Text generation auxiliary processing method and system based on machine learning

Abstract

The invention discloses a text generation auxiliary processing method and system based on machine learning, and relates to the technical field of auxiliary processing, and the method comprises the steps: traversing a user emotion sequence, and recognizing at least one emotion inflection point user node from the user node sequence; aiming at each identified emotion inflection point user node, executing inflection point correlation analysis: taking the position of the emotion inflection point user node in the user node sequence as a starting point, performing backtracking search in dialogue log data based on a preset backtracking rule, and positioning a key triggering customer service node; and taking the position of the emotion inflection point user node in the user node sequence as a starting point, determining an appeal evolution mode based on an appeal feature set of the emotion inflection point user node and a predetermined number of user nodes behind the emotion inflection point user node, and generating dialogue abstract data based on all the emotion inflection point user nodes and a correlation analysis result thereof. The method has the effect of improving the text generation auxiliary processing efficiency.

Description

Technical Field

[0001] This application relates to the field of auxiliary processing technology, and in particular to a text generation auxiliary processing method and system based on machine learning. Background Technology

[0002] In the field of existing intelligent customer service systems and dialogue analysis technology, the automatic generation of summaries from multi-turn customer service conversations is a common requirement. Traditional methods mainly rely on text compression, keyword extraction, or generative summarization techniques based on sequence-to-sequence models. Their core objective is to extract the theme or key factual information of the conversation and generate a condensed narrative text. However, these technical approaches essentially treat dialogue as a static collection of information or a linear flow of text, making it difficult to deeply reveal the essential contradictions that dynamically evolve during service interactions. In particular, they cannot automatically identify and highlight the key service nodes and their causal chains that lead to a deterioration in user experience, thus limiting the value of the generated summaries in supporting service quality review, agent training, and process optimization.

[0003] Specifically, existing technical solutions, when processing dialogue summaries, often treat each round of dialogue content equally or only weight importance based on superficial features such as word frequency and position. This results in summaries that are usually a "chronicle" of the dialogue facts or a list of key points. While retaining basic information, they completely lose sight of the two core threads inherent in the dialogue: the "changes in emotional tension" and the "evolution of core demands." For example, when a user's dissatisfaction gradually builds up until it explodes due to repeated unresolved issues or ineffective responses, existing summaries may only record the user's final complaint, completely ignoring the inflection point of emotional escalation and the specific customer service response that triggered it. These are precisely the "critical moments" that most need to be analyzed and reflected upon in service improvement. Furthermore, user demands are not static during interactions; they may shift, escalate, or evolve as the service progresses. Traditional static summarization methods cannot capture and present this dynamic evolution process, making it impossible for managers to gain insight into the complete lifecycle of the user's true intentions from the summaries.

[0004] Furthermore, due to a lack of structured analytical capabilities regarding the inherent causal logic of dialogue interactions, the summaries generated by existing technologies cannot automatically pinpoint the root cause of service defects. They can only answer "what was said in the dialogue," but not deeper questions such as "why the user is dissatisfied" and "how dissatisfaction arises and evolves." This forces customer service personnel to still invest significant time manually listening to complete recordings or reading lengthy logs, relying on personal experience to subjectively identify service shortcomings—a process that is inefficient and inconsistent in its standards.

[0005] To address the aforementioned issues, there is an urgent need in this field for an intelligent summary generation method that can go beyond superficial information aggregation, deeply analyze the dynamic process of dialogue interaction, and especially automatically identify key points of emotional shifts, locate service triggering reasons, and track the evolution path of user demands, so as to produce structured analysis reports that truly serve to improve quality and efficiency and have action-oriented guidance significance. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this application provides a text generation auxiliary processing method and system based on machine learning.

[0007] In a first aspect, this application provides a machine learning-based text generation auxiliary processing method, comprising the following steps: Obtain the dialogue log data of the target customer service conversation. The dialogue log data contains multiple rounds of dialogue records arranged in chronological order. Each round of dialogue records includes user speech text and customer service speech text. Based on the dialogue log data, a user node sequence is constructed, wherein each user node corresponds to the user's speech text in one round of dialogue record, and the user nodes in the user node sequence are arranged in chronological order; For each current user node in the user node sequence, perform the following processing: The user's utterance text corresponding to the current user node is input into the sentiment analysis model to obtain the sentiment state value of the current user node. The sentiment state value comes from a predefined sentiment state label set. The user speech text corresponding to the current user node is processed to extract the request, and the request feature set of the current user node is obtained. The request feature set contains at least one request feature, and each request feature includes a request type identifier and a confidence level. Based on the processed emotional state values ​​of all user nodes, a user emotional sequence is generated; Traverse the user emotion sequence and, based on a preset emotion transition judgment rule, identify at least one emotion inflection point user node from the user node sequence; For each identified user node at an emotional inflection point, perform inflection point correlation analysis: Starting from the position of the emotional inflection point user node in the user node sequence, based on the predetermined backtracking rules, backtracking search is performed in the dialogue log data to locate the key triggering customer service node. The key triggering customer service node is the customer service node before the emotional inflection point user node and whose customer service text is determined to meet the preset triggering conditions. Starting from the position of the emotional inflection point user node in the user node sequence, the demand evolution pattern is determined based on the demand feature set of the emotional inflection point user node and a predetermined number of subsequent user nodes. The demand evolution pattern is used to describe the change type of the core demand. Based on the analysis results of all user nodes at emotional inflection points and their correlations, dialogue summary data is generated. The dialogue summary data includes at least the identifier of each user node at an emotional inflection point, the customer service text of the corresponding key triggering customer service node, and the corresponding demand evolution pattern.

