Artificial intelligence communication assistance method, device and equipment adaptive to multiple scenarios

Abstract

The application discloses an artificial intelligence communication assistance method and device suitable for multiple scenes, and relates to the technical field of information processing. The method comprises the following steps: obtaining dialogue audio or importing dialogue text; performing scene recognition, emotion temperature calculation and interaction stage determination on the dialogue audio or dialogue text to obtain structured context data containing a scene, an emotion and an interaction stage; performing dependency syntax analysis on the structured context data by using a BERT multi-task regression model and a keyword regular rule engine to obtain multi-dimensional problem positioning; analyzing the root cause of the multi-dimensional problem positioning from four dimensions of behavior motivation, object characteristics, interaction habits and external environment to obtain a multi-dimensional root cause set; and generating an instant response technique based on the identified scene, emotion stage and the multi-dimensional root cause set. The application can improve communication efficiency and the practicability of an intelligent agent.

Description

Technical Field

[0001] This application relates to the field of information processing technology, and in particular to an artificial intelligence communication assistance method, device, and equipment adapted to multiple scenarios. Background Technology

[0002] Existing intelligent communication assistance tools are highly fragmented, capable of performing only single tasks such as audio transcription, emotion recognition, and script recommendation. Each module operates independently, and data cannot be automatically transferred. Users must repeatedly switch between multiple software programs and manually export and import dialogue data during communication analysis, resulting in a lengthy and cumbersome process that significantly reduces the efficiency of communication assistance analysis. Furthermore, during data transfer between tools, key contextual information such as emotional fluctuations, implicit demands from the other party, and accumulated past conflicts is easily overlooked or lost. In subsequent stages, the lack of a complete dialogue context as a foundation for analysis leads to communication solutions that are disconnected from the actual dialogue scenario, rendering them of limited reference value. Summary of the Invention

[0003] In view of the above-mentioned defects or deficiencies in the related technologies, the purpose of this application is to provide an artificial intelligence communication assistance method, device, or equipment that is adaptable to multiple scenarios, which can improve communication efficiency and the practicality of intelligent agents.

[0004] To achieve the above objectives, this application provides the following solution: Firstly, this application provides an AI-assisted communication method adaptable to multiple scenarios. The method includes: acquiring dialogue audio or importing dialogue text; performing scene recognition, emotion temperature calculation, and interaction stage determination on the dialogue audio or dialogue text to obtain structured context data containing scene, emotion, and interaction stage; using a BERT multi-task regression model and a keyword regularization rule engine to perform dependency parsing on the structured context data to obtain multi-dimensional problem localization; analyzing the root causes of the multi-dimensional problem localization from four dimensions: behavioral motivation, object characteristics, interaction habits, and external environment to obtain a multi-dimensional root cause set; and generating an instant response script based on the identified scene, emotion stage, and the multi-dimensional root cause set.

[0005] Optionally, the step of performing scene recognition, emotion temperature calculation, and interaction stage determination on the dialogue audio or dialogue text to obtain structured context data containing scene, emotion, and interaction stage includes: transcribing the acquired dialogue audio into a first standardized dialogue text; cleaning and preprocessing the dialogue text to obtain a second standardized dialogue text; inputting the first standardized dialogue text and / or the second standardized dialogue text into a pre-trained scene classification model, outputting the confidence score corresponding to each candidate scene; selecting the scene with the highest confidence score above the confidence threshold as the recognition result, and marking all scenes with confidence scores below the threshold as unknown scenes; and employing sentiment analysis. The model calculates the emotion score for each line of dialogue, and calculates the average emotion score and emotion fluctuation variance based on all emotion scores within the dialogue window. Combining the average emotion score and emotion fluctuation variance, the overall emotion state of the current dialogue is obtained. Multi-dimensional observation features of each round of dialogue are extracted to form a temporal observation feature sequence. The emission probability of the observation feature vector in the hidden state of each interaction stage is fitted using a Gaussian mixture model. The state transition probability between interaction stages is obtained using a hidden Markov model. Using the Viterbi decoding algorithm, the optimal hidden state path is solved for the temporal observation feature sequence based on the emission probability and the state transition probability to obtain the interaction stage of each round of dialogue.

[0006] Optionally, the step of using a BERT multi-task regression model and a keyword regularization rule engine to perform dependency parsing on the structured context data to obtain multi-dimensional problem localization includes: parsing each dialogue statement in the structured context data syntactically, inputting the parsed dialogue text into the BERT multi-task regression model, and simultaneously outputting predicted scores for the execution end, demand end, and relationship end; starting the keyword regularization rule engine to match the accusatory sentences, implicit demands, and historical contradiction-type feature words and phrases corresponding to the execution end, demand end, and relationship end, and obtaining rule scores for each dimension respectively; if the confidence level of the predicted score output by the BERT multi-task regression model is higher than a preset threshold, the predicted score and the rule scores for each dimension are weighted and fused to obtain the final problem score for each dimension; if the confidence level of the predicted score output by the BERT multi-task regression model is lower than a preset threshold, the rule scores for each dimension are used as the final problem score for each dimension.

