Emotion behavior visual analysis method, system and device based on artificial intelligence

Abstract

The application discloses an emotion behavior visual analysis method, system and device based on artificial intelligence, and particularly relates to the technical field of emotion behavior analysis, and is used for solving the problem that the existing task flow analysis technology cannot quantitatively identify the inertial conduction effect of the operation subject emotion between cross nodes; by acquiring a multi-source asynchronous interaction log containing text content, the emotion dynamic characteristics of the operation subject on the task node are extracted by analyzing the text, the emotion dynamic characteristic sequence is used as observation data, a hidden Markov model is trained to obtain a state transition probability matrix, and then the appearance frequency and state conversion type of each task node in the optimal state transition path are analyzed, the key hub node of the emotion dynamic conduction is identified, the emotion dynamic inertia disturbance intensity of the task node is evaluated according to the state transition probability matrix and the key hub node, and is mapped to a task flow topology diagram and rendered, so that the quantitative analysis and intuitive presentation of the emotion dynamic inertia effect in the task flow are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of emotional behavior analysis, more particularly, the present application relates to an emotional behavior visualization analysis method, system and device based on artificial intelligence. BACKGROUND

[0002] In the field of collaboration process and task execution analysis, log analysis based on information systems is a key technical means to optimize execution efficiency and identify collaboration bottlenecks. Existing technologies mainly collect and analyze interaction logs generated by various collaboration platforms, teaching management systems or project collaboration tools, model, simulate and visualize task processes, focus on the timing, time consumption, path compliance and resource consumption of task activities, and the core goal is to find process bottlenecks, improve automation level and ensure system compliance. Whether it is a rule-based analysis engine or an application of machine learning for abnormal pattern recognition, the observation dimension of existing solutions is fixed on the behavior records of the operation entity itself.

[0003] However, the quality of any collaboration or task execution process is ultimately driven by the participants, and their psychological state, especially the emotional factors that persist, has a potential and important influence on judgment and selection. The existing technical system cannot identify and quantify the continuous influence of the subjective emotional state triggered by the previous event on the subsequent associated or unrelated task decision nodes in the cross-task and cross-link scenario, i.e. the emotional inertia effect, which makes the current analysis results have systematic bias, can only reflect what happened but cannot explain why inefficient decisions or risks will appear here, resulting in that the analysis conclusion stays on the surface in supporting process optimization and behavior intervention. SUMMARY

[0004] In order to overcome the above-mentioned defects of the prior art, the present application provides an emotional behavior visualization analysis method, system and device based on artificial intelligence to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0006] The emotional behavior visualization analysis method based on artificial intelligence comprises:

[0007] S1, acquiring a multi-source asynchronous interaction log set of a target task process, the interaction log comprising a timestamp, an operation subject identifier, a text content and a task node identifier;

[0008] S2, analyzing the text content of the interaction log and extracting emotional dynamic features of the operation subject on the task node;

[0009] S3, taking the emotional dynamic feature of the operation subject on the task node sequence as an observation sequence, and training a state transition probability matrix through a hidden Markov model;

[0010] S4, analyzing the occurrence frequency and state conversion type of each task node in the optimal state transition path based on the state transition probability matrix, and identifying a key hub node of emotional dynamic conduction;

[0011] S5, evaluating the emotional dynamic inertia disturbance intensity of each task node according to the state transition probability matrix and the key hub node;

[0012] S6, mapping the emotional dynamic inertia disturbance intensity of the task node to a task flow topology graph and visualizing rendering.

[0013] Further, S1 includes:

[0014] According to the definition of the target task flow, raw logs are collected from the operation history of the project management system and the collaboration communication platform participating in the target task flow;

[0015] The collected raw logs are cleaned, and aligned and merged according to the unified timestamp, operation subject identifier, text content and task node identifier, to generate a multi-source asynchronous interaction log set.

[0016] Further, S2 includes:

[0017] The text content of the interaction log is segmented and tagged with parts of speech, and the emotional state expressed in the text is identified based on pre-constructed sentiment analysis rules;

[0018] In combination with the context semantics of the emotional state and the task node associated with it in the process, the direction and strength of the emotional state on the task flow promotion or blockage are determined;

[0019] The direction and strength are comprehensively quantified to generate a quantized vector as an emotional dynamic feature.

[0020] Further, S3 includes:

[0021] The emotional dynamic feature quantized vector of the operation subject arranged in time sequence is organized as an observation sequence;

[0022] The number of hidden states of the hidden Markov model is set, and the state transition probability matrix and the observation probability matrix are initialized;

[0023] The hidden Markov model is iteratively trained using the forward-backward algorithm to optimize the parameters of the state transition probability matrix and the observation probability matrix until the model converges, obtaining the trained state transition probability matrix.

[0024] Further, S4 includes:

[0025] For each operation subject, the emotion motive feature observation sequence is decoded based on the state transition probability matrix using the Viterbi algorithm to generate a corresponding optimal state transition path;

[0026] The total frequency of each task node appearing in the optimal state transition path of all operation subjects is counted, and the hidden state transition type occurring at the corresponding task node in the path is recorded;

[0027] Based on the appearance frequency and state transition type of the task node, a comprehensive feature value representing its conduction importance is calculated;

[0028] According to the comprehensive feature value, the task nodes are sorted, and the pre-set number of task nodes sorted in the front are identified as key nodes of emotion motive conduction.

[0029] Further, based on the appearance frequency and state transition type of the task node, the comprehensive feature value representing its conduction importance is calculated, including: normalizing the total frequency of the task node appearing in all optimal state transition paths to obtain a standardized frequency; counting the types of different hidden state transition types occurring at the task node, and calculating the proportion of each transition type in the total transition events; the standardized frequency and the number of state transition types and the proportion of main transition types are weighted and summed to generate a comprehensive feature value for sorting.

[0030] Further, S5 includes:

[0031] Based on the state transition probability matrix, the state residence expectation of each task node in all possible state transition paths of the hidden Markov model is calculated, and the state residence expectation is combined with the observation emotion motive feature intensity associated with the task node itself to generate a basic strength representing the emotion motive inertia of the corresponding task node itself;

[0032] According to the high-probability outward transition characteristics of the key node represented in the state transition probability matrix, the radiation range and strength attenuation coefficient of its emotion motive influence are calculated, and according to the connection relationship between the task nodes and the radiation range in the task flow topological graph, the emotion motive conduction strength of the key node to the task nodes is calculated;

[0033] The basic strength of the task node itself and the emotion motive conduction strength received from one or more key nodes are weighted and superimposed, wherein the weight is dynamically adjusted according to the order of the task node being accessed in the optimal state transition path, to generate the final emotion motive inertia disturbance strength of the task node.

[0034] Further, S6 includes:

[0035] Obtain a task flow topology graph representing a logical connection relationship between task nodes.

[0036] Convert the emotional dynamic inertia disturbance intensity value of the task node into a corresponding visual attribute value according to a preset segmentation mapping rule;

[0037] Render the corresponding task node graph element in the task flow topology graph based on the visual attribute value, and generate a color gradient or size gradient visualization layer according to the difference in the emotional dynamic inertia disturbance intensity value.

[0038] In another aspect, the present application provides an emotional behavior visualization analysis system based on artificial intelligence, comprising:

[0039] A log acquisition module is configured to acquire a multi-source asynchronous interaction log set of a target task flow, and the interaction log includes a timestamp, an operation subject identifier, text content, and a task node identifier;

[0040] A feature extraction module is configured to analyze the text content of the interaction log and extract emotional dynamic features of the operation subject on the task node;

[0041] A matrix training module is configured to use the emotional dynamic features of the operation subject on the task node sequence as an observation sequence, and train a state transition probability matrix through a hidden Markov model;

[0042] A node identification module is configured to analyze the occurrence frequency and state conversion type of each task node in the optimal state transition path based on the state transition probability matrix, and identify a key hub node of emotional dynamic conduction;

[0043] An intensity evaluation module is configured to evaluate the emotional dynamic inertia disturbance intensity of each task node according to the state transition probability matrix and the key hub node;

[0044] A visual rendering module is configured to map the emotional dynamic inertia disturbance intensity of the task node to the task flow topology graph and perform visual rendering.