[0008] Preferably, the user's utterance text corresponding to the current user node is input into the sentiment analysis model to obtain the sentiment state value of the current user node, including: The user speech text corresponding to the current user node is segmented and vectorized to obtain a text vector. The text vector is input into the sentiment analysis model; Obtain the classification results output by the sentiment analysis model, wherein the classification results include the probability of belonging to each sentiment state label in the predefined sentiment state label set; Emotional state labels with probability values ​​exceeding a first preset threshold are identified as the current user node's emotional state value.

[0009] Preferably, the user's utterance text corresponding to the current user node is processed to extract the request, thereby obtaining the request feature set of the current user node, including: Domain keyword entities are extracted from the user's spoken text using named entity recognition technology; The extracted domain keyword entities are matched with a preset demand type knowledge base, which stores the mapping relationship between keyword entities and standard demand types. For each successfully matched standard appeal type, the confidence level is calculated based on the standard appeal type and the user's utterance text, and appeal features containing the standard appeal type and corresponding confidence level are generated. Combine at least one generated claim feature into the claim feature set of the current user node.

[0010] Preferably, based on a preset emotion transition determination rule, at least one emotion inflection point user node is identified from the user node sequence, including: The user emotion sequence is smoothed and filtered, and the emotion state values ​​of two adjacent user nodes in the smoothed and filtered user emotion sequence are compared sequentially. When the difference between the emotional state values ​​of two adjacent user nodes exceeds the second preset threshold, the latter user node is marked as a candidate emotional inflection point user node. Verify all marked candidate emotional inflection point user nodes, filter out candidate points that do not conform to the preset context logic, and determine the remaining candidate points as emotional inflection point user nodes.

[0011] Preferably, based on predetermined backtracking rules, a backtracking search is performed on the dialogue log data to locate key triggering customer service nodes, including: Starting from the dialogue round corresponding to the user node at the emotional inflection point, examine the customer service text of each round backwards; For each round of customer service dialogue text detected, at least one of the following triggering conditions shall be applied sequentially for judgment: First trigger condition: Determine whether the customer service text does not respond to the request type corresponding to the request feature with the highest confidence in the directly adjacent previous round of user nodes; Second trigger condition: Determine whether the customer service text contains more than a preset proportion of general template statements and does not contain specific action information; The third triggering condition is to determine whether the key factual information in the customer service text contradicts the key factual information in an earlier customer service text. The customer service node corresponding to the dialogue round in which the customer service message text that first meets any of the aforementioned triggering conditions is located is positioned as the key triggering customer service node.

[0012] Preferably, based on the emotional inflection point user node and the demand feature set of a predetermined number of subsequent user nodes, the demand evolution pattern is determined, including: Extract the first demand type with the highest confidence in the demand feature set of the emotional inflection point user node, and use it as the core demand. Analyze the changes in the most confident type of appeal among a predetermined number of user nodes after the emotional inflection point user node; If there is a demand type that is the same as the aforementioned core demand and whose confidence level continues to increase, then the demand evolution mode is determined to be demand escalation. If the confidence level of the aforementioned core appeal decreases significantly and a dominant appeal type different from the aforementioned core appeal emerges, then the appeal evolution pattern is determined to be appeal shift. If the confidence level of the aforementioned core appeal remains unchanged, and at least one new appeal type with a confidence level exceeding the third preset threshold is added, then the appeal evolution mode is determined to be an appeal addition.

[0013] Preferably, based on the analysis results of all user nodes at emotional inflection points and their correlations, dialogue summary data is generated, including: For each user node at an emotional inflection point, an independent data structure is generated. The data structure includes at least the inflection point location information, the description of the emotional jump, the text references and trigger reason classifications of the associated key trigger customer service nodes, and the demand evolution pattern. Arrange the data structures corresponding to all user nodes at emotional inflection points according to their position in the user node sequence; Add dialogue overview information to the beginning of the arranged data structure sequence and service evaluation information based on all triggering reasons to the end to form dialogue summary data.

[0014] Preferred options also include: Based on the service evaluation information, service improvement suggestion data is generated, which includes optimization suggestion items categorized by high-frequency triggering reasons; The service improvement suggestion data is associated with the dialogue summary data, stored, and then output.

[0015] Secondly, this application provides a machine learning-based text generation auxiliary processing system, comprising: The text acquisition module is used to acquire the dialogue log data of the target customer service conversation. The dialogue log data contains multiple rounds of dialogue records arranged in chronological order. Each round of dialogue records includes user speech text and customer service speech text. The node sequence construction module is used to construct a user node sequence based on the dialogue log data, wherein each user node corresponds to the user's speech text in one round of dialogue record, and the user nodes in the user node sequence are arranged in chronological order; The processing module is configured to perform the following processing for each current user node in the user node sequence: The user's utterance text corresponding to the current user node is input into the sentiment analysis model to obtain the sentiment state value of the current user node. The sentiment state value comes from a predefined sentiment state label set. The user speech text corresponding to the current user node is processed to extract the request, and the request feature set of the current user node is obtained. The request feature set contains at least one request feature, and each request feature includes a request type identifier and a confidence level. The emotion sequence generation module is used to generate user emotion sequences based on the processed emotion state values ​​of all user nodes; The identification module is used to traverse the user emotion sequence and identify at least one emotional inflection point user node from the user node sequence based on a preset emotion transition judgment rule. The correlation analysis module is used to perform inflection point correlation analysis for each identified user node at an emotional inflection point: The positioning module is used to locate the key triggering customer service node by taking the position of the emotional inflection point user node in the user node sequence as the starting point and performing backtracking search in the dialogue log data based on the predetermined backtracking rules. The key triggering customer service node is the customer service node before the emotional inflection point user node and whose customer service text is determined to meet the preset triggering conditions. The pattern determination module is used to determine the demand evolution pattern based on the demand feature set of the emotional inflection point user node and a predetermined number of subsequent user nodes, starting from the position of the emotional inflection point user node in the user node sequence. The demand evolution pattern is used to describe the change type of the core demand. The dialogue summary generation module is used to generate dialogue summary data based on all emotional inflection point user nodes and their correlation analysis results. The dialogue summary data includes at least the identifier of each emotional inflection point user node, the customer service text of the corresponding key triggering customer service node, and the corresponding demand evolution pattern.