[0007] Optionally, the step of analyzing the root causes of the multi-dimensional problem location from four dimensions—behavioral motivation, object characteristics, interaction habits, and external environment—to obtain a multi-dimensional root cause set includes: extracting communication bottlenecks, unmet needs, and accumulated conflict labels of the multi-dimensional problem location as root cause inference input fragments; inferring and matching potential root causes from the four dimensions—behavioral motivation, object characteristics, interaction habits, and external environment—and outputting the inference confidence score corresponding to each root cause; calling a preset keyword regularization rule library to supplement and verify the model inference results, and for categories with model confidence scores lower than a preset threshold, matching corresponding feature statements to supplement and generate root cause entries; sorting all root cause entries in descending order according to inference confidence scores, and selecting the top-ranked root causes as valid root causes; the valid root causes include the dimension they belong to, textual descriptions, and confidence score information.

[0008] Optionally, the step of generating an immediate response script based on the identified scenario, emotional stage, and the multi-dimensional root cause set includes: extracting the current scenario tag, the emotional stage of the dialogue, and the question type corresponding to the effective root cause to form an index retrieval condition; searching a preset script template library based on the index retrieval condition; if a matching template is found, extracting the slot placeholders within the template; extracting entity information from the dialogue context and root cause description to fill the template slots and generate preliminary script content; if no matching template is found, loading standardized prompt words, inputting the scenario, emotional stage, root cause, and dialogue text into a large language model to dynamically generate multiple candidate scripts; retrieving a taboo word library corresponding to the current scenario, screening and eliminating illegal expressions in the preliminary script content or the multiple candidate scripts, and generating the immediate response script.

[0009] Optionally, the method further includes: matching the root cause-suggestion mapping templates based on the two root causes with the highest confidence to generate structured long-term improvement suggestions containing daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content; adding explicit identifiers and implicit source metadata to the instant response scripts and the structured long-term improvement suggestions to obtain historical instant response scripts and historical structured long-term improvement suggestions with content traceability.

[0010] Optionally, the step of matching the root cause-suggestion mapping template with the two root causes with the highest confidence to generate structured long-term improvement suggestions containing daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content includes: sorting the multi-dimensional root cause set in descending order of confidence and extracting the top two root causes; reading a pre-configured root cause-suggestion mapping template library and retrieving corresponding improvement suggestion template entries based on the root cause type and root cause description of the two extracted root causes to obtain the root cause-suggestion mapping template; extracting contextual information related to people, events, and business entities from the dialogue text and root cause descriptions; substituting the contextual information into the root cause-suggestion mapping template to generate daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content respectively; and assigning corresponding type tags to the daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content to assemble them into structured long-term improvement suggestions.

[0011] Secondly, this application provides an AI-powered communication assistance device adapted to multiple scenarios, the device comprising: The acquisition module is used to acquire dialogue audio or import dialogue text. The processing module is used to perform scene recognition, emotion temperature calculation and interaction stage determination on the dialogue audio or dialogue text to obtain structured context data containing scene, emotion and interaction stage. The localization module is used to perform dependency parsing on the structured context data using the BERT multi-task regression model and keyword regularization rule engine to obtain multi-dimensional problem localization. The analysis module is used to analyze the root causes of the multi-dimensional problem from four dimensions: behavioral motivation, object characteristics, interaction habits, and external environment, and obtain a multi-dimensional root cause set. The generation module is used to generate instant response scripts based on the identified scenario, emotional stage, and the multi-dimensional root cause set.

[0012] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the AI-assisted communication method adapted to multiple scenarios described in the first aspect above.

[0013] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides an AI-powered communication assistance method, device, and equipment adaptable to multiple scenarios. It obtains structured contextual data through scene recognition, emotion temperature calculation, and interaction stage determination using dialogue audio or text. This provides unified and complete basic dialogue information for end-to-end analysis, reducing context loss caused by switching between multiple tools. It leverages dependency parsing combined with a BERT multi-task regression model and a keyword regularization rule engine to achieve multi-dimensional problem localization, balancing the model's generalization and reasoning capabilities with the rule's precise fallback advantage. It can identify various communication bottlenecks at the execution, demand, and relationship levels. Furthermore, it analyzes behavioral motivation and target... This approach employs a layered root cause analysis across four dimensions: characteristics, interaction habits, and external environment, penetrating surface-level communication conflicts to pinpoint deeper underlying causes. Finally, it combines scenario, emotional stage, and multi-dimensional root cause set to generate targeted, real-time response scripts. The entire process eliminates the need for manual data transfer between software programs, improving dialogue analysis efficiency. Furthermore, relying on multi-model fusion and multi-dimensional layered reasoning mechanisms, the analysis results closely align with real-world dialogue contexts, and the output response scripts are highly adapted to the current communication scenario and the root cause of the conflict. This addresses the shortcomings of existing intelligent communication assistance tools, such as fragmented functionality, one-sided analysis, and empty, unrealistic output solutions, thereby improving the timeliness and practicality of dialogue assistance. Attached Figure Description

[0014] 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 of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart illustrating an AI-assisted communication method adapted to multiple scenarios, provided as an embodiment of this application; Figure 2 A schematic diagram of the functional modules of an AI-powered communication assistance device adapted to multiple scenarios, provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0018] In one exemplary embodiment, such as Figure 1 As shown, this application provides an AI-assisted communication method adaptable to multiple scenarios, which includes the following steps S101 to S105. Wherein: Step S101: Obtain the dialogue audio or import the dialogue text.

