An intelligent customer service interaction sentence semantic understanding method, system and medium

Abstract

The application discloses an intelligent customer service interaction sentence semantic understanding method and system and a medium, which comprises the following steps: collecting interaction sentences and response data, performing preprocessing and generating a multi-round customer service interaction sentence sequence; inputting an improved SimCSE model, enhancing context semantic representation and generating an enhanced semantic representation set; extracting a round semantic change relationship, calculating semantic correlation and offset degree, and constructing a semantic evolution track; analyzing the semantic state change process, locating abnormal nodes and identifying implicit intention migration state; correcting the historical semantic state, fusing the current semantic information to generate a target user intention representation; and matching business intention information and outputting a dynamic semantic understanding result combined with the evolution track. Through the improved SimCSE model combined with semantic migration perception contrast learning and a historical semantic suppression mechanism, the application realizes accurate identification and dynamic semantic understanding of implicit intention migration in the multi-round interaction process of intelligent customer service.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and in particular to a method, system, and medium for semantic understanding of intelligent customer service interactive statements. Background Technology

[0002] With the development of artificial intelligence technology, intelligent customer service systems have been widely applied in service scenarios such as finance, e-commerce, healthcare, and government affairs. Existing intelligent customer service systems typically utilize natural language processing (NLP) technology to perform semantic analysis on user interactions, extracting keywords, semantic features, and intent information from user statements through text encoding models, and combining this with a business knowledge base to classify user questions, provide automatic responses, and match service processes. For multi-turn customer service interaction scenarios, some technical solutions further incorporate contextual information, improving the ability to identify user needs by analyzing the correlation between historical and current interaction statements.

[0003] However, existing intelligent customer service semantic understanding methods mainly rely on the static correlation between historical interaction information and current statement features. During multiple rounds of interaction, they typically assume that historical semantic states remain valid, failing to dynamically adjust the influence of historical information based on semantic changes during user communication. When a user shifts from their original business question to a new business need, such as from order inquiry to after-sales processing, existing methods are easily influenced by previous dialogue content, continuing to use historical intent, resulting in an inaccurate identification of the user's current true needs.

[0004] Existing methods based on semantic representation learning typically focus on improving the semantic similarity judgment ability between single-turn statements. They lack targeted analysis mechanisms for interactive statements in intelligent customer service scenarios that appear semantically similar but have different actual intentions. Furthermore, they lack the ability to dynamically identify semantic evolution processes, intent transfer nodes, and historical semantic failures, making it difficult to effectively handle complex interactive situations such as user omissions, topic switching, and changes in implicit needs. Therefore, existing technologies still suffer from insufficient semantic understanding accuracy, weak intent transfer recognition capabilities, and unreasonable utilization of contextual information in multi-turn intelligent customer service interactions.

[0005] Therefore, how to provide a method, system, and medium for semantic understanding of intelligent customer service interaction statements is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method, system, and medium for semantic understanding of intelligent customer service interaction statements. This invention fully utilizes natural language processing, contrastive learning, and semantic representation learning techniques. By improving the SimCSE model, it performs context-enhanced encoding, semantic transfer perception analysis, and historical semantic influence suppression on multi-turn customer service interaction statements. Combined with the construction of user interaction semantic evolution trajectory, implicit intent transfer detection, and target user intent correction, it achieves dynamic understanding of users' real needs in complex multi-turn customer service scenarios. It can effectively reduce the problem of intent misjudgment caused by historical semantic inertia and has the advantages of high semantic understanding accuracy, strong intent transfer recognition capability, and good adaptability to multi-turn interactions.

[0007] A method for semantic understanding of intelligent customer service interaction statements according to an embodiment of the present invention includes:

[0008] Collect user interaction statements, customer service response statements, interaction round information and time sequence information during the intelligent customer service interaction process, perform preprocessing on user interaction statements and customer service response statements, and generate multi-round customer service interaction statement sequences according to the interaction round arrangement;

[0009] The sequence of multi-round customer service interaction statements is input into the improved SimCSE model, which includes a context interaction enhancement coding layer, a semantic transfer awareness contrastive learning layer, and a historical semantic inhibition gating layer to generate an enhanced semantic representation set.

[0010] Based on the enhanced semantic representation set, the semantic change relationship between consecutive interaction rounds is extracted. The semantic association degree, semantic offset degree and intention change degree between the current user interaction statement and the historical interaction statement are calculated. The semantic evolution trajectory of user interaction is constructed according to the direction of semantic change.

