NLP-AI intelligent analysis method for customer acquisition marketing

Abstract

The application discloses an NLP-AI intelligent analysis method for customer acquisition marketing, and relates to the technical field of customer acquisition marketing, and comprises the following steps: inputting the continuous expression of the user's context into expression sequence arrangement, matching the expression direction by constructing a turning expression matching rule set, identifying the position relationship before and after the sentiment polarity word and the negative modifier, and judging whether the user produces intention position reversal in the continuous expression of the context; for the expression content judged as having intention position reversal, the emotional verbs, evaluation objects, turning guide words and semantic target words are extracted by establishing a semantic jump component marking list, and the subjective expression mutation points in the sentence and between sentences are marked to determine the semantic change clues. The application solves the problem that the intention position reversal in the continuous expression of the context is difficult to identify, and realizes real intention dynamic analysis and scoring regulation effect driven by the semantic change clues.

Description

Technical Field

[0001] This invention relates to the field of customer acquisition marketing technology, and more specifically to an NLP-AI intelligent analysis method for customer acquisition marketing. Background Technology

[0002] NLP-AI intelligent analytics for customer acquisition and marketing refers to the automated parsing and intelligent mining of unstructured text data related to customers from multiple channels (such as user consultation records, social media comments, e-commerce reviews, questionnaires, customer service dialogues, etc.) by integrating Natural Language Processing (NLP) and Artificial Intelligence (AI) technologies. This enables key functions such as customer needs identification, interest point extraction, user intent recognition, sentiment analysis, and customer value prediction, thereby driving precise marketing and efficient customer acquisition. The existing NLP-AI intelligent analysis technology in customer acquisition marketing typically involves the following steps: First, data collection and integration, acquiring customer text information from multiple platforms and formats through web crawlers or APIs; next, text preprocessing, including word segmentation, part-of-speech tagging, noise reduction, entity recognition, and sentiment extraction to ensure the accuracy of subsequent analysis; then, feature extraction and modeling, using deep learning (such as BERT, Transformer, etc.) or machine learning algorithms to model customer language behavior features and extract potential information such as purchase intention, behavioral preferences, and conversion probability; next, through intent recognition and customer profiling modules, mapping individual users into multi-dimensional feature vectors and segmenting users based on profile tags; finally, through intelligent recommendation and marketing decision-making mechanisms, accurately pushing personalized products or services to target customers, achieving an automated, high-conversion marketing loop. This entire process emphasizes data-driven and semantic understanding as its foundation, with intelligent models at its core, forming a complete closed loop from understanding customer language behavior to executing marketing actions. It represents a crucial technological path for upgrading traditional marketing to intelligent marketing.

[0003] The existing technology has the following shortcomings:

[0004] In the implementation of NLP-AI intelligent analysis methods for customer acquisition and marketing, when a user's intent shifts within a continuous contextual expression—for example, initially expressing a negative attitude and then a positive one—their true conversion intent dynamically adjusts with the semantic change. In such expressions, the semantic direction typically relies on the contextual semantic relationship guided by transitional words (such as "however," "later," "actually," etc.). Because this type of stance change is semantically leaping and context-dependent, NLP analysis methods based on static sentence processing struggle to accurately analyze the semantic shift process. Furthermore, existing AI scoring mechanisms rely on semantic consistency to build conversion models, lacking the ability to identify and judge stance reversal trends. Therefore, existing NLP-AI intelligent analysis technologies for customer acquisition and marketing cannot analyze the true intent through NLP structural analysis and AI intent scoring mechanisms based on semantic shift cues when users exhibit intent reversal within a continuous contextual expression. This leads to delayed or erroneous judgments of the user's current conversion intent, affecting the system's accuracy in precise customer acquisition, resulting in decreased conversion rates, reduced user response quality, and even model training bias, impacting the reliability of subsequent intelligent marketing decisions.

[0005] The information disclosed in the background section is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The purpose of this invention is to provide an NLP-AI intelligent analysis method for customer acquisition and marketing, in order to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an NLP-AI intelligent analysis method for customer acquisition marketing, specifically including the following steps:

[0008] S1. Arrange the user's continuous contextual input into an expression sequence, match the expression direction by constructing a set of rules for expression direction matching, identify the positional relationship between sentiment polarity words and negative modifiers, and determine whether the user has reversed their intention stance in the continuous contextual expression.

[0009] S2. For expressions judged to have an intentional reversal of stance, extract emotional verbs, evaluation objects, transition introductory words and semantic target words by establishing a list of semantic jump markers, and mark subjective expression abrupt change points within and between sentences to determine semantic shift clues;

[0010] S3. The extracted semantic transformation clues are used as nodes to reorganize the paths according to the expression order, differences in intention direction and semantic repulsion, generating a continuous semantic map with directionality and jump density to represent the evolution trajectory of the true intention.

[0011] S4. Input the transition nodes, jump direction labels and semantic segment strength indicators in the continuous semantic graph into the natural language processing parsing process and the artificial intelligence scoring and judgment process, respectively. By intervening in the parsing path of natural language processing and guiding the adjustment of scoring weights, the analysis of true intent is completed.

[0012] S5. Perform time series comparison between the natural language processing parsing results and the artificial intelligence scoring results, construct a scoring fluctuation model, perform confidence fitting based on the jump segments in the continuous expression semantic map, and dynamically adjust the analysis output of the true intent based on the fitting results.

[0013] Preferably, S1 specifically includes the following steps:

[0014] S101. Arrange the continuous contextual expression inputs formed by the user in the same semantic session cycle into an expression sequence according to the actual input order, and construct an expression unit index based on time tags and input intervals to ensure that each expression unit in the expression sequence has a definite semantic position in the context.

[0015] S102. Construct a set of turning expression matching rules for expression direction matching, perform sequential matching on adjacent expression units in the expression sequence, and determine whether there is an expression direction jump between each pair of expression units consisting of transition words, sentiment polarity words, and subject consistency.

[0016] S103. After the expression direction matching is completed, the sentiment polarity words and negative modifiers in each expression unit are identified based on the syntactic dependency relationship, and the positional relationship between the negative modifier and the sentiment polarity words in the syntactic path is determined. When the expression direction changes from forward to reverse or reverse to forward, and the relative positions of the negative modifiers are swapped, it is determined that the user has reversed the intention stance in the continuous expression in the context.

[0017] Preferably, S102 is as follows:

[0018] Based on a sentiment analysis dictionary and a grammar rule base, a set of rules for matching turning expressions is constructed, which includes transition words, sentiment polarity words, semantic boundary markers and subjective expression components. Each rule in the rule set consists of a triple, defining the transition relationship, polarity direction and subject continuation features.

[0019] The adjacent expression units in the expression sequence are sequentially matched, and the rules in the turning expression matching rule set are applied in turn. Keyword extraction and subject continuation judgment are performed on each pair of expression units, and the existence of expression direction jump candidate features is marked.

[0020] When candidate features simultaneously satisfy the following conditions: the presence of a transitional introductory word, a reversal of the emotional polarity direction, and consistency of the subject in adjacent expression units, it is determined that there is a change in the expression direction between the expression unit pairs.

[0021] Preferably, S2 is as follows:

[0022] Semantic extraction is performed on expressions that are judged to have an intentional reversal of stance. A list of semantic jump component tags containing emotional verbs, evaluation objects, transition introductory words and semantic target words is constructed. The emotional verbs and evaluation objects in the expression are identified by dependency parsing combined with part-of-speech tags. Semantic target words and transition introductory words related to their semantics are extracted by using contextual clues.

[0023] The extracted emotional verbs, evaluation objects, transition introductory words and semantic target words are semantically arranged according to the expression order and syntactic structure. Based on the emotional contrast relationship and semantic transition signal in the expression content, the position of semantic change in the expression is identified. Syntactic structure change points are marked within sentences and intentional change boundaries are marked between sentences to form a set of subjective expression change point markers.

[0024] By performing correlation analysis between the set of subjective expression mutation point markers and the list of semantic jump component markers, and based on the path of emotional stance change guided by the transition introductory words, combined with the semantic tension distribution between syntactic mutation points, directional jump nodes are extracted as semantic transformation clues for subsequent construction and identification of true intentions.

