Personalized communication training content generation method based on big data

Abstract

The invention discloses a personalized communication training content generation method based on big data. The method comprises the following steps: S1, constructing a multi-modal feature sequence; s2, constructing a user expression brain map; s3, performing cognitive causal alignment on the time axis of the user expression brain map; s4, identifying ambiguous expressions in user input, expanding a plurality of semantic interpretation paths, forming a semantic trajectory tree, and generating a primary semantic path and a secondary semantic path; s5, generating conventional communication content based on the primary semantic path, constructing an evolution expression path based on the secondary semantic path, generating challenging expression content, and integrating and outputting a double-track fusion training corpus; s6, identifying a negative expression fragment in the double-track fusion training corpus, and generating an evolution training sequence based on an annealing control function; and S7, performing content reconstruction on the double-track fusion training corpus based on the evolution training sequence, and generating personalized communication training content. According to the method, multi-modal modeling and a semantic evolution control method are fused, and personalized communication training content generation is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, and particularly relates to a personalized communication training content generation method based on big data. BACKGROUND

[0002] At present, communication ability training gradually evolves from the traditional classroom mode to the digital and intelligent direction, and the existing system is mostly based on static text analysis or shallow dialogue interaction, lacking comprehensive perception and deep analysis of multi-modal data in actual communication behavior of users, and it is difficult to realize system modeling of individual communication style, expression bias and emotional mode, and it is unable to support efficient, dynamic and sustainable evolution of personalized communication training scheme.

[0003] In the prior art, the content generation method for communication ability improvement generally uses single modal input, such as constructing a semantic model based on language text or performing simple emotion classification through voice features, which has obvious deficiencies in processing complex expression behaviors and cannot capture the cooperative evolution relationship between multi-modal clues such as semantics, voice and facial expression. For communication content with ambiguity, emotional jump or ambiguous expression, the existing system lacks recognition and explanation mechanism, and cannot support multi-path expression understanding and judgment, thereby reducing the pertinence and controllability of the training content.

[0004] The prior art introduces a deep learning model for semantic recognition and dialogue generation, but lacks the embedding of cognitive mechanism and does not effectively model the sequence and causal relationship of information flow between modes. For example, in the combination processing of voice and text, only feature-level splicing or attention weight fusion is performed, without considering the time delay and cognitive driving relationship between voice response and semantic expression, making it difficult to build a user model reflecting the real expression mechanism.

[0005] In addition, the existing training system lacks adaptability to communication style, ignores the difference in psychological state and expression evolution trajectory of users, and the output content mode is single, the feedback lacks challenge and guidance, and it is difficult to realize gradual expression migration. Especially when facing users with negative expression tendency or expression anxiety, the traditional system lacks dynamic control mechanism for content generation, which is easy to cause mismatch between output feedback and user psychological state, thereby affecting the training effect.

[0006] In the training content generation process, the current method lacks path-level modeling and semantic evolution strategy support, and cannot realize the gradual guidance from inefficient expression to high-quality expression. Some systems use pre-trained models to generate multiple answers, but lack path scoring mechanism and semantic evolution criteria, and the content selection is random, which is difficult to support structured training process. More importantly, the existing method cannot effectively fuse the user's regular expression path and deviated expression path, lacks a unified training corpus organization mechanism, and it is difficult to generate double-track fusion content with stability and challenge.

[0007] There is also a blank for the generation intervention of negative expression. Current communication training systems usually only provide evasive suggestions, lack of expression path mechanism guiding users to gradually evolve from negative language to positive language, lack of semantic annealing strategy to control the change range of expression, resulting in lack of coherence and psychological adaptability of the generated content, and unable to achieve cognitive load controllable expression training.

[0008] Therefore, how to provide a personalized communication training content generation method based on big data is a problem that those skilled in the art need to solve. SUMMARY

[0009] An object of the present application is to provide a personalized communication training content generation method based on big data. The present application combines multi-modal behavior modeling, cognitive causal alignment, semantic trajectory analysis and annealing expression evolution control method, and constructs a personalized communication training content generation mechanism driven by user psychological state. It can accurately identify user expression characteristics and cognitive rhythm, dynamically generate an evolution training sequence gradually transitioning from negative expression to positive expression, and finally output a dual-track fusion training corpus with coherent semantics, consistent style and moderate challenge. It has the technical advantages of strong adaptability, high intervention, and clear expression optimization path, significantly improving the personalized intelligence level and practical effectiveness of communication training.

[0010] The personalized communication training content generation method based on big data according to the embodiment of the present application comprises the following steps: S1, collecting multi-modal communication behavior data of a user, extracting semantic features, voice features and facial features, and constructing a multi-modal feature sequence; S2, constructing a user expression brain map based on the multi-modal feature sequence, the user expression brain map comprising semantic nodes, non-semantic nodes and associated edges, the semantic nodes being labeled as negative expression nodes or positive expression nodes according to historical expression contexts; S3, performing cognitive causal alignment on the time axis of the user expression brain map according to the inter-modal cognitive response delay in the multi-modal feature sequence, to generate a cognitive-aligned user expression brain map; S4, identifying ambiguous expressions in user input based on the cognitive-aligned user expression brain map, expanding multiple semantic interpretation paths to form a semantic trajectory tree, and scoring each semantic interpretation path in combination with user psychological state information to generate a primary semantic path and a secondary semantic path; S5, generating regular communication content based on the primary semantic path, constructing an evolution expression path based on the secondary semantic path, generating challenging expression content, and integrating and outputting a dual-track fusion training corpus; S6, identify the negative expression segment in the double-track fusion training corpus, guide the generation of content based on the annealing control function from the negative expression node, gradually transition to the positive expression node along the semantic evolution path, and generate an evolution training sequence; S7, based on the evolution training sequence, reconstruct the content of the double-track fusion training corpus to generate personalized communication training content.