[0011] Based on the semantic evolution trajectory of user interaction, the influence relationship between historical semantic states and current semantic states is analyzed, abnormal change nodes in the semantic evolution process are located, and the implicit intent migration state of users is identified by combining entity change information, intent offset information and semantic conflict information, and intent migration detection results are generated.

[0012] Based on the intent migration detection results, the historical semantic state and the current semantic state are correlated and corrected to reduce the impact of invalid historical semantic states on the current intent judgment, and the effective semantic information in the current interaction statement is integrated to generate the target user intent representation;

[0013] The system matches the target user's intent representation with the business intent information in the intelligent customer service business knowledge base, combines the semantic evolution trajectory of user interaction to determine the current user's real needs, and outputs dynamic semantic understanding results.

[0014] Optionally, the preprocessing of user interaction statements and customer service response statements includes performing text cleaning, entity extraction, and statement segmentation on the user interaction statements and customer service response statements.

[0015] Optionally, generating a multi-round customer service interaction statement sequence based on the interaction rounds includes:

[0016] Obtain the interaction time information, user identification information, and interaction round identification information corresponding to the user interaction statements. Sort the user interaction statements corresponding to the same user identification according to the order of interaction time to determine the relationship of consecutive interaction rounds.

[0017] Based on the interaction round identifier information, the user interaction statement is divided into multiple interaction statement fragments, and associated with the customer service response statement of the corresponding round, forming an interaction statement pair containing user input information and customer service feedback information;

[0018] Based on the sequential relationships between pairs of interactive statements, semantic continuity, topic continuation, and topic change relationships between adjacent rounds of interaction are extracted to construct a multi-round interactive context chain.

[0019] The interaction statement fragments and their corresponding context chains are combined according to the interaction rounds to generate a multi-round customer service interaction statement sequence that includes historical interaction information and current interaction information.

[0020] Optionally, the generation of the enhanced semantic representation set includes:

[0021] An improved SimCSE model is constructed based on the sentence encoding layer, semantic projection layer, and contrast learning layer of the SimCSE model. A context interaction enhancement encoding layer is embedded at the input of the SimCSE model, and a historical semantic suppression gating layer is embedded between the semantic projection layer and the contrast learning layer. The contrast learning layer is replaced with a semantic transfer-aware contrast learning layer, forming an improved SimCSE model for multi-turn customer service interaction statements.

[0022] Obtain the current user interaction statement, historical user interaction statement, and corresponding customer service response statement from a multi-round customer service interaction statement sequence. Associate and arrange the current user interaction statement, historical user interaction statement, and customer service response statement according to the interaction round, and input them into the context interaction enhancement coding layer to extract interaction round association features, context semantic association features, and semantic change features to generate a context-enhanced semantic representation.

[0023] The context-enhanced semantic representation is input into the sentence encoding layer and the semantic projection layer. Feature mapping is performed on the context-enhanced semantic representation to generate the basic semantic embedding representation corresponding to each round of interaction. The training sample set of the semantic transfer perception contrast learning layer is constructed based on the intent association relationship in the sequence of customer service interaction statements in multiple rounds.

[0024] The basic semantic embedding represents the input semantic transfer perception contrastive learning layer, based on the business meaning in the multi-round customer service interaction statement sequence. Figure 1 The system generates a consensus graph representation sample based on the relationship. It generates similar but different intentional hard negative samples based on semantically similar but different business intents. It performs comparative optimization based on the consensus graph representation sample and the similar but different intentional hard negative samples to generate an intent-distinguishing semantic representation.

[0025] The intent-discriminative semantic representation is input into the historical semantic suppression gating layer. The relationship between the semantic representations corresponding to historical user interaction statements and the semantic representations corresponding to current user interaction statements is analyzed. The proportion of historical semantic information retention is determined based on the degree of semantic change. The influence of historical semantic state on current semantic state is adjusted to generate dynamic semantic adjustment representation.

[0026] By integrating context-enhanced semantic representation, intent-discriminative enhanced semantic representation, and dynamic semantic adjustment representation, and arranging them according to the interaction rounds corresponding to the multi-round customer service interaction statement sequence, an enhanced semantic representation set containing semantic preservation features, semantic shift features, and intent transfer features is generated.