[0025] Preferably, S3 is as follows:

[0026] The extracted semantic transformation clues are constructed into a set of nodes. Each node in the set is bound to the timestamp and intent direction attribute of the corresponding expression content, and arranged according to the expression order of the context to form an expression order structure with temporal markers and stance orientation.

[0027] The difference in intentional direction attributes between adjacent nodes in the expression sequence structure is calculated. Combining the semantic orientation features of each node, a semantic repulsion scoring system is established based on the level of semantic conflict and the degree of emotional tension. Based on this, a jump path is constructed between nodes to achieve path reorganization.

[0028] The nodes after path reorganization are integrated with the jump paths to generate a continuous semantic graph. In the continuous semantic graph, the path direction attribute and jump density distribution are set. The jump density is used to characterize the concentration trend of semantic transformation cues in the expression sequence and to represent the evolutionary trajectory of the true intention.

[0029] Preferably, S4 specifically includes the following steps:

[0030] S401. Extract transition nodes, jump direction labels and semantic segment strength identifiers from the continuous expression semantic graph, and establish an index mapping relationship between the transition nodes and the corresponding original expression segments. At the same time, bind the jump direction labels and semantic segment strength identifiers to the corresponding semantic segments to form a parsing input set.

[0031] S402. Input the parsing input set into the natural language processing parsing process. During the parsing process, limit the starting position of semantic parsing based on the transformation node, constrain the direction of semantic dependency expansion with the jump direction label, and assign parsing priority to the semantic branches in the parsing path according to the semantic fragment strength identifier, so as to realize the intervention of the natural language processing parsing path.

[0032] S403. Input the parsed input set into the artificial intelligence scoring and judgment process, take the transition node as the scoring benchmark position, set the scoring offset direction according to the jump direction label, and adjust the weight distribution in the scoring and judgment process according to the semantic fragment strength identifier to complete the analysis of the true intent.

[0033] Preferably, S401 is as follows:

[0034] Identify all locations marked as semantic transition points from the continuous semantic graph, and combine the degree of change in intent direction of nodes in the expression sequence structure with the semantic repulsion value. Select expression nodes that simultaneously satisfy the polarity reversal of intent direction and semantic repulsion score exceeding the set threshold as transition nodes, and extract the direction label of the path connected to each transition node as the transition direction label.

[0035] Each extracted transition node is mapped one-to-one with its corresponding original expression fragment in the expression order structure, and a node position index is attached to the transition direction label to form a transition pointing relationship chain.

[0036] The semantic intensity label value in each semantic segment is used as the semantic segment intensity identifier, and combined with the jump direction label bound to it to form a data group containing directional and semantic weight information. This data group is then integrated with the number of the transition node to form the parsing input set.

[0037] Preferably, S5 is as follows:

[0038] The natural language processing parsing results and the artificial intelligence scoring results are aligned according to the time sequence of the transition nodes. The intent judgment value in the parsing results and the score value in the scoring results are uniformly mapped to the same time axis. A time series comparison structure containing the parsing sequence and the scoring sequence is constructed to characterize the synchronous relationship between intent judgment and score changes.

[0039] Based on the time series comparison structure, the fluctuation features formed by the change of rating values ​​over time are extracted, and combined with the position of the jump segment marked in the continuous expression semantic graph, the fluctuation features are segmented and modeled to construct a rating fluctuation model to describe the amplitude, direction of change and stable interval of rating fluctuations.

[0040] Based on the fluctuation characteristics of the scoring fluctuation model in the jump range, confidence level fitting is performed. The fitted confidence level results are correlated with the analysis results of the corresponding time period. The analysis output of true intent is dynamically adjusted according to the trend of confidence level changes, so as to realize the time-series correction of intent judgment results.

[0041] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0042] 1. This invention constructs a dynamic semantic modeling mechanism for context-dependent continuous expression, achieving accurate identification and semantic trajectory tracking of user intent reversal scenarios in expression. Compared to traditional NLP methods that extract intent solely based on static intra-sentence features, this approach introduces key steps such as expression direction shift recognition, semantic shift component extraction, and subjective expression mutation point annotation into the expression sequence. This effectively captures complex linguistic phenomena such as emotional polarity reversals and intent mutations driven by transitional introductory phrases in user expression. Furthermore, by constructing a continuous expression semantic map with directionality and shift density, the semantic evolution process is expressed in a structured manner, thereby enhancing the ability to analyze the changing trends of true intent.

[0043] 2. This invention introduces a parsed input set to intervene in the NLP parsing process and the AI ​​scoring process, establishing a dynamic connection between structured semantic cues and the model scoring mechanism. Combined with scoring fluctuation modeling and confidence fitting mechanisms, it achieves temporal correction and dynamic control of intent judgment results. This series of methods not only enhances the AI ​​model's perception and response capabilities when processing nonlinear semantic expressions but also effectively avoids scoring bias caused by semantic jumps. In customer acquisition and marketing scenarios, it exhibits higher intent recognition accuracy and user response prediction stability, providing a more reliable decision-making basis for subsequent intelligent marketing strategy formulation. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0045] Figure 1 This is a flowchart illustrating the NLP-AI intelligent analysis method for customer acquisition and marketing according to the present invention. Detailed Implementation

[0046] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, they are provided so that the description of this disclosure will be more complete and fully convey the concept of the exemplary embodiments to those skilled in the art.

[0047] This invention provides, for example Figure 1 The NLP-AI intelligent analysis method shown for customer acquisition and marketing includes the following steps:

[0048] S1. Arrange the user's continuous contextual input into an expression sequence, match the expression direction by constructing a set of rules for expression direction matching, identify the positional relationship between sentiment polarity words and negative modifiers, and determine whether the user has reversed their intention stance in the continuous contextual expression.

[0049] In this embodiment, S1 specifically includes the following steps:

[0050] S101. Arrange the continuous contextual expression inputs formed by the user in the same semantic session cycle into an expression sequence according to the actual input order, and construct an expression unit index based on time tags and input intervals to ensure that each expression unit in the expression sequence has a definite semantic position in the context.

[0051] Arranging a sequence of expressions formed by a user within the same semantic conversation cycle according to their actual input order can be achieved by capturing all text input content from a user in an interactive session and sorting it by time based on the timestamp of each input. To ensure that each input has a traceable contextual position, each piece of input text can be labeled with a time tag. The presence of expressions belonging to the same conversation cycle is determined by checking whether the time interval between adjacent inputs exceeds a preset semantic breakpoint threshold. The expression unit index is constructed by generating an independent index number for each input and binding it to a time tag, thus forming a sequence of expressions arranged chronologically. For example, if a user inputs "I think this price is a bit high" and "But the service is not bad" consecutively within one minute, these two inputs, due to their short time interval, belong to the same semantic conversation cycle and will be labeled as expression unit 1 and expression unit 2 respectively. Arranging them in the order of input constitutes an expression sequence, facilitating subsequent semantic progression analysis and intent determination.

[0052] A semantic session cycle refers to the set of text inputs generated by a user during a continuous expression of intent. It typically consists of a series of inputs completed within a tight timeframe, with semantic content exhibiting coherence, progression, or transitions. The semantic session cycle can be determined using a dual standard: setting a time interval threshold and the similarity of the input content's theme. When the time interval between user inputs is too large or the content theme shifts significantly, it is classified as a different session cycle. An expression unit index is a unique identifier for each input segment in the expression sequence. It can be constructed based on timestamp numbers, auto-incrementing sequences, or hash combinations to achieve multi-dimensional location management. Each expression unit index not only contains sequential numbering information but can also include semantic feature labels and contextual position information. This allows for accurate identification of the logical relationships between expression segments during subsequent direction determination, intent recognition, or structural reorganization of the expression sequence, improving the processing accuracy of the expression chain and the reliability of recognizing changes in expression stance.

[0053] S102. Construct a set of turning expression matching rules for expression direction matching, perform sequential matching on adjacent expression units in the expression sequence, and determine whether there is an expression direction jump between each pair of expression units consisting of transition words, sentiment polarity words, and subject consistency.