[0011] Optionally, the multi-modal communication behavior data includes language text, speech audio and video image; The language text is encoded by a BERT model for context semantic vector, and a CR model is used for named entity recognition, sentiment polarity classification and syntax dependency analysis to extract semantic features; The speech audio is extracted by an openSMILE acoustic analysis tool for pitch, intensity, speech rate, formant frequency and Mel frequency cepstrum coefficient, and a bidirectional long short-term memory network is used for sequence modeling of frame-level features to extract speech features; The video image is detected and tracked by an OpenFace tool to extract facial action units, eye movement and micro-expression change rate, and a ResNet-50 convolutional neural network is used to encode the facial region image to generate facial features.

[0012] Optionally, the S2 specifically includes: S21, obtaining the multi-modal feature sequence, wherein the semantic features are used to construct semantic nodes, and the speech features and facial features are used to construct non-semantic nodes; S22, constructing the connection edges between the semantic nodes based on the syntax dependency relationship, and the edge weight is the cosine similarity between the words; S23, combining the K-Means clustering algorithm to cluster facial expressions, eye movements and speech prosody patterns to form non-semantic node categories, and constructing the cosine similarity edge structure of non-semantic nodes; S24, constructing a user expression brain map based on the semantic nodes, non-semantic nodes and associated edges; S25, based on the user historical communication interaction data, counting the number of positive feedback and negative feedback of each semantic node in different situations, and calculating the expression feedback score of the semantic node: ; Wherein, The expression feedback score of the i-th semantic node is represented by and represent the feedback weight, The number of positive feedback of the i-th semantic node is represented by The number of negative feedback of the i-th semantic node is represented by​​ the number of negative feedbacks of the semantic node; S26, compare the expression feedback score with a preset score threshold, if the expression feedback score is greater than or equal to the preset score threshold, mark the semantic node as a positive expression node, otherwise mark it as a negative expression node, and write the marking information into the user expression brain map.

[0013] Optionally, the S3 specifically comprises: S31, based on the timestamp information of each modality data in the multi-modal feature sequence, extract the response start time and peak time of the semantic feature, the voice feature and the facial feature, and construct a modality event time sequence; S32, combine any two modality event sequences in pairs, and calculate the average response delay between each modality event by using a sliding time window method , wherein represents the modality number, and the average response delay is the average start time difference between the modality event relative to the modality event. S33, represent the semantic feature, the voice feature and the facial feature as modality event sequences respectively. , and set a response delay threshold ; when the Granger causality test result is true and , set the corresponding element in the cognitive causal alignment mapping matrix, otherwise set it to ; S34, adjust the time axis of the semantic node and the non-semantic node in the user expression brain map according to the cognitive causal alignment mapping matrix : when , it means that the modality is prior to the modality in the cognitive processing sequence, adjust the time position of the corresponding node, and set the causal dependence edge; when , keep the original time relationship between the corresponding nodes unchanged, and do not set the causal dependence edge; Finally, a time axis structure reflecting the cognitive sequence relationship between modalities is generated, and a cognitive aligned user expression brain map is obtained.

[0014] Optionally, the S4 specifically comprises: S41, based on the cognitive aligned user expression brain map, perform text preprocessing and semantic ambiguity detection on the user input semantic node, and identify the semantic node with ambiguity; S42, using the semantic node with ambiguity as a starting node, using a BERT model to expand the starting node to generate multiple rounds of prediction to form multiple semantic interpretation paths, each semantic interpretation path consisting of consecutive semantic nodes; S43, constructing a tree-shaped semantic trajectory tree according to the generated multiple semantic interpretation paths, wherein the starting node is the root node, and each semantic interpretation path is a different branch of the semantic trajectory tree; S44, establishing a psychological state feature vector according to user historical communication data and combining user psychological state information, the psychological state feature vector including emotional stability, anxiety level and communication style tendency; S45, using the psychological state feature vector to score the psychological fit degree of each semantic interpretation path in the semantic trajectory tree, using cosine similarity to evaluate the fit degree of the semantic interpretation path and the current psychological state of the user, and obtaining a matching score corresponding to each semantic interpretation path; S46, according to the matching score, the semantic interpretation paths in the semantic trajectory tree are sorted, and the semantic interpretation path with the highest matching score is determined as the main semantic path, and the remaining semantic interpretation paths are determined as secondary semantic paths.

[0015] Optionally, the S5 specifically includes: S51, extracting the semantic representation of each semantic node of the main semantic path, combining the set context language template and the user historical language style to generate regular communication content; S52, identifying the semantic node of the secondary semantic path whose semantic direction deviation from the main semantic path exceeds the preset deviation threshold as the evolution starting point, constructing the evolution expression path by gradually approaching the main semantic path, so that the semantic distance between nodes gradually converges; S53, expressing the semantic nodes in the evolution expression path with emotional turning points, attitude comparison sentence patterns and viewpoint reconstruction to generate challenging expression content for expressing challenges and changes; S54, integrating the regular communication content and the challenging expression content in a nested manner, the nested manner including sequential embedding, parallel comparison and emotional progression, and outputting a double-track fusion training corpus.

