior in the basic behavior set to obtain an extended behavior set. If the count of the negation words is zero, then the basic set of behaviors is used as the extended set of behaviors.

[0060] The specific implementation method is as follows: First, we obtain the pre-constructed mapping relationships of behavioral superordinate terms. These mapping relationships are a knowledge base that maps specific behavioral verb phrases to chains of superordinate terms arranged in ascending order of abstraction. From the beginning to the end of the superordinate term chain, the level of abstraction of the superordinate terms gradually increases, and the semantic coverage gradually expands. For example, the superordinate term chain associated with "paying capital contribution" is "paying capital contribution → fulfilling capital contribution obligations → shareholder contribution → capital replenishment." Here, "paying capital contribution" is the most specific behavioral description, "fulfilling capital contribution obligations" is slightly more abstract, "shareholder contribution" is even more abstract, and "capital replenishment" is the highest-level legal concept, which has a legal inclusive relationship with "paying capital contribution." These superordinate term chains are compiled by legal experts based on the legal and regulatory framework.

[0061] Furthermore, text fragments containing behavioral qualitative labels are extracted from deterministic intent anchors. Using these text fragments as query terms, a matching behavioral verb phrase is searched within the behavioral hypernym mapping relationship, and the corresponding hypernym chain is extracted. If a match is successful, the hypernym chain of the behavioral qualitative label is obtained; if a match fails, meaning there is no entry for the behavioral verb phrase in the behavioral hypernym mapping relationship, the text fragment of the behavioral qualitative label itself is used as the unique hypernym, and no expansion is performed.

[0062] The expansion quantity is calculated based on the logical expansion level value determined in the above steps. The expansion quantity is equal to the logical expansion level value minus 1. When the logical expansion level value is 1, the expansion quantity is 0, that is, no expansion is performed; when the level value is 4, the expansion quantity is 3, that is, 3 hypernyms are selected from the hypernym chain for expansion.

[0063] Furthermore, select the first few extended superordinate terms from the starting position in the superordinate term chain. For example, if the logical expansion level is 3, the extension number is 2, and the superordinate term chain is "pay capital contribution → fulfill capital contribution obligation → shareholder capital contribution → capital replenishment", then select the first two superordinate terms "fulfill capital contribution obligation" and "shareholder capital contribution". Merge the selected superordinate terms with the text fragments of the behavioral qualitative labels into a set, and after deleting any duplicate terms in the set, form the basic behavior set. For example, the merged basic behavior set is {"pay capital contribution", "fulfill capital contribution obligation", "shareholder capital contribution"}.

[0064] Finally, the case of negative words is handled. The count of negative words contained in the behavioral qualitative tags is obtained from the deterministic intent anchor. If the negative word count is greater than zero, it indicates that the user query contains a negative expression. In this case, a negative prefix is ​​added to each behavior in the basic behavior set; the negative prefix is ​​the negative word extracted in step S100. For example, if the negative word is "not", then the basic behavior set {"Pay capital contribution", "Fulfill capital contribution obligation", "Shareholder capital contribution"} becomes the extended behavior set {"Not paid capital contribution", "Not fulfilled capital contribution obligation", "Not shareholder capital contribution"} after adding the negative prefix. By adding the negative prefix, the extended behavior set not only covers the negative behaviors directly queried by the user but also covers related higher-level behaviors, while preserving the negative semantics. If the negative word count is zero, the basic behavior set is directly used as the extended behavior set.

[0065] The following technical effects were achieved through this step: First, by comparing the demand fuzzy entropy with the first entropy threshold determined based on the count of negative words and the length of behavioral phrases, and by comparing the semantic compactness with the first compactness threshold determined based on the length of the subject name, a two-dimensional quantitative evaluation of the fuzziness of the user's query intent and the semantic compactness is achieved, providing an adaptive decision-making mechanism for the dynamic determination of the logical expansion level value.

[0066] Second, by mapping relationships between behavioral superordinate terms and logically expanding hierarchical values, a hierarchical expansion of behavioral qualitative labels is achieved. When a user query is vague or loosely structured, the search scope is automatically expanded along the superordinate term chain to cover relevant legal situations; when a user query is clear and compactly structured, the original behavioral description is preserved for accurate retrieval. Simultaneously, negative word processing retains the negative semantics of the user query, avoiding the loss of semantic information during the expansion process.

[0067] S400: Extract the associated legal provision index number of each behavior from the extended behavior set, filter it by the subject qualification type identifier and reorder it by the text overlap rate of the object constraint conditions and the constituent elements of the legal provision to obtain the rearranged legal provision index sequence; The above steps generate an extended behavior set through logical expansion of hierarchical values ​​and hypernym chain expansion. This set contains multiple expressions of user query behavior at different levels of abstraction. However, each behavior in the extended behavior set only describes the behavioral dimension of the query and has not yet been associated with a specific legal provision. In legal retrieval, the applicability of a legal provision is closely related to the qualification type of the querying entity; the same behavior by different entities may be subject to completely different legal provisions. Furthermore, the degree of textual matching between the constituent elements of the legal provision and the object constraints in the user query directly reflects the relevance of the search results to the user's needs.