[0054] S103. After the expression direction matching is completed, the sentiment polarity words and negative modifiers in each expression unit are identified based on the syntactic dependency relationship, and the positional relationship between the negative modifier and the sentiment polarity words in the syntactic path is determined. When the expression direction changes from forward to reverse or reverse to forward, and the relative positions of the negative modifiers are swapped, it is determined that the user has reversed the intention stance in the continuous expression in the context.

[0055] To determine whether a user's intention has reversed within a continuous context, syntactic dependency analysis can be used to identify sentiment polarity words and negative modifiers in each expression unit, and further determine the positional relationship of the negative modifier relative to the sentiment polarity word in the syntactic path. Syntactic dependency analysis is a technique for identifying the modification, subject-verb, and verb-object relationships between words within a sentence. By constructing a dependency tree structure, the grammatical subordination relationship between sentiment polarity words and negative modifiers can be clearly located. For example, in the expression "I don't like this design," the analysis results will show that "like" is the sentiment polarity word, and "not" is its negative modifier, with "like" as the head word and "not" as the modifier preceding it. If the relative position of the negative modifier relative to the sentiment polarity word is reversed in the two expression units, for example, the first expression is "I don't think it's appropriate," and the second expression is "I think it's pretty good," that is, the negation moves from the front to disappears, then combined with the trend of the change in expression direction, it can be judged that the intention has reversed from negative to positive, thus making a more detailed semantic confirmation of the change in expression direction.

[0056] Syntactic dependency relations refer to the structural connections between words within a sentence. They reveal the combinational rules of modifiers and modified words and are a core basis for deep sentence understanding. Sentiment polarity words are those that carry emotional inclinations, such as "like," "annoy," "satisfied," and "regretful," and they are often used as the core basis for judging user attitudes. Negative modifiers, such as "no," "not," "none," and "never," are used to reverse the meaning of sentiment polarity words. In expression, the collocation and order of these words significantly affect semantic understanding. By comparing the changes in expression direction with whether the syntactic dependency positions of negative modifiers have been interchanged, the accuracy of judging changes in the user's true intention can be enhanced. This operation not only compensates for the lack of judgment on subtle semantic reversals in sentiment analysis but also provides a more reliable basis for intention recognition in the continuous structure of contextual expressions, helping to accurately capture changes in stance in intelligent analysis.

[0057] In this embodiment, S102 specifically refers to:

[0058] Based on a sentiment analysis dictionary and a grammar rule base, a set of rules for matching turning expressions is constructed, which includes transition words, sentiment polarity words, semantic boundary markers and subjective expression components. Each rule in the rule set consists of a triple, defining the transition relationship, polarity direction and subject continuation features.

[0059] Constructing a set of rules for matching shifting expressions, including transition words, sentiment polarity words, semantic boundary markers, and subjective expression components, can be achieved by integrating a sentiment analysis dictionary and a grammar rule base. The sentiment analysis dictionary is a collection of words labeled with emotional attributes, where entries are assigned positive, negative, or neutral polarity weights to identify keywords expressing emotional tendencies in text; for example, "satisfied" is a positive sentiment polarity word, and "disappointed" is a negative sentiment polarity word. The grammar rule base is used to parse dependency relationships between sentence components, identifying conjunctions such as "but," "however," and "although" that have semantic transition functions—that is, transition words—used to determine whether there is a shift in stance between expressions. Semantic boundary markers are identifying components used to define semantic shifts in paragraphs or sentences, usually represented by punctuation marks or connecting phrases; for example, "however" or "actually" are often used to mark the starting point of semantic jumps. Subjective expression components include the user's subjective emotional evaluation of the target object such as a product, service, or experience, such as "I think it's too expensive" or "I feel it's good," where "I think" and "I feel" are signals of subjective stance. The rule set for expression reversal is built upon the above-mentioned elements, using a combination of manual rule induction and automatic mining. It consists of rule triplets composed of three elements: "transitional conjunctions, polarity direction, and subject-related features." Each triplet rule represents a language pattern of directional shifts in expression. For example, if the transitional conjunction "but" is present, and the preceding and following sentences contain sentiment words whose polarity direction changes from negative to positive, and the subject remains consistent, then this expression structure can be identified as a potential intention reversal pattern. By constructing this rule set, directional shifts in user expression can be efficiently identified, providing a structured basis for subsequent modeling and judgment of true intent.

[0060] The adjacent expression units in the expression sequence are sequentially matched, and the rules in the turning expression matching rule set are applied in turn. Keyword extraction and subject continuation judgment are performed on each pair of expression units, and the existence of expression direction jump candidate features is marked.

[0061] When performing sequential matching of adjacent expression units in an expression sequence, each pair of consecutive expression units needs to be selected according to the input order, and the rules in the transition expression matching rule set are applied one by one. Each rule consists of three linguistic features: transition words, polarity direction changes, and subject continuation status, forming the criteria for judging language expression transitions. Keyword extraction can be performed by combining regular expression matching and part-of-speech tagging to extract words with emotional polarity or structural guidance functions from each expression unit, such as "but," "on the contrary," and "although" as transition words, and "like" and "dissatisfied" as emotional polarity words. Subject continuation judgment is performed by using syntactic dependency analysis tools to detect whether the subjects in two expression units are consistent. For example, if the subject in both expression units is "I," then the subject continuation is considered valid. When two or more of the three conditions in a rule are met, the current expression unit pair is marked as a candidate feature for a change in expression direction. Candidate features refer to expression structure identifiers that have not yet been fully confirmed but have the possibility of transition, and are used to subsequently determine whether the intention and stance have changed. For example, if the first sentence is "I originally didn't plan to buy it," and the second sentence is "But now I think it's worth it," then keyword extraction identifies "planned" and "thinks" as polar words, and "but" as a transitional conjunction. Furthermore, the subject "I" is detected to be consistent in both sentences. This pair of expressive units is then marked as candidate features for a shift in expression direction, providing a basis for further identification of intentional stance reversals. This process effectively transforms scattered input into parsable structural cues, thereby improving the semantic accuracy and response judgment capabilities of subsequent analysis.

[0062] When candidate features simultaneously satisfy the following conditions: the presence of a transitional introductory word, a reversal of the emotional polarity direction, and consistency of the subject in adjacent expression units, it is determined that there is a change in the expression direction between the expression unit pairs.

[0063] When candidate features simultaneously satisfy the following conditions: the presence of a transitional conjunction, a reversal of sentiment polarity, and consistency of subject in adjacent expression units, a change in expression direction can be determined between the two expression units. The presence of transitional conjunctions can be automatically identified by constructing a connective lexicon and combining it with dependency parsing tools. For example, words like "but," "however," and "however" typically appear at the beginning or middle of a sentence, serving as connectors for semantic transitions. Determining a reversal of sentiment polarity requires combining sentiment polarity words extracted from the two expression units and comparing sentiment values ​​in a sentiment lexicon. For instance, if the first expression unit contains "very bad" and the next contains "quite like," a polarity reversal can be identified. Determining subject consistency can be achieved using subject-verb structure recognition tools. By extracting the subject from the sentence and performing string comparison or entity reference resolution, it can be confirmed whether the subject refers to the same object in both expression units. For example, if both expression units use "I" as the subject or there are cases where "I" and "myself" can be merged referents, then subject consistency can be determined. When a transitional conjunction appears, and the emotional expression shifts from negative to positive or vice versa, while the subject remains unchanged, it indicates a change in the speaker's stance, constituting a complete semantic feature of a shift in expression direction. For example, in the expression, "I originally thought this product was useless, but it's actually quite good," the conjunction "but" is a transitional conjunction, "useless" and "quite good" are emotional polarity reversal words, and "I" is the continuous subject, satisfying all the judgment conditions. Therefore, a shift in expression direction should be confirmed. This judgment process can effectively identify the user's true intention change during the expression process, which is of key value for subsequent intention analysis and precise marketing strategy decisions.

[0064] S2. For expressions judged to have an intentional reversal of stance, extract emotional verbs, evaluation objects, transition introductory words and semantic target words by establishing a list of semantic jump markers, and mark subjective expression abrupt change points within and between sentences to determine semantic shift clues;

[0065] In this embodiment, S2 specifically refers to:

[0066] Semantic extraction is performed on expressions that are judged to have an intentional reversal of stance. A list of semantic jump component tags containing emotional verbs, evaluation objects, transition introductory words and semantic target words is constructed. The emotional verbs and evaluation objects in the expression are identified by dependency parsing combined with part-of-speech tags. Semantic target words and transition introductory words related to their semantics are extracted by using contextual clues.