[0019] In the example embodiment, if the dialogue is text, no processing is required and it can proceed directly to subsequent processing; if it is audio, the audio data in 16kHz PCM format is acquired in real time through the recording device SDK and transcribed into text. In other embodiments, voiceprint recognition technology can also be used to separate speakers and identify the speaker of each message; for example, in a sales scenario, it can distinguish between a customer and a salesperson; in a family communication scenario, it can distinguish between conversations between spouses, parents and children, and between generations; in a workplace meeting scenario, it can distinguish between discussion, reporting, and review conversations; and in a customer service consultation scenario, it can distinguish between a consultant and a customer service representative.

[0020] Step S102: Perform scene recognition, emotion temperature calculation, and interaction stage determination on the dialogue audio or dialogue text to obtain structured context data containing scene, emotion, and interaction stage.

[0021] Optionally, step S102 may include: transcribing the acquired dialogue audio into a first standardized dialogue text; cleaning and preprocessing the dialogue text to obtain a second standardized dialogue text; inputting the first standardized dialogue text and / or the second standardized dialogue text into a pre-trained scene classification model, outputting the confidence score corresponding to each candidate scene; selecting the scene with the highest confidence score above the confidence threshold as the recognition result, and marking all scenes with confidence scores below the threshold as unknown scenes; using a sentiment analysis model to calculate the sentiment score of each dialogue sentence by sentence, calculating the average sentiment score and the sentiment fluctuation variance based on all sentiment scores within the dialogue window, and combining the average sentiment score and the sentiment fluctuation variance to obtain the overall sentiment state of the current dialogue; extracting multi-dimensional observation features of each round of dialogue to form a temporal observation feature sequence; fitting the emission probability of the observation feature vector in the hidden state of each interaction stage using a Gaussian mixture model; obtaining the state transition probability between interaction stages using a hidden Markov model; and using the Viterbi decoding algorithm to solve for the optimal hidden state path based on the emission probability and the state transition probability in the temporal observation feature sequence to obtain the interaction stage of each round of dialogue.

[0022] It should be noted that the pre-trained scene classification model in this application embodiment is a pre-trained classifier built based on BERT; specifically, using ert-base-chinese as the base, a 4-class fully connected layer is added to the output layer. The output layer of the pre-trained classifier outputs the confidence scores for all candidate scenes; then, a threshold filtering strategy is used to determine the final scene category label based on the confidence scores. In this application embodiment, the pre-trained classifier outputs the probability of each scene label, and the label with the highest probability and a confidence score > 0.7 is taken as the final recognition result; if the confidence scores are all below 0.7, they are marked as unknown scenes and processed by a subsequent keyword regularization rule engine or manual intervention. All candidate scenes in this application include common dialogue scenarios in daily life, such as family scenarios, workplace business scenarios, and public service scenarios.

[0023] Preferably, the sentiment analysis model can be the SnowNLP model. The SnowNLP model takes the segmented dialogue text as input and outputs a raw sentiment score ranging from [0,1]. Through a linear transformation formula: Transformed sentiment score = 2 × Original sentiment score - 1, the raw sentiment score is mapped to the interval [-1,1] to obtain the sentiment score corresponding to a single dialogue. A negative sentiment score represents negative sentiment, a sentiment score of 0 represents neutral sentiment, and a sentiment score of positive sentiment represents positive sentiment. After obtaining the sentiment score for each message, the overall sentiment index of the current dialogue window is calculated, i.e.: Average sentiment score: ; Variance of mood fluctuation: ; The overall emotional state of the current conversation is determined based on the average emotion score and the variance of emotion fluctuation. For example, if the average emotion score is... If the average emotional score is less than -0.3 and the variance of emotional fluctuation is greater than or equal to 0.15, then the current conversation is considered to be in a "conflict escalation phase"; if ... If the average emotional score is less than -0.3 and the variance of emotional fluctuation is less than 0.15, the conversation is considered to be in a "suppressed and negative period"; if the average emotional score is less than or equal to -0.3, the conversation is considered to be in a "suppressed and negative period". If the average emotional score is ≤0.5 and the variance of emotional fluctuation is ≥0.15, then the current conversation is considered to be in an "emotional turbulence period"; if the average emotional score is -0.3 ≤ If the average emotional score is ≤0.5 and the variance of emotional fluctuation is <0.15, then the current conversation is considered to be in a "stable and neutral period"; if the average emotional score is ≤0.5, then the current conversation is considered to be in a "stable and neutral period". If the average emotional score is greater than 0.5 and the variance of emotional fluctuation is less than 0.15, then the current conversation is considered to be in a "harmonious period"; if the average emotional score is less than 0.5, then the conversation is considered to be in a "harmonious period". If the value is greater than 0.5 and the variance of emotional fluctuation is greater than or equal to 0.15, then the current conversation is considered to be in a "hot and active period".

[0024] In this application, during the process of determining the current interaction stage of a dialogue, each round of dialogue corresponds to a set of multi-dimensional observation features; the interaction stage is a hidden state that cannot be directly read from the dialogue text. A Gaussian mixture model is used to fit the emission probability of the observed features at each interaction stage, i.e., the probability of that set of observed features appearing given a certain interaction stage; a Hidden Markov Model represents the state transition probability of interaction stages between adjacent dialogue rounds; after obtaining the emission probability and state transition probability, a Viterbi decoding algorithm is used to solve for the optimal path in the temporal observation feature sequence, outputting the most likely interaction stage of the current dialogue. The interaction stages include the initial statement stage, the appeal expression stage, and the conflict escalation stage.