[0016] Optionally, the S6 specifically includes: S61, performing fragment-level semantic analysis on the double-track fusion training corpus to identify the content corresponding to the negative expression node, and marking it as a negative expression fragment; S62, locating all semantic evolution paths leading from the negative expression node to the positive expression node, and constructing an annealing control function: ; wherein, represents the annealing temperature of the current round, represents an initial annealing temperature, represents a current generation round, represents a semantic distance between the current generated content and the target positive expression node, represents a user cognitive adaptation rate, and represents a regulation coefficient; The annealing control function is used to dynamically adjust the semantic migration amplitude in the generation process, wherein the higher the annealing temperature, the greater the expression jump amplitude allowed, and when the annealing temperature decreases, the generated content tends to be stable; S63, taking the negative expression segment as the initial input, combining the current round annealing temperature, guiding the generated content to gradually transition to the positive expression node along the semantic evolution path, and realizing the gradual evolution of language expression in the semantic and emotional directions; S64, setting an evolution criterion for judging whether the generation process meets the evolution termination condition, wherein the evolution criterion is that the current generated content has reached the positive expression node in the user expression brain map, and the cosine similarity between the semantic representation of the current generated content and the target positive expression node exceeds a set similarity threshold; S65, when the evolution termination condition is met, the generation process is terminated and the evolution training sequence is output, otherwise the current generated content is taken as the input of the next generation, and the next round of annealing control is entered until the evolution termination condition is met or the maximum generation round is reached.

[0017] Optionally, the S7 specifically comprises: S71, sentence processing and semantic feature extraction are performed on the evolution training sequence; S72, semantic feature extraction is performed on the regular communication content and the challenging expression content in the double-track fusion training corpus; S73, the semantic features in the evolution training sequence are matched with the semantic features in the double-track fusion training corpus, and sentence pairs with a semantic correlation degree higher than a preset correlation degree threshold are marked as reconstruction areas; S74, based on the sentence structure and emotional expression mode in the evolution training sequence, the sentences in the reconstruction area are replaced, rearranged and inserted; S75, after the content reconstruction is completed, language fluency detection and semantic consistency verification are performed, and personalized communication training content is generated.

[0018] The beneficial effects of the present application are: (1) The user semantic, speech and facial multi-modal communication behavior is uniformly modeled, an expression brain map with time causal alignment characteristics is constructed, and the real communication state and cognitive rhythm of the user can be accurately reflected.

[0019] (2) Introduce semantic trajectory tree and psychological state matching scoring mechanism, support multi-path explanation and ordering of ambiguous expression, improve the adaptability of training content to user's current psychological characteristics and expression preference.

[0020] (3) Based on the double-track fusion mechanism, the subject semantic path and the challenging expression path are fused, which effectively improves the diversity and challenge of the training corpus, while maintaining semantic consistency and style stability, and enhances the training guidance effect.

[0021] (4) Through the dynamic annealing control function, the expression evolves from the negative node to the positive node, and the semantic convergence criterion is used to generate the evolution training sequence, realizing the controllable migration of individual expression ability and the change of behavior mode.

[0022] (5) Based on the evolution training sequence, the training corpus is reconstructed to generate personalized communication training content with psychological adaptability, content progression and expression specificity, improving the individualized intervention ability and generation quality of the communication training system. BRIEF DESCRIPTION OF DRAWINGS

[0023] The accompanying drawings are included to provide a further understanding of the present application, and constitute a part of the specification, which together with the embodiments of the present application, serve to explain the present application, and do not constitute a limitation on the present application. In the drawings: Figure 1 The overall flowchart of the personalized communication training content generation method based on big data proposed by the present application; Figure 2 The double-track fusion training corpus flowchart of the primary semantic path and the secondary semantic path of the personalized communication training content generation method based on big data proposed by the present application; Figure 3 The expression evolution generation process flowchart of the personalized communication training content generation method based on big data proposed by the present application based on the annealing control function. DETAILED DESCRIPTION

[0024] The present application will now be further described in detail in conjunction with the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the present application in a schematic manner, and therefore only show the components related to the present application.

[0025] Reference Figures 1-3 The personalized communication training content generation method based on big data includes the following steps: S1, collect the multi-modal communication behavior data of the user, extract the semantic features, voice features and facial features, and construct a multi-modal feature sequence; S2, construct a user expression brain map based on the multi-modal feature sequence, the user expression brain map comprising semantic nodes and non-semantic nodes and associated edges, the semantic nodes being labeled as negative expression nodes or positive expression nodes according to historical expression contexts; S3, cognitively align a time axis of the user expression brain map according to inter-modal cognitive response delays in the multi-modal feature sequence, to generate a cognitively aligned user expression brain map; S4, based on the cognitively aligned user expression brain map, identify ambiguous expressions in user input, develop multiple semantic interpretation paths, form a semantic trajectory tree, and score each semantic interpretation path in combination with user psychological state information, to generate a primary semantic path and a secondary semantic path; S5, generate regular communication content based on the primary semantic path, construct an evolving expression path based on the secondary semantic path, generate challenging expression content, and integrate output dual-track fusion training corpus; S6, identify negative expression segments in the dual-track fusion training corpus, and based on an annealing control function, guide generation of content to start from a negative expression node and gradually transition to a positive expression node along a semantic evolution path, to generate an evolution training sequence; S7, based on the evolution training sequence, reconstruct content of the dual-track fusion training corpus, to generate personalized communication training content.

[0026] The present application proposes a personalized communication training content generation process oriented to multi-modal input, with clear structure and path evolution driven, covering the whole process of data collection, expression modeling, semantic path identification, training corpus fusion and expression migration, an expression brain map is constructed through semantic and non-language nodes, dynamic evolution of communication features is accurately captured, with the help of a semantic trajectory tree, multi-path evaluation and annealing type language control mechanism, training content forms a complete closed loop from cognitive understanding to expression evolution, a controllable training path from negative expression to positive communication is realized, and content adaptability, behavior intervention accuracy and user expression ability growth efficiency are effectively improved.