[0067] To perform semantic extraction on expressions deemed to contain a shift in intent, the first step is to structurally analyze continuous expressions using a syntactic analysis framework and establish a semantic dependency network between words using dependency syntax tools. The verb center and its associated object and subject components are extracted using a dependency tree. Furthermore, a pre-constructed sentiment lexicon is used to distinguish sentiment verbs from evaluation objects. For example, in the expression "I didn't like this feature at first, but after using it for a while, I think it's pretty good," "like" and "pretty good" can be identified as sentiment verbs, and "feature" as the evaluation object. Using the contextual clue of the conjunction "but" and "after using it for a while," "after using it for a while" can be further extracted as the semantic target word, reflecting the basis of the user's shift in intent. Finally, these elements are organized into a list of semantic shift markers, providing a foundation for subsequent semantic shift modeling. This process, through fine-grained structural analysis and contextual element fusion, can reveal the directional sentiment change paths hidden in user expressions.

[0068] In this process, semantic extraction refers to the precise extraction of core emotional structures from complex expressions using syntactic analysis and semantic recognition techniques. The semantic shift marker list is a dataset composed of multiple semantically sensitive elements, including sentiment verbs (such as "like" and "disappointed"), evaluation objects (such as "service" and "interface"), transitional introductory words (such as "but," "however," and "however"), and semantic target words (such as "after trying" and "used for a few days"). These elements constitute key signals of semantic reversal. Dependency syntax is used to construct a dependency relationship structure graph between words, and auxiliary part-of-speech tagging tools can identify the functional roles of words, thereby accurately identifying sentiment verbs and their evaluation objects. Contextual cues refer to linguistic units that provide a reversal trend or attitude change in the logical relationship before and after the expression, thereby inferring potential shift intentions. By combining these elements, not only can the semantic direction of the expression be reconstructed, but key inputs can also be provided for subsequent modeling of true intentions.

[0069] The extracted emotional verbs, evaluation objects, transition introductory words and semantic target words are semantically arranged according to the expression order and syntactic structure. Based on the emotional contrast relationship and semantic transition signal in the expression content, the position of semantic change in the expression is identified. Syntactic structure change points are marked within sentences and intentional change boundaries are marked between sentences to form a set of subjective expression change point markers.

[0070] To semantically arrange the extracted affective verbs, evaluation objects, transitional introductory phrases, and semantic target words according to their expression order and syntactic structure, these components first need to be linearly reconstructed in the context according to their temporal order and linguistic organization. Simultaneously, they are embedded into the semantic level using dependency relation structures. For example, for the expression "This function was originally useless, but now I think it's okay," by classifying "useless" and "okay" as negative and neutral affective verbs respectively, identifying the evaluation object as "function," the transitional introductory phrase as "but," and the semantic target word as "now," an affective contrast relationship can be constructed within the sentence, and transitional signal points can be marked. Next, by comparing the jumps in affective polarity, temporal adverbs, and grammatical structure between the two paragraphs, the semantic reverse trend between "useless" and "okay" is identified. Syntactic structural mutation points are marked at the points where the syntactic structure switches, and points of semantic evolution direction change are determined in cross-sentence expressions, defining them as boundaries of intentional mutation. Finally, these two types of mutation locations are unified into a set of subjective expression mutation point markers, providing support for subsequent expression map generation.

[0071] In this process, semantic arrangement refers to the logical organization of extracted keywords according to the input sequence of the content in the context and the nesting relationship of grammatical structures, ensuring their sequential and structural nature in the analysis. Emotional contrast refers to the opposite or conflicting relationship in emotional polarity between consecutive expressions, such as the shift from "bad" to "good," representing a reversal of emotional polarity. Semantic transition signals are usually triggered by transitional introductory phrases, used to indicate a change in the direction of the expression's meaning. Syntactic structural abrupt change points refer to the locations where the main structure of a sentence switches, such as the subject-verb-object framework jumping from one evaluative verb to another. Intent transition boundary refers to the boundary line in cross-sentence expressions where subjective intent shifts from negative to positive, or vice versa. The subjective expression abrupt change point marker set is a set of semantic markers used to record the locations and feature labels of the above two types of abrupt changes, serving as an important basis for subsequent analysis of true intent. In this way, not only can the temporal trajectory of user attitude changes be captured, but a data foundation can also be laid for building semantic transition models.

[0072] By performing correlation analysis between the set of subjective expression mutation point markers and the list of semantic jump component markers, and based on the path of emotional stance change guided by the transition introductory words, combined with the semantic tension distribution between syntactic mutation points, directional jump nodes are extracted as semantic transformation clues for subsequent construction and identification of true intentions.

[0073] In practice, the association analysis between the set of subjective expression mutation point markers and the list of semantic jump component markers first requires constructing a time alignment mechanism and a grammatical dependency graph to map the positions of markers inside and outside sentences to jump components. Next, for each transitional introductory phrase, the affected emotional verbs and evaluation objects are extracted along the preceding and following expression paths, and the semantic polarity is examined to detect whether there is a significant reversal. For example, in the statement "I thought it was meaningless, but later I found it quite valuable," the transition point in the marker set corresponds precisely to the position of the transitional introductory phrase "but," showing an evolutionary path of emotion from negative to positive. By calculating the semantic tension along this path, a clear evaluation tension is found between "meaningless" and "quite valuable," allowing the extraction of jump nodes and recording their directional labels and semantic strength. Ultimately, this node is identified as a semantic transition clue, serving as the core basis for constructing the true intent and guiding path selection and judgment weight setting in subsequent natural language analysis, thereby achieving accurate restoration of the user's true intent.

[0074] In this process, association analysis refers to cross-validating and path-connecting the relationship between two independent tag sets through logical mapping and contextual semantic modeling. Its goal is to capture the transition points in the expression that truly reflect changes in user attitude. The semantic transition component tag list records the keyword components of the expression, such as affective verbs and evaluation objects, while the subjective expression mutation point tag set defines the nodes where the expression structure changes. The transition introductory phrase is used here as a signal to switch the direction of expression, defining the boundary between preceding and following affective stances. The affective stance change path refers to the evolution of affective expression between the expression units on both sides of the transition introductory phrase. The semantic tension distribution reflects the degree of imbalance in the intensity and direction of affective expression between syntactic mutation points; when there is an evaluation conflict or a large difference in semantic intensity between two expression units, it will be identified as a transition node. The semantic transition cues are ultimately a set of expression transition indicators confirmed through the above calculations and matching, serving as the core markers driving the user intent recognition process forward.

[0075] S3. The extracted semantic transformation clues are used as nodes to reorganize the paths according to the expression order, differences in intention direction and semantic repulsion, generating a continuous semantic map with directionality and jump density to represent the evolution trajectory of the true intention.

[0076] In this embodiment, S3 specifically refers to:

[0077] The extracted semantic transformation clues are constructed into a set of nodes. Each node in the set is bound to the timestamp and intent direction attribute of the corresponding expression content, and arranged according to the expression order of the context to form an expression order structure with temporal markers and stance orientation.

[0078] In constructing a node set from extracted semantic shift cues, a natural language processing-based expression segmentation approach can be employed. Each semantic abrupt change is treated as an independent node, and a corresponding timestamp is appended to each node to indicate its order of appearance in the user's continuous expressions. Each node also needs to have its intent direction attribute calculated using a semantic parsing model, typically a positive intent, a negative intent, or a neutral state. By concatenating these nodes chronologically and combining them with the intent direction attribute, a serialized data structure is formed—the expression order structure. For example, if the expression first states "I think this product is no good," followed by "But later I decided to give it a try," two nodes can be extracted. The first is bound to a negative intent direction, and the second to a positive intent direction. Arranging them according to their timestamps creates an expression order structure with temporal markers and stance orientation. This structure helps in subsequently judging the evolution of the user's intent and provides a basis for semantic shifts.