[0025] The core parameters of the hidden Markov model in this application include: The number of hidden states K=6, corresponding to: S1 for the opening, S2 for the statement of the problem, S3 for the emotional outburst and / or stalemate, S4 for clarifying the needs, S5 for discussing solutions, and S6 for the closing.

[0026] The initial probability vector π = [0.9, 0.05, 0.02, 0.01, 0.01, 0.01] (90% from the start).

[0027] The state transition matrix is ​​determined by: analyzing 2,000 labeled real dialogues (covering four scenarios), counting the frequency of transitions from each stage to the next, and then using maximum likelihood estimation to obtain the transition probabilities. For example: S1→S1(0.3), S1→S2(0.7); S2→S2(0.4), S2→S3(0.5), S2→S4(0.1); S3→S3(0.5), S3→S4(0.4), S3→S6(0.1); S4→S4(0.4), S4→S5(0.6); S5→S5(0.5), S5→S6(0.5); It should be noted that reverse jumps are prohibited, such as returning from S4 to S2.

[0028] The emission probability matrix B is fitted with the features of each round of dialogue (round length, emotion change, keyword hit count) using a Gaussian mixture model (3 Gaussian components per state). All parameters are learned on labeled data using the Baum-Welch algorithm, and the current interaction stage is decoded using Viterbi during actual prediction.

[0029] The scene recognition results, the overall emotional state of the current dialogue, and the interaction stage of each round of dialogue are packaged to generate structured context data.

[0030] Step S103: The BERT multi-task regression model and keyword regularization rule engine are used to perform dependency parsing on the structured context data to obtain multi-dimensional problem localization.

[0031] In the example embodiment, the BERT multi-task regression model includes a bottom-level BERT semantic encoding layer and three parallel and independent regression output heads in the upper layer; the dialogue text after dependency parsing is input to the BERT semantic encoding layer; the BERT semantic encoding layer encodes the semantics of the text and outputs the global semantic feature vector corresponding to the CLS position; the global semantic feature vector is sent to the three regression output heads respectively; the three regression output heads are operated independently to predict the execution end score, demand end score, and relation end score respectively.

[0032] The keyword regular expression rule engine includes a dimension-keyword dictionary, a regular expression rule library, and a score mapping module. The dimension-keyword dictionary is divided into three independent sub-dictionaries: an execution-side keyword library, a demand-side keyword library, and a relationship-side keyword library. Each dictionary stores feature words for the corresponding dimension. The regular expression rule library stores preset sentence matching rules to identify typical discourse patterns such as accusations, demands, opposition, and rehashing old grievances. The score mapping module counts the number of times the text hits keywords and regular expression rules for each dimension, and calculates the rule scores for the execution, demand, and relationship sides according to pre-designed scoring rules. Furthermore, based on the predicted distribution of rule scores for the execution, demand, and relationship dimensions, the model prediction confidence corresponding to the current prediction result is calculated. The judgment method of the keyword regular expression rule engine is shown in Table 1 below.

[0033] Table 1

[0034] Specifically, the dialogue text after dependency parsing is fed into the keyword regular expression rule engine; the keyword regular expression rule engine scans the dialogue text in sequence, matching the keyword dictionary and regular expression rules corresponding to the execution end, demand end, and relationship end respectively; the number of hits in each dimension is counted, and the rule scores for the execution end, demand end, and relationship end are calculated based on the number of hits.

[0035] To perform weighted fusion with the model score output by the BERT multi-task regression model, the range of rule scores for each dimension obtained by the keyword regularization rule engine is limited to [-1, 0], which is consistent with the score range of the BERT multi-task regression model and facilitates weighted calculation. For example, the scoring logic is: +0.1 for hitting a basic keyword; +0.3 for hitting a high-weight conflict regularization sentence; and the upper limit of the single-dimensional score is 1.

[0036] Optionally, step S103 may include: performing syntactic parsing on each dialogue statement in the structured context data, inputting the parsed dialogue text into the BERT multi-task regression model, and simultaneously outputting the predicted scores for the execution side, demand side, and relationship side; starting the keyword regularization rule engine to match the accusatory sentences, implicit demands, and historical contradiction-type feature words and phrases corresponding to the execution side, demand side, and relationship side, and obtaining the rule scores for each dimension respectively; if the confidence level of the predicted score output by the BERT multi-task regression model is higher than a preset threshold, the predicted score and the rule scores for each dimension are weighted and fused to obtain the final problem score for each dimension; if the confidence level of the predicted score output by the BERT multi-task regression model is lower than a preset threshold, the rule scores for each dimension are used as the final problem scores for each dimension.

[0037] For example, if the confidence level is ≥0.6, then the final question score for each dimension = predicted score × 0.7 + rule score for each dimension × 0.3; if the confidence level is <0.6, then the final question score for each dimension = predicted score × 0.3 + rule score for each dimension × 0.7.

[0038] The BERT multi-task regression model possesses deep semantic understanding capabilities, enabling it to uncover implicit or unspoken communication contradictions in dialogue texts. The keyword regularization engine outputs highly interpretable results, serving to verify and correct the predictions of the BERT multi-task regression model. By dynamically allocating weights based on the model's prediction confidence, it balances semantic generalization and recognition stability, thereby improving the accuracy of multi-dimensional problem localization results.