[0045] In another aspect, the present application provides an emotional behavior visualization analysis device based on artificial intelligence, comprising:

[0046] One or more processors;

[0047] A storage device configured to store one or more programs;

[0048] When the one or more programs are executed by the one or more processors, the one or more processors implement an emotional behavior visualization analysis method based on artificial intelligence.

[0049] Compared with the prior art, the present application has the following beneficial effects:

[0050] 1. By introducing emotion analysis on the content of interaction logs, traditional process analysis is upgraded from simple behavior tracking to quantitative modeling of the psychological driving force of the operation subject. Through the extracted emotional driving features and the constructed hidden Markov model and inertia disturbance evaluation model, the emotional conduction inertia effect across task nodes is captured and measured, enabling process analysis to explain the deep subjective cause of low efficiency or decision risk at a specific task node, thus pushing the analysis conclusion from describing what happened to explaining why it happened, providing accurate intervention targets for subsequent behavior guidance and process optimization.

[0051] 2. A process visualization layer reflecting the emotional driving disturbance intensity is generated, providing a direct decision support tool for collaboration management and task execution optimization, which can intuitively identify key hub nodes of emotional inertia accumulation or conduction. These nodes are often potential risk areas or efficiency bottleneck areas that are ignored by traditional analysis. By taking appropriate guidance or process intervention measures for these nodes, decision bias or collaboration blockage caused by negative emotional inertia can be effectively alleviated, thereby improving the robustness and execution quality of the overall collaboration process, realizing a closed loop from deep perception to intelligent decision-making. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 Flowchart of the emotion behavior visualization analysis method based on artificial intelligence of the present application;

[0053] Figure 2 Structure diagram of the emotion behavior visualization analysis system based on artificial intelligence of the present application. DETAILED DESCRIPTION

[0054] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the present application.

[0055] Embodiment 1: Figure 1 The emotion behavior visualization analysis method based on artificial intelligence of the present application is given, including:

[0056] S1, a multi-source asynchronous interaction log set of a target task process is obtained, and the interaction log includes a timestamp, an operation subject identifier, text content and a task node identifier;

[0057] S2, the text content of the interaction log is analyzed and the emotional driving features of the operation subject at the task node are extracted;

[0058] S3, taking the emotional dynamic characteristics of the operation subject on the task node sequence as an observation sequence, and training a state transition probability matrix through a hidden Markov model;

[0059] S4, analyzing the occurrence frequency and state conversion type of each task node in the optimal state transition path based on the state transition probability matrix, and identifying a key hub node of emotional dynamic conduction;

[0060] S5, evaluating the emotional dynamic inertia disturbance intensity of each task node according to the state transition probability matrix and the key hub node;

[0061] S6, mapping the emotional dynamic inertia disturbance intensity of the task node to a task flow topology graph and visualizing rendering.

[0062] S1, obtaining a multi-source asynchronous interaction log set of a target task flow, the interaction log including a timestamp, an operation subject identifier, text content, and a task node identifier, and the implementation is as follows:

[0063] First, according to the definition of the target task flow, the original logs are collected from the operation history of the project management system and the collaboration communication platform participating in the target task flow. The definition of the target task flow refers to, before the analysis is implemented, a specific work flow composed of a series of ordered task activities is determined according to the task or collaboration management rules, the work flow has a clear starting node, a terminal node and intermediate processing nodes, and each node corresponds to a unique task node identifier. After the target task flow is determined, the software systems participating in the execution of the flow are identified and confirmed. Generally, the collaboration communication platform that records task communication and collaboration records, and the project management system that records task state changes and flow records are determined as the systems that need to collect logs. Collecting original logs from the project management system means accessing the operation audit table or transaction log table in the database of the system, extracting all data rows with the task or activity type code field matching the target task flow code within the preset analysis time range through structured query language, and each row of data constitutes an original log. Collecting original logs from the collaboration communication platform means calling the application programming interface provided by the system or directly querying the service record database table, extracting entries containing the target task flow or related task subject keywords in the record content summary field within the same preset analysis time range, and each entry constitutes an original log. The preset analysis time range is set by the implementer according to historical data analysis requirements, for example, it can be set to the past twelve months.

[0064] Next, the collected original logs are cleaned and aligned and merged according to unified timestamps, operation subject identifiers, text contents, and task node identifiers to generate a multi-source asynchronous interaction log set. Cleaning the collected original logs specifically refers to performing data preprocessing operations on the original log entries obtained from two independent systems respectively. The cleaning operation mainly includes removing duplicate records, filling in missing values of key fields, and correcting format errors. Removing duplicate records is achieved by comparing the contents of all fields of the log entries, and only the first entry is retained for completely identical entries. Filling in missing values of key fields, for example, when a log is missing the timestamp of the operation, the timestamps of the adjacent records before and after the log in the original data table are parsed, and a linear interpolation method is used to estimate and fill in the missing values. Correcting format errors, for example, converting various date and time representations in the timestamp field to the year-month-day hour-minute-second format in coordinated universal time. After completing the preliminary cleaning, the alignment and merging operation is performed. According to the unified timestamp, it refers to converting the time information in all original logs to the same time zone reference timestamp representation, and unifying their precision, for example, converting them all to coordinated universal time and unifying their precision to seconds. According to the unified operation subject identifier, it refers to establishing a cross-system user identity mapping table. The mapping table records the correspondence between different system user accounts and a unified operation subject identifier. By querying this table, the user identifier from the project management system and the account identifier from the collaboration communication platform are both mapped to the same unified operation subject identifier, thereby generating a unified operation subject identifier. According to the text content, it refers to retaining or generating text information describing the operation behavior. If the original log contains a description field, its content is directly retained; if not, a piece of text content is generated according to the operation type name and the task or entity name involved. According to the task node identifier, it refers to associating each original log to a specific node of the target task process according to its specific operation description through rule matching. Rule matching is based on a predefined dictionary that associates a set of feature keywords and operation modes to each task node identifier. For example, when the text content contains keywords such as task help and delay and the operation type is creating a task, the task node identifier of this log is assigned to A001. The construction of this dictionary can be based on statistical analysis of historical logs or domain expert knowledge.

[0065] After the four dimensions are unified, all log entries from different systems are globally sorted in ascending order according to the unified time stamps. After sorting, multiple log entries belonging to the same operation subject on the same task node and occurring continuously in time are merged. The merging condition is that the operation subject identifiers of the continuous entries are the same, the task node identifiers are the same, and the time stamp difference between adjacent entries is less than a preset merging time window threshold. The merging time window threshold is set according to the average processing time of the task flow link, for example, it can be set to five minutes. When the merging condition is met, the text contents of these entries are connected into a paragraph, the earliest time stamp is kept as the time stamp of the merged record, and the operation subject identifier and the task node identifier remain unchanged, thereby forming a more complete interaction record. Finally, all the records processed by cleaning, alignment and merging are arranged in chronological order to form a structured multi-source asynchronous interaction log set. Each interaction log in the set strictly contains four data fields of time stamp, operation subject identifier, text content and task node identifier, providing a standardized data basis for subsequent analysis.