[0079] Intent direction attribute is a feature value that characterizes the semantic attitude tendency of an expression. It is usually determined based on the semantic direction of sentiment polarity words, syntactic subject sequence, and contextual consistency, and is divided into three categories: positive intent, negative intent, and neutral intent. It is used to indicate the user's stance at a specific expression node. The expression sequence structure with temporal markers and stance orientation is a semantic sequence doubly labeled in both the time and intent dimensions. It not only preserves the natural chronological order of the user's expression content but also clarifies the semantic direction of each expression unit in the stance transition path. This structure supports the construction of transition paths across expression units, providing a precise basis for subsequent calculation of intent trends and analysis of transition density, effectively avoiding the problem of traditional static sentiment determination ignoring the semantic evolution process.

[0080] The difference in intentional direction attributes between adjacent nodes in the expression sequence structure is calculated. Combining the semantic orientation features of each node, a semantic repulsion scoring system is established based on the level of semantic conflict and the degree of emotional tension. Based on this, a jump path is constructed between nodes to achieve path reorganization.

[0081] In the expressive sequence structure, the intentional direction attribute of adjacent nodes can be mapped to numerical vectors through encoding, such as +1 for positive, -1 for negative, and 0 for neutral. A directional difference function is then constructed to measure the degree of change in the emotional stance of two nodes. Combining the semantic orientation of each node (e.g., pointing to a behavior, product, service, or result), semantic relevance and conflict level are calculated using a word vector model, thus establishing a semantic conflict level. For example, "give up" and "try" have a high-level conflict, while "consider" and "try" have a low-level conflict. Simultaneously, the intensity of emotional words in the context, such as "very dissatisfied" and "may consider," is used to calculate the emotional tension level. Combining the semantic conflict level and emotional tension level, a semantic repulsion scoring system is constructed. Each pair of adjacent nodes is scored; if the score exceeds a custom threshold, a jump path is marked between the two nodes, forming a non-linear emotional jump sequence. For example, if a user's expression changes from "I don't trust it at all" to "I thought about it later and maybe I could give it a try," then the difference is high and the semantics are significantly contradictory. This can be identified as a high-jump path through the scoring system, which is suitable for reorganizing the user's intent evolution path.

[0082] Difference calculation measures the degree of shift in emotional stance between adjacent nodes in an expression sequence structure, typically achieved through numerical comparison of intention direction attributes. Semantic orientation features are the emotional or behavioral target categories associated with nodes, used to determine whether semantic changes exhibit object continuity or topic consistency. Semantic conflict level reflects whether there is an opposing relationship between core words in two expressions in the semantic space; this level is determined based on the semantic distance between word vectors and negation markers. Emotional tension refers to the intensity of emotional fluctuations in an expression, assessed through the semantic depth of emotional words and the contextual emotional gradient. The semantic repulsion scoring system is a weighted scoring model built upon semantic conflict and emotional tension, used to quantify the intensity of jumps between expressions. Jump paths are path connections with semantic fragmentation and intentional shift characteristics, used to extract key attitude-shifting paragraphs from the original sequence, providing a structural foundation with directional and abrupt features for modeling true intentions.

[0083] The nodes after path reorganization are integrated with the jump paths to generate a continuous semantic graph. In the continuous semantic graph, the path direction attribute and jump density distribution are set. The jump density is used to characterize the concentration trend of semantic transformation cues in the expression sequence and to represent the evolutionary trajectory of the true intention.

[0084] After path reorganization, each semantic shift clue in the node set can be mapped to an independent node in the graph, and the transition paths can be connected to the corresponding nodes as edges to construct a continuous semantic graph. During generation, each edge needs to be bound to a direction attribute, which is determined based on the difference in intent direction between the source and target nodes. For example, a shift from a negative attitude to a positive attitude can be labeled as positive, and vice versa. The transition density is calculated based on the number of transition paths within a unit expression interval. A sliding window is used to scan the entire expression sequence, counting the number of transition paths within each window, and the results are normalized to form a transition density distribution map. For example, if in a user expression, "I initially thought it wouldn't work," "But then I thought I could try," and "Actually, it's quite suitable," multiple transition paths appear in the middle window, and their transition density will be higher than other paragraphs, thus allowing identification of concentrated areas of the user's true intent. The final continuous semantic graph not only shows the direction of semantic evolution but also quantifies the frequency and intensity of changes, providing a structured basis for subsequent intent judgment.

[0085] A continuous semantic graph is a structured graphical representation for semantic evolution modeling. Each node represents a semantic shift cue, and each side represents the emotional shift connection between two cues. Path direction attributes identify the trend of the content expressed between nodes in terms of intent stance; these are discrete values ​​labeled based on changes in emotional polarity direction, used to determine the dynamic trajectory of user attitude shifts. Jump density distribution is a statistical indicator reflecting the concentration of emotional stance changes. By statistically analyzing the frequency of jump paths through a sliding window of the expression sequence, it can locate expression regions with frequent emotional fluctuations. The jump density distribution not only provides spatial location characteristics of emotional changes but also reveals potentially key segments in the user's psychological transformation process. Integrating path direction attributes and jump density distribution into the semantic graph helps improve the discrimination accuracy and reasoning ability in the real intent modeling process, making the entire analysis process more interpretable and structurally complete.

[0086] S4. Input the transition nodes, jump direction labels and semantic segment strength indicators in the continuous semantic graph into the natural language processing parsing process and the artificial intelligence scoring and judgment process, respectively. By intervening in the parsing path of natural language processing and guiding the adjustment of scoring weights, the analysis of true intent is completed.

[0087] In this embodiment, S4 specifically includes the following steps:

[0088] S401. Extract transition nodes, jump direction labels and semantic segment strength identifiers from the continuous expression semantic graph, and establish an index mapping relationship between the transition nodes and the corresponding original expression segments. At the same time, bind the jump direction labels and semantic segment strength identifiers to the corresponding semantic segments to form a parsing input set.

[0089] S402. Input the parsing input set into the natural language processing parsing process. During the parsing process, limit the starting position of semantic parsing based on the transformation node, constrain the direction of semantic dependency expansion with the jump direction label, and assign parsing priority to the semantic branches in the parsing path according to the semantic fragment strength identifier, so as to realize the intervention of the natural language processing parsing path.

[0090] When inputting the parsed input set into the natural language processing (NLP) parsing process, each transition node must first be identified and used as the starting point for semantic parsing. This limitation of the starting point breaks the traditional static linear processing flow of analyzing sequentially from the first word, allowing for a re-examination of the expression path starting from the semantic abrupt change point, thereby focusing on the core fragments of the intention change. Next, the system controls the direction of semantic dependency expansion based on each transition direction label. For example, when the transition direction label points to an intention reversal, semantic dependency analysis will prioritize expanding towards the fragments after the expression transition, constructing the main dependency path. During the construction of the semantic dependency graph, the semantic fragment strength label of each semantic fragment is used as the basis for adjusting branch priority. High-intensity fragments will be given higher weight and receive higher priority participation in parsing. For example, if the expression "Although I don't intend to buy it, this configuration does attract me," then the expression "attract" will be parsed first in the semantic dependency graph because the transition direction points to a positive stance and has high emotional intensity. This process achieves deep intervention in the NLP parsing path, making parsing no longer limited by syntactic structure, but able to focus on the main theme of the expression around semantic abrupt changes.

[0091] Natural Language Processing (NLP) parsing refers to the process of constructing intra- and inter-sentence semantic structures through steps such as word segmentation, part-of-speech tagging, syntactic dependency relation identification, and semantic role labeling. The starting point of semantic parsing is the initial anchor point, usually determined by important semantic events, emotional shifts, or logical transitions; its accurate location determines the parsing approach. The semantic dependency expansion direction is the analysis path determined during the semantic dependency tree construction process, used to guide the determination of subsequent dependency relationships between words and the direction of branch extension. Semantic branches in the parsing path are extension chains in different directions within the semantic dependency graph. Assigning parsing priorities to these branches is done through a preset semantic strength weight function, used to determine which branch is more likely to represent the core of the true intent. The NLP parsing path is the traversal trajectory of the entire dependency graph construction process, reflecting how the system constructs a semantic understanding structure based on input information. Path selection directly affects the accuracy of semantic understanding and the final intent recognition result. By adjusting the starting point, direction, and weights, this processing mode can more accurately capture the evolution of the user's true stance in their expression.