[0068] Step S400 in the method provided in this embodiment of the invention includes: Traverse each behavior in the extended behavior set, match legal provisions containing behavior text fragments from the legal provision index library of the legal retrieval platform, extract the index numbers of the successfully matched legal provisions, and merge and remove duplicates to obtain the initial index set; Iterate through each legal provision index number in the initial index set to obtain the subject scope field of the legal provision corresponding to the legal provision index number; The text fragment identifying the subject qualification type is matched with the subject scope field for substring matching, and the legal provision index number that does not contain the text fragment is removed to obtain the first filtering index set; Extract the text fragments of the object constraint conditions from the deterministic intent anchor points, and segment the text fragments of the object constraint conditions by characters to obtain the object word sequence; Iterate through each legal provision index number in the first filtering index set to obtain the constituent element field of the corresponding legal provision; The constituent element field is segmented by character, the number of words appearing in the constituent element field in the object word sequence is counted, and the text overlap rate corresponding to the legal provision index number is obtained by combining the total length of the object word sequence. According to the text overlap rate from high to low, the legal provision index numbers in the first filtering index set are sorted in descending order, and the sorted sequence is used as the rearranged legal provision index sequence.

[0069] The specific implementation method is as follows: First, initial index matching is performed. The legal provision index database of the regulatory retrieval platform is a database that structures and indexes current legal provisions. Each legal provision has a unique index number and is associated with key behavioral text fragments. Each behavioral term in the extended behavioral set generated in step S300 is traversed. Using this behavioral term as the query term, the legal provisions whose key behavioral text fragments contain this term are retrieved from the legal provision index database, and the index numbers of the successfully matched legal provisions are extracted. When multiple behavioral terms match the same legal provision, the index number will be duplicated. All extracted index numbers are merged and duplicates are removed to obtain the initial index set. For example, the extended behavior set is {"failed to pay capital contribution", "failed to fulfill capital contribution obligation", "failed to contribute capital as a shareholder"}. "failed to pay capital contribution" matches legal provision index numbers FL001 and FL002, "failed to fulfill capital contribution obligation" matches legal provision index numbers FL001 and FL003, and "failed to contribute capital as a shareholder" matches legal provision index number FL004. After merging and deduplicating, the initial index set is {FL001, FL002, FL003, FL004}.

[0070] Further, subject qualification screening is performed. Each legal provision index number in the initial index set is traversed, and the subject application scope field corresponding to that legal provision index number is retrieved from the legal provision index database. The subject application scope field is the part of the legal provision text that explicitly defines the type of subject to which the provision applies, such as "shareholder of a limited liability company," "founder of a joint-stock company," or "partner of a partnership." The text fragment of the subject qualification type identifier in the deterministic intent anchor point generated in step S100 is used as the matching object and subjected to substring matching with the subject application scope field; that is, it checks whether the text fragment of the subject qualification type identifier appears completely in the text of the subject application scope field. If it is not included, that is, the text fragment of the subject qualification type identifier does not appear in the subject application scope field, it indicates that the legal provision is not applicable to the subject type of the current query user, and the legal provision index number is removed from the set. After traversing all legal provision index numbers, the retained legal provision index numbers constitute the first filtering index set. For example, the subject qualification type identifier is "shareholder of a limited liability company". The subject scope field of legal provision FL001 includes "shareholder of a limited liability company", so the match is successful. The subject scope field of legal provision FL003 is "founder of a joint-stock company", which does not include "shareholder of a limited liability company", so FL003 is excluded. The first filtering index set is {FL001, FL002, FL004}.

[0071] Further, the text is reordered based on the overlap rate of the object constraint text. Text fragments of object constraint conditions are extracted from the deterministic intent anchor points generated in step S100. These text fragments are segmented by character, using the same segmentation method as the original query text in step S200, resulting in an object term sequence. Each term in the object term sequence is a unit independent term. Each legal provision index number in the first filtering index set is traversed, and the corresponding legal provision's constituent element field is obtained from the legal provision index library. The constituent element field is the core part of the legal provision text describing the applicable conditions and constituent elements of the provision, such as "shareholders failed to pay their capital contributions on time and in full," "capital contribution amount," and "capital contribution period." The constituent element field is segmented using the same character segmentation method to obtain a list of constituent element terms.

[0072] The number of terms appearing in the constituent element field of the object term sequence is counted; that is, for each term in the object term sequence, it is checked whether it has appeared in the constituent element term list. The number of terms appearing in the constituent element field is divided by the total length of the object term sequence (the total number of object terms). The quotient is the text overlap rate corresponding to the legal provision index number. The text overlap rate ranges from 0 to 1. The higher the value, the greater the overlap between the constituent elements of the legal provision and the object constraints queried by the user in terms of text description, and the stronger the relevance of the legal provision to the user's needs. The legal provision index numbers in the first filtering index set are sorted in descending order according to the text overlap rate from high to low. The sorted sequence is used as the rearranged legal provision index sequence. The earlier the legal provision index number ranks in the rearranged legal provision index sequence, the higher the matching degree between its corresponding legal provision and the user's query needs, and the preferred legal provision for generating search results.

[0073] For example, the object constraint text fragment is "investment", the object term sequence is {"investment"}, and the total length is 1. Legal provision FL001 contains the word "investment" in its constituent element field, with 1 matching term and a text overlap rate of 1 ÷ 1 = 1.0; legal provision FL002 contains "investment" in its constituent element field, with a text overlap rate of 1.0; legal provision FL004 does not contain "investment" in its constituent element field, with a text overlap rate of 0. After sorting the text overlap rates from high to low, FL001 and FL002 are tied for first place, and FL004 is at the end. The rearranged legal provision index sequence is {FL001, FL002, FL004}. After initial matching, subject qualification screening, and object constraint overlap rate reordering, the legal provision matching and reordering process is as follows: Figure 3 As shown.