[0066] S2, analyze the text content of the interaction log and extract the emotional and dynamic characteristics of the operation subject on the task node, which is specifically implemented as:

[0067] First, the text content of the interaction log is segmented and tagged with parts of speech, and the emotional state expressed in the text is identified based on pre-built sentiment analysis rules. The text content of the interaction log operated here is directly from the text content field in the multi-source asynchronous interaction log set obtained through the aforementioned data preparation steps. In implementation, mature segmentation and part-of-speech tagging techniques are used to process the text content, specifically by calling the corresponding function functions in the trained Chinese language processing library. The function takes a complete text content as input, and its internal running mechanism is based on statistical models and dictionary matching to automatically cut continuous character sequences into independent, semantically meaningful word units, and tag each word unit with its grammatical category, such as noun, verb, or adjective. Subsequently, the results of the segmentation are identified based on pre-built sentiment analysis rules. The core components of the pre-built sentiment analysis rules are a domain-enhanced sentiment dictionary and a set of emotion score calculation and mapping rules. The domain-enhanced sentiment dictionary is formed by supplementing a large number of professional terms and their sentiment attributes closely related to the target task process scene based on a general sentiment dictionary through manual review and annotation. Each entry in the dictionary contains at least the word itself, sentiment polarity classification, and a basic sentiment intensity score. Sentiment polarity classification includes positive, negative, and neutral. The basic sentiment intensity score is an integer that quantifies the emotional energy contained in the word in isolation. Its numerical range can be pre-set, for example, the basic sentiment intensity score of a positive word is positive, and that of a negative word is negative. For example, the word "satisfied" can be tagged as positive polarity and assigned a basic sentiment intensity score of +2, and the word "fault" can be tagged as negative polarity and assigned a basic sentiment intensity score of -2. The emotion score calculation and mapping rules define how to calculate a comprehensive sentiment tendency score from a piece of text and classify the score into a specific emotional state. The calculation process starts with traversing all the word units in the text and querying the domain-enhanced sentiment dictionary for matching. The basic sentiment intensity scores of all matching words are algebraically added to obtain an initial sentiment total score. Then, the initial sentiment total score is corrected according to the modifying words in the text. The correction rules include: when a negation word such as "not" or "failed" is identified, the sign of the first sentiment word after it is reversed, i.e. positive to negative and negative to positive; when a degree adverb such as "very" or "slightly" is identified, the score of the modified sentiment word is multiplied by a corresponding intensity adjustment coefficient, which is pre-set according to the strength of the adverb, for example, the intensity adjustment coefficient corresponding to "very" is 1.5, and the intensity adjustment coefficient corresponding to "slightly" is 0.5. After correction, the final sentiment tendency score is obtained.

[0068] To map the continuous sentiment tendency scores to discrete emotion state labels, emotion state classification thresholds need to be set. Specifically, two thresholds, positive and negative sentiment classification thresholds, are set. The values of these two thresholds are not fixed, and they are determined by performing the above sentiment analysis on a large number of historical interaction logs, calculating the sentiment tendency scores of all texts, and drawing a distribution graph. According to the requirements of task or process analysis, select the appropriate quantile points on the score distribution graph as the thresholds. For example, select the score point that makes a certain proportion of texts classified as positive state as the positive sentiment classification threshold, and select the score point that makes a certain proportion of texts classified as negative state as the negative sentiment classification threshold. After obtaining the thresholds, the mapping rule is: if the sentiment tendency score of a text is higher than the positive sentiment classification threshold, its emotion state label is “positive”; if the score is lower than the negative sentiment classification threshold, the emotion state label is “negative”; if the score is between the negative sentiment classification threshold and the positive sentiment classification threshold, the emotion state label is “neutral”.

[0069] Next, the direction and strength of the effect of the emotion state on the task flow are determined by combining the emotion state with the context semantics of the task node associated with it in the flow. In this step, the emotion state comes from the output of the previous step, and the task node associated with it is uniquely determined by the task node identifier contained in the interaction log record. The context semantics refers to the responsibilities, usual operation actions and position in the flow sequence of the task node in the overall predefined target task flow. The judgment process relies on a pre-configured task influence rule knowledge base. The construction of this knowledge base is based on the summary of expert experience of collaboration or task flow, which takes task node identifier and emotion state label as combined input, and outputs the expected emotion effect direction. The effect direction is a binary classification, which is used to represent whether the emotion produces a net effect of driving the flow forward, i.e. “evolution”, or hindering its progress, i.e. “blockage”. For example, the rule can be defined as: for the task node identifier “result review”, when the emotion state is “positive”, the effect direction is “promotion”; when the emotion state is “negative”, the effect direction is “blockage”.

[0070] The quantitative judgment of the action strength is realized through a calculation model. The first input of the model is the absolute value of the sentiment tendency score obtained in the previous step, which represents the original expression intensity of the emotion. The second input is the influence weight coefficient of the current task node. The influence weight coefficient of each task node reflects the importance of the node in the stable operation of the entire process, and its assignment can be obtained by collecting the opinions of multiple experts in the field (such as instructors and project managers) through the Delphi method and then averaging the opinions. For example, core review or decision nodes in the process may be assigned a higher influence weight coefficient, while general information recording or distribution nodes may be assigned a lower influence weight coefficient. The calculation model multiplies the emotion expression intensity and the node influence weight coefficient to obtain a preliminary action strength value. In order to enable the strength values from different task nodes and different texts to be compared on the same scale, the preliminary action strength value needs to be normalized. Normalization requires a reference benchmark, which is determined by the global maximum and global minimum of the preliminary action strength values obtained by calculating all historical interaction log data. Using these two statistical values, each preliminary action strength value is linearly transformed into a standard strength representation interval, such as between 0 and 1, to obtain the final action strength value.

[0071] Finally, the action direction and strength are comprehensively quantified to generate a quantitative vector as an emotional dynamic feature. Comprehensive quantification is an information encoding process aimed at encapsulating discrete action direction classification, continuous action strength values, and optional other emotional dimension information into a fixed-length numerical array, i.e., a quantitative vector. In specific implementation, first, the action direction is numerically encoded, such as assigning the "evolution" direction to the numerical value +1 and the "block" direction to the numerical value -1. Then, a quantitative vector is constructed. A basic vector construction method is a two-dimensional vector, with the first element storing the direction encoding value and the second element storing the final action strength value. To provide more information in subsequent analysis, an extended vector construction method can be used. For example, a five-dimensional quantitative vector is constructed, with the five dimensions being: direction encoding value, final action strength value, binary indication value indicating whether the emotional state is "positive", binary indication value indicating whether the emotional state is "negative", and original sentiment tendency score. The binary indication value takes the value 1 or 0, representing "yes" or "no". The original sentiment tendency score is directly taken from the calculation result of the first step. After this step, the text content of each interaction log is converted into a corresponding emotional dynamic feature quantitative vector. The quantitative vector and the operation subject identifier and task node identifier recorded in the interaction log that generates it maintain a strict association relationship and together serve as key feature data representing the emotional behavior of the operation subject at a specific task node, preparing for subsequent sequence analysis and model training steps.

[0072] S3, taking the emotion-motivation feature of the operation subject on the task node sequence as an observation sequence, training a state transition probability matrix through a hidden Markov model, and specifically implemented as:

[0073] First, the emotion-motivation feature quantitative vector of the operation subject arranged in time sequence is organized into an observation sequence. The input data of this step directly comes from the emotion-motivation feature quantitative vector generated in the foregoing step S2. Each emotion-motivation feature quantitative vector is associated with the operation subject identifier, timestamp and task node identifier contained in the interaction log record in which it is generated. In implementation, all emotion-motivation feature quantitative vectors are grouped according to the operation subject identifier, ensuring that each operation subject has an independent data subset. Within each operation subject data subset, all emotion-motivation feature quantitative vectors belonging to the operation subject are strictly arranged in ascending order according to the time indicated by the timestamp field, thereby forming an emotion-motivation feature quantitative vector sequence arranged in time sequence. This sequence constitutes the observation sequence defined in this step. One operation subject corresponds to one observation sequence, and the entire training set contains multiple such observation sequences generated by multiple different operation subjects. The length of each observation sequence is equal in value to the number of interaction log records generated by the operation subject in the target task flow, and the specific value is different due to the difference in activity of different operation subjects. The emotion-motivation feature quantitative vector as the basic element in the observation sequence has been defined and generated in step S2, for example, a 5-dimensional vector, so each observation sequence can be mathematically represented as a matrix with the number of rows as the time step and the number of columns as the dimension of the emotion-motivation feature quantitative vector.