[0092] S403. Input the parsed input set into the artificial intelligence scoring and judgment process, take the transition node as the scoring benchmark position, set the scoring offset direction according to the jump direction label, and adjust the weight distribution in the scoring and judgment process according to the semantic fragment strength identifier to complete the analysis of the true intent.

[0093] In the process of inputting the parsed input set into the AI ​​scoring and judgment process, the first step is to select a transition node as the scoring benchmark position to ensure that the scoring calculation focuses on the location where the semantic direction of the expression changes significantly. Then, the scoring offset direction is set according to the transition direction label. If the direction label points to a semantic reversal or a change in stance, the scoring model needs to give higher attention to the subsequent expression at the transition node, achieving a dynamic shift in the scoring focus. Based on this, the semantic strength identifier of each semantic segment is used as a weight parameter in the scoring and judgment process to adjust the contribution ratio of different semantic segments to the overall score. For example, when a user expression contains the statement "I didn't want to consider it, but this design moved me," the scoring model will use the transition node corresponding to "moved" as the starting point for scoring, extending towards a positive stance, and assigning a higher scoring influence due to its higher semantic strength, thus concluding that the user's intent has shifted to a positive direction. Through the offset path and weight enhancement strategy based on transition guidance, the scoring process can break free from the constraints of linear logic and more flexibly assess the impact of semantic evolution on the true intent.

[0094] The AI-powered scoring process refers to the computational process of quantitatively assessing the tendency, intent, or conversion potential of text expression based on deep learning models or semantic weighted models. The scoring baseline position is the starting point of the scoring function; its setting determines the model's focus in processing complex expressions and is typically tied to key nodes of semantic shifts. The scoring offset direction indicates which direction the scoring model should prioritize in its evaluation, addressing the non-linear trends caused by semantic shifts or position reversals. For example, when shifting from a negative expression to a positive statement, the offset direction will point towards the emotionally uplifting paragraph. The semantic segment strength identifier assigns differentiated weights to each scoring unit, calculated based on factors such as emotional polarity, subjective intensity, and verb emotional coloring. This adjusts the internal weight distribution of the overall score, making the scoring results more reflective of the true trend of intent evolution. Combining these three elements, the scoring judgment not only possesses structural flexibility but also dynamically adapts to the instability of the expressed content, significantly improving the ability to identify user conversion intentions.

[0095] In this embodiment, S401 specifically refers to:

[0096] Identify all locations marked as semantic transition points from the continuous semantic graph, and combine the degree of change in intent direction of nodes in the expression sequence structure with the semantic repulsion value. Select expression nodes that simultaneously satisfy the polarity reversal of intent direction and semantic repulsion score exceeding the set threshold as transition nodes, and extract the direction label of the path connected to each transition node as the transition direction label.

[0097] In continuous semantic graphs, the identification of semantic transition points can be achieved by detecting abrupt changes in the direction of intent between nodes. First, each node in the expression sequence needs to be assigned an intent direction value, which can be determined as positive or negative by analyzing the emotional polarity of words in the semantic segment through contextual analysis. When the intent direction between two adjacent nodes exhibits a polarity reversal from positive to negative or vice versa, and the semantic repulsion value between the two nodes exceeds a preset conflict threshold, the latter node can be identified as a semantic transition point. The semantic repulsion value can be calculated based on a weighted sum of factors such as the degree of semantic contradiction, contextual semantic tension, emotional antagonism, and subject consistency. For example, if a user expresses "I think this company is unreliable" in the first paragraph of an expression sequence, and "However, their recent service has been very satisfactory" in the second, there is a reversal of emotional polarity between the two paragraphs, and the word "however" in the context forms a clear semantic transition connector. The overall semantic tension is high, thus meeting the set threshold and being identified as a transition point. In the path connected by the transition node, the jump direction label "from negative to positive" can be extracted along the polarity change direction. The label will be bound to the path and used as a guide for subsequent analysis.

[0098] Semantic shift points refer to nodes in an expression sequence where a significant jump in subjective attitude occurs due to a reversal of semantic polarity and a change in contextual tension. These points typically involve a clear shift in emotional tone and are accompanied by a change in intentional attitude. The degree of change in the intentional direction of a node measures whether there has been a significant shift in subjective stance between preceding and following expressions; its value is derived from a directional analysis of contextual emotional preferences and subjective evaluation directions. Semantic repulsion measures the degree of semantic antagonism between two expression nodes, usually based on a fusion score of semantic opposition, emotional polarity conflict, and differences in expression intensity. A polarity reversal in the intentional direction reflects a clear reversal trend in the user's attitude during the expression process, while setting a threshold value aims to filter out weak shifts or non-critical nodes with unclear semantic changes. Shift nodes are core clues in semantic analysis, and shift direction labels are used to mark the stance evolution trends guided by these nodes, serving as an important basis for structural reconstruction and intention judgment. Through this process, the system can establish dynamic change points in the semantic decision-making path, providing stable anchor points for subsequent intention recognition.

[0099] Each extracted transition node is mapped one-to-one with its corresponding original expression fragment in the expression order structure, and a node position index is attached to the transition direction label to form a transition pointing relationship chain.

[0100] After identifying the transition nodes, a one-to-one index mapping relationship needs to be constructed between these transition nodes and their corresponding original expression segments in the expression sequence structure. This is achieved by pre-setting a unique index number for each expression segment in the expression sequence structure, which is automatically generated based on the user's input time order or semantic parsing order. Each transition node is associated with a specific expression segment, so by matching the node's timestamp or contextual position label, the node can be accurately bound to its source original expression segment, achieving a one-to-one mapping relationship. Then, a node position index is attached to each transition direction label, so that each direction label not only represents the direction of semantic polarity change but also carries the specific location where the transition occurred. Through this binding method, a transition pointing relationship chain can be generated. This chain consists of several structured data sets containing transition node numbers, transition direction labels, and their corresponding original expression segment indices. The establishment of this transition pointing relationship chain enables the subsequent natural language parsing and scoring process to accurately locate semantic mutation points, improving the path guidance capability of intent recognition.

[0101] In this process, "original expression fragments" refer to raw language units formed by users in continuous contextual expression, which have not been broken down or structured. These are typically sentences, phrases, or paragraphs, carrying specific subjective stances or emotional expressions. The "one-to-one index mapping relationship" is a precise mapping mechanism that uses unique identifiers to pinpoint nodes in the semantic graph to the raw expression language units, ensuring the traceability and accuracy of semantic processing. The "node position index" refers to the specific position number marking the occurrence of transformation nodes in the entire expression sequence, usually represented by an integer sequence number or timestamp. The "jump-pointing relationship chain" is a chain structure that links the directional changes and positional order of multiple transformation nodes, describing the evolutionary path of semantic direction and serving as the foundational set of clues for subsequent analysis path intervention and scoring guidance. Through the construction of this structure, a two-way alignment bridge can be established between the structured representation of semantics and the raw expression.

[0102] The semantic intensity label value in each semantic segment is used as the semantic segment intensity identifier, and combined with the jump direction label bound to it to form a data group containing directional and semantic weight information. This data group is then integrated with the number of the transition node to form the parsing input set.

[0103] To accurately measure the impact of different semantic segments on user intent judgment when processing expression fragments, a semantic intensity label value needs to be assigned to each semantic segment. This semantic intensity label value is obtained by comprehensively calculating the sentiment polarity, word intensity, subjective expression level, and contextual influence factors in the semantic segment using a natural language feature analysis model. The calculation method includes using a word vector model to score the intensity of sentiment verbs and evaluative words, combining high-frequency stance expression patterns in the corpus, and further identifying semantic components that enhance expression intensity based on a subjective lexicon and rhetorical patterns. For example, in the expression "I wasn't very interested before, but now I find it increasingly good," "increasingly good" has a high semantic intensity, and the system will assign a high semantic intensity label value to this expression. Subsequently, the semantic intensity label value of each semantic segment is used as a semantic segment intensity identifier, and it is combined with the bound transition direction label to form a data set containing directional information and semantic weight dimensions. Then, each data set is associated with the transition node number to which it belongs, constructing a complete set of parsed inputs. This parsed input set serves as the foundational data structure for the natural language processing parsing process and the artificial intelligence scoring and judgment process. It ensures that the scoring and judgment not only considers the semantic direction but also incorporates the influence of semantic weights, thereby achieving a more accurate characterization and judgment of the user's true intentions.