[0074] The following technical effects were achieved through this step: First, by sequentially performing three layers of filtering—behavior matching, subject qualification screening, and object constraint overlap rate reordering—the scope of legal provision candidates is gradually narrowed from the expanded behavior set. This ensures that the final search results are highly consistent with the user's query needs in three dimensions: behavior description, subject applicability, and object relevance, thus solving the problem of the single dimension of legal provision screening in traditional keyword matching methods.

[0075] Second, by calculating the text overlap rate between the object term sequence and the constituent elements of the legal provisions, the semantic consistency between the query object constraints and the constituent elements of the legal provisions is quantified into a comparable numerical indicator, and then sorted in descending order to prioritize the legal provisions that best match the user's needs, thereby improving the accuracy and usability of the search results.

[0076] S500: The content of the legal provision corresponding to the first legal provision index number in the rearranged legal provision index sequence is combined with the required fuzzy entropy to adjust the display length of the legal provision content and generate a search result with intent parsing feature code.

[0077] By reordering based on behavioral matching, subject qualification screening, and text overlap rate of object constraints, a reordered legal provision index sequence was obtained, arranged in descending order of relevance. The legal provision corresponding to the first legal provision index number has the highest matching degree with the user's query requirements and is the core content of the search results. However, in legal retrieval, the main text of legal provisions is usually quite long, containing multiple clauses and details. If the user's query is vague or the intent is not clear, directly presenting a concise legal provision fragment may cause the user to miss key information; if the user's query intent is clear and specific, displaying the full text may reduce the efficiency of information retrieval. In addition, the search results need to contain traceable intent parsing feature codes to facilitate the review and auditing of subsequent search processes.

[0078] Step S500 in the method provided in this embodiment of the invention includes: Obtain the full text of the legal provision corresponding to the first legal provision index number, wherein the full text of the legal provision includes the legal provision number, the legal provision title, and the legal provision body; Calculate the total character length of the legal text and, based on the required fuzzy entropy, obtain the truncation length; Extract the specified number of characters from the beginning of the main text of the legal provision to obtain the corrected text segment; Extract the verification accumulation value from the deterministic intent anchor point, and concatenate the legal provision number, the legal provision title, the corrected text fragment, and the verification accumulation value in sequence to generate the retrieval result with intent parsing feature code.

[0079] The specific implementation method is as follows: First, obtain the full text of the legal provision corresponding to the first index number. Extract the first index number from the rearranged legal provision index sequence obtained in step S400. Retrieve the full text of the corresponding legal provision from the legal provision database of the legal retrieval platform based on this index number. The full text of the legal provision consists of three parts: the legal provision number, the legal provision title, and the legal provision body. The legal provision number is a unique identifier for the legal provision in the legal system, usually composed of the abbreviation of the legal name and the clause number; the legal provision title is a brief summary statement of the legal provision; and the legal provision body is the specific legal text content of the legal provision.

[0080] Further, the truncated length of the legal provision text is calculated. The total character length of the legal provision text is obtained, i.e., the number of all characters in the text. The truncated length is calculated as follows: multiply the total character length of the legal provision text by 1 and subtract the difference in demand fuzzy entropy obtained in step S200. The product is then rounded down to the nearest integer as the truncated length. The logic for calculating the truncated length is as follows: when the demand fuzzy entropy is large, it indicates that the user's query contains many vague expressions and the intent is not clear. A smaller difference between 1 and the demand fuzzy entropy results in a shorter truncated length, displaying a more concise legal provision text that highlights core clauses to prevent users from getting lost in a sea of ​​information. When the demand fuzzy entropy is small, it indicates that the user's query is clear and the intent is explicit. A larger difference between 1 and the demand fuzzy entropy, close to 1, results in a truncated length close to the full text length, or even displaying the complete legal provision text, to meet the user's comprehensive need for accurate legal information.

[0081] For example, the total character length of the legal text is 2000, the demand fuzzy entropy obtained in step S200 is 0.046, and the difference between 1 and the demand fuzzy entropy is 1 minus 0.046, which equals 0.954. The truncated length is 2000 multiplied by 0.954, which equals 1908, and rounded down to 1908.

[0082] Furthermore, starting from the beginning of the legal provision's main text, a length of characters is extracted as the revised text fragment. The extraction begins with the first character of the legal provision's main text, retaining the first consecutive characters of the extracted length and discarding any subsequent excess characters. If the extracted length is greater than or equal to the total character length of the legal provision's main text (i.e., the difference between 1 and the required fuzzy entropy is greater than or equal to 1), then the entire legal provision's main text is extracted, and the revised text fragment is the full text of the legal provision. For example, if the total character length of the legal provision's main text is 2000, and the extracted length is 1908, then the first 1908 characters are retained as the revised text fragment, and the last 92 characters are discarded.

[0083] Finally, search results with intent parsing feature codes are generated. The accumulated verification value is extracted from the deterministic intent anchor points generated in step S100. The accumulated verification value is the verification value obtained in step S100 by adding the four statistical parameters (subject name length, title count, negative word count, and quantifier count) and converting them into a fixed-length string. The legal provision number, legal provision title, revised text fragment, and accumulated verification value are concatenated in sequence, separated by a preset delimiter, to generate structured search results with intent parsing feature codes.