[0074] Then, the number of hidden states of the hidden Markov model is set, and the state transition probability matrix and the observation probability matrix are initialized. The hidden Markov model is a statistical model used to describe the process of generating an observation sequence randomly by a hidden Markov chain. In this specific application scenario, the hidden state is interpreted as the intrinsic, not directly observed psychological state or behavior pattern stage of the operating subject when performing the task process. Setting the number of hidden states is a key hyperparameter that must be determined before model training. The determination of this value is not arbitrary, and its setting can be based on the understanding of the domain knowledge of the target task process, or it can also be obtained by a data-driven method. One method is to set the number of hidden states based on the domain or process knowledge, for example, set the number of hidden states to be associated with the number of recognized key decision stages or emotional transition stages in the target task process, which can be set to a value similar to but slightly less than the total number of task nodes, for example, when there are 8 task nodes, the number of hidden states can be set to 5. Another method is to select the number of hidden states by using a data-driven grid search strategy. In specific implementation, a candidate range of the number of hidden states is set in advance, for example, from 3 to 8. Within this range, each candidate value is taken as the number of hidden states in turn, a part of the historical observation sequence data is used as the training set to train a hidden Markov model, and another part of the data is used as the validation set to calculate the log-likelihood value of the model as the evaluation index. Finally, the candidate value that makes the log-likelihood value of the validation set reach the maximum is selected as the formally set number of hidden states. After setting the number of hidden states, the two core parameter matrices of the model need to be initialized. The first is the state transition probability matrix, which is a square matrix with the number of rows and columns equal to the number of hidden states set. Each element in the matrix represents the probability of transitioning from one hidden state to another. When initializing the matrix, the uniform initialization method is usually used, that is, it is assumed that at the initial time, the possibility of transitioning from one hidden state to any other hidden state is equal. The specific operation is to assign all elements in each row of the matrix to 1 divided by the number of hidden states, so as to ensure that the sum of all elements in each row is 1. The second is the observation probability matrix. Since the emotion dynamic feature quantization vector is a multi-dimensional continuous numerical vector, a continuous probability distribution is usually used instead of a discrete probability matrix to represent the observation probability. Correspondingly, a probability distribution parameter describing the generation of the observation vector needs to be initialized for each hidden state. Gaussian distribution is usually used as the probability distribution model. Therefore, the initialization work includes initializing a mean vector and a covariance matrix of a Gaussian distribution for each hidden state. The mean vector can be initialized as a sample vector randomly selected from all emotion dynamic feature quantization vectors of the observation sequence. The covariance matrix can be initialized as a diagonal matrix, and the element values on the diagonal can be obtained by calculating the variance of all emotion dynamic feature quantization vectors in each dimension. This initialization process provides a determined starting point for the subsequent iterative optimization algorithm.

[0075] Finally, the forward-backward algorithm is used to iteratively train the HMM, optimizing the parameters of the state transition probability matrix and the observation probability matrix until the model converges, obtaining the trained state transition probability matrix. The forward-backward algorithm is the standard method for parameter estimation of HMMs, aiming to maximize the total probability of generating all observation sequences given the model parameters. A single iteration of the algorithm consists of three core stages: forward calculation, backward calculation, and parameter re-estimation. In the forward calculation stage, a value called the forward probability is calculated for each observation sequence in the training set, each time step, and each possible hidden state. This value represents the probability of observing the data from the beginning of the sequence up to the current time step given the model parameters at the current iteration, with the model being in the specific hidden state at the current time step. The calculation is performed recursively from the previous time step to the current time step. In the backward calculation stage, a value called the backward probability is calculated, representing the conditional probability of observing the remaining data from the next time step to the end of the sequence given the model parameters at the current iteration and the specific hidden state at the current time step. The calculation is performed recursively from the next time step to the previous time step. After completing the forward and backward probability calculations for an observation sequence, two key intermediate statistics can be derived. The first is the probability of being in a specific hidden state at a specific time step given the entire observation sequence and the current model parameters. The second is the probability of transitioning from a specific hidden state to another specific hidden state at adjacent time steps. These statistics are calculated and accumulated separately for all observation sequences in the training set. Based on the global statistics accumulated from all observation sequences, the model parameters are updated. For the state transition probability matrix, the new estimate of each element is calculated using the corresponding accumulated statistics. Specifically, the state transition probability matrix is a square matrix with the number of rows and columns equal to the pre-set number of hidden states N. The variable i represents the index of the starting hidden state, and the variable j represents the index of the target hidden state, where i and j take integer values from 1 to N. The new probability estimate of transitioning from hidden state i to hidden state j is calculated as the accumulated expected number of transitions from state i to state j divided by the accumulated expected total number of transitions from state i. For the observation probability parameters, which are the parameters of the Gaussian distribution corresponding to each hidden state, the new estimate of the mean vector is calculated as the probability-weighted sum of all observation vectors in that hidden state divided by the accumulated probability expectation of that hidden state. The new estimate of the covariance matrix is obtained by calculating the probability-weighted outer product of the deviations between the observation vectors and their new mean vector in that hidden state. Completing the re-estimation of all parameters once completes one iteration. The next step is to determine whether the model has converged.The judgment basis is the change amplitude of the corresponding element values of the state transition probability matrix obtained after this iteration re-estimation and the state transition probability matrix of the last iteration. Calculate the sum of the absolute values of all corresponding element differences. If the sum is less than a pre-set model convergence judgment threshold, it is considered that the model parameters have tended to be stable, and the iteration training can be terminated. At this time, the state transition probability matrix obtained by the latest re-estimation is the trained state transition probability matrix. If the sum is greater than or equal to the model convergence judgment threshold, replace the old parameters with all the new parameters obtained by this re-estimation, and start a new round of iteration calculation. The specific value of the model convergence judgment threshold is determined according to the requirement of the model training accuracy. Usually, it is set to a small positive number, for example, 0.0001. The setting of the model convergence judgment threshold is based on the fact that when the parameter change is very small, further iteration has no significant meaning for the improvement of the model effect. Through the above iteration process, the model starts from the initial parameters, and automatically learns and optimizes the state transition mode that best explains the dynamic change rule of the emotional dynamic feature sequence under the driving of the observation sequence data. Finally, this mode is condensed in the trained state transition probability matrix, which accurately describes the possibility of the transition between hidden states.