[0039] Step S104: Analyze the root causes of multidimensional problem localization from four dimensions: behavioral motivation, object characteristics, interaction habits, and external environment, and obtain a multidimensional root cause set.

[0040] Optionally, step S104 may include: extracting communication bottlenecks, unmet needs, and accumulated conflict labels from multi-dimensional problem localization as root cause inference input fragments; inferring and matching potential root causes from four dimensions—behavioral motivation, object characteristics, interaction habits, and external environment—based on the root cause inference input fragments, and outputting the inference confidence score for each root cause; calling a preset keyword regularization rule library to supplement and verify the model inference results, and for categories with model confidence scores lower than a preset threshold, matching corresponding feature statements to supplement and generate root cause entries; sorting all root cause entries in descending order according to the inference confidence scores, and selecting the top-ranked root causes as valid root causes; valid root causes include their respective dimensions, textual descriptions, and confidence scores.

[0041] Understandably, the multi-dimensional root cause set in this application includes four dimensions: behavioral motivation, object characteristics, interaction habits, and external environment. Based on the multi-dimensional problem localization results, it further explores the deeper, underlying causes of communication conflicts. First, structured label information is extracted from the multi-dimensional problem localization results to form a root cause inference input fragment. This input fragment is then fed into a BERT causal inference model fine-tuned from an external common sense knowledge base. The model infers potential root causes and their corresponding inference confidence scores from the four preset dimensions: behavioral motivation, object characteristics, interaction habits, and external environment. Next, an independent keyword regularization rule library is used for fallback verification. When the model's inference confidence score in a certain dimension falls below a preset threshold, feature statement matching is used to supplement and generate root cause entries, compensating for the shortcomings of the deep learning model's inference inaccuracies. All root cause entries are uniformly arranged in descending order of inference confidence score, and several entries with higher confidence scores are selected as valid root causes. Each valid root cause records its corresponding dimension, textual description, and inference confidence score, ultimately forming a multi-dimensional root cause set.

[0042] Step S105: Based on the identified scenario, emotional stage, and multi-dimensional root cause set, generate an instant response script.

[0043] Optionally, step S105 may include: extracting the current scene tags, the emotional stage of the dialogue, and the question type corresponding to the effective root cause to form index retrieval conditions; searching the preset dialogue template library based on the index retrieval conditions; if a matching template is found, extracting the slot placeholders in the template; extracting entity information from the dialogue context and root cause description to fill the template slots and generate preliminary dialogue content; if no matching template is found, loading standardized prompt words, inputting the scene, emotional stage, root cause, and dialogue text into the large language model to dynamically generate multiple candidate dialogues; retrieving the taboo word library corresponding to the current scene, screening and eliminating illegal expressions in the preliminary dialogue content or multiple candidate dialogues, and generating an immediate response dialogue.

[0044] By extracting current scene tags, dialogue emotional stages, and the question types corresponding to effective root causes to form index retrieval conditions, a preset dialogue template library can be retrieved based on multi-dimensional conditions. When a matching template is found, placeholders in the template slots are extracted, and entity information is extracted from the dialogue context and root cause descriptions to fill the template slots and generate preliminary dialogue content, realizing the instantiation of template dialogue output. When no matching template is found, standardized prompt words are loaded, and the scene, emotional stage, root cause, and dialogue text are input into the large language model to dynamically generate multiple candidate dialogues, realizing the backup dialogue generation in the case where the template is not covered. The taboo word library corresponding to the current scene is retrieved to screen and remove illegal expressions from the preliminary dialogue content or multiple candidate dialogues. The output of the template path and the large model generation path are uniformly filtered for violations, and finally, an instant response dialogue is generated, realizing the combination of template retrieval generation and large model dynamic generation, and the response output is screened by the taboo word library.

[0045] In other embodiments, the method may further include steps S201 to S202, wherein: Step S201: Based on the two root causes with the highest confidence, match the root cause-suggestion mapping template to generate structured long-term improvement suggestions that include daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content.

[0046] Optionally, step S201 may include: sorting the multi-dimensional root cause set in descending order of confidence level, and extracting the top two root causes; reading the pre-configured root cause-suggestion mapping template library, and retrieving corresponding improvement suggestion template entries to obtain the root cause-suggestion mapping template based on the root cause type and root cause description of the two extracted root causes; extracting contextual information related to people, events, and business entities from the dialogue text and root cause descriptions; substituting the contextual information into the root cause-suggestion mapping template to generate daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content; and assigning corresponding type tags to the daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content to assemble structured long-term improvement suggestions.

[0047] Step S202: Add explicit identifiers and implicit source metadata to the instant response scripts and structured long-term improvement suggestions to obtain historical instant response scripts and historical structured long-term improvement suggestions with content traceability.