[0084] For example, the legal provision number is "Article 28 of the Company Law," the title is "Shareholders' Obligation to Pay Capital Contributions in Full and on Time," and the revised text fragment is the extracted content of the legal provision. The cumulative verification value is "00000010." After being concatenated in sequence, the generated search result is: "Article 28 of the Company Law | Shareholders' Obligation to Pay Capital Contributions in Full and on Time | [Revised Text Fragment] | 00000010." The cumulative verification value is embedded at the end of the search result as an intent parsing feature code. During subsequent review, the sum of each statistical parameter can be recalculated and compared with the feature code to verify whether the search result was generated by the current intent parsing process, preventing the search result from being accidentally tampered with during transmission or storage.

[0085] The following technical effects were achieved through this step: First, by dynamically adjusting the length of the legal provisions displayed through the fuzzy entropy of demand, concise core clauses are provided when the user's query intent is vague, and more comprehensive legal provisions are provided when the user's query intent is clear, thus achieving adaptive matching between the information density of the search results and the clarity of the user's needs.

[0086] Second, by attaching the accumulated verification value as an intent parsing feature code to the search results, a traceable and verifiable search result identification mechanism is established, which facilitates subsequent review of the search process and improves the credibility of regulatory search results.

[0087] Example 2, as Figure 4 As shown, based on the same inventive concept as the user intent intelligent analysis method based on a large language model provided in Embodiment 1, this embodiment of the invention also provides a user intent intelligent analysis system based on a large language model. The system includes an intent anchor parsing module 11, a demand fuzzy quantification module 12, a dynamic hierarchical expansion module 13, a legal provision matching and sorting module 14, and a retrieval result generation module 15. The system includes: The intent anchor parsing module 11 is used to obtain the original query text input by the user in the legal retrieval platform, and parse the subject qualification type identifier, behavior qualitative label and object constraint conditions in the original query text to generate a deterministic intent anchor. The demand fuzzy quantification module 12 is used to calculate the weighted coverage density of the word sequence in the original query text relative to the preset fuzzy vocabulary, obtain the demand fuzzy entropy, and calculate the semantic compactness based on the co-occurrence distance of the subject qualification type identifier, behavior qualitative label and object constraint in the original query text. The dynamic hierarchical expansion module 13 is used to dynamically determine the logical expansion level value based on the required fuzzy entropy and the semantic compactness, and to perform hierarchical expansion on the behavioral qualitative labels in the deterministic intent anchor point according to the logical expansion level value to generate an expanded behavior set. The legal provision matching and sorting module 14 is used to extract the associated legal provision index number of each behavior from the extended behavior set, and re-sort it according to the text overlap rate of the subject qualification type identifier and the object constraint conditions and the constituent elements of the legal provision to obtain a rearranged legal provision index sequence. The retrieval result generation module 15 is used to combine the content of the legal provision corresponding to the first legal provision index number in the rearranged legal provision index sequence with the required fuzzy entropy, correct the display length of the legal provision content, and generate retrieval results with intent parsing feature codes.

[0088] In one embodiment, the intent anchor parsing module 11 is further configured to obtain the original query text input by the user in the legal retrieval platform, and parse it to obtain the subject qualification type identifier, behavior qualitative label and object constraint conditions in the original query text, generating a deterministic intent anchor, including: The original query text is input into a pre-trained large language model, and the output is the label type sequence corresponding to each word. Extract consecutive word fragments labeled as subject type from the tag type sequence, including at least natural person name fragments, legal person name fragments, and non-legal person organization name fragments; The total character length, the number of appellations contained in the word fragments of the subject type, and the first occurrence position number in the original query text are counted and used as the subject name length, appellation count, and subject position index, respectively. Concatenate the word fragments labeled as subject type into a text fragment of subject qualification type identifier in the original order, and record the starting character position and ending character position of the subject qualification type identifier; Extract consecutive word segments labeled as behavioral characteristics from the tag type sequence, wherein the behavioral characteristics include at least verb phrase segments and verb-object phrase segments; The occurrence frequency of negative words, total character length, and number of adverbs in the lexical segments used to characterize the behavior are counted and used as the count of negative words, the length of the behavior phrase, and the count of adverb modifications, respectively. The word fragments of the behavior qualitative analysis are concatenated in their original order to form the text fragments of the behavior qualitative labels, and the starting character position and ending character position of the behavior qualitative labels are recorded. Extract consecutive word segments labeled as object constraints from the tag type sequence, wherein the object constraints include at least noun phrase segments and quantity phrase segments; The number of times quantifiers appear, the total length of characters, and the number of adjectives appear in the lexical fragments constrained by the object are counted, and these are used as the quantifier count, the object phrase length, and the adjective modification count, respectively. The word segments marked as object constraints are concatenated in their original order to form the text segment of the object constraint condition, and the starting character position and ending character position of the object constraint condition are recorded. The text fragments of the subject qualification type identifier, the text fragments of the behavior qualitative label, and the text fragments of the object constraint condition are associated and stored with their respective start character positions and end character positions, as well as various statistical parameters, to generate the deterministic intent anchor point.