[0027] In the embodiment, the multi-modal communication behavior data comprises language text, speech audio and video image; The language text is subjected to context semantic vector encoding through a BERT model, and is subjected to named entity recognition, sentiment polarity classification and syntax dependency relationship analysis in combination with a CR model, to extract semantic features; The speech audio is subjected to pitch, intensity, speech rate, formant frequency and Mel frequency cepstrum coefficient extraction through an openSMILE acoustic analysis tool, and is subjected to sequence modeling of frame-level features by using a bidirectional long short-term memory network, to extract speech features; The video image is subjected to face detection and tracking by the OpenFace tool, facial action units, eye movement and micro-expression change rate are extracted, and a ResNet-50 convolutional neural network is used to encode the facial region image to generate facial features.

[0028] In this step, multi-modal perception and deep feature extraction technology are introduced to comprehensively obtain language, speech and facial expression data of users in the communication process, and BERT, BiLSTM and ResNet neural network models are used to construct feature vectors, ensuring the accuracy and stability of the communication behavior modeling. Each modality information has time sequence continuity and expression semantic correlation, providing a solid data foundation for user expression brain mapping construction and semantic migration, and significantly improving the comprehensive performance of the training system in recognition ability, reaction sensitivity and detail capture precision.

[0029] In this embodiment, S2 specifically includes: S21, acquiring the multi-modal feature sequence, wherein the semantic features are used to construct semantic nodes, and the speech features and facial features are used to construct non-semantic nodes; S22, constructing connection edges between the semantic nodes based on the syntactic dependency relationship, and the edge weight is the cosine similarity between words; S23, combining the K-Means clustering algorithm to cluster facial expressions, eye movements and speech prosody patterns to form non-semantic node categories, and constructing the cosine similarity edge structure of the non-semantic nodes; S24, constructing a user expression brain map based on the semantic nodes, non-semantic nodes and associated edges; S25, based on the user historical communication interaction data, counting the number of positive feedback and the number of negative feedback of each semantic node corresponding to the expression in different situations, and calculating the expression feedback score of the semantic node: ; wherein, represents the expression feedback score of the i-th semantic node, and represent the feedback weight, represents the number of positive feedback of the i-th semantic node, represents the number of negative feedback of the i-th semantic node. S26, comparing the expression feedback score with a preset score threshold, if the expression feedback score is greater than or equal to the preset score threshold, marking the semantic node as a positive expression node, otherwise marking it as a negative expression node, and writing the marking information into the user expression brain map.

[0030] ​​​The step forms an expression brain map data structure by establishing a structured connection relationship between semantic and non-semantic nodes, and reflects user historical behaviors into expression labels by using an expression feedback score mechanism, realizes mapping from behavior records to expression cognitive state, and can clearly identify a tendency expression by quantifying the number of positive and negative feedbacks and setting a feedback score threshold, thereby laying a logical foundation for evolution path construction and personalized feedback generation, and improving the emotion adaptability and expression diagnosis capability of communication content training.

[0031] In the embodiment, the S3 specifically includes: S31, based on the timestamp information of each modality data in the multi-modal feature sequence, extracting response start time and peak time of semantic features, speech features and facial features, and constructing a modality event time sequence; S32, combining any two modality event sequences in pairs, and calculating the average response delay between each modality event by using a sliding time window method , wherein represents the modality number, and the average response delay is the average start time difference between the events in the modality and the events in the modality . S33, representing the semantic features, speech features and facial features as modality event sequences , respectively. S34, performing Granger causality test on each modality pair , and setting a response delay threshold ; when the causality test result is true and , setting the corresponding element in the cognitive causal alignment mapping matrix, otherwise setting it as ; S34, adjusting the time axis of the semantic nodes and non-semantic nodes in the user expression brain map according to the cognitive causal alignment mapping matrix ; when , it means that the modality is prior to the modality in the cognitive processing sequence, adjusting the time position of the corresponding nodes, and setting a causal dependence edge; when , the original time relationship between the corresponding nodes is kept unchanged, and no causal dependence edge is set; Finally, a time axis structure reflecting the cognitive sequence relationship between modalities is generated, and a cognitive aligned user expression brain map is obtained.

[0032] The step can effectively eliminate the cognitive misplacement problem of multi-modal features in time distribution by analyzing the response delay and sequence of multi-modal data in actual expression, combining Granger causality test and time axis rearrangement strategy, completing cognitive causal alignment of user expression brain map, making the expression sequence more truly reflect the user information processing order, and thus more accurately reasoning the logical path of user language generation in generation, improving the semantic reasoning accuracy and the explainability of generation behavior.

[0033] In the embodiment, the S4 specifically includes: S41, based on the cognitive aligned user expression brain map, performing text preprocessing and semantic ambiguity detection on the semantic nodes input by the user, and identifying semantic nodes with ambiguity; S42, taking the semantic node with ambiguity as a starting node, using a BERT model to expand multiple rounds of prediction generation for the starting node, forming multiple semantic interpretation paths, each semantic interpretation path being composed of continuous semantic nodes; S43, constructing a tree-shaped semantic trajectory tree according to the generated multiple semantic interpretation paths, wherein the starting node is a root node, and each semantic interpretation path is a different branch of the semantic trajectory tree; S44, establishing a psychological state feature vector according to user historical communication data and combining user psychological state information, the psychological state feature vector including emotional stability, anxiety level and communication style tendency; In actual application, when the user interacts with the system through language, the voice and text input of the user will be the basic data source for psychological state evaluation. Taking emotional stability as an example, the system will call an open source emotion recognition tool (such as the TensorFlow model of Google or the emotion analysis API in Microsoft Azure cognitive service) to analyze the tone change, speech speed fluctuation and emotional vocabulary in the user's voice, and obtain the emotional classification probability distribution such as “angry”, “sad”, “calm” and the like. Through statistical analysis of the frequency and intensity of emotional fluctuations, it can be mapped to a continuous numerical value between 0 and 1, and the lower the value, the more unstable the emotion.