[0048] Implementing steps S101 to S105 above, structured contextual data is obtained through scene recognition, emotion temperature measurement, and interaction stage determination using dialogue audio or text. This provides unified and complete basic dialogue information for the entire process analysis, reducing the context loss problem caused by switching between multiple tools. Relying on dependency parsing combined with the BERT multi-task regression model and keyword regularization rule engine, multi-dimensional problem localization is achieved, balancing the model's generalization reasoning ability with the rule's precise fallback advantage. It can identify various communication bottlenecks at the execution, demand, and relationship levels. Furthermore, it analyzes behavioral motivation, object characteristics, and interaction habits... This approach employs a multi-dimensional approach to root cause analysis, exploring four key dimensions: external environment, and internal environment, to penetrate surface-level communication conflicts and pinpoint deeper underlying causes. Finally, it generates targeted, real-time response scripts by combining scenario, emotional stage, and multi-dimensional root cause set. The entire process eliminates the need for manual data transfer between software programs, improving dialogue analysis efficiency. Furthermore, leveraging multi-model fusion and multi-dimensional layered reasoning mechanisms, the analysis results closely align with real-world dialogue contexts, and the output response scripts are highly adapted to the current communication scenario and the root cause of the conflict. This addresses the shortcomings of existing intelligent communication assistance tools, such as fragmented functionality, one-sided analysis, and empty, unrealistic output solutions, thereby enhancing the timeliness and practicality of dialogue assistance.

[0049] Based on the same inventive concept, this application also provides an AI communication assistance device adapted to multiple scenarios for implementing the aforementioned AI communication assistance method adapted to multiple scenarios. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the AI ​​communication assistance device adapted to multiple scenarios provided below can be found in the limitations of the AI ​​communication assistance method adapted to multiple scenarios described above, and will not be repeated here.

[0050] In one exemplary embodiment, such as Figure 2 As shown, an AI communication assistance device adapted to multiple scenarios is provided. The AI ​​communication assistance device 200 adapted to multiple scenarios includes: Module 201 is used to acquire dialogue audio or import dialogue text. Processing module 202 is used to perform scene recognition, emotion temperature calculation and interaction stage determination on dialogue audio or dialogue text, and obtain structured context data containing scene, emotion and interaction stage. The localization module 203 is used to perform dependency parsing on structured context data using the BERT multi-task regression model and keyword regularization rule engine to obtain multi-dimensional problem localization. Analysis module 204 is used to analyze the root causes of multi-dimensional problem localization from four dimensions: behavioral motivation, object characteristics, interaction habits, and external environment, and obtain a multi-dimensional root cause set; The generation module 205 is used to generate instant response scripts based on the identified scenario, emotional stage, and multi-dimensional root cause set.

[0051] As one possible implementation, the processing module 202 is specifically used to: transcribe the acquired dialogue audio into a first standardized dialogue text; perform cleaning and preprocessing on the dialogue text to obtain a second standardized dialogue text; input the first standardized dialogue text and / or the second standardized dialogue text into a pre-trained scene classification model, outputting the confidence score corresponding to each candidate scene; select the scene with the highest confidence score above the confidence threshold as the recognition result, and mark all scenes with confidence scores below the threshold as unknown scenes; use a sentiment analysis model to calculate the sentiment score of each dialogue sentence by sentence, calculate the average sentiment score and sentiment fluctuation variance based on all sentiment scores within the dialogue window, and combine the average sentiment score and sentiment fluctuation variance to obtain the overall sentiment state of the current dialogue; extract multi-dimensional observation features of each round of dialogue to form a temporal observation feature sequence; fit the emission probability of the observation feature vector in the hidden state of each interaction stage using a Gaussian mixture model; obtain the state transition probability between interaction stages using a hidden Markov model; and use the Viterbi decoding algorithm to solve for the optimal hidden state path based on the emission probability and state transition probability of the temporal observation feature sequence to obtain the interaction stage of each round of dialogue.

[0052] As one possible implementation, the aforementioned localization module 203 is specifically used to: perform syntactic parsing on each dialogue statement in the structured context data; input the syntactically parsed dialogue text into the BERT multi-task regression model; and simultaneously output predicted scores for the execution end, demand end, and relationship end. It then activates a keyword regularization rule engine to match accusatory phrases, implicit demands, and historical contradiction-type feature words and phrases corresponding to the execution end, demand end, and relationship end, obtaining rule scores for each dimension. If the confidence level of the predicted score output by the BERT multi-task regression model is higher than a preset threshold, the predicted score is weighted and fused with the rule scores for each dimension to obtain the final problem score for each dimension. If the confidence level of the predicted score output by the BERT multi-task regression model is lower than a preset threshold, the rule scores for each dimension are used as the final problem scores for each dimension.

[0053] As one possible implementation, the analysis module 204 is specifically used to: extract communication bottlenecks, unmet needs, and accumulated conflict labels from multi-dimensional problem localization as root cause inference input fragments; infer and match potential root causes from four dimensions: behavioral motivation, object characteristics, interaction habits, and external environment, and output the inference confidence score corresponding to each root cause; call a preset keyword regularization rule library to supplement and verify the model inference results, and for categories with model confidence scores lower than a preset threshold, match corresponding feature statements to supplement and generate root cause entries; sort all root cause entries in descending order according to inference confidence scores, and select the top-ranked root causes as valid root causes; valid root causes include their respective dimensions, textual descriptions, and confidence score information.