[0089] The pre-training steps of the large language model include: Original query text samples were collected from the historical query logs of the legal retrieval platform. Each word in each sample was manually labeled with a tag type. The tag types included at least subject type, behavior characterization, object constraint, and irrelevant type, forming a labeled sample set. Construct a pre-trained network based on the Transformer architecture, wherein the pre-trained network includes a shared semantic encoding layer and three parallel output branches; The original query text samples in the labeled sample set are input into the shared semantic coding layer to obtain the contextual semantic representation of each word. The contextual semantic representation is then input into the three parallel output branches. Each output branch outputs the probability distribution of each word belonging to the corresponding label type through a fully connected mapping and a normalized exponential function. For each output branch, the cross-entropy loss function is used to calculate the branch loss value between the predicted probability distribution and the manually labeled label type. The three branch loss values ​​are added together to obtain the total loss value. The internal parameters of the pre-trained network are then iteratively updated through backpropagation. When the decrease in the total loss value on the independent validation sample set is less than the convergence threshold for multiple consecutive iterations, training is stopped, and a pre-trained large language model is obtained.

[0090] The text fragment of the subject qualification type identifier, the start and end character positions of the subject qualification type identifier, the subject name length, the title count, and the subject position index are encapsulated into a first feature area; The text fragments of the behavioral qualitative labels, the starting and ending character positions of the behavioral qualitative labels, the count of negative words, the length of behavioral phrases, and the count of adverb modifications are encapsulated into a second feature region; The text fragment of the object constraint, the starting and ending character positions of the object constraint, the quantifier count, the object phrase length, and the adjective modification count are encapsulated into a third feature region; Separators are inserted between adjacent feature areas in the order of the first feature area, the second feature area, and the third feature area to form an anchor content sequence. The length of the subject name, the number of titles, the number of negative words, and the number of quantifiers are added together to obtain a verification cumulative value. The verification cumulative value is converted into a fixed-length string and appended to the end of the anchor content sequence to generate the deterministic intent anchor.

[0091] In one embodiment, the demand fuzzy quantification module 12 is further configured to calculate the weighted coverage density of the word sequence in the original query text relative to a preset fuzzy vocabulary, obtain the demand fuzzy entropy, and calculate the semantic compactness based on the co-occurrence distance of the subject qualification type identifier, behavioral qualitative label, and object constraint conditions in the original query text, including: Obtain a pre-constructed fuzzy word list, wherein the fuzzy word list contains three categories: degree fuzzy words, range fuzzy words, and condition fuzzy words, and each category of fuzzy words corresponds to a fuzzy contribution coefficient; The original query text is segmented by character to obtain a word sequence; Traverse each word in the word sequence, and when the word matches any fuzzy word in the fuzzy word list, extract the fuzzy contribution coefficient of the category to which the fuzzy word belongs; The unit contribution value of a lexical is obtained by dividing the fuzzy contribution coefficient by the absolute distance from each lexical to the starting character position of the behavior qualitative label. The unit contribution values ​​corresponding to all lexical units that hit the fuzzy word in the lexical sequence are summed, and combined with the total character length of the lexical sequence, the required fuzzy entropy is obtained. The starting character positions of the subject qualification type identifier, the starting character positions of the behavior qualitative label, and the starting character positions of the object constraint conditions are extracted from the deterministic intent anchor point and denoted as the first position, the second position, and the third position, respectively. Calculate the first absolute difference between the second position and the first position, calculate the second absolute difference between the third position and the second position, and combine the total character length of the original query text to calculate the normalized co-occurrence span; Subtracting the normalized co-occurrence span from the value 1 yields the semantic compactness.

[0092] In one embodiment, the dynamic hierarchical expansion module 13 is further configured to dynamically determine the logical expansion level value based on the demand fuzzy entropy and the semantic compactness, and to perform hierarchical expansion on the behavioral qualitative labels in the deterministic intent anchor point according to the logical expansion level value, generating an expanded behavior set, including: The demand fuzzy entropy is compared with a first entropy threshold to determine whether the demand fuzzy entropy exceeds the first entropy threshold, and a first judgment result is obtained, wherein the first entropy threshold is determined based on the count of negative words and the length of the behavior phrase; The semantic compactness is compared with a first compactness threshold to determine whether the semantic compactness is lower than the first compactness threshold, and a second determination result is obtained, wherein the first compactness threshold is determined based on the length of the subject name; When the first judgment result is yes and the second judgment result is yes, the logical expansion level value is set to the maximum value; When the first judgment result is negative and the second judgment result is negative, the logical expansion level value is set to the minimum value; When the first judgment result is inconsistent with the second judgment result, obtain the count of negative words contained in the behavior qualitative label. If the count of negative words is greater than zero, set the logical expansion level value to a value that is relatively large in the middle; otherwise, set it to a value that is relatively small in the middle. The logical expansion level value is an integer between 1 and 4.

[0093] Based on the logical expansion hierarchy value, the behavioral qualitative tags in the deterministic intent anchor point are hierarchically expanded to generate an expanded behavior set, including: Obtain a pre-constructed behavior hypernym mapping relationship, wherein each behavior verb phrase in the behavior hypernym mapping relationship is associated with a hypernym chain arranged in ascending order of abstraction level; Extract the hypernym chain that matches the text fragment of the behavior qualitative label from the behavior hypernym mapping relationship. If the match fails, the text fragment of the behavior qualitative label itself is taken as the unique hypernym. The expansion quantity is determined based on the logical expansion level value; Select the first extended number of hypernyms starting from the starting position in the hypernym chain, merge them with the text fragments of the behavior qualitative labels, and remove duplicate terms to form a basic behavior set; Obtain the count of negative words contained in the behavioral qualitative label. If the count of negative words is greater than zero, add a negative prefix to each behavior in the basic behavior set to obtain an extended behavior set. If the count of the negation words is zero, then the basic set of behaviors is used as the extended set of behaviors.