[0034] The anxiety level is evaluated by analyzing the negative word frequency (such as “worried”, “afraid”, “stress” and the like) in the user input content and the hesitation performance in the syntax structure (such as frequent use of “maybe”, “possibly”, “don't know” and the like), and calibrated by combining a lightweight questionnaire data training model to output an anxiety index in the range of 0 to 1.

[0035] The communication style tendency is analyzed by analyzing the style characteristics of the user's language use in the continuous conversation, including the proportion of direct command sentences, the frequency of emotional expression language, and the use of empathetic language (such as "I understand your feelings"). The feature is classified by K-Means clustering method (such as "direct type", "empathetic type", "logical type"), and its tendency is numerically encoded, for example, by principal component mapping to represent the style tendency degree between 0 and 1.

[0036] Finally, these quantitative parameters form a fixed-dimensional psychological state feature vector, for example: [emotional stability = 0.35, anxiety level = 0.62, communication style tendency-empathetic = 0.78], which is used to score the psychological fit degree of each semantic interpretation path in the semantic trajectory tree, so as to select the most suitable language output path for the current user's psychological state.

[0037] S45, scoring the psychological fit degree of each semantic interpretation path in the semantic trajectory tree using the psychological state feature vector, using cosine similarity to evaluate the fit degree of the semantic interpretation path and the current psychological state of the user, and obtaining the matching score corresponding to each semantic interpretation path; S46, sorting the semantic interpretation paths in the semantic trajectory tree according to the matching score, determining the semantic interpretation path with the highest matching score as the main semantic path, and determining the remaining semantic interpretation paths as secondary semantic paths.

[0038] This step provides a joint mechanism for ambiguity recognition and path evaluation, which expands multiple semantic interpretation paths with the help of the BERT model, and combines the psychological state feature vector for matching and scoring, realizes individual adaptability optimization of the semantic selection process, and through semantic trajectory tree modeling and main and secondary path division, can effectively avoid random jumping of model output, ensure the balance of training content in cognitive fit and expression extension, and improve the effectiveness of communication feedback content in user expression bias correction, thought clarity reconstruction, etc.

[0039] In this embodiment, the S5 specifically includes: S51, extracting the semantic representation of each semantic node of the main semantic path, combining the set context language template and the user's historical language style, and generating regular communication content; S52, identifying the semantic node of the secondary semantic path whose semantic direction deviation from the main semantic path exceeds the preset deviation threshold as the evolution starting point, constructing the evolution expression path by gradually approaching the main semantic path, so that the semantic distance between nodes gradually converges; S53, expressing emotional turning points, attitude comparison sentence patterns, and viewpoint reconstruction for the semantic nodes in the evolution expression path, generating challenging expression content for expressing challenges and changes; S54, integrate the conventional communication content and the challenging expression content in a nested manner, including sequential embedding, parallel comparison and emotional progression, and output a double-track fusion training corpus.

[0040] This step builds a training corpus system with conventional and challenging expressions coexisting by introducing semantic deviation recognition, challenging expression construction and dynamic fusion strategy. Not only does it retain the familiar communication mode of the user, but also gradually guides them into a new path outside the expression comfort zone, forming a corpus structure with confrontation and stability coexisting. The content style is continuous and the semantic tension is controllable during the training process. The fusion strategy ensures that the output corpus is logically coherent and has appropriate psychological carrying capacity, enhancing the educational guidance value and practical adaptability of training feedback.

[0041] In this embodiment, S6 specifically includes: S61, perform semantic analysis on the double-track fusion training corpus at the segment level, identify the content corresponding to the negative expression node semantics, and mark it as a negative expression segment; S62, locate all semantic evolution paths leading from the negative expression node to the positive expression node, and construct an annealing control function: ; wherein, T represents the current round annealing temperature, T0 represents the initial annealing temperature, N represents the current generation round, D represents the semantic distance between the current generated content and the target positive expression node, V represents the user's cognitive adaptation rate, and K represents the regulation coefficient; The annealing control function is used to dynamically adjust the semantic migration amplitude in the generation process. The higher the annealing temperature, the greater the expression jump amplitude allowed. When the annealing temperature decreases, the generated content tends to be stable. S63, taking the negative expression segment as the initial input, combining the current round annealing temperature, guiding the generated content to gradually transition to the positive expression node along the semantic evolution path, realizing the gradual evolution of language expression in the semantic and emotional direction; S64, set an evolution criterion to determine whether the generation process meets the evolution termination condition. The evolution criterion is that the current generated content has reached the positive expression node in the user's expression brain map, and the cosine similarity between the semantic representation of the current generated content and the target positive expression node exceeds the set similarity threshold. S65, when the evolution termination condition is met, terminate the generation process and output the evolution training sequence. Otherwise, continue to use the current generated content as the input for the next generation, enter the next round of annealing control until the evolution termination condition is met or the maximum generation round is reached.

[0042] This step realizes the continuous evolution control mechanism of expression content from negative nodes to positive nodes by constructing a dynamic multi-factor annealing control function combining semantic distance and user cognitive adaptation rate. The annealing control function finely adjusts the semantic migration amplitude in the generation process, and combines semantic path convergence and behavior index to achieve the criterion, ensuring that the output content meets the target training direction, significantly improving the guidance accuracy and behavior conversion efficiency of the model generated content, and meeting the dual requirements of gradualness and individualized response in expression training.