[0016] Thirdly, this application provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform any of the above-described machine learning-based text generation auxiliary processing methods.

[0017] In summary, this application includes at least one of the following beneficial technical effects: 1. This application provides a machine learning-based text generation auxiliary processing method. By elevating summary generation from describing facts to explaining cause and effect and evolution, compared with traditional methods that only statically summarize dialogue content, this invention identifies emotional inflection points, traces back to locate key trigger customer service nodes, and analyzes the evolution pattern of demands. It not only reveals what happened in the dialogue, but also deeply analyzes why it happened and how it will develop in the future. The generated dialogue summary data is a structured analysis report that directly points to the key conflict points, responsible links, and user intent lifecycle in service interaction. It provides insights with direct action guidance for service review and optimization, rather than simply listing information. It also realizes the automation and structured location of the critical path of service experience deterioration. Traditional methods rely on human experience to subjectively find problems from lengthy logs, which is inefficient. This invention, through preset emotional jump judgment rules and trigger conditions, can automatically and objectively identify the "critical moment" that leads to the negative turning point of user emotions and its directly related customer service response, accurately locate service shortcomings, and greatly improve the efficiency and accuracy of service quality assessment. 2. By extracting demands and constructing a set of demand features in each round, and analyzing the evolution pattern of demands after the emotional inflection point, it is possible to clearly present the flow and change trajectory of users' core demands in the service process. This allows managers to go beyond the limitations of single-round dialogue, understand the complete life cycle of user demands from a global perspective, identify whether the service process has effectively resolved the problem or triggered the spread or escalation of demands, and thus evaluate the effectiveness of the service strategy. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a text generation auxiliary processing method based on machine learning, according to an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of a machine learning-based text generation auxiliary processing system according to an embodiment of this application. Detailed Implementation

[0021] The following description, in conjunction with the implementation of this invention, is merely an example and illustration of the concept of this invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention.

[0022] Overview of this application: In existing technologies, customer service dialogue summary generation largely relies on text compression or statistical keyword extraction, making it difficult to balance information completeness and service insight depth. Traditional methods are prone to homogenization issues in summaries when users experience significant emotional fluctuations during the dialogue, leading to the omission of critical service errors. Existing methods cannot simultaneously analyze the dynamic relationship between user emotional state and core demands. Especially when inappropriate service responses escalate user dissatisfaction, static text summarization models exhibit structural biases, failing to meet the decision-making needs for accurate service quality review and improvement.

[0023] To address the aforementioned issues, the inventors discovered a correlation between user emotional inflection points and specific customer service response behaviors. They implemented summary value compensation by constructing an emotion-appeal coupling analysis sequence. During the research, it was found that drastic negative emotional shifts are highly sensitive to service errors, while the evolution path of core appeals, though delayed in response, is highly targeted. Therefore, they proposed a dual-path analysis approach that dynamically initiates backtracking and forward analysis based on emotional inflection points. Further business validation incorporated the mapping relationship between emotional shift characteristics, key trigger sentences, and appeal evolution patterns into the summary generation logic, forming a structured output mechanism focused on service quality and effectiveness.

[0024] Specifically, the processing system first sequentially acquires customer service dialogue logs and constructs a sequence of user nodes. By analyzing the semantic content of each user node in the sequence, it calculates its emotional state value and extracts its set of request features in parallel. When the difference in emotional state values ​​between adjacent user nodes exceeds a set threshold, the system automatically marks the node as an emotional inflection point and initiates correlation analysis: it traces back the dialogue history to locate customer service responses that meet preset trigger conditions as key trigger sentences, and it tracks changes in the user's request feature set forward to determine the evolution pattern of the request. When generating the final summary, the system uses all identified emotional inflection points as the core framework, integrates the analysis of their corresponding key trigger sentences and the conclusions of the request evolution, and generates a structured report using a pre-set text template. For dialogues where no emotional inflection points are detected, a conventional summary model based on statistical analysis of the request feature set is used to output the results.

[0025] Compared to existing technologies, traditional methods rely on weighted compression or surface feature extraction of dialogue text, lacking in-depth analysis of interactive dynamics and causal logic, and are prone to losing key information in complex service conflict scenarios. This solution innovatively integrates temporal emotion state tracking and discrete demand feature extraction, establishing a dual-path analysis mechanism driven by emotion inflection points to deepen the value of the summary content. Unlike existing static summarization models, this solution can intelligently focus the analysis based on real-time interactive emotional tension and present the causal chain of the service process through structured association mapping, significantly improving the effectiveness of the summary results in supporting operational optimization decisions.

[0026] Through the above technical solutions, this application effectively overcomes the problem that traditional dialogue summarization cannot reveal the inherent contradictions and evolutionary processes of service interactions, improving the decision support accuracy of the summarization while ensuring the efficiency of automated processing. The dynamic sentiment inflection point detection and dual-path analysis mechanism take into account both the immediacy of service error response and the process of user demand evolution, and the structured output ensures the efficiency of information acquisition in front-line management and training scenarios.

[0027] The following is in conjunction with the appendix Figure 1-2 This application will be described in further detail.

[0028] Example 1 This application discloses a machine learning-based text generation auxiliary processing method.