[0094] In one embodiment, the legal provision matching and sorting module 14 is further configured to extract the associated legal provision index number of each behavior from the extended behavior set, and reorder it based on the text overlap rate of the subject qualification type identifier and the object constraint conditions and the constituent elements of the legal provision, to obtain a rearranged legal provision index sequence, including: Traverse each behavior in the extended behavior set, match legal provisions containing behavior text fragments from the legal provision index library of the legal retrieval platform, extract the index numbers of the successfully matched legal provisions, and merge and remove duplicates to obtain the initial index set; Iterate through each legal provision index number in the initial index set to obtain the subject scope field of the legal provision corresponding to the legal provision index number; The text fragment identifying the subject qualification type is matched with the subject scope field for substring matching, and the legal provision index number that does not contain the text fragment is removed to obtain the first filtering index set; Extract the text fragments of the object constraint conditions from the deterministic intent anchor points, and segment the text fragments of the object constraint conditions by characters to obtain the object word sequence; Iterate through each legal provision index number in the first filtering index set to obtain the constituent element field of the corresponding legal provision; The constituent element field is segmented by character, the number of words appearing in the constituent element field in the object word sequence is counted, and the text overlap rate corresponding to the legal provision index number is obtained by combining the total length of the object word sequence. According to the text overlap rate from high to low, the legal provision index numbers in the first filtering index set are sorted in descending order, and the sorted sequence is used as the rearranged legal provision index sequence.

[0095] In one embodiment, the retrieval result generation module 15 is further configured to combine the content of the legal provision corresponding to the first legal provision index number in the rearranged legal provision index sequence with the required fuzzy entropy, correct the display length of the legal provision content, and generate retrieval results with intent parsing feature codes, including: Obtain the full text of the legal provision corresponding to the first legal provision index number, wherein the full text of the legal provision includes the legal provision number, the legal provision title, and the legal provision body; Calculate the total character length of the legal text and, based on the required fuzzy entropy, obtain the truncation length; Extract the specified number of characters from the beginning of the main text of the legal provision to obtain the corrected text segment; Extract the verification accumulation value from the deterministic intent anchor point, and concatenate the legal provision number, the legal provision title, the corrected text fragment, and the verification accumulation value in sequence to generate the retrieval result with intent parsing feature code.

Claims

1. A user intent intelligent analysis method based on a large language model, characterized in that, The method includes: Obtain the original query text input by the user in the legal retrieval platform, and parse the subject qualification type identifier, behavior qualitative label and object constraint conditions in the original query text to generate a deterministic intent anchor point; Calculate the weighted coverage density of the word sequence in the original query text relative to the preset fuzzy vocabulary to obtain the demand fuzzy entropy, and calculate the semantic compactness based on the co-occurrence distance of the subject qualification type identifier, behavior qualitative label and object constraint conditions in the original query text; Based on the required fuzzy entropy and the semantic compactness, the logical expansion level value is dynamically determined, and the behavioral qualitative labels in the deterministic intent anchor point are expanded hierarchically according to the logical expansion level value to generate an expanded behavior set; The associated legal provision index number of each behavior is extracted from the extended behavior set, and after being filtered by the subject qualification type identifier and reordered by the text overlap rate of the object constraint conditions and the constituent elements of the legal provision, a rearranged legal provision index sequence is obtained. The content of the legal provision corresponding to the first legal provision index number in the rearranged legal provision index sequence is combined with the required fuzzy entropy to adjust the display length of the legal provision content and generate a search result with intent parsing feature code.

2. The user intent intelligent analysis method based on a large language model according to claim 1, characterized in that, Obtain the original query text input by the user from the legal retrieval platform, and parse it to obtain the subject qualification type identifier, behavior qualitative label, and object constraint conditions from the original query text, generating a deterministic intent anchor point, including: The original query text is input into a pre-trained large language model, and the output is the label type sequence corresponding to each word. Extract consecutive word fragments labeled as subject type from the tag type sequence, including at least natural person name fragments, legal person name fragments, and non-legal person organization name fragments; The total character length, the number of appellations contained in the word segment of the subject type, and the first occurrence position number in the original query text are counted and used as the subject name length, appellation count, and subject position index, respectively. Concatenate the word fragments labeled as subject type into a text fragment of subject qualification type identifier in the original order, and record the starting character position and ending character position of the subject qualification type identifier; Extract consecutive word segments labeled as behavioral characteristics from the tag type sequence, wherein the behavioral characteristics include at least verb phrase segments and verb-object phrase segments; The occurrence frequency of negative words, total character length, and number of adverbs in the lexical segments used to characterize the behavior are counted and used as the count of negative words, the length of the behavior phrase, and the count of adverb modifications, respectively. The word fragments of the behavior qualitative analysis are concatenated in their original order to form the text fragments of the behavior qualitative labels, and the starting character position and ending character position of the behavior qualitative labels are recorded. Extract consecutive word segments labeled as object constraints from the tag type sequence, wherein the object constraints include at least noun phrase segments and quantity phrase segments; The number of times quantifiers appear, the total length of characters, and the number of adjectives appear in the lexical fragments constrained by the object are counted, and these are used as the quantifier count, the object phrase length, and the adjective modification count, respectively. The word segments marked as object constraints are concatenated in their original order to form the text segment of the object constraint condition, and the starting character position and ending character position of the object constraint condition are recorded. The text fragments of the subject qualification type identifier, the text fragments of the behavior qualitative label, and the text fragments of the object constraint condition are associated and stored with their respective start character positions and end character positions, as well as various statistical parameters, to generate the deterministic intent anchor point.