[0054] As one possible implementation, the aforementioned generation module 205 is specifically used to: extract the current scene tags, the emotional stage of the dialogue, and the question type corresponding to the effective root cause to form index retrieval conditions; search the preset dialogue template library based on the index retrieval conditions; if a matching template is found, extract the slot placeholders within the template; extract entity information from the dialogue context and root cause description to fill the template slots and generate preliminary dialogue content; if no matching template is found, load standardized prompt words, input the scene, emotional stage, root cause, and dialogue text into the large language model to dynamically generate multiple candidate dialogues; retrieve the taboo word library corresponding to the current scene, screen and eliminate illegal expressions from the preliminary dialogue content or multiple candidate dialogues, and generate an immediate response dialogue.

[0055] As one possible implementation, the aforementioned AI-powered communication assistance device 200, adaptable to multiple scenarios, also includes a suggestion and tracing module. The suggestion and tracing module is used to: match the root cause-suggestion mapping template based on the two root causes with the highest confidence, generate structured long-term improvement suggestions that include daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content; and add explicit identifiers and implicit tracing metadata to the instant response scripts and structured long-term improvement suggestions to obtain historical instant response scripts and historical structured long-term improvement suggestions with content traceability.

[0056] As one possible implementation, the aforementioned suggestion and tracing module is specifically used to: sort the multi-dimensional root cause set in descending order of confidence level, and extract the top two root causes; read the pre-configured root cause-suggestion mapping template library, and retrieve the corresponding improvement suggestion template entries to obtain the root cause-suggestion mapping template based on the root cause type and root cause description of the two extracted root causes; extract contextual information related to people, events, and business entities from the dialogue text and root cause descriptions; substitute the contextual information into the root cause-suggestion mapping template to generate daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content; and assign corresponding type tags to the daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content to assemble them into structured long-term improvement suggestions.

[0057] This approach utilizes dialogue audio or text for scene recognition, emotion temperature measurement, and interaction stage determination to obtain structured contextual data. This provides unified and complete basic dialogue information for the entire process analysis, reducing context loss caused by switching between multiple tools. It leverages dependency parsing combined with BERT multi-task regression models and keyword regularization rule engines to achieve multi-dimensional problem localization, balancing the model's generalization reasoning ability with the rule's precise fallback advantage. This allows for the identification of various communication bottlenecks at the execution, demand, and relationship levels. Furthermore, it conducts layered root cause mining from four dimensions: behavioral motivation, object characteristics, interaction habits, and external environment, penetrating surface communication conflicts to pinpoint deeper causes. Finally, it generates targeted, real-time response scripts based on the scene, emotion stage, and multi-dimensional root cause set. The entire process eliminates the need for manual data transfer across software, improving dialogue analysis efficiency. Simultaneously, relying on multi-model fusion and multi-dimensional layered reasoning mechanisms, the analysis results closely match real-world dialogue contexts, and the output response scripts are highly adapted to the current communication scenario and the root cause of conflict. This addresses the shortcomings of existing intelligent communication assistance tools, such as fragmented functionality, one-sided analysis, and empty, unrealistic output solutions, thus improving the timeliness and practicality of dialogue assistance.

[0058] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 3As shown, the computer device includes a processor, memory, input and / or output interfaces (I and / or O), and a communication interface. The processor, memory, and I and / or O interfaces are connected via a device bus, and the communication interface is connected to the device bus via the I and / or O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores operating devices, computer programs, and a database. The internal memory provides an environment for the operation of the operating devices and computer programs stored in the non-volatile storage media. The database stores AI-powered communication assistance data adapted to various scenarios. The I and / or O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements an AI-powered communication assistance method adapted to various scenarios.

[0059] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0060] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0061] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0062] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0063] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0064] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0065] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0066] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0067] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. An AI-assisted communication method adaptable to multiple scenarios, characterized in that, The AI-assisted communication method adapted to multiple scenarios includes: Get the audio of the conversation or import the text of the conversation; Scene recognition, emotion temperature calculation, and interaction stage determination are performed on the dialogue audio or dialogue text to obtain structured context data containing scene, emotion, and interaction stage. The BERT multi-task regression model and keyword regularization rule engine are used to perform dependency parsing on the structured context data to obtain multi-dimensional problem localization. The root causes of the multidimensional problem are analyzed from four dimensions: behavioral motivation, object characteristics, interaction habits, and external environment, resulting in a multidimensional root cause set. Based on the identified scenario, emotional stage, and the multi-dimensional root cause set, an instant response script is generated.

2. The AI-assisted communication method adaptable to multiple scenarios according to claim 1, characterized in that, The process of performing scene recognition, emotion temperature calculation, and interaction stage determination on the dialogue audio or dialogue text yields structured context data containing scene, emotion, and interaction stage, including: The acquired dialogue audio is transcribed into the first standardized dialogue text; The dialogue text is cleaned and preprocessed to obtain the second standardized dialogue text; Input the first standardized dialogue text and / or the second standardized dialogue text into the pre-trained scene classification model, output the confidence score corresponding to each candidate scene; and select the scene with the highest confidence score above the confidence score threshold as the recognition result, and mark all scenes with confidence scores below the threshold as unknown scenes; The sentiment analysis model is used to calculate the sentiment score of each sentence in a dialogue. The average sentiment score and the variance of sentiment fluctuation are calculated based on all sentiment scores in the dialogue window. The average sentiment score and the variance of sentiment fluctuation are combined to obtain the overall sentiment state of the current dialogue. Extract multi-dimensional observation features from each round of dialogue to form a time-series observation feature sequence; The emission probability of observed feature vectors in the hidden state of each interaction stage is fitted by a Gaussian mixture model. Obtain the state transition probabilities between interaction stages using a hidden Markov model; Using the Viterbi decoding algorithm, the optimal hidden state path is solved for the time-series observation feature sequence based on the emission probability and state transition probability, thus obtaining the interaction stage of each round of dialogue.