[0043] In this embodiment, the S7 specifically includes: S71, sentence processing and semantic feature extraction of the evolution training sequence; S72, semantic feature extraction of the regular communication content and challenging expression content in the double-track fusion training corpus; S73, matching the semantic features in the evolution training sequence with the semantic features in the double-track fusion training corpus, finding sentence pairs with a semantic correlation degree higher than a preset correlation degree threshold, and marking them as reconstruction areas; S74, replacing, rearranging and inserting the sentences in the reconstruction area based on the sentence structure and emotional expression pattern in the evolution training sequence; S75, after content reconstruction, language fluency detection and semantic consistency verification are performed to generate personalized communication training content.

[0044] This step realizes the reorganization of expression content structure and style by reconstructing the content of the double-track training corpus with the evolution training sequence. Based on semantic feature matching and high correlation area positioning, the communication content is replaced, inserted and rearranged in sequence, combined with language fluency and semantic consistency verification technology, to ensure that the output result maintains logical integrity and has language naturalness. Finally, personalized communication training content with emotional guidance, cognitive adaptation and expression evolution characteristics is formed, significantly improving the educational targeting and practical acceptability of the output content of the training system.

[0045] Example 1: In order to verify the feasibility of the application in implementation, the application is applied to the individualized communication training task in the language communication course of a certain university. For the interactive performance of 35 test students in the simulation communication practice system, personalized communication training content conforming to cognitive characteristics and communication habits is generated.

[0046] In practical applications, the system first collects multi-modal communication behavior data of students in the interaction process, including speech audio, text input and video image, and synchronously records in each round of conversation window. Based on the BERT model, the text is encoded into semantic vectors, and the conditional random field model is used to complete the named entity recognition, sentiment polarity judgment and syntax structure analysis; openSMILE is used for audio pitch, speech rate, formant and MFCC feature extraction, and then sent to BiLSTM network to extract speech sequence features; facial images are extracted through OpenFace to extract eye movement, micro-expression rate and facial action unit, and ResNet-50 is used to complete facial encoding.

[0047] All modal features form a multi-modal feature sequence under a unified timeline, and then the expression window is divided, and the user expression brain map is constructed in each expression window. The graph includes semantic nodes and non-semantic nodes, and the connection edges are formed through syntactic relationships, and the structure is generated by clustering non-verbal modalities using K-means. Then based on the historical interaction record, the response times of each semantic node in the positive feedback and negative feedback scenarios are counted, and the expression feedback score is used to mark positive and negative expressions, and the positive feedback score is set to 0.7, the negative feedback score is set to 0.3, and the preset score threshold is 0.15.

[0048] After that, the modal response delay is extracted through the sliding time window, and the Granger causality test is used to judge the cause-effect relationship between modalities, and a cognitive causal alignment mapping matrix is constructed. For example, the average delay of speech to semantics is 180ms, the average delay of semantics to face is 220ms, and the average delay of speech to face is 120ms, and all of them pass the Granger significance test, so in the matrix =1.

[0049] After completing the cognitive causal alignment, the system identifies ambiguous semantic nodes on the expression brain map and expands the semantic trajectory tree. Taking "how can you say this again" as an example, BERT generates three semantic interpretation paths, which are directed to the explanation type, accusation type and concern type semantic sequence respectively. The system combines the user's historical psychological state vector (anxiety score 3.8 / 5, emotional fluctuation score 4.2 / 5) to match the scores of the three paths and sort them, and finally selects the concern type as the main semantic path, and the other two as the secondary paths.

[0050] Then, the system generates regular communication content based on the main semantic path; for the nodes with large deviation in the secondary paths, it constructs the evolution expression path, and generates challenging expression content by using viewpoint reversal, attitude comparison and language progression. Finally, through the semantic nesting mechanism, the double-track training corpus is fused and output, and each training sample contains regular and challenging expression double-channel content.

[0051] In the expression optimization phase, the system identifies negative expression fragments in the training corpus, for example, "I don't think you're right," and sets up the path evolution process between it and the positive expression node "I understand your point of view." The initial temperature T0 in the annealing control function is 1.0, the regulation coefficient =0.05, =0.3, =0.6, =0.1, the semantic distance dinitial is 0.68, and the user cognitive adaptation rate r=0.35. The first round generates the content "I understand your position, but I have different views," and the semantic distance decreases to 0.42; the second round "your idea has some reason," and the distance is 0.21; the third round converges to 0.09 and exceeds the preset threshold 0.1, and the generated sequence is marked as an evolution training sequence.

[0052] The system finally reconstructs the training corpus, inserts the evolution sequence into the high semantic correlation fragment, and generates personalized content with natural language style, clear logic, and rich expression training value through semantic replacement and sequential rearrangement.

[0053] The average scores of expression positivity, semantic clarity, and communication confidence before and after training are improved by 26.3%, 31.7%, and 24.5%, respectively. The students using the method of the present application have an average positive expression proportion of 71.2% in the classroom simulation communication test, which is significantly higher than the baseline of 38.4%. The average training content satisfaction score is 4.65, which is higher than the baseline of 3.72, and the full score is 5, which is significantly better than the traditional static template training method.

[0054] Table 1 Change of key indicators of personalized communication content generation system before and after training Student Number Forward Expression Proportion (Before) Forward Expression Proportion (After) Average Semantic Distance Convergence Turns Content Satisfaction Score Psychological Fit Path Number Modal Response Delay Difference (ms) U01 42.1% 74.8% 3 4.7 2 180 U02 36.9% 68.5% 4 4.6 3 205 U03 40.4% 70.2% 2 4.9 2 178 U04 31.8% 66.7% 3 4.5 2 192 U05 39.2% 73.0% 3 4.8 3 186

[0055] The data in the above table 1 shows the changes of key performance indicators of 5 students in the sample before and after training using the personalized communication content generation system of the present application. Before training, the positive expression proportion of the 5 students was at a low level, with the lowest being 31.8% (U04) and the highest being 42.1% (U01), indicating that they had a certain negative expression tendency or unstable emotional expression in communication. After training with the system, the positive expression proportion was significantly improved, with U01 increasing to 74.8% and U04 increasing to 66.7%, reflecting the positive guiding effect of the evolution training sequence on the user's expression method.