[0076] S4, based on the state transition probability matrix, analyzing the appearance frequency and state conversion type of each task node in the optimal state transition path to identify the key hub node of emotional dynamic conduction, which is specifically implemented as:

[0077] First, for each operation subject, the emotion dynamic feature observation sequence is decoded based on the state transition probability matrix using the Viterbi algorithm to generate the corresponding optimal state transition path. The input of this step includes two key parts, the first part is the state transition probability matrix trained in the preceding step S3, and the second part is the emotion dynamic feature observation sequence generated in the preceding step S2 and has been grouped by operation subjects and arranged in chronological order. The Viterbi algorithm is a dynamic programming algorithm used to find the hidden state sequence that is most likely to produce the observed sequence given the observation sequence and the parameters of the hidden Markov model. In the specific implementation of decoding, the algorithm needs to be executed independently for each operation subject corresponding to the emotion dynamic feature observation sequence. Before the algorithm runs, two matrices related to the length of the sequence and the number of hidden states need to be initialized. The first is the path probability matrix, which has the time steps of the emotion dynamic feature observation sequence as the number of rows and the number of hidden states of the hidden Markov model as the number of columns, used to store the maximum cumulative probability of each hidden state reached at each time step in the recursive calculation process. The second is the path backtracking matrix, which has the same dimensions as the path probability matrix, used to store the hidden state index corresponding to the previous time step when the current path probability is reached. The calculation starts from the first time step of the sequence. The initial state probability distribution of the hidden Markov model and the observation probability calculated according to the observation vector of the current time step are used to initialize the first row of the path probability matrix. For each subsequent time step in the sequence starting from the second time step, and for each possible current hidden state, the algorithm performs the core recursive calculation. This calculation needs to refer to the probability values of all hidden states in the path probability matrix at the previous time step, the probability values of the state transition from the previous state to the current state in the state transition probability matrix, and the observation probability value calculated according to the observation vector of the current time step, which produces the observation in the current state. The algorithm calculates the path probability of each hidden state transition from the previous state to the current state and produces the current observation, i.e., the path probability of the previous state multiplied by the corresponding transition probability and multiplied by the current observation probability. From all possible previous states, the one that maximizes the calculated path probability is selected, and its path probability is recorded as the path probability matrix value of the current time step reaching the current hidden state, and the number of the selected previous state is recorded in the corresponding position of the path backtracking matrix. This recursive process continues until the last time step of the observation sequence is processed. At the last time step, the algorithm checks the last row of the path probability matrix and selects the hidden state with the maximum path probability value as the terminal state of the optimal state transition path. Subsequently, according to the path backtracking matrix, the previous state of each time step is found in reverse from the terminal state until the starting point of the sequence is reached, thereby obtaining a complete sequence composed of hidden state numbers, which is the optimal state transition path decoded and generated for the operation subject. This path represents the internal state transition trajectory that is most likely to produce the observed emotion dynamic feature sequence.

[0078] Then, the total frequency of each task node appearing in all optimal state transition paths of all operation agents is counted, and the hidden state transition types occurring on the corresponding task node in the path are recorded. After obtaining the optimal state transition paths of all operation agents, these paths need to be associated with specific physical task nodes. The association method relies on the original data records. For each time step in the optimal state transition path, its decoding basis is the emotional dynamic feature observation value corresponding to the time step, which comes from an original interaction log record with a clear task node identification. Therefore, through data indexing, for each hidden state position in the optimal state transition path, its corresponding task node identification can be associated. After the association is completed, statistics are started. The total frequency of each task node appearing refers to counting each task node identification appearing in the path sequence by traversing all optimal state transition paths of all operation agents. Each time a task node identification appears once, no matter how many time steps it lasts, it is counted as appearing once. Finally, the total number of times each task node is accessed in all paths is accumulated, which is the total frequency of its appearance. Recording the hidden state transition type requires analyzing the continuous hidden state sequence associated with each task node identification on each optimal state transition path. In the specific analysis, for the continuous time step segment associated with the same task node identification in the path, the hidden state sequence is checked. Within the segment, if the hidden state numbers of two adjacent time steps are different, it is determined that a state transition from the former to the latter occurs on the task node, which is recorded as a specific transition type. The transition type is represented by an ordered pair, such as from hidden state number 2 to hidden state number 3. For each task node, all transition types occurring on the node in all path segments are collected to form a transition type list. This list records all specific state change events observed on the node.

[0079] Then, based on the frequency of occurrence of the task nodes and the state transition type, a comprehensive characteristic value representing its conduction importance is calculated. The calculation process includes three sub-steps performed in sequence. The first sub-step is to normalize the total frequency of occurrence of the task nodes to obtain the standardized frequency. In implementation, first, collect the total frequency of occurrence of all the task nodes to be counted, and find the maximum and minimum values among them. For any task node, the calculation method of its standardized frequency is to subtract the global minimum frequency value from the original total frequency of occurrence of the node, and then divide the difference by the difference between the global maximum frequency value and the global minimum frequency value. Through this minimum and maximum normalization, the frequency indicators of all task nodes are linearly mapped to the interval of 0 to 1. The second sub-step is to analyze the hidden state transition type occurring at the task node. First, for the transition type list of each task node, the number of categories of non-repeated transition types is counted. For example, a list contains 5 transitions, but only involves two different ordered pairs, so the category number is 2. Second, the proportion of the main transition type is calculated. To this end, the transition type with the highest occurrence frequency needs to be determined in the transition type list, which is defined as the main transition type. Then, the proportion of the occurrence frequency of the main transition type to the total number of transition events of the node is calculated, and this proportion value is the proportion of the main transition type. The third sub-step is to perform weighted summation of the three intermediate indicators obtained above to generate the final comprehensive characteristic value. Weighted summation means that each indicator is assigned a weight coefficient. The three weight coefficients correspond to the standardized frequency, the number of transition type categories, and the proportion of the main transition type, respectively. The specific value of the weight coefficient needs to be set in advance, which can be based on the experience of experts in the relevant field, or determined by analyzing the correlation between historical data and the effect target of the process or task. For example, if it is believed that frequent access to the node is important, a higher weight, such as 0.5, can be given to the standardized frequency; if it is believed that the diversity of state transitions better reflects the complex conduction characteristics of the node, a moderate weight, such as 0.3, can be given to the number of transition type categories; if it is believed that the concentration of state transition direction represents a clear impact pattern, a certain weight, such as 0.2, can be given to the proportion of the main transition type. The sum of the three weight coefficients must be 1. In actual application, the weight coefficients can be adjusted according to the specific analysis scenario. In the specific calculation of the comprehensive characteristic value of a task node, multiply the standardized frequency of the node by the first weight coefficient, multiply the number of transition type categories of the node by the second weight coefficient, multiply the proportion of the main transition type of the node by the third weight coefficient, then add the three product results, and the sum value is the comprehensive characteristic value of the task node. This value is a dimensionless comprehensive score.

[0080] Finally, the task nodes are sorted according to the comprehensive characteristic values, and the task nodes with the top pre-set number of the sorted order are identified as the key hub nodes of emotional power conduction. After calculating the comprehensive characteristic values of all task nodes, all nodes are arranged in descending order according to the values from large to small, and a sorted list is generated. Then, a specific pre-set number needs to be determined to select the corresponding number of task nodes from the top of the list. The determination method of the pre-set number can be flexibly selected according to the actual analysis requirements. One method is to directly specify a fixed value, for example, to specify that the top 5 nodes need to be found. Another method is to determine based on the distribution of the comprehensive characteristic values, for example, to draw the distribution curve of the comprehensive characteristic values of all nodes, observe the inflection point where the slope of the curve changes significantly, and take the number of nodes before the inflection point as the pre-set number. Another method is to set a key threshold, for example, to set the comprehensive characteristic value greater than or equal to 0.75 as the selection standard, then all nodes with a comprehensive characteristic value greater than or equal to 0.75 after sorting will be selected, and the number is the pre-set number at this time. The setting of the key threshold 0.75 can be based on historical analysis experience value, or can be obtained by calculating a certain statistical quantile of the comprehensive characteristic values of all nodes, for example, taking the value at the 80th percentile as the threshold. After determining the pre-set number by any one or combination of the above methods, the corresponding number of task nodes at the top of the descending order list is selected. These selected task nodes, that is, the key hub nodes officially identified as playing a core hub role in the transmission and evolution of emotional power in the entire target task process. The identification list of these key hub nodes will be an important input for the subsequent evaluation step. The entire identification process objectively filters the key nodes with concentrated influence from a large number of task nodes through algorithm decoding, statistical quantification and comprehensive sorting.