[0029] Reference Figure 1 A machine learning-based text generation auxiliary processing method includes the following steps: Obtain the dialogue log data of the target customer service conversation. The dialogue log data contains multiple rounds of dialogue records arranged in chronological order. Each round of dialogue records includes user speech text and customer service speech text. Based on the dialogue log data, a user node sequence is constructed, wherein each user node corresponds to the user's speech text in one round of dialogue record, and the user nodes in the user node sequence are arranged in chronological order; For each current user node in the user node sequence, perform the following processing: The user's utterance text corresponding to the current user node is input into the sentiment analysis model to obtain the sentiment state value of the current user node. The sentiment state value comes from a predefined sentiment state label set. The user speech text corresponding to the current user node is processed to extract the request, and the request feature set of the current user node is obtained. The request feature set contains at least one request feature, and each request feature includes a request type identifier and a confidence level. Based on the processed emotional state values ​​of all user nodes, a user emotional sequence is generated; Traverse the user emotion sequence and, based on a preset emotion transition judgment rule, identify at least one emotion inflection point user node from the user node sequence; For each identified user node at an emotional inflection point, perform inflection point correlation analysis: Starting from the position of the emotional inflection point user node in the user node sequence, based on the predetermined backtracking rules, backtracking search is performed in the dialogue log data to locate the key triggering customer service node. The key triggering customer service node is the customer service node before the emotional inflection point user node and whose customer service text is determined to meet the preset triggering conditions. Starting from the position of the emotional inflection point user node in the user node sequence, the demand evolution pattern is determined based on the demand feature set of the emotional inflection point user node and a predetermined number of subsequent user nodes. The demand evolution pattern is used to describe the change type of the core demand. Based on the analysis results of all user nodes at emotional inflection points and their correlations, dialogue summary data is generated. The dialogue summary data includes at least the identifier of each user node at an emotional inflection point, the customer service text of the corresponding key triggering customer service node, and the corresponding demand evolution pattern.

[0030] In this embodiment of the application, the dialogue log data refers to the structured conversation records exported from the customer service system database. Specifically, it includes dialogue rounds sorted by interaction timestamps. Each round clearly distinguishes between the text of the user's initiated speech and the text of the customer service personnel's reply, and is used as the original data source for subsequent time series analysis and feature extraction.

[0031] Among them, the user node sequence refers to the sequence formed by abstracting the user's speech text in each round of the dialogue log into an independent data node and arranging it in chronological order. Specifically, it is achieved by assigning a unique index to each user's speech text and associating it with its original context, serving as the basic data structure for carrying out emotion and appeal analysis.

[0032] Among them, the emotion state value refers to the discrete classification label output by the emotion analysis model after calculating the user's utterance text. Specifically, it comes from a predefined set of labels containing categories such as calm, confused, dissatisfied, and angry, and is used to quantitatively represent the user's emotional state in a single round of interaction.

[0033] The appeal feature set refers to one or more structured feature sets that express user intent, which are parsed from a single round of user discourse text. Each feature specifically includes a standardized appeal type identifier and a confidence score that indicates the degree of certainty about the existence of the intent, and is used to objectively describe the user's core needs in this round.

[0034] Among them, the user emotion sequence refers to the ordered label sequence formed by arranging all user nodes in chronological order and extracting their corresponding emotion state values. Specifically, it is constructed by sequentially reading the emotion state value of each node in the user node sequence, and is used to observe the continuous change trend of user emotions in the dialogue process as a whole.

[0035] Among them, the emotional inflection point user node refers to a specific user node in the user's emotional sequence whose emotional state value is different from the emotional state value of the previous user node and meets the preset jump judgment rule. Specifically, it is identified by comparing whether the ordinal difference between adjacent emotional state values ​​in the predefined label set exceeds a threshold, which is used to locate the key moment when the user's emotions deteriorate significantly.

[0036] Among them, the key trigger customer service node refers to the customer service response node that is identified during the retrospective search process as potentially directly triggering the subsequent emotional inflection point. The determination is based on the rule analysis of whether the customer service text meets the preset trigger conditions (such as failure to respond to core demands, use of evasive language, or provision of contradictory information), which is used to associate and reveal the error points in the service interaction.

[0037] Among them, the demand evolution pattern refers to the core demand change type summarized by comparing and analyzing the demand feature sets of user nodes at the emotional inflection point and several subsequent user nodes. Specifically, it includes categories such as demand escalation, demand transfer or demand addition, which is used to depict the dynamic development path of the user's core intention after the emotional inflection point occurs.

[0038] Among them, the dialogue summary data refers to the structured report data with the aforementioned analysis results as the core content. Specifically, it is generated by integrating the corresponding key trigger customer service node text references and the conclusions of the evolution pattern of the request, with each emotional inflection point user node as the organizational unit. It is used to replace the traditional chronological summary and directly serve the review of service quality and efficiency.

[0039] The core innovation of this solution lies in the construction of a dialogue summary generation mechanism based on temporal emotion inflection point detection and dual-path correlation analysis. By dynamically locating key points of user emotional deterioration and intelligently associating and tracing the causes of service errors and the evolution path of demands, it achieves a leap from surface information compression to in-depth insight into service interaction contradictions, and solves the technical bottleneck of traditional summarization methods that cannot reveal the causal logic and dynamic evolution of the service process.

[0040] The working process and principle of this solution are as follows: First, acquire time-sorted dialogue log data of the target customer service conversations, and construct a user node sequence based on this data as the basis for analysis. Next, for each user node in the sequence, perform parallel sentiment analysis to obtain its sentiment state value, and extract its request feature set. Generate a user sentiment sequence based on the sentiment state values ​​of all nodes, and identify user nodes that indicate significant sentiment deterioration through preset rules. Then, for each identified sentiment inflection point, initiate dual-path correlation analysis: one path uses predetermined backtracking rules to locate the key triggering customer service node leading to sentiment deterioration, while the other path analyzes the evolution pattern of the core requests based on the request feature set. Finally, integrate all sentiment inflection points and their correlation analysis results to generate structured dialogue summary data centered on service error points and the evolution of requests. In this way, the system can automatically focus on and deeply analyze key conflicts and their evolution in the service process, providing direct and powerful decision-making basis for service quality assessment and optimization.