[0056] From the average semantic distance convergence round, most students complete the semantic evolution transition from negative expression nodes to positive expression nodes within 3 rounds, indicating that the annealing control function and evolution path constructed by the system have good stability and efficiency, and can achieve gradual correction of expression direction in a short training period.

[0057] The content satisfaction score is maintained at 4.5 or above, and the U03 score reaches 4.9, indicating that the generated training content is widely recognized by users in terms of language naturalness, expression fit and semantic coherence, and reflects the outstanding advantages of the application in personalized content generation.

[0058] In terms of the number of psychological fit paths, each student can match 2 to 3 high-fit semantic paths, which effectively enhances the ability to analyze ambiguous expressions and improves the relevance and diversity of semantic interpretation paths.

[0059] The modal response delay difference is maintained between 178ms and 205ms, verifying the feasibility and precision of the inter-modal cognitive causal alignment method in the application, and ensuring the synchronization and logical consistency of each modal node in the expression brain map.

[0060] As can be seen from the embodiment, the personalized communication training content generation method based on big data proposed by the application can construct a user expression brain map with a cognitive causal structure based on deep modeling of multi-modal behavior data, and through semantic trajectory tree generation, evolution path design and annealing guide mechanism, realize the gradual migration from negative expression to positive expression, and finally output high-quality personalized communication content that fits individual psychological state and expression style. Through actual application data verification, the method has significant effect in improving expression positivity, enhancing content fit, and optimizing training efficiency, and has good practicality and promotional value.

[0061] The above describes only the preferred specific embodiments of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes to the technical solutions and inventive concepts of the application within the technical scope disclosed by the application, which should be covered within the protection scope of the application.

Claims

1. A method for generating personalized communication training content based on big data, characterized in that, Includes the following steps: S1. Collect users' multimodal communication behavior data, extract semantic features, voice features and facial features, and construct a multimodal feature sequence; S2. Construct a user expression mind map based on the multimodal feature sequence. The user expression mind map includes semantic nodes, non-semantic nodes, and related edges. The semantic nodes are labeled as negative expression nodes or positive expression nodes according to the historical expression context. S3. Based on the intermodal cognitive response delay in the multimodal feature sequence, calculate the average response delay between each modal event, perform Granger causality test, and perform cognitive causal alignment on the time axis of the user expression mind map to generate a cognitively aligned user expression mind map. S4. Based on the user expression mind map after cognitive alignment, identify ambiguous expressions in user input, expand multiple semantic interpretation paths, form a semantic trajectory tree, and score each semantic interpretation path in combination with user psychological state information to generate the main semantic path and secondary semantic path. S5. Generate regular communication content based on the main semantic path, construct an evolutionary expression path based on the secondary semantic path, generate challenging expression content, and integrate and output dual-track fusion training corpus. S6. Identify negative expression segments in the dual-track fusion training corpus, and guide the generated content from the negative expression node to the positive expression node along the semantic evolution path based on the annealing control function to generate an evolutionary training sequence. S7. Based on the evolutionary training sequence, the content of the dual-track fusion training corpus is reconstructed to generate personalized communication training content.

2. The method for generating personalized communication training content based on big data according to claim 1, characterized in that, The multimodal communication behavior data includes spoken text, audio, and video images; The language text is encoded into contextual semantic vectors using the BERT model, and then combined with the CR model to perform named entity recognition, sentiment polarity classification, and syntactic dependency analysis to extract semantic features. The speech audio was analyzed using the openSMILE acoustic analysis tool to extract pitch, intensity, speech rate, formant frequencies, and Mel frequency cepstral coefficients. A bidirectional long short-term memory network was used to perform sequence modeling of frame-level features to extract speech features. The video images are used for face detection and tracking using the OpenFace tool to extract facial action units, eye movements, and micro-expression change rates. The facial region images are then encoded using a ResNet-50 convolutional neural network to generate facial features.

3. The method for generating personalized communication training content based on big data according to claim 1, characterized in that, S2 specifically includes: S21. Obtain the multimodal feature sequence, wherein semantic features are used to construct semantic nodes, and speech features and facial features are used to construct non-semantic nodes; S22. Construct connection edges between semantic nodes based on syntactic dependency relations, with the edge weights being the cosine similarity between words; S23. Combine the K-Means clustering algorithm to cluster facial expressions, eye movements and speech prosody patterns to form non-semantic node categories, and construct the cosine similarity edge structure of non-semantic nodes. S24. Construct a user expression mind map based on semantic nodes, non-semantic nodes, and associated edges; S25. Based on users' historical communication and interaction data, count the number of positive and negative feedbacks generated by the expression corresponding to each semantic node in different contexts, and calculate the expression feedback score of the semantic node: ； in, Indicates the first The expression feedback score of each semantic node. and Indicates the feedback weight. Indicates the first The number of positive feedbacks for each semantic node. Indicates the first The number of negative feedbacks for each semantic node; S26. Compare the expression feedback score with a preset score threshold. If the expression feedback score is greater than or equal to the preset score threshold, the semantic node is marked as a positive expression node; otherwise, it is marked as a negative expression node, and the labeling information is written into the user expression mind map.