[0081] S5, according to the state transition probability matrix and the key hub node, evaluating the emotional power inertia disturbance intensity suffered by each task node, specifically implemented as:

[0082] First, based on the state transition probability matrix, the state residence expectation of each task node in all possible state transition paths of the hidden Markov model is calculated, and the state residence expectation is combined with the observation emotion dynamic feature strength associated with the task node itself to generate the basic strength representing the emotion dynamic inertia of the corresponding task node. The core input of this step is the state transition probability matrix trained in the previous step S3. The calculation of the state residence expectation is essentially to solve the stationary distribution of the Markov chain corresponding to the state transition probability matrix. The stationary distribution is a probability vector, and each component of the vector represents the long-term probability of the system being in each hidden state after infinite state transitions. This vector is obtained by solving a system of linear equations, with the specific constraints being that the product of the probability vector and the state transition probability matrix is equal to the probability vector itself, and the sum of all components of the vector is 1. In the specific numerical calculation implementation, the power iteration method is used for solving. The power iteration method starts with a randomly initialized positive probability vector whose components sum to 1, and repeatedly performs multiplication operations on the vector and the state transition probability matrix. After each multiplication operation, the obtained vector is normalized, i.e., each component of the vector is divided by the sum of all components, to ensure that the sum of its components remains 1. This iteration process is repeated until the difference between the vectors obtained in the adjacent two iterations is small enough to be considered as convergence. The difference is usually measured as the sum of the absolute values of the differences between the corresponding components of the two vectors. When the sum is less than a preset convergence threshold for calculating the stationary distribution, for example, less than 0.00001, the iteration is stopped, and the vector obtained in the last iteration is identified as the stationary distribution vector. Each component value in the vector is the state residence expectation of the corresponding hidden state. Next, the state residence expectation needs to be combined with the observation emotion dynamic feature strength associated with the task node itself. The observation emotion dynamic feature strength comes from the specific dimension value in the emotion dynamic feature quantization vector generated in step S2. For each task node, first determine its typical hidden state set. The determination method is to analyze all the optimal state transition paths in step S4 and count the one or several hidden states that appear most frequently on the task node. For example, if a task node is most commonly associated with hidden state number 3 and hidden state number 7 in the path, then its typical hidden state set is the set containing state 3 and state 7. Then, extract the state residence expectation values of these typical hidden states from the stationary distribution vector. At the same time, extract all the emotion dynamic feature quantization vectors associated with the task node from the historical data, take out the dimension values representing the emotion strength, and calculate the arithmetic mean of these strength values as the typical observation emotion dynamic feature strength of the task node. The combination operation is to sum the extracted state residence expectation values, and then multiply them with the typical observation emotion dynamic feature strength value to obtain a preliminary combined value.Finally, the preliminary combined values of all task nodes are globally normalized, for example, by dividing each preliminary combined value by the maximum value among all preliminary combined values, so that the result falls within the interval of 0 to 1. The normalized value is defined as the base strength of the emotional momentum inertia of the task node itself.

[0083] Next, according to the high-probability outward transition characteristics of the key hub node represented in the state transition probability matrix, the radiation range and the intensity decay coefficient of the emotional dynamic influence are calculated, and according to the connection relationship between the task nodes and the radiation range in the task flow topology graph, the emotional dynamic conduction intensity of the key hub node to the task node is calculated. The list of key hub nodes directly comes from the identification result of step S4. The analysis object of the high-probability outward transition characteristics is the row vector corresponding to the typical hidden state mapped by the key hub node in the state transition probability matrix. When calculating the radiation range, for each key hub node, the row vector corresponding to its typical hidden state is checked, and all target hidden states with transition probability values greater than a preset outward transition probability threshold are selected. The task nodes mapped by these target hidden states constitute the direct radiation range of the emotional dynamic influence of the key hub node. The setting of the outward transition probability threshold needs to be based on the overall distribution of all non-zero transition probability values in the state transition probability matrix. One setting method is to calculate the average μ and standard deviation σ of all transition probability values, and set the outward transition probability threshold as μ + n x σ, where n is a positive coefficient, for example n = 1. Another method is to pre-set a fixed proportion ρ according to experience, for example ρ = 0.3, that is, only the top thirty percent of the transition paths are retained. The intensity decay coefficient is used to simulate the decay of the influence with the topological distance. The task flow topology graph defines the logical connection relationship between the task nodes, and the edge between two adjacent nodes in the graph represents the direct task flow or dependency relationship. The topological distance d between two task nodes is defined as the number of edges in the shortest path from one node to another node in the task flow topology graph. The intensity decay coefficient is a function of the topological distance d, and an exponential decay model is usually used. For example, define a basic decay rate parameter λ, whose value range is between 0 and 1, for example λ = 0.6. Then, for the task node with a topological distance of d from the key hub node, the decay coefficient is equal to the dth power of λ. The basic decay rate parameter λ can be calibrated according to the decay law of emotional influence in historical data. The emotional dynamic conduction intensity I{K->B} of the key hub node K to the task node B is calculated. First, it is judged whether the task node B is within the radiation range of the key hub node K. If not, I{K->B} is recorded as 0. If yes, I{K->B} is determined by three factors: the basic intensity SK of the key hub node K itself, the decay coefficient δ{K->B} from K to B, and the connection strength C{K->B} from K to B. The connection strength C{K->B} is calculated based on historical behavior data, and the number of times of directly entering the task node B immediately after the task node corresponding to the key hub node K is counted, accounting for the proportion of all direct transition times from the task node corresponding to the key hub node. Finally, the calculation of the conduction intensity is expressed as: I{K->B} = SK x δ{K->B} x C{K->B}.

[0084] Finally, the base strength of the task node itself is weighted and superimposed with the received emotional power conduction strength from one or more key hub nodes, where the weight is dynamically adjusted according to the order of the task node being visited in the optimal state transition path, to generate the final emotional power inertia disturbance strength of the task node. For each task node B to be evaluated, its own base strength SB and the conduction strength I{K->B} from all key hub nodes K need to be summarized. The weighted superimposition process assigns a weight to each strength component. The weight of the base strength of the task node itself is called the internal inertia weight ωB, and the weight of the conduction strength from each key hub node K is called the external influence weight γ{K->B}. The dynamic adjustment of these weights is based on the relative order of the task node B and the key hub node K appearing in the optimal state transition path. In specific implementation, first, based on all the optimal state transition paths decoded in step S4, the average time step sequence number T of each task node and each key hub node appearing in all paths is calculated. The smaller the average time step sequence number, the more likely the node is to be visited early in the process. The principle of dynamic adjustment is that for a task node B and a key hub node K, if the average sequence number TB of the task node B is significantly greater than the average sequence number TK of the key hub node K, i.e. TB>TK, it is considered that the influence of K on B is more direct and more likely to occur, and therefore a higher value of the external influence weight γ{K->B} should be assigned; otherwise, a lower value should be assigned. The specific weight value can be set by a function based on the sequence number difference Δ=TB-TK. When Δ>0, γ{K->B} can be set to be positively correlated with Δ, for example γ{K->B}=γ0×(1+α×Δ), where γ0 is a base influence weight and α is an adjustment coefficient less than 1. When Δ≤0, then γ{K->B}=γ0', where γ0' is a lower base weight value. The setting of the internal inertia weight ωB can be considered complementary to the sum of all external influence weights, for example, ωB=1-Σγ{K->B}, to ensure that ωB+Σγ{K->B}=1. After calculating the initial weights of all strength components, normalization processing is performed to make the sum of all weights equal to 1. Finally, the calculation of the emotional power inertia disturbance strength PB of the task node B is expressed as: PB=(SB×ωB)+Σ(I{K->B}×γ{K->B}). The strength value PB comprehensively reflects the inertia of the node itself and the external conduction influence, and through the dynamic weighting of the path order, it realizes the quantitative evaluation of the emotional power disturbance level.