[0027] Optionally, constructing the semantic evolution trajectory of user interaction based on the direction of semantic change includes:

[0028] Obtain semantic preservation features, semantic offset features, and intent transfer features corresponding to each interaction round in the enhanced semantic representation set. Establish continuous semantic state nodes according to the interaction round order in the multi-round customer service interaction statement sequence, and associate each semantic state node with the corresponding interaction statement to generate a semantic state node set.

[0029] Calculate the semantic association degree between semantic state nodes corresponding to adjacent interaction rounds, determine the semantic continuity relationship between consecutive interaction rounds based on the feature matching relationship between the semantic representation corresponding to the current user interaction statement and the semantic representation corresponding to the historical interaction statement, and generate a semantic association feature set.

[0030] Based on the semantic association feature set, the feature changes between adjacent semantic state nodes are analyzed, and the semantic preservation feature change, semantic offset feature change, and intent migration feature change are extracted. The semantic offset degree between the current user interaction statement and the historical interaction statement is determined according to the feature change, and a semantic offset feature set is generated.

[0031] Based on the degree of semantic association and semantic offset, the semantic state nodes corresponding to each interaction round are connected. A positive association relationship is established between continuous semantic state nodes with the same direction of semantic change. Semantic state nodes whose direction of semantic change changes are marked as semantic change nodes, thus generating the semantic evolution trajectory of user interaction.

[0032] Based on the semantic evolution trajectory analysis of user interaction, the evolution path between each semantic state node is analyzed, and the degree of intent change corresponding to different interaction rounds is determined by combining the change amount of intent migration features. This generates a semantic evolution trajectory of user interaction that includes semantic state change path and intent change information.

[0033] Optionally, the generation of intent migration detection results includes:

[0034] Obtain each semantic state node and its corresponding evolution path in the semantic evolution trajectory of user interaction, extract the historical semantic state information, current semantic state information, semantic preservation features, semantic offset features and intent migration features corresponding to each semantic state node, and construct a semantic state influence analysis set;

[0035] Based on the semantic state influence analysis set, the relationship between historical semantic states and current semantic states is analyzed. The influence relationship of historical semantic states is determined according to the degree of semantic association and semantic preservation of historical semantic states and the degree of semantic shift of current semantic states, and historical semantic influence features are generated.

[0036] Based on the historical semantic influence characteristics, change detection is performed on continuous semantic state nodes in the semantic evolution trajectory of user interaction, identifying semantic state nodes whose semantic deviation exceeds the preset change range, and combining the corresponding interaction round information to determine abnormal change nodes in the semantic evolution process.

[0037] Obtain the current user interaction statement and historical user interaction statement corresponding to the abnormal change node, extract the entity change information, intent offset information and semantic conflict information between the two, perform correlation analysis on the entity change information, intent offset information and semantic conflict information, and generate implicit intent migration features.

[0038] Based on the implicit intent migration feature analysis, the direction of semantic state change before and after the abnormal change node is analyzed to determine whether the intent corresponding to the current user interaction statement is separated from the intent corresponding to the historical semantic state. The semantic state node where the intent changes is marked as intent migration node, and intent migration detection results containing intent migration node, intent migration direction and intent migration state are generated.

[0039] Optionally, generating the target user intent representation includes:

[0040] Obtain the intent migration node, intent migration direction and intent migration status from the intent migration detection results, read the historical semantic status and current semantic status associated with the corresponding intent migration node, extract the historical semantic representation, current semantic representation and semantic offset features, and construct a semantic state correction set;

[0041] Based on the semantic state correction set analysis, the semantic association between historical semantic state and current semantic state is analyzed. The effective association range of historical semantic state is determined according to the intent migration state. The effective historical semantic information and invalid historical semantic information related to the current user interaction statement are identified, and historical semantic state correction information is generated.

[0042] Based on historical semantic state correction information, the influence of failed historical semantic states on the current semantic state is adjusted, suppression processing is performed on failed historical semantic information, and contextual features corresponding to valid historical semantic information are retained to generate corrected historical semantic representations;

[0043] The modified historical semantic representation is fused with the current semantic representation. The entity change information, intent offset information and semantic transfer features in the current interaction statement are combined to supplement the intent features corresponding to the current user interaction statement and generate a fused semantic intent representation.