3. The user intent intelligent analysis method based on a large language model according to claim 2, characterized in that, The pre-training steps of the large language model include: Original query text samples were collected from the historical query logs of the legal retrieval platform. Each word in each sample was manually labeled with a tag type. The tag types included at least subject type, behavior characterization, object constraint, and irrelevant type, forming a labeled sample set. Construct a pre-trained network based on the Transformer architecture, wherein the pre-trained network includes a shared semantic encoding layer and three parallel output branches; The original query text samples in the labeled sample set are input into the shared semantic coding layer to obtain the contextual semantic representation of each word. The contextual semantic representation is then input into the three parallel output branches. Each output branch outputs the probability distribution of each word belonging to the corresponding label type through a fully connected mapping and a normalized exponential function. For each output branch, the cross-entropy loss function is used to calculate the branch loss value between the predicted probability distribution and the manually labeled label type. The three branch loss values ​​are added together to obtain the total loss value. The internal parameters of the pre-trained network are then iteratively updated through backpropagation. When the decrease in the total loss value on the independent validation sample set is less than the convergence threshold for multiple consecutive iterations, training is stopped, and a pre-trained large language model is obtained.

4. The user intent intelligent analysis method based on a large language model according to claim 2, characterized in that, The text fragments of the subject qualification type identifier, the behavior qualitative label, and the object constraint condition are associated and stored with their respective start and end character positions, as well as various statistical parameters, to generate the deterministic intent anchor point, including: The text fragment of the subject qualification type identifier, the start and end character positions of the subject qualification type identifier, the subject name length, the title count, and the subject position index are encapsulated into a first feature area; The text fragments of the behavioral qualitative labels, the starting and ending character positions of the behavioral qualitative labels, the count of negative words, the length of behavioral phrases, and the count of adverb modifications are encapsulated into a second feature region; The text fragment of the object constraint, the starting and ending character positions of the object constraint, the quantifier count, the object phrase length, and the adjective modification count are encapsulated into a third feature region; Separators are inserted between adjacent feature areas in the order of the first feature area, the second feature area, and the third feature area to form an anchor content sequence. The length of the subject name, the number of titles, the number of negative words, and the number of quantifiers are added together to obtain a verification cumulative value. The verification cumulative value is converted into a fixed-length string and appended to the end of the anchor content sequence to generate the deterministic intent anchor.

5. The user intent intelligent analysis method based on a large language model according to claim 1, characterized in that, Calculate the weighted coverage density of the word sequence in the original query text relative to a preset fuzzy vocabulary to obtain the demand fuzzy entropy, and calculate the semantic compactness based on the co-occurrence distance of the subject qualification type identifier, behavior qualitative label, and object constraint conditions in the original query text, including: Obtain a pre-constructed fuzzy word list, wherein the fuzzy word list contains three categories: degree fuzzy words, range fuzzy words, and condition fuzzy words, and each category of fuzzy words corresponds to a fuzzy contribution coefficient; The original query text is segmented by character to obtain a word sequence; Traverse each word in the word sequence, and when the word matches any fuzzy word in the fuzzy word list, extract the fuzzy contribution coefficient of the category to which the fuzzy word belongs; The unit contribution value of a lexical is obtained by dividing the fuzzy contribution coefficient by the absolute distance from each lexical to the starting character position of the behavior qualitative label. The unit contribution values ​​corresponding to all lexical units that hit the fuzzy word in the lexical sequence are summed, and combined with the total character length of the lexical sequence, the required fuzzy entropy is obtained. The starting character positions of the subject qualification type identifier, the starting character positions of the behavior qualitative label, and the starting character positions of the object constraint conditions are extracted from the deterministic intent anchor point and denoted as the first position, the second position, and the third position, respectively. Calculate the first absolute difference between the second position and the first position, calculate the second absolute difference between the third position and the second position, and combine the total character length of the original query text to calculate the normalized co-occurrence span; Subtracting the normalized co-occurrence span from the value 1 yields the semantic compactness.

6. The user intent intelligent analysis method based on a large language model according to claim 1, characterized in that, Based on the stated requirement fuzzy entropy and the stated semantic compactness, the logical expansion level value is dynamically determined, including: The demand fuzzy entropy is compared with a first entropy threshold to determine whether the demand fuzzy entropy exceeds the first entropy threshold, and a first judgment result is obtained, wherein the first entropy threshold is determined based on the count of negative words and the length of the behavior phrase; The semantic compactness is compared with a first compactness threshold to determine whether the semantic compactness is lower than the first compactness threshold, and a second determination result is obtained, wherein the first compactness threshold is determined based on the length of the subject name; When the first judgment result is yes and the second judgment result is yes, the logical expansion level value is set to the maximum value; When the first judgment result is negative and the second judgment result is negative, the logical expansion level value is set to the minimum value; When the first judgment result is inconsistent with the second judgment result, obtain the count of negative words contained in the behavior qualitative label. If the count of negative words is greater than zero, set the logical expansion level value to a value that is relatively large in the middle; otherwise, set it to a value that is relatively small in the middle. The logical expansion level value is an integer between 1 and 4.