3. The AI-assisted communication method adaptable to multiple scenarios according to claim 2, characterized in that, The BERT multi-task regression model and keyword regularization rule engine are used to perform dependency parsing on the structured context data to obtain multi-dimensional problem localization, including: Syntactically parse each dialogue statement in the structured context data, input the parsed dialogue text into the BERT multi-task regression model, and simultaneously output the prediction scores for the execution end, demand end, and relationship end. The keyword regular expression rule engine is launched to match the accusatory sentences, implicit requirements, and historical contradictions of the execution end, demand end, and relationship end, and obtain the rule scores for each dimension. If the confidence level of the predicted score output by the BERT multi-task regression model is higher than the preset threshold, the predicted score and the rule scores of each dimension are weighted and fused to obtain the final question score for each dimension; if the confidence level of the predicted score output by the BERT multi-task regression model is lower than the preset threshold, the rule scores of each dimension are used as the final question score for each dimension.

4. The AI-assisted communication method adaptable to multiple scenarios according to claim 3, characterized in that, The root causes of the multi-dimensional problem are analyzed from four dimensions: behavioral motivation, object characteristics, interaction habits, and external environment, resulting in a multi-dimensional root cause set, including: Extract the communication bottlenecks, unmet needs, and accumulated conflict tags from the multi-dimensional problem location as input fragments for root cause reasoning; The root cause reasoning input fragments are used to infer and match potential root causes from four dimensions: behavioral motivation, object characteristics, interaction habits, and external environment, and the reasoning confidence of each root cause is output. The system calls a preset keyword regular expression rule library to supplement and verify the model inference results. For categories where the model confidence is lower than a preset threshold, it matches the corresponding feature statements to supplement and generate root cause entries. All root cause entries are sorted in descending order based on inference confidence, and the root causes with the highest confidence are selected as valid root causes; the valid root causes include the dimension to which they belong, the textual description, and the confidence information.

5. The AI-assisted communication method adaptable to multiple scenarios according to claim 4, characterized in that, The generation of immediate response scripts based on the identified scenario, emotional stage, and the multi-dimensional root cause set includes: Extract current scene tags, dialogue emotional stage, and question types corresponding to effective root causes to form index retrieval conditions; Based on the index retrieval conditions, a preset dialogue template library is retrieved. If a matching template is found, the slot placeholders in the template are extracted. Entity information is extracted from the dialogue context and root cause description to fill the template slots and generate preliminary dialogue content. If no matching template is found, standardized prompt words are loaded, and the scene, emotional stage, root cause, and dialogue text are input into the large language model to dynamically generate multiple candidate dialogues. The prohibited word library corresponding to the current scenario is retrieved, and the preliminary wording content or the multiple candidate wordings are screened to remove illegal expressions, thereby generating the instant response wording.

6. The AI-assisted communication method adaptable to multiple scenarios according to claim 5, characterized in that, The method further includes: Based on the two root causes with the highest confidence, a root cause-suggestion mapping template is matched to generate structured long-term improvement suggestions that include daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content. Explicit identifiers and implicit source metadata are added to the instant response scripts and the structured long-term improvement suggestions to obtain historical instant response scripts and historical structured long-term improvement suggestions with content traceability.

7. The AI-assisted communication method adaptable to multiple scenarios according to claim 6, characterized in that, Based on the two root causes with the highest confidence levels, a root cause-suggestion mapping template is matched to generate structured long-term improvement suggestions that include daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content. Sort the multidimensional root cause set in descending order of confidence, and extract the two root causes that appear at the top of the sort; Read the pre-configured root cause-suggestion mapping template library, and retrieve the corresponding improvement suggestion template entries based on the root cause type and root cause description of the two extracted root causes to obtain the root cause-suggestion mapping template; Extract contextual information related to people, events, and business entities from dialogue text and root cause descriptions; Substitute the context information into the root cause-suggestion mapping template to generate daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content, respectively. The daily action tasks, tool usage practice items, relationship repair measures, and knowledge learning content are labeled with corresponding types, and assembled into structured long-term improvement suggestions.

8. An AI-powered communication assistance device adaptable to multiple scenarios, characterized in that, The AI-powered communication assistance device adapted to multiple scenarios: The acquisition module is used to acquire dialogue audio or import dialogue text. The processing module is used to perform scene recognition, emotion temperature calculation and interaction stage determination on the dialogue audio or dialogue text to obtain structured context data containing scene, emotion and interaction stage. The localization module is used to perform dependency parsing on the structured context data using the BERT multi-task regression model and keyword regularization rule engine to obtain multi-dimensional problem localization. The analysis module is used to analyze the root causes of the multi-dimensional problem from four dimensions: behavioral motivation, object characteristics, interaction habits, and external environment, and obtain a multi-dimensional root cause set. The generation module is used to generate instant response scripts based on the identified scenario, emotional stage, and the multi-dimensional root cause set.

9. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the AI ​​communication assistance method adapted to multiple scenarios as described in any one of claims 1-7.

Images