[0085] S6, map the emotional power inertia disturbance strength received by the task node to the task flow topology graph and perform visual rendering, which is specifically implemented as:

[0086] First, a task flow topology graph representing the logical connection relationship between task nodes is obtained. The task flow topology graph is a graphical model used to describe the sequence and logical flow relationship between each task node in the target task flow. There are two feasible implementation methods for obtaining this graph. The first method is to directly parse and extract from the electronic file carrying the definition of the target task flow. For example, if the task flow is managed by a workflow or project management engine, the node elements representing task activities and the sequential flow elements connecting these nodes can be read from the process modeling file stored by the engine, thereby constructing the corresponding topology graph. The second method is to statistically reconstruct based on the task node identification sequences recorded in the historical interaction log data. Specifically, the time-ordered task node identification sequences generated by all operation subjects in the multi-source asynchronous interaction log set generated in step S1 are collected, and then the frequency of each pair of adjacent task node identification combinations occurring in the entire historical sequence is counted. For each pair of node combinations, if the frequency of occurrence is higher than a preset connection relationship judgment threshold, a directed edge is used to connect the two nodes in the topology graph to reflect the high-frequency task flow path that occurs in actual collaboration. The setting of the connection relationship judgment threshold is based on the statistical distribution of the transition frequency of all node pairs in the historical sequence. For example, the average value of the transition frequency of all node pairs can be calculated, and this average value is used as the connection relationship judgment threshold. Through either of the above methods, a task flow topology graph composed of a node set and an edge set is finally obtained, where each node corresponds to a unique task node identification, which is completely consistent with the task node identification used in all the above steps, ensuring seamless data association.

[0087] Next, the emotional dynamic inertia disturbance intensity value of the task node is converted into the corresponding visual attribute value according to the preset segmented mapping rule. The input data of this step is the emotional dynamic inertia disturbance intensity value strictly corresponding to each task node identification calculated and generated in step S5, denoted as the numerical value I. The preset segmented mapping rule is a set of conversion rules for systematically mapping a continuous intensity numerical value interval to a group of discrete attribute values with clear visual significance. The primary step in establishing this rule is to perform a global statistical analysis on the emotional dynamic inertia disturbance intensity values of all task nodes to determine their numerical range, including the minimum value Imin and the maximum value Imax, and to understand their distribution characteristics. Based on the analysis results and the established visualization target, the number of segments M is determined. For example, if 5 different visual styles are planned to be used for differentiation, then M = 5. Subsequently, M-1 segment boundary thresholds need to be determined to divide the entire intensity value range into M continuous intervals. The determination method of the segment boundary threshold is not fixed. One method is to use equal interval division, that is, to divide the numerical range Imax-Imin into M parts, thereby calculating M-1 equally spaced boundary thresholds. Another method is to perform quantile division based on the statistical distribution of intensity values, for example, a set of quantile points are pre-set, such as the 20% quantile, the 40% quantile, the 60% quantile, and the 80% quantile. After arranging the intensity value data in ascending order, the numerical values at these quantile points are used as the segment boundary thresholds. This method ensures that the number of nodes included in each interval is approximately equal. After the segmentation rule is determined, a visual attribute value needs to be pre-defined for each intensity interval. Visual attributes mainly involve color and size. For color attributes, a specific color code can be specified for each interval in the selected color space model. For example, in the RGB color model, 5 RGB color values from light blue to deep red can be defined for the 5 intervals, for example, interval 1 corresponds to RGB (200, 230, 255), and interval 5 corresponds to RGB (255, 0, 0). For size attributes, a graphic element size value can be specified for each interval, such as the radius value of a circular graphic element, interval 1 corresponds to a smaller radius, and interval 5 corresponds to a larger radius. The mapping process is, for any task node and its intensity value I, it is compared with M-1 segment boundary thresholds in turn to determine which intensity interval it falls into, and then it is mapped to the pre-set color attribute value C and size attribute value S of that interval.

[0088] Finally, the corresponding task node graph elements in the task flow topology graph are rendered based on the visual attribute values, and a color gradient or size gradient visualization layer is generated according to the high and low differences in the emotional dynamic inertia disturbance intensity values. The rendering operation is performed in a computing environment with graphics rendering capability. The specific implementation process is that, first, the obtained task flow topology graph is loaded into the environment, and each task node graph element in the graph has an initial coordinate position on a two-dimensional plane. Subsequently, each task node graph element in the topology graph is traversed, and the visual attribute values of the node, i.e., the color attribute value C and the size attribute value S, are retrieved according to the task node identifier of the node obtained through the mapping process in the previous step. Then, the graph elements are redrawn in visual style according to these attribute values. For color rendering, the fill color attribute of the graph element is set to the color represented by the color attribute value C. For size rendering, the geometric size parameters of the graph element are adjusted according to the size attribute value S, for example, the radius of a circular graph element is set to S. This rendering process acts on all task node graph elements in the topology graph. After rendering is completed, all node graph elements present regular visual changes according to the high and low of the emotional dynamic inertia disturbance intensity values, thereby forming a visualization layer overlaid on the original topology structure. The color gradient is reflected in the continuous change of the color of the node graph element, the node with high intensity value presents warm color or dark tone, the node with low intensity value presents cold color or light tone, and the overall color gradient effect from low to high is formed. The size gradient is reflected in the regular difference in the size of the node graph element, the node graph element with high intensity value is larger in size, and the node graph element with low intensity value is smaller in size, and the overall size gradient effect from large to small is formed. The finally generated visualization layer and the topology connection line of the bottom layer together form a complete task flow graph, and the significant difference in the visual features of the nodes enables the analyst to intuitively locate the high-intensity area and low-intensity area of the emotional dynamic disturbance, thereby providing an intuitive visual tool for understanding the emotional stress distribution and conduction hotspots in the collaboration process. The entire step realizes the conversion from quantitative analysis results to intuitive visual perception, and completes the final output link of the method chain.

[0089] Embodiment 2: Figure 2 The structure diagram of the emotional behavior visualization analysis system based on artificial intelligence is given, and the emotional behavior visualization analysis system based on artificial intelligence comprises:

[0090] A log acquisition module is configured to acquire a multi-source asynchronous interaction log set of a target task flow, and the interaction log comprises a timestamp, an operation subject identifier, text content, and a task node identifier.

[0091] A feature extraction module is configured to analyze the text content of the interaction log and extract emotional dynamic features of the operation subject on the task node.

[0092] A matrix training module is configured to take the emotional dynamic features of the operation subject on the sequence of task nodes as an observation sequence, and train a state transition probability matrix through a hidden Markov model;

[0093] A node identification module is configured to analyze the appearance frequency and state conversion type of each task node in the optimal state transition path based on the state transition probability matrix, and identify a key hub node of emotional dynamic conduction;

[0094] A strength evaluation module is configured to evaluate the emotional dynamic inertia disturbance strength of each task node according to the state transition probability matrix and the key hub node;

[0095] A visual rendering module is configured to map the emotional dynamic inertia disturbance strength of the task node to a task flow topology diagram and perform visual rendering.

[0096] Embodiment 3: This embodiment is to introduce an emotional behavior visual analysis device based on artificial intelligence. The emotional behavior visual analysis device based on artificial intelligence comprises one or more processors; further comprises a storage device for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement an emotional behavior visual analysis method based on artificial intelligence.

[0097] The calculations involved in the embodiments are all dimensionless numerical calculations. The preset parameters and threshold values in the calculations are set by a person skilled in the art according to actual conditions.

[0098] It should be noted that the present application can be deployed in the device itself to realize embedded application, or run on a PC or other terminal with a user interface, thereby meeting various hardware environments and use requirements.