7. The user intent intelligent analysis method based on a large language model according to claim 1, characterized in that, Based on the logical expansion hierarchy value, the behavioral qualitative tags in the deterministic intent anchor point are hierarchically expanded to generate an expanded behavior set, including: Obtain a pre-constructed behavior hypernym mapping relationship, wherein each behavior verb phrase in the behavior hypernym mapping relationship is associated with a hypernym chain arranged in ascending order of abstraction level; Extract the hypernym chain that matches the text fragment of the behavior qualitative label from the behavior hypernym mapping relationship. If the match fails, the text fragment of the behavior qualitative label itself is taken as the unique hypernym. The expansion quantity is determined based on the logical expansion level value; Select the first extended number of hypernyms starting from the starting position in the hypernym chain, merge them with the text fragments of the behavior qualitative labels, and remove duplicate terms to form a basic behavior set; Obtain the count of negative words contained in the behavioral qualitative label. If the count of negative words is greater than zero, add a negative prefix to each behavior in the basic behavior set to obtain an extended behavior set. If the count of the negation words is zero, then the basic set of behaviors is used as the extended set of behaviors.

8. The user intent intelligent analysis method based on a large language model according to claim 1, characterized in that, The associated legal provision index numbers of each behavior are extracted from the extended behavior set. After filtering by the subject qualification type identifier and reordering based on the text overlap rate between the object constraints and the constituent elements of the legal provision, a rearranged legal provision index sequence is obtained, including: Traverse each behavior in the extended behavior set, match legal provisions containing behavior text fragments from the legal provision index library of the legal retrieval platform, extract the index numbers of the successfully matched legal provisions, and merge and remove duplicates to obtain the initial index set; Iterate through each legal provision index number in the initial index set to obtain the subject scope field of the legal provision corresponding to the legal provision index number; The text fragment identifying the subject qualification type is matched with the subject scope field for substring matching, and the legal provision index number that does not contain the text fragment is removed to obtain the first filtering index set; Extract the text fragments of the object constraint conditions from the deterministic intent anchor points, and segment the text fragments of the object constraint conditions by characters to obtain the object word sequence; Iterate through each legal provision index number in the first filtering index set to obtain the constituent element field of the corresponding legal provision; The constituent element field is segmented by character, the number of words appearing in the constituent element field in the object word sequence is counted, and the text overlap rate corresponding to the legal provision index number is obtained by combining the total length of the object word sequence. According to the text overlap rate from high to low, the legal provision index numbers in the first filtering index set are sorted in descending order, and the sorted sequence is used as the rearranged legal provision index sequence.

9. The user intent intelligent analysis method based on a large language model according to claim 1, characterized in that, The content of the legal provision corresponding to the first legal provision index number in the rearranged legal provision index sequence is combined with the required fuzzy entropy to adjust the display length of the legal provision content, generating retrieval results with intent parsing feature codes, including: Obtain the full text of the legal provision corresponding to the first legal provision index number, wherein the full text of the legal provision includes the legal provision number, the legal provision title, and the legal provision body; Calculate the total character length of the legal text and, based on the required fuzzy entropy, obtain the truncation length; Extract the specified number of characters from the beginning of the main text of the legal provision to obtain the corrected text segment; Extract the verification accumulation value from the deterministic intent anchor point, and concatenate the legal provision number, the legal provision title, the corrected text fragment, and the verification accumulation value in sequence to generate the retrieval result with intent parsing feature code.

10. A user intent intelligent analysis system based on a large language model, characterized in that, The system is used to implement the user intent intelligent analysis method based on a large language model as described in any one of claims 1-9, the system comprising: The intent anchor parsing module is used to obtain the original query text input by the user in the legal retrieval platform, and parse it to obtain the subject qualification type identifier, behavior qualitative label and object constraint conditions in the original query text, and generate a deterministic intent anchor. The demand fuzzy quantification module is used to calculate the weighted coverage density of the word sequence in the original query text relative to the preset fuzzy vocabulary, obtain the demand fuzzy entropy, and calculate the semantic compactness based on the co-occurrence distance of the subject qualification type identifier, behavior qualitative label and object constraint conditions in the original query text. The dynamic hierarchical expansion module is used to dynamically determine the logical expansion level value based on the required fuzzy entropy and the semantic compactness, and to expand the behavioral qualitative labels in the deterministic intent anchor point hierarchically according to the logical expansion level value to generate an expanded behavior set. The legal provision matching and sorting module is used to extract the associated legal provision index number of each behavior from the extended behavior set, and re-sort it according to the text overlap rate of the subject qualification type identifier and the object constraint conditions and the constituent elements of the legal provision to obtain a rearranged legal provision index sequence. The retrieval result generation module is used to combine the content of the legal provision corresponding to the first legal provision index number in the rearranged legal provision index sequence with the required fuzzy entropy, correct the display length of the legal provision content, and generate retrieval results with intent parsing feature codes.

Images