[0041] Furthermore, the user's utterance text corresponding to the current user node is input into the sentiment analysis model to obtain the sentiment state value of the current user node, including: The user speech text corresponding to the current user node is segmented and vectorized to obtain a text vector. The text vector is input into the sentiment analysis model; Obtain the classification results output by the sentiment analysis model, wherein the classification results include the probability of belonging to each sentiment state label in the predefined sentiment state label set; Emotional state labels with probability values ​​exceeding a first preset threshold are identified as the current user node's emotional state value.

[0042] In one embodiment, the above steps specifically illustrate the application details of the sentiment analysis model. Word segmentation and vectorization are mature preliminary steps in natural language processing, converting unstructured text into numerical features (i.e., text vectors) that the model can process. Alternative techniques include using different word segmentation tools or word vector models. After inputting the text vectors into the sentiment analysis model, the core computational logic of the model is typically based on the Softmax function, outputting the probability distribution of each emotional state label. For example, for a four-class classification model (calm, confused, dissatisfied, angry), its output might be a four-dimensional vector P=[p1,p2,p3,p4], where each term represents the probability of belonging to the corresponding category, and the sum is 1. A first preset threshold (e.g., 0.6) plays a crucial filtering role here: if the probability of a certain category exceeds this threshold, it is directly classified as belonging to that category, which improves the certainty and reliability of the judgment; if all probabilities do not exceed the threshold, the category with the highest probability is usually used as the output. This dynamic adjustment logic ensures a clear output even when the model's confidence level is low, while simultaneously providing a stricter threshold for high-confidence judgments. Its technical effect is a balance between classification accuracy and stability. The results of this step provide the most basic quantitative data for subsequently constructing user emotion sequences and identifying emotion inflection points.

[0043] Furthermore, the user's discourse text corresponding to the current user node is processed to extract the request, thereby obtaining the request feature set of the current user node, including: Domain keyword entities are extracted from the user's spoken text using named entity recognition technology; The extracted domain keyword entities are matched with a preset demand type knowledge base, which stores the mapping relationship between keyword entities and standard demand types. For each successfully matched standard appeal type, the confidence level is calculated based on the standard appeal type and the user's utterance text, and appeal features containing the standard appeal type and corresponding confidence level are generated. Combine at least one generated claim feature into the claim feature set of the current user node.

[0044] In one embodiment, the above steps detail how to structurally extract requests from user discourse. The role of named entity recognition (NID) technology in extracting domain-specific keyword entities is to identify core elements in the discourse that are strongly related to the business, such as "refund," "logistics," and "complaint." Alternative techniques include dictionary-based matching or deep learning sequence labeling models. Subsequently, matching is performed with a pre-defined request type knowledge base. This maps specific entities to standardized, limited request categories (e.g., "inquire about refund policy," "complain about logistics delays"), which is crucial for achieving request generalization and comparability. The knowledge base can be constructed based on historical data mining or defined by business experts. For each matched standard request type, calculating the confidence level quantifies the strength of the user's intention in this round. A feasible calculation logic is: Confidence Level = (Number of Keyword Occurrences / Total Number of Keywords) * Emotional Intensity Factor. The emotional intensity factor can be mapped based on the current emotional state value (e.g., requests in an angry state have higher confidence weighting). Finally, a request feature set is generated, which technically provides a lightweight, computable, structured intention description for each round of user discourse. This feature set serves as a direct data source for subsequently determining the evolution pattern of the appeal. For example, the user's statement, "How much longer do I have to wait for my refund? It's been three days already!" can be analyzed using named entity recognition technology to extract the keywords "refund," "how long," and "three days," matching them to the standard appeal type "urging refund progress." Combined with the user's emotional state of "dissatisfaction," a high confidence level (e.g., 0.85) can be calculated.

[0045] Furthermore, based on preset emotion transition determination rules, at least one emotion inflection point user node is identified from the user node sequence, including: The user emotion sequence is smoothed and filtered, and the emotion state values ​​of two adjacent user nodes in the smoothed and filtered user emotion sequence are compared sequentially. When the difference between the emotional state values ​​of two adjacent user nodes exceeds the second preset threshold, the latter user node is marked as a candidate emotional inflection point user node. Verify all marked candidate emotional inflection point user nodes, filter out candidate points that do not conform to the preset context logic, and determine the remaining candidate points as emotional inflection point user nodes.

[0046] In one embodiment, the above steps constitute the core logic for identifying key emotional inflection points. First, smoothing the user's emotional sequence eliminates random errors or instantaneous fluctuations that may exist in single-round emotional analysis. For example, an isolated "anger" point is corrected to the mean of its preceding and following emotions (such as "dissatisfaction"), ensuring the robustness of subsequent detection. This can be achieved using simple algorithms such as moving averages. Next, sequentially comparing the emotional state values ​​of adjacent nodes detects drastic changes in emotion using a sliding window approach. The difference calculation here can be simply performed by treating a predefined set of emotional state labels as ordered variables and calculating their ordinal differences (e.g., "confusion" is 2, "anger" is 4, the difference is 2). The second preset threshold (e.g., set to 2) is crucial here, directly defining what constitutes a "drastic" jump. If the threshold is set too high, some important emotional deterioration points may be missed; if the threshold is too low, a large number of false alarms may occur, interfering with subsequent analysis. Marking the next node that meets the criteria as a candidate point is based on the reasonable assumption that "emotional deterioration is triggered by an event in the previous moment and manifests in the next moment." The final verification step is another innovative aspect of this solution. Its function is to filter out illogical candidate points using contextual logic. For example, if several rounds of dialogue preceding a candidate point consisted of automated robot responses without any substantial interaction, then the emotional shift at that point might not be due to a service issue and should be filtered out. The technical effect of this step is to accurately pinpoint the significant negative emotional turning point for the user, truly triggered by the service interaction process, providing high-quality input for subsequent root cause analysis.