[0104] Semantic strength annotation is a quantitative indicator used to measure the intensity of semantic expression, typically calculated by a language model combining semantic depth and sentiment concentration. Semantic segment strength identifiers are uniquely encoded representations of semantic strength within each unit of expression, used for rapid retrieval and differentiation of semantic segment importance in subsequent processes. The parsing input set is a structured set composed of several data units containing semantic segment strength identifiers, transition direction labels, and transition node numbers, fed into the subsequent natural language processing flow as input. The construction of this set not only improves semantic analysis efficiency but also enhances the schedulability of the parsing path and the controllability of the scoring process at the structural level. Through this process, multi-dimensional semantic information can be effectively integrated, providing accurate support for subsequent intent evolution modeling and intelligent judgment.

[0105] S5. Perform time series comparison between the natural language processing parsing results and the artificial intelligence scoring results, construct a scoring fluctuation model, perform confidence fitting based on the jump segments in the continuous expression semantic map, and dynamically adjust the analysis output of the true intent based on the fitting results.

[0106] In this embodiment, S5 specifically refers to:

[0107] The natural language processing parsing results and the artificial intelligence scoring results are aligned according to the time sequence of the transition nodes. The intent judgment value in the parsing results and the score value in the scoring results are uniformly mapped to the same time axis. A time series comparison structure containing the parsing sequence and the scoring sequence is constructed to characterize the synchronous relationship between intent judgment and score changes.

[0108] In the implementation process, the intent judgment value corresponding to each transition node in the natural language processing parsing results can be used as a key variable. This intent judgment value reflects the stance or tendency of the current expression segment in the semantic graph (e.g., positive, negative, or neutral labels). Simultaneously, the score values ​​at corresponding time points in the AI ​​scoring results are extracted as matching evaluation indicators (e.g., rating levels or conversion intention values). By retrieving the timestamp of each transition node in the continuous expression semantic graph, the results from these two sources are aligned along the time dimension, mapping each intent judgment value and its corresponding time point score value onto the same time axis, constructing a set of paired data point sequences. Subsequently, based on these paired data, parsing sequences and scoring sequences are constructed, forming a time series comparison structure, which can reveal the synchronicity or offset between intent judgment and scoring judgment in different time periods. For example, in a certain expression, if the intent judgment value changes significantly first, while the score value adjusts with a lag, a potential scoring delay response interval can be identified.

[0109] The time-series alignment structure refers to treating the parsed sequence and the scoring sequence as two parallel time vector structures, with each time node corresponding to a set of intent judgment values ​​and score values. This structure achieves precise linkage and comparison between the two sequences by mapping the semantic labels output by natural language processing (NLP) with the numerical labels of the AI ​​scoring results. The intent judgment value is an expression intention classification label obtained in the NLP process based on semantic dependency path and sentiment jump analysis, while the score value is a conversion evaluation quantity output based on sentiment intensity and semantic trend calculations in the scoring judgment process. The purpose of constructing this structure is to align the two result sequences on a time basis and to annotate key events at jump nodes, enabling the subsequent modeling process to accurately identify the correlation between intent changes and score fluctuations. Through this structure, a data foundation and structural support can be provided for subsequent score fluctuation modeling, confidence fitting, and intent correction.

[0110] Based on the time series comparison structure, the fluctuation features formed by the change of rating values ​​over time are extracted, and combined with the position of the jump segment marked in the continuous expression semantic graph, the fluctuation features are segmented and modeled to construct a rating fluctuation model to describe the amplitude, direction of change and stable interval of rating fluctuations.

[0111] To construct a rating fluctuation model, the first step is to extract the rating value sequence of AI rating judgment results over time from the time series comparison structure. Then, using the sliding window method or time slicing technique, local statistical analysis is performed on this rating sequence to obtain the amplitude, directional trend, and stability of the rating curve across different time periods. Specifically, the rising, falling, or flattening trend of the rating curve can be determined by setting rating increase / decrease thresholds. These rating change points are then marked and segmented based on the positions of abrupt transitions in the continuous semantic graph, forming a set of time segments that match the semantic structure of the expression. Within each abrupt transition segment, statistical measures such as the mean fluctuation, extreme value amplitude, and slope of change are extracted. These are then combined with the time point information of semantic abrupt changes in the context to construct a rating change feature vector. Finally, the features from each time period are combined to form a rating fluctuation model. This model not only depicts the dynamic changes in rating values ​​during the expression process but also identifies which abrupt transitions correspond to abnormal rating fluctuation intervals, thus providing input for subsequent confidence fitting and dynamic adjustment. For example, if the rating rises sharply after a turning point expression while the semantic direction changes synchronously, it can be determined that the rating fluctuation is highly consistent with the change in intent.

[0112] The rating fluctuation model is a structured analysis model generated based on statistical features extracted from rating time series. It quantifies the changing trends and fluctuation characteristics of ratings as semantic expression evolves. Rating fluctuation amplitude refers to the difference between the maximum and minimum rating values ​​within a local time window, reflecting changes in expression intensity. The direction of change is determined by calculating the slope of the linear regression of the rating sequence within the transition segment, indicating whether the rating is rising towards conversion or falling towards resistance. Stable intervals are identified using indicators such as standard deviation or mean shift, marking continuous regions with high semantic consistency in the expression. Transition segments are key expression mutation fragments extracted from the continuous semantic map; their locations define the start and end boundaries of the fluctuation analysis, giving the rating fluctuation model semantic structure-related attributes. Through joint modeling of these structures, the rating fluctuation model can support dynamic confidence assessment, rating anomaly identification, and intent reversal detection.

[0113] Based on the fluctuation characteristics of the scoring fluctuation model in the jump range, confidence level fitting is performed. The fitted confidence level results are correlated with the analysis results of the corresponding time period. The analysis output of true intent is dynamically adjusted according to the trend of confidence level changes, so as to realize the time-series correction of intent judgment results.

[0114] When performing confidence level fitting, the first step is to extract key fluctuation features within the transition segments based on the rating fluctuation model, including the amplitude of rating fluctuations, the duration of fluctuations, and the stability of the direction of rating changes. This can be achieved by segmenting the rating sequence within the transition segments and calculating the consistency level of local rating changes, thus obtaining the initial confidence value for the corresponding time interval. Subsequently, this initial confidence value is correlated with the natural language processing (NLP) parsing results within the same time period, assigning a confidence weight to each parsing result. By comparing the built-in confidence change trends over consecutive time periods, it can be determined whether the parsing results exhibit short-term instability or sustained reliability. For example, when users' expressed attitudes gradually become more consistent after a semantic shift, and the rating fluctuation model shows a gradual and stable increase in ratings, the confidence level fitting results strengthen over time. In this case, the analysis output of the true intent is dynamically adjusted towards a stable direction. Conversely, if the rating oscillates repeatedly within the transition segments, the confidence weight of the corresponding parsing results is reduced, delaying or smoothing the intent determination results, thereby completing the temporal-level correction.

[0115] Confidence fitting refers to the quantitative matching of the stability characteristics reflected in the rating fluctuation model with the analysis results to generate a continuous numerical value or level label representing the credibility of the current intent judgment. The rating fluctuation model provides the information basis for fluctuation intensity and stable intervals, distinguishing between transient semantic noise and genuine intent changes. The confidence result is a quantitative description of the reliability of the analysis output, and its trend reflects the credible evolution of intent judgment over time. The analysis output of genuine intent refers to the intent conclusion formed by integrating the natural language processing analysis results and the artificial intelligence rating judgment results. By introducing the confidence change trend for dynamic adjustment, it is possible to avoid the excessive influence of a single expression anomaly on the overall judgment. Temporal correction emphasizes adjusting the temporal consistency of intent results in the context of continuous expression, making the final intent judgment more consistent with the gradual changes in the user's true stance.

[0116] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions according to the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means (e.g., infrared, wireless, microwave, etc.). A computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. Available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media. Semiconductor media can be solid-state drives.