4. The method for generating personalized communication training content based on big data according to claim 1, characterized in that, S3 specifically includes: S31. Based on the timestamp information of each modality data in the multimodal feature sequence, extract the response start time and peak time of semantic features, speech features and facial features, and construct a modal event time series; S32. For any two modal event sequences, combine them pairwise and use the sliding time window method to calculate the average response delay between each modal event. ,in Indicates the mode number, and the average response delay is the mode number. Events relative to modality The average start time difference of events in the middle; S33. Represent semantic features, speech features, and facial features as modal event sequences respectively. For each modal pair Perform Granger causality testing and set a response delay threshold. When the causality test result is true and At that time, the corresponding elements in the cognitive causal alignment mapping matrix are set. Otherwise set to ; S34. Align the mapping matrix according to the cognitive causality. Adjust the timeline of semantic and non-semantic nodes in the user's expression mind map: when When, it indicates a mode. Prior to modality in cognitive processing sequence Adjust the time position of the corresponding node and set causal dependency edges; when At the same time, the original temporal relationship between the corresponding nodes remains unchanged, and no causal dependency edges are set; Finally, a timeline structure reflecting the cognitive sequence between modalities is generated, and a cognitively aligned user expression mind map is obtained.

5. The method for generating personalized communication training content based on big data according to claim 1, characterized in that, S4 specifically includes: S41. Based on the cognitively aligned user expression mind map, perform text preprocessing and semantic ambiguity detection on the semantic nodes input by the user, and identify the semantic nodes with ambiguity. S42. Using the ambiguous semantic node as the starting node, the BERT model is used to perform multiple rounds of prediction and generation on the starting node to form multiple semantic interpretation paths, each semantic interpretation path consisting of consecutive semantic nodes. S43. Construct a tree-structured semantic trajectory tree based on the generated multiple semantic interpretation paths, where the starting node is the root node and each semantic interpretation path is a different branch of the semantic trajectory tree. S44. Based on the user's historical communication data and combined with the user's psychological state information, establish a psychological state feature vector, which includes emotional stability, anxiety level and communication style tendency. S45. The psychological fit of each semantic interpretation path in the semantic trajectory tree is scored using the psychological state feature vector. The cosine similarity method is used to evaluate the fit between the semantic interpretation path and the user's current psychological state, and the matching score corresponding to each semantic interpretation path is obtained. S46. Sort the semantic interpretation paths in the semantic trajectory tree according to the matching score, and determine the semantic interpretation path with the highest matching score as the primary semantic path, and the remaining semantic interpretation paths as secondary semantic paths.

6. The method for generating personalized communication training content based on big data according to claim 1, characterized in that, S5 specifically includes: S51. Extract the semantic representation of each semantic node in the main semantic path, and generate regular communication content by combining the set context language template and the user's historical language style. S52. Take the semantic nodes whose semantic direction deviation from the main semantic path exceeds a preset deviation threshold as the starting point of evolution, and construct the evolutionary expression path by gradually moving closer to the main semantic path, so that the semantic distance between nodes gradually converges. S53. Reconstruct the semantic nodes in the evolutionary expression path by expressing emotional transitions, contrasting attitudes, and reconstructing viewpoints to generate challenging expression content for expressing challenges and changes. S54. Integrate routine communication content with challenging expression content in a nested manner, including sequential embedding, parallel comparison and emotional progression, and output dual-track fusion training corpus.

7. The method for generating personalized communication training content based on big data according to claim 1, characterized in that, S6 specifically includes: S61. Perform segment-level semantic analysis on the dual-track fusion training corpus, identify the content that corresponds to the semantics of the negative expression node, and mark it as a negative expression segment; S62. Locate all semantic evolution paths from negative expression nodes to positive expression nodes, and construct the annealing control function: ； in, This indicates the annealing temperature for the current round. Indicates the initial annealing temperature. Indicates the current generation round. This represents the semantic distance between the currently generated content and the target positive expression node. Indicates the user's cognitive adaptation rate. and Indicates the control coefficient; The annealing control function is used to dynamically adjust the semantic transfer amplitude during the generation process. The higher the annealing temperature, the greater the allowable expression jump amplitude, and the generated content tends to be stable when the annealing temperature decreases. S63. Using negative expression fragments as initial input, and combined with the annealing temperature of the current round, guide the generated content to gradually transition to positive expression nodes along the semantic evolution path, so as to realize the gradual evolution of language expression in terms of semantics and emotion. S64. An evolutionary criterion is set to determine whether the generation process meets the evolution termination condition. The evolutionary criterion is that the current generated content has reached the positive expression node in the user's expression mind map, and the cosine similarity between the semantic representation of the current generated content and the target positive expression node exceeds the set similarity threshold. S65. When the evolution termination condition is met, terminate the generation process and output the evolution training sequence; otherwise, continue to use the current generated content as the input for the next round of generation and enter the next round of annealing control until the evolution termination condition is met or the maximum number of generation rounds is reached.

8. The method for generating personalized communication training content based on big data according to claim 1, characterized in that, Specifically, S7 includes: S71. Perform sentence segmentation and semantic feature extraction on the evolutionary training sequence; S72. Extract semantic features from the regular communication content and challenging expression content in the dual-track fusion training corpus; S73. Match the semantic features in the evolutionary training sequence with the semantic features in the dual-track fusion training corpus, find sentence pairs with semantic relevance higher than a preset relevance threshold, and mark them as reconstructed regions. S74. Based on the sentence structure and emotion expression patterns in the evolutionary training sequence, replace, rearrange, and insert sentences in the reconstructed region; S75. After completing the content reconstruction, conduct language fluency testing and semantic consistency verification to generate personalized communication training content.

Images