[0047] Furthermore, based on predetermined backtracking rules, a backtracking search is performed on the dialogue log data to locate key triggering customer service nodes, including: Starting from the dialogue round corresponding to the user node at the emotional inflection point, examine the customer service text of each round backwards; For each round of customer service dialogue text detected, at least one of the following triggering conditions shall be applied sequentially for judgment: First trigger condition: Determine whether the customer service text does not respond to the request type corresponding to the request feature with the highest confidence in the directly adjacent previous round of user nodes; Second trigger condition: Determine whether the customer service text contains more than a preset proportion of general template statements and does not contain specific action information; The third triggering condition is to determine whether the key factual information in the customer service text contradicts the key factual information in an earlier customer service text. The customer service node corresponding to the dialogue round in which the customer service message text that first meets any of the aforementioned triggering conditions is located is positioned as the key triggering customer service node.

[0048] In one embodiment, the above steps specifically define how to trace back to find the root cause of service errors. Checking backwards from the emotional inflection point simulates the logic of manually reviewing a scenario to find the source of a problem. The three triggering conditions applied are a summary of experience regarding common service error types in this solution, with clear business orientation. The first triggering condition (ignoring demands) checks whether customer service directly ignored the user's strongest demands, which is the most direct cause of dissatisfaction. The second triggering condition (vague evasion) checks whether customer service used ineffective appeasement language without providing substantive solutions, which exacerbates the user's sense of powerlessness and anger. The third triggering condition (information contradiction) checks whether the customer service team provided inconsistent information, which seriously damages trust. The preset percentage (e.g., 70%) and the judgment of "key factual information" (e.g., time, policy terms) can be achieved through rules or simple text matching. The innovation of this step lies in transforming qualitative service error standards into calculable, automated judgment rules. Its technical effect is to automatically and objectively locate the specific customer service response round and its error type that led to the emotional inflection point, providing conclusive "evidence" for generating actionable summaries. For example, when it is identified that a user is angry because they did not receive a response to "urging a refund", the customer service node found through backtracking may contain a template phrase such as "Please rest assured, we will expedite the process", which will then be captured by the second trigger condition.

[0049] Furthermore, based on the emotional inflection point user node and the demand feature set of a predetermined number of subsequent user nodes, the demand evolution pattern is determined, including: Extract the first demand type with the highest confidence in the demand feature set of the emotional inflection point user node, and use it as the core demand. Analyze the changes in the most confident type of appeal among a predetermined number of user nodes after the emotional inflection point user node; If there is a demand type that is the same as the aforementioned core demand and whose confidence level continues to increase, then the demand evolution mode is determined to be demand escalation. If the confidence level of the aforementioned core appeal decreases significantly and a dominant appeal type different from the aforementioned core appeal emerges, then the appeal evolution pattern is determined to be appeal shift. If the confidence level of the aforementioned core appeal remains unchanged, and at least one new appeal type with a confidence level exceeding the third preset threshold is added, then the appeal evolution mode is determined to be an appeal addition.

[0050] In one application embodiment, the above steps define how to analyze the trend of user intent after the emotional inflection point occurs. Extracting the initial core demands helps identify the focal issue before the emotional outburst. Analyzing the changes in demands at subsequent predetermined number of nodes (e.g., 3-5 rounds) helps capture the evolution trend of demands within a reasonable observation window, avoiding interference from more distant and irrelevant conversations. The judgment rules for the three evolution modes clearly distinguish different service impact consequences: "Demand escalation" means the original problem intensifies, with users demanding more urgency or adopting a stronger attitude; "Demand transfer" means users may lose confidence in resolving the original problem and raise new, or even irrelevant, complaints; "Added demands" means that on top of the unresolved original problem, users have added new dissatisfaction. "Continuously increasing confidence" can be judged by calculating whether the slope of the confidence level for this type of demand is greater than zero in subsequent rounds; "significant decline" can be defined by judging whether its confidence level falls out of the top few or below a certain threshold; the third preset threshold is used to define the importance of "added" demands. The technical effect of this step is to quantify the business impact of service failures, enabling the summary to not only explain "where the mistake was made" but also "what the user wants to do next," greatly enriching the decision support value of the summary.

[0051] Furthermore, based on the analysis results of all user nodes at emotional inflection points and their correlations, dialogue summary data is generated, including: For each user node at an emotional inflection point, an independent data structure is generated. The data structure includes at least the inflection point location information, the description of the emotional jump, the text references and trigger reason classifications of the associated key trigger customer service nodes, and the demand evolution pattern. Arrange the data structures corresponding to all user nodes at emotional inflection points according to their position in the user node sequence; Add dialogue overview information to the beginning of the arranged data structure sequence and service evaluation information based on all triggering reasons to the end to form dialogue summary data.

[0052] In one embodiment, the above steps define a structured method for generating the final summary data. An independent data structure is generated for each inflection point, encapsulating all the analytical results from previous steps (inflection point, triggering cause, evolution of demands) into a complete story unit. This facilitates modular organization and presentation of information. Arranging these structures chronologically reconstructs the natural timeline of the dialogue conflict, facilitating understanding of the escalation process. Adding a dialogue overview (such as dialogue ID, total duration, and final result) at the beginning and service evaluation information (such as "three evasive responses occurred in this round of dialogue") at the end provides readers with a global perspective and summarizing insights. The innovation of this step lies in the structural design of the output; it is not a continuous prose, but an analytical report assembled from standardized modules according to a timeline. Its technical effect is that the generated summary is highly structured, information-dense, and highlights key points, and is easily parsed and reused by other systems (such as quality inspection systems and training systems), achieving seamless integration from analysis to application.

[0053] Furthermore, it also includes: Based on the service evaluation information, service improvement suggestion data is generated, which includes optimization suggestion items categorized by high-frequency triggering reasons; The service improvement suggestion data is associated with the dialogue summary data, stored, and then output.