[0117] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0118] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0119] In the several embodiments provided in this application, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0120] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0121] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0122] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An NLP-AI intelligent analysis method for customer acquisition and marketing, characterized in that, Specifically, the following steps are included: S1. Arrange the user's continuous contextual input into an expression sequence, match the expression direction by constructing a set of rules for expression direction matching, identify the positional relationship between sentiment polarity words and negative modifiers, and determine whether the user has reversed their intention stance in the continuous contextual expression. S2. For expressions judged to have an intentional reversal of stance, extract emotional verbs, evaluation objects, transition introductory words and semantic target words by establishing a list of semantic jump markers, and mark subjective expression abrupt change points within and between sentences to determine semantic shift clues; S3. The extracted semantic transformation clues are used as nodes to reorganize the paths according to the expression order, differences in intention direction and semantic repulsion, generating a continuous semantic map with directionality and jump density to represent the evolution trajectory of the true intention. S4. Input the transition nodes, jump direction labels and semantic segment strength indicators in the continuous semantic graph into the natural language processing parsing process and the artificial intelligence scoring and judgment process, respectively. By intervening in the parsing path of natural language processing and guiding the adjustment of scoring weights, the analysis of true intent is completed. S5. Perform time series comparison between the natural language processing parsing results and the artificial intelligence scoring results, construct a scoring fluctuation model, perform confidence fitting based on the jump segments in the continuous expression semantic map, and dynamically adjust the analysis output of the true intent based on the fitting results.

2. The NLP-AI intelligent analysis method for customer acquisition marketing according to claim 1, characterized in that, S1 specifically includes the following steps: S101. Arrange the continuous contextual expression inputs formed by the user in the same semantic session cycle into an expression sequence according to the actual input order, and construct an expression unit index based on time tags and input intervals to ensure that each expression unit in the expression sequence has a definite semantic position in the context. S102. Construct a set of turning expression matching rules for expression direction matching, perform sequential matching on adjacent expression units in the expression sequence, and determine whether there is an expression direction jump between each pair of expression units consisting of transition words, sentiment polarity words, and subject consistency. S103. After the expression direction matching is completed, the sentiment polarity words and negative modifiers in each expression unit are identified based on the syntactic dependency relationship, and the positional relationship between the negative modifier and the sentiment polarity words in the syntactic path is determined. When the expression direction changes from forward to reverse or reverse to forward, and the relative positions of the negative modifiers are swapped, it is determined that the user has reversed the intention stance in the continuous expression in the context.

3. The NLP-AI intelligent analysis method for customer acquisition marketing according to claim 2, characterized in that, S102 specifically refers to: Based on a sentiment analysis dictionary and a grammar rule base, a set of rules for matching turning expressions is constructed, which includes transition words, sentiment polarity words, semantic boundary markers and subjective expression components. Each rule in the rule set consists of a triple, defining the transition relationship, polarity direction and subject continuation features. The adjacent expression units in the expression sequence are sequentially matched, and the rules in the turning expression matching rule set are applied in turn. Keyword extraction and subject continuation judgment are performed on each pair of expression units, and the existence of expression direction jump candidate features is marked. When candidate features simultaneously satisfy the following conditions: the presence of a transitional introductory word, a reversal of the emotional polarity direction, and consistency of the subject in adjacent expression units, it is determined that there is a change in the expression direction between the expression unit pairs.

4. The NLP-AI intelligent analysis method for customer acquisition marketing according to claim 1, characterized in that, S2 specifically refers to: Semantic extraction is performed on expressions that are judged to have an intentional reversal of stance. A list of semantic jump component tags containing emotional verbs, evaluation objects, transition introductory words and semantic target words is constructed. The emotional verbs and evaluation objects in the expression are identified by dependency parsing combined with part-of-speech tags. Semantic target words and transition introductory words related to their semantics are extracted by using contextual clues. The extracted emotional verbs, evaluation objects, transition introductory words and semantic target words are semantically arranged according to the expression order and syntactic structure. Based on the emotional contrast relationship and semantic transition signal in the expression content, the position of semantic change in the expression is identified. Syntactic structure change points are marked within sentences and intentional change boundaries are marked between sentences to form a set of subjective expression change point markers. By performing correlation analysis between the set of subjective expression mutation point markers and the list of semantic jump component markers, and based on the path of emotional stance change guided by the transition introductory words, combined with the semantic tension distribution between syntactic mutation points, directional jump nodes are extracted as semantic transformation clues for subsequent construction and identification of true intentions.

5. The NLP-AI intelligent analysis method for customer acquisition marketing according to claim 1, characterized in that, S3 specifically refers to: The extracted semantic transformation clues are constructed into a set of nodes. Each node in the set is bound to the timestamp and intent direction attribute of the corresponding expression content, and arranged according to the expression order of the context to form an expression order structure with temporal markers and stance orientation. The difference in intentional direction attributes between adjacent nodes in the expression sequence structure is calculated. Combining the semantic orientation features of each node, a semantic repulsion scoring system is established based on the level of semantic conflict and the degree of emotional tension. Based on this, a jump path is constructed between nodes to achieve path reorganization. The nodes after path reorganization are integrated with the jump paths to generate a continuous semantic graph. In the continuous semantic graph, the path direction attribute and jump density distribution are set. The jump density is used to characterize the concentration trend of semantic transformation cues in the expression sequence and to represent the evolutionary trajectory of the true intention.

6. The NLP-AI intelligent analysis method for customer acquisition marketing according to claim 1, characterized in that, S4 specifically includes the following steps: S401. Extract transition nodes, jump direction labels and semantic segment strength identifiers from the continuous expression semantic graph, and establish an index mapping relationship between the transition nodes and the corresponding original expression segments. At the same time, bind the jump direction labels and semantic segment strength identifiers to the corresponding semantic segments to form a parsing input set. S402. Input the parsing input set into the natural language processing parsing process. During the parsing process, limit the starting position of semantic parsing based on the transformation node, constrain the direction of semantic dependency expansion with the jump direction label, and assign parsing priority to the semantic branches in the parsing path according to the semantic fragment strength identifier, so as to realize the intervention of the natural language processing parsing path. S403. Input the parsed input set into the artificial intelligence scoring and judgment process, take the transition node as the scoring benchmark position, set the scoring offset direction according to the jump direction label, and adjust the weight distribution in the scoring and judgment process according to the semantic fragment strength identifier to complete the analysis of the true intent.

7. The NLP-AI intelligent analysis method for customer acquisition marketing according to claim 6, characterized in that, S401 specifically refers to: Identify all locations marked as semantic transition points from the continuous semantic graph, and combine the degree of change in intent direction of nodes in the expression sequence structure with the semantic repulsion value. Select expression nodes that simultaneously satisfy the polarity reversal of intent direction and semantic repulsion score exceeding the set threshold as transition nodes, and extract the direction label of the path connected to each transition node as the transition direction label. Each extracted transition node is mapped one-to-one with its corresponding original expression fragment in the expression order structure, and a node position index is attached to the transition direction label to form a transition pointing relationship chain. The semantic intensity label value in each semantic segment is used as the semantic segment intensity identifier, and combined with the jump direction label bound to it to form a data group containing directional and semantic weight information. This data group is then integrated with the number of the transition node to form the parsing input set.

8. The NLP-AI intelligent analysis method for customer acquisition marketing according to claim 1, characterized in that, S5 specifically refers to: The natural language processing parsing results and the artificial intelligence scoring results are aligned according to the time sequence of the transition nodes. The intent judgment value in the parsing results and the score value in the scoring results are uniformly mapped to the same time axis. A time series comparison structure containing the parsing sequence and the scoring sequence is constructed to characterize the synchronous relationship between intent judgment and score changes. Based on the time series comparison structure, the fluctuation features formed by the change of rating values ​​over time are extracted, and combined with the position of the jump segment marked in the continuous expression semantic graph, the fluctuation features are segmented and modeled to construct a rating fluctuation model to describe the amplitude, direction of change and stable interval of rating fluctuations. Based on the fluctuation characteristics of the scoring fluctuation model in the jump range, confidence level fitting is performed. The fitted confidence level results are correlated with the analysis results of the corresponding time period. The analysis output of true intent is dynamically adjusted according to the trend of confidence level changes, so as to realize the time-series correction of intent judgment results.

Images