[0044] Based on the analysis of the target business intent corresponding to the current user interaction statement, the fused semantic intent representation is verified according to the intent change path in the semantic evolution trajectory to determine the user's true intent corresponding to the current interaction stage and generate the target user intent representation.

[0045] Optionally, the output dynamic semantic understanding result includes:

[0046] Obtain the current user intent features, intent change path and semantic evolution trajectory information from the target user intent representation, extract the business semantic features, business entity features and business association features corresponding to each business intent in the intelligent customer service business knowledge base, and construct a business intent matching set;

[0047] Based on the business intent matching set, the semantic matching relationship between the target user intent representation and each business intent information is calculated. Based on the current user intent features, business semantic features and business entity features, the candidate business intent set is determined and sorted according to the degree of matching.

[0048] The candidate business intent set is correlated with the semantic evolution trajectory of user interaction. The intent migration nodes, intent migration directions and intent change paths in the semantic evolution trajectory are read. The consistency relationship between the candidate business intent and the current user intent representation is verified, and the business intent verification result is generated.

[0049] Based on the business intent verification results, target business intents corresponding to the current user's actual needs are filtered, and customer service processing flow information, service response information and interaction status information corresponding to the target business intents are associated to generate dynamic semantic understanding information.

[0050] The dynamic semantic understanding information is organized according to the current interaction round, intent migration state, and business processing type to generate dynamic semantic understanding results that include the user's current intent, semantic change process, and business processing direction.

[0051] A semantic understanding system for intelligent customer service interaction statements according to an embodiment of the present invention includes:

[0052] The interactive data acquisition module is used to collect user interaction statements, customer service response statements, and interaction information, and generate a multi-round customer service interaction statement sequence after preprocessing.

[0053] The semantic enhancement representation module is used to generate an enhanced semantic representation set by fusing contextual information, intent transfer information and historical semantic information based on the improved SimCSE model;

[0054] The semantic trajectory construction module is used to analyze the semantic change relationship between rounds based on the enhanced semantic representation set, calculate semantic association, offset and intent change features, and construct the semantic evolution trajectory of user interaction.

[0055] The intent migration analysis module is used to analyze semantic state changes based on the semantic evolution trajectory of user interaction, locate abnormal change nodes and identify implicit intent migration states, and generate intent migration detection results.

[0056] The intent state correction module is used to adjust the influence of historical semantic states based on intent migration detection results and integrate current interaction semantic information to generate a target user intent representation.

[0057] The semantic result generation module is used to match the target user's intent representation with business intent information, combine semantic evolution trajectory to determine the user's real needs, and output dynamic semantic understanding results.

[0058] Optionally, the computer-readable storage medium stores computer program instructions, which, when executed by a processor, perform a method for semantic understanding of intelligent customer service interaction statements.

[0059] The beneficial effects of this invention are:

[0060] This invention constructs an intelligent customer service interaction statement semantic understanding method based on an improved SimCSE model. The method inputs a multi-round customer service interaction statement sequence into an improved SimCSE model that includes a context interaction enhancement coding layer, a semantic transfer perception contrast learning layer, and a historical semantic suppression gating layer. The method dynamically analyzes the semantic relationship between the current interaction statement and the historical interaction statement, and generates an enhanced semantic representation set that includes semantic preservation features, semantic shift features, and intent transfer features, thereby improving the completeness and accuracy of semantic information extraction during customer service interaction.

[0061] This invention constructs a semantic evolution trajectory of user interaction, analyzes the semantic change relationships in consecutive interaction rounds, locates abnormal semantic evolution change nodes by combining the degree of semantic association, the degree of semantic offset, and the degree of intent change, and identifies the user's implicit intent migration state based on entity change information, intent offset information, and semantic conflict information. This can avoid the semantic inertia problem caused by the over-reliance on historical context in traditional methods and improve the accuracy of identifying the user's true intent in complex multi-turn dialogue scenarios.

[0062] This invention corrects the association between historical and current semantic states based on intent migration detection results, reducing the interference of invalid historical semantic information on the current intent judgment. It also integrates effective semantic information in the current interaction statement to generate a target user intent representation, and then combines it with the intelligent customer service business knowledge base to output dynamic semantic understanding results. This enables the intelligent customer service system to adapt to complex interaction situations such as changes in user needs, topic switching, and implicit expressions, thereby improving the stability, real-time performance, and business adaptability of customer service semantic understanding. Attached Figure Description

[0063] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0064] Figure 1 This is a flowchart of a method for semantic understanding of intelligent customer service interaction statements proposed in this invention;

[0065] Figure 2 This is a graph showing the semantic change index of continuous interaction rounds in a semantic understanding method for intelligent customer service interaction statements proposed in this invention.