[0054] In one embodiment, the above steps represent a further value extension of the core solution. Generating improvement suggestion data based on service assessment information directly translates analytical conclusions into actionable improvement recommendations, achieving a closed loop from diagnosis to prescription. For example, if the service assessment identifies "vague evasion" as a frequent trigger, the improvement suggestion data might include specific items such as "for urging-related issues, the processing timeframe and current progress should be clearly stated in the response." Linking, storing, and outputting these two data points not only informs the customer service team "where things are not going well" but also guides them on "how to improve," significantly enhancing the practical application value of this patented method in actual customer service operations management, upgrading it from an analytical tool into a component of a complete service quality optimization solution.

[0055] Example 2 This application also discloses a machine learning-based text generation auxiliary processing system.

[0056] Reference Figure 2 A machine learning-based text generation auxiliary processing system includes: The text acquisition module is used to acquire the dialogue log data of the target customer service conversation. The dialogue log data contains multiple rounds of dialogue records arranged in chronological order. Each round of dialogue records includes user speech text and customer service speech text. The node sequence construction module is used to construct a user node sequence based on the dialogue log data, wherein each user node corresponds to the user's speech text in one round of dialogue record, and the user nodes in the user node sequence are arranged in chronological order; The processing module is configured to perform the following processing for each current user node in the user node sequence: The user's utterance text corresponding to the current user node is input into the sentiment analysis model to obtain the sentiment state value of the current user node. The sentiment state value comes from a predefined sentiment state label set. The user speech text corresponding to the current user node is processed to extract the request, and the request feature set of the current user node is obtained. The request feature set contains at least one request feature, and each request feature includes a request type identifier and a confidence level. The emotion sequence generation module is used to generate user emotion sequences based on the processed emotion state values ​​of all user nodes; The identification module is used to traverse the user emotion sequence and identify at least one emotional inflection point user node from the user node sequence based on a preset emotion transition judgment rule. The correlation analysis module is used to perform inflection point correlation analysis for each identified user node at an emotional inflection point: The positioning module is used to locate the key triggering customer service node by taking the position of the emotional inflection point user node in the user node sequence as the starting point and performing backtracking search in the dialogue log data based on the predetermined backtracking rules. The key triggering customer service node is the customer service node before the emotional inflection point user node and whose customer service text is determined to meet the preset triggering conditions. The pattern determination module is used to determine the demand evolution pattern based on the demand feature set of the emotional inflection point user node and a predetermined number of subsequent user nodes, starting from the position of the emotional inflection point user node in the user node sequence. The demand evolution pattern is used to describe the change type of the core demand. The dialogue summary generation module is used to generate dialogue summary data based on all emotional inflection point user nodes and their correlation analysis results. The dialogue summary data includes at least the identifier of each emotional inflection point user node, the customer service text of the corresponding key triggering customer service node, and the corresponding demand evolution pattern.

[0057] The above content is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention, they should all fall within the protection scope of the present invention.

[0058] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0059] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention.

Claims

1. A text generation auxiliary processing method based on machine learning, characterized in that, Includes the following steps: Obtain the dialogue log data of the target customer service conversation. The dialogue log data contains multiple rounds of dialogue records arranged in chronological order. Each round of dialogue records includes user speech text and customer service speech text. Based on the dialogue log data, a user node sequence is constructed, wherein each user node corresponds to the user's speech text in one round of dialogue record, and the user nodes in the user node sequence are arranged in chronological order; For each current user node in the user node sequence, perform the following processing: The user's utterance text corresponding to the current user node is input into the sentiment analysis model to obtain the sentiment state value of the current user node. The sentiment state value comes from a predefined sentiment state label set. The user speech text corresponding to the current user node is processed to extract the request, and the request feature set of the current user node is obtained. The request feature set contains at least one request feature, and each request feature includes a request type identifier and a confidence level. Based on the processed emotional state values ​​of all user nodes, a user emotional sequence is generated; Traverse the user emotion sequence and, based on a preset emotion transition judgment rule, identify at least one emotion inflection point user node from the user node sequence; For each identified user node at an emotional inflection point, perform inflection point correlation analysis: Starting from the position of the emotional inflection point user node in the user node sequence, based on the predetermined backtracking rules, backtracking search is performed in the dialogue log data to locate the key triggering customer service node. The key triggering customer service node is the customer service node before the emotional inflection point user node and whose customer service text is determined to meet the preset triggering conditions. Starting from the position of the emotional inflection point user node in the user node sequence, the demand evolution pattern is determined based on the demand feature set of the emotional inflection point user node and a predetermined number of subsequent user nodes. The demand evolution pattern is used to describe the change type of the core demand. Based on the analysis results of all user nodes at emotional inflection points and their correlations, dialogue summary data is generated. The dialogue summary data includes at least the identifier of each user node at an emotional inflection point, the customer service text of the corresponding key triggering customer service node, and the corresponding demand evolution pattern.

2. The text generation auxiliary processing method based on machine learning according to claim 1, characterized in that, The user's utterance text corresponding to the current user node is input into the sentiment analysis model to obtain the sentiment state value of the current user node, including: The user speech text corresponding to the current user node is segmented and vectorized to obtain a text vector. The text vector is input into the sentiment analysis model; Obtain the classification results output by the sentiment analysis model, wherein the classification results include the probability of belonging to each sentiment state label in the predefined sentiment state label set; Emotional state labels with probability values ​​exceeding a first preset threshold are identified as the current user node's emotional state value.

3. The text generation auxiliary processing method based on machine learning according to claim 1, characterized in that, The user's utterance text corresponding to the current user node is processed to extract the request, thereby obtaining the request feature set of the current user node, including: Domain keyword entities are extracted from the user's spoken text using named entity recognition technology; The extracted domain keyword entities are matched with a preset demand type knowledge base, which stores the mapping relationship between keyword entities and standard demand types. For each successfully matched standard appeal type, the confidence level is calculated based on the standard appeal type and the user's utterance text, and appeal features containing the standard appeal type and corresponding confidence level are generated. Combine at least one generated claim feature into the claim feature set of the current user node.