[0099] The above-described embodiments can be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented by software, the above-described embodiments can be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, the processes or functions described in the embodiments of the present application are wholly or partially generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transferred from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center through wireless or wired transmission. The wired transmission includes optical fiber, twisted pair, coaxial cable, etc. The wireless transmission includes infrared, microwave, etc. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. containing one or more available medium collections. The available medium can be a magnetic medium (for example, a floppy disk, a hard disk, a magnetic tape), an optical medium (for example, a DVD), or a semiconductor medium. The semiconductor medium can be a solid-state disk.

[0100] Those skilled in the art can clearly understand that, for the convenience and brevity of the description, the specific working processes of the above-described system, device and module can refer to the corresponding processes in the foregoing method embodiments, which will not be described here.

[0101] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented in other ways. For example, the above-described device embodiments are only schematic, for example, the division of the modules is only a logical function division, and actual implementation can have another division manner, for example, a plurality of modules or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the displayed or discussed each other can be indirect coupling or communication connection through some interfaces, devices or modules, which can be electrical, mechanical or other forms.

[0102] The modules described as separate components can or can not be physically separated, and the components displayed as modules can or can not be physical modules, which can be located in one place or distributed on a plurality of network modules. Some or all of the modules can be selected according to actual needs to achieve the purpose of the embodiments.

[0103] In addition, each functional module in the various embodiments of the present application can be integrated in one processing module, or each module can exist physically independently, or two or more modules can be integrated in one module.

[0104] If the functions are implemented in the form of software function modules and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to make a computer device (which can be a personal computer, a server or a network device, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various media that can store program codes.

[0105] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

[0106] Finally: the above is only the preferred embodiment of the present application, and is not used to limit the present application. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application should be included in the protection scope of the present application.

Claims

1. An emotion behavior visualizing analysis method based on artificial intelligence, characterized by, Comprise: S1, acquire the multi-source asynchronous interaction log set of the target task flow, the interaction log includes timestamp, operation subject identification, text content and task node identification; S2, analyze the text content of the interaction log and extract the emotional dynamic characteristics of the operation subject on the task node; S3, take the emotional dynamic characteristics of the operation subject on the task node sequence as the observation sequence, and train the state transition probability matrix through the hidden Markov model; S4, based on the state transition probability matrix, analyze the frequency and state conversion type of each task node in the optimal state transition path, and identify the key node of emotional dynamic conduction; S5, according to the state transition probability matrix and the key node, evaluate the emotional dynamic inertia disturbance intensity of each task node, including: Based on the state transition probability matrix, calculate the state residence expectation of each task node in all possible state transition paths of the hidden Markov model, and combine the state residence expectation with the observation emotional dynamic characteristic intensity associated with the task node itself to generate the basic strength of the emotional dynamic inertia of the corresponding task node; According to the high probability outward transfer characteristics of the key node in the state transition probability matrix, calculate the radiation range and strength attenuation coefficient of its emotional dynamic influence, and according to the connection relationship between the task node and the radiation range in the task flow topology graph, calculate the emotional dynamic conduction intensity of the key node to the task node; The basic strength of the task node itself and the emotional dynamic conduction strength received from one or more key nodes are weighted and superimposed, wherein the weight is dynamically adjusted according to the order of the task node being accessed in the optimal state transition path, to generate the final emotional dynamic inertia disturbance intensity of the task node; S6, map the emotional dynamic inertia disturbance intensity of the task node to the task flow topology graph and render it visually. 2.The artificial intelligence-based emotion behavior visual analysis method of claim 1, wherein, S1 includes: According to the definition of the target task flow, collect the original log from the operation history of the project management system and the collaboration communication platform participating in the target task flow; Clean the collected original log, and align and merge according to the unified timestamp, operation subject identification, text content and task node identification, to generate a multi-source asynchronous interaction log set. 3.The artificial intelligence-based emotional behavior visual analysis method of claim 1, wherein S2 Comprise: Carry out word segmentation and part-of-speech tagging on the text content of the interaction log, and identify the emotional state expressed in the text based on the pre-constructed sentiment analysis rules; Combine the emotional state with the context semantics of the task node in the process to determine the direction and strength of the emotional state on the promotion or blocking of the task flow; Comprehensively quantify the direction and strength to generate a quantitative vector as the emotional dynamic characteristic. 4.The artificial intelligence-based emotion behavior visual analysis method of claim 1, wherein S3 includes: Organize the emotional dynamic characteristic quantitative vectors of the operation subject in time sequence as observation sequences; Set the number of hidden states of the hidden Markov model, and initialize the state transition probability matrix and the observation probability matrix; Use the forward-backward algorithm to iteratively train the hidden Markov model, optimize the parameters of the state transition probability matrix and the observation probability matrix, until the model converges, and obtain the trained state transition probability matrix. 5.The artificial intelligence-based emotion behavior visual analysis method of claim 1, wherein S4 Comprise: For each operation subject, the emotion dynamic feature observation sequence is decoded based on the state transition probability matrix using the Viterbi algorithm to generate a corresponding optimal state transition path; The total frequency of each task node appearing in the optimal state transition path of all operation subjects is counted, and the hidden state transition type occurring at the corresponding task node in the path is recorded; Based on the appearance frequency and state transition type of the task node, a comprehensive feature value representing its transmission importance is calculated; According to the comprehensive feature value, the task nodes are sorted, and the pre-set number of task nodes ranked in the front are identified as the key hub nodes of emotion dynamic transmission. 6.The artificial intelligence-based emotion behavior visual analysis method of claim 5, wherein, Based on the appearance frequency and state transition type of the task node, a comprehensive feature value representing its transmission importance is calculated, including: normalizing the total frequency of the task node appearing in all optimal state transition paths to obtain a standardized frequency; counting the types of different hidden state transitions occurring at the task node, and calculating the proportion of each transition type in the total transition events; the standardized frequency and the number of state transition types and the proportion of the main transition type are weighted and summed to generate a comprehensive feature value for sorting. 7.The artificial intelligence-based emotion behavior visual analysis method of claim 1, wherein S6 It includes: obtaining a task flow topology graph representing the logical connection relationship between task nodes; The emotional dynamic inertia disturbance intensity value of the task node is converted into the corresponding visual attribute value according to the preset segmentation mapping rule; Based on the visual attribute value, the corresponding task node graph element in the task flow topology graph is rendered, and a color gradient or size gradient visualization layer is generated according to the difference in the emotional dynamic inertia disturbance intensity value.

8. An emotion behavior visualizing analysis system based on artificial intelligence for implementing the emotion behavior visualizing analysis method based on artificial intelligence according to any one of claims 1 to 7, characterized in that, It includes: a log acquisition module for acquiring a multi-source asynchronous interaction log set of a target task flow, the interaction log including a timestamp, an operation subject identifier, a text content, and a task node identifier; a feature extraction module for analyzing the text content of the interaction log and extracting the emotional dynamic features of the operation subject on the task node; a matrix training module for training a state transition probability matrix based on the emotional dynamic features of the operation subject on the task node sequence as an observation sequence through a hidden Markov model; a node identification module for analyzing the appearance frequency and state transition type of each task node in the optimal state transition path based on the state transition probability matrix, and identifying the key hub nodes of emotion dynamic transmission; an intensity evaluation module for evaluating the emotional dynamic inertia disturbance intensity of each task node based on the state transition probability matrix and the key hub nodes; a visual rendering module for mapping the emotional dynamic inertia disturbance intensity of the task node to the task flow topology graph and visualizing the rendering. 9.An emotion behavior visualizing analysis device based on artificial intelligence, characterized by, It includes: one or more processors; a storage device for storing one or more programs; when one or more programs are executed by one or more processors, one or more processors implement an artificial intelligence-based emotion behavior visualization analysis method as claimed in any one of claims 1-7.

Images