4. The text generation auxiliary processing method based on machine learning according to claim 1, characterized in that, Based on preset emotion transition determination rules, at least one user node at an emotion inflection point is identified from the user node sequence, including: The user emotion sequence is smoothed and filtered, and the emotion state values ​​of two adjacent user nodes in the smoothed and filtered user emotion sequence are compared sequentially. When the difference between the emotional state values ​​of two adjacent user nodes exceeds the second preset threshold, the latter user node is marked as a candidate emotional inflection point user node. Verify all marked candidate emotional inflection point user nodes, filter out candidate points that do not conform to the preset context logic, and determine the remaining candidate points as emotional inflection point user nodes.

5. The text generation auxiliary processing method based on machine learning according to claim 1, characterized in that, Based on predetermined backtracking rules, a backtracking search is performed on the dialogue log data to locate key triggering customer service nodes, including: Starting from the dialogue round corresponding to the user node at the emotional inflection point, examine the customer service text of each round backwards; For each round of customer service dialogue text detected, at least one of the following triggering conditions shall be applied sequentially for judgment: First trigger condition: Determine whether the customer service text does not respond to the request type corresponding to the request feature with the highest confidence in the directly adjacent previous round of user nodes; Second trigger condition: Determine whether the customer service text contains more than a preset proportion of general template statements and does not contain specific action information; The third triggering condition is to determine whether the key factual information in the customer service text contradicts the key factual information in an earlier customer service text. The customer service node corresponding to the dialogue round in which the customer service message text that first meets any of the aforementioned triggering conditions is located is positioned as the key triggering customer service node.

6. The text generation auxiliary processing method based on machine learning according to claim 1, characterized in that, Based on the emotional inflection point user node and the demand feature set of a predetermined number of subsequent user nodes, the demand evolution pattern is determined, including: Extract the first demand type with the highest confidence in the demand feature set of the emotional inflection point user node, and use it as the core demand. Analyze the changes in the most confident type of appeal among a predetermined number of user nodes after the emotional inflection point user node; If there is a demand type that is the same as the aforementioned core demand and whose confidence level continues to increase, then the demand evolution mode is determined to be demand escalation. If the confidence level of the aforementioned core appeal decreases significantly and a dominant appeal type different from the aforementioned core appeal emerges, then the appeal evolution pattern is determined to be appeal shift. If the confidence level of the aforementioned core appeal remains unchanged, and at least one new appeal type with a confidence level exceeding the third preset threshold is added, then the appeal evolution mode is determined to be an appeal addition.

7. The text generation auxiliary processing method based on machine learning according to claim 1, characterized in that, Based on the analysis results of all user nodes at emotional inflection points and their correlations, dialogue summary data is generated, including: For each user node at an emotional inflection point, an independent data structure is generated. The data structure includes at least the inflection point location information, the description of the emotional jump, the text references and trigger reason classifications of the associated key triggering customer service nodes, and the demand evolution pattern. Arrange the data structures corresponding to all user nodes at emotional inflection points according to their position in the user node sequence; Add dialogue overview information to the beginning of the arranged data structure sequence and service evaluation information based on all triggering reasons to the end to form dialogue summary data.

8. The text generation auxiliary processing method based on machine learning according to claim 7, characterized in that, Also includes: Based on the service evaluation information, service improvement suggestion data is generated, which includes optimization suggestion items categorized by high-frequency triggering reasons; The service improvement suggestion data is associated with the dialogue summary data, stored, and then output.

9. A machine learning-based text generation auxiliary processing system, applied to the machine learning-based text generation auxiliary processing method described in any one of claims 1-8, characterized in that, include: The text acquisition module is used to acquire the dialogue log data of the target customer service conversation. The dialogue log data contains multiple rounds of dialogue records arranged in chronological order. Each round of dialogue records includes user speech text and customer service speech text. The node sequence construction module is used to construct a user node sequence based on the dialogue log data, wherein each user node corresponds to the user's speech text in one round of dialogue record, and the user nodes in the user node sequence are arranged in chronological order; The processing module is configured to perform the following processing for each current user node in the user node sequence: The user's utterance text corresponding to the current user node is input into the sentiment analysis model to obtain the sentiment state value of the current user node. The sentiment state value comes from a predefined sentiment state label set. The user speech text corresponding to the current user node is processed to extract the request, and the request feature set of the current user node is obtained. The request feature set contains at least one request feature, and each request feature includes a request type identifier and a confidence level. The emotion sequence generation module is used to generate user emotion sequences based on the processed emotion state values ​​of all user nodes; The identification module is used to traverse the user emotion sequence and identify at least one emotional inflection point user node from the user node sequence based on a preset emotion transition judgment rule. The correlation analysis module is used to perform inflection point correlation analysis for each identified user node at an emotional inflection point: The positioning module is used to locate the key triggering customer service node by backtracking and searching the dialogue log data based on the position of the emotional inflection point user node in the user node sequence, starting from the position of the emotional inflection point user node. The key triggering customer service node is the customer service node before the emotional inflection point user node and whose customer service text is determined to meet the preset triggering conditions. The pattern determination module is used to determine the demand evolution pattern based on the demand feature set of the emotional inflection point user node and a predetermined number of subsequent user nodes, starting from the position of the emotional inflection point user node in the user node sequence. The demand evolution pattern is used to describe the change type of the core demand. The dialogue summary generation module is used to generate dialogue summary data based on all emotional inflection point user nodes and their correlation analysis results. The dialogue summary data includes at least the identifier of each emotional inflection point user node, the customer service text of the corresponding key triggering customer service node, and the corresponding demand evolution pattern.

10. A computer-readable storage medium, characterized in that: The system stores instructions that, when executed on a computer, cause the computer to perform a machine learning-based text generation auxiliary processing method as described in any one of claims 1 to 8.

Images