[0066] Figure 3 This is a comparative learning diagram of the semantic embedding space distribution before and after learning of the intelligent customer service interactive statement semantic understanding method proposed in this invention.

[0067] Figure 4 This is a schematic diagram of the structure of an intelligent customer service interactive statement semantic understanding system proposed in this invention. Detailed Implementation

[0068] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0069] refer to Figures 1-3 A method for semantic understanding of intelligent customer service interaction statements, comprising:

[0070] Collect user interaction statements, customer service response statements, interaction round information and time sequence information during the intelligent customer service interaction process, perform preprocessing on user interaction statements and customer service response statements, and generate multi-round customer service interaction statement sequences according to the interaction round arrangement;

[0071] The sequence of multi-round customer service interaction statements is input into the improved SimCSE model, which includes a context interaction enhancement coding layer, a semantic transfer awareness contrastive learning layer, and a historical semantic inhibition gating layer to generate an enhanced semantic representation set.

[0072] Based on the enhanced semantic representation set, the semantic change relationship between consecutive interaction rounds is extracted. The semantic association degree, semantic offset degree and intention change degree between the current user interaction statement and the historical interaction statement are calculated. The semantic evolution trajectory of user interaction is constructed according to the direction of semantic change.

[0073] Based on the semantic evolution trajectory of user interaction, the influence relationship between historical semantic states and current semantic states is analyzed, abnormal change nodes in the semantic evolution process are located, and the implicit intent migration state of users is identified by combining entity change information, intent offset information and semantic conflict information, and intent migration detection results are generated.

[0074] Based on the intent migration detection results, the historical semantic state and the current semantic state are correlated and corrected to reduce the impact of invalid historical semantic states on the current intent judgment, and the effective semantic information in the current interaction statement is integrated to generate the target user intent representation;

[0075] The system matches the target user's intent representation with the business intent information in the intelligent customer service business knowledge base, combines the semantic evolution trajectory of user interaction to determine the current user's real needs, and outputs dynamic semantic understanding results.

[0076] In this embodiment, the preprocessing of user interaction statements and customer service response statements includes performing text cleaning, entity extraction, and statement segmentation on the user interaction statements and customer service response statements.

[0077] In this embodiment, generating a multi-round customer service interaction statement sequence based on the interaction rounds includes:

[0078] Obtain the interaction time information, user identification information, and interaction round identification information corresponding to the user interaction statements. Sort the user interaction statements corresponding to the same user identification according to the order of interaction time to determine the relationship of consecutive interaction rounds.

[0079] Based on the interaction round identifier information, the user interaction statement is divided into multiple interaction statement fragments, and associated with the customer service response statement of the corresponding round, forming an interaction statement pair containing user input information and customer service feedback information;

[0080] Based on the sequential relationships between pairs of interactive statements, semantic continuity, topic continuation, and topic change relationships between adjacent rounds of interaction are extracted to construct a multi-round interactive context chain.

[0081] The interaction statement fragments and their corresponding context chains are combined according to the interaction rounds to generate a multi-round customer service interaction statement sequence that includes historical interaction information and current interaction information.

[0082] In this embodiment, generating the enhanced semantic representation set includes:

[0083] An improved SimCSE model is constructed based on the sentence encoding layer, semantic projection layer, and contrast learning layer of the SimCSE model. A context interaction enhancement encoding layer is embedded at the input of the SimCSE model, and a historical semantic suppression gating layer is embedded between the semantic projection layer and the contrast learning layer. The contrast learning layer is replaced with a semantic transfer-aware contrast learning layer, forming an improved SimCSE model for multi-turn customer service interaction statements.

[0084] Obtain the current user interaction statement, historical user interaction statements, and corresponding customer service response statements from a multi-round customer service interaction statement sequence. Associate and arrange the current user interaction statement, historical user interaction statements, and customer service response statements according to the interaction round, and input them into a context-enhanced coding layer. Extract interaction round association features, context semantic association features, and semantic change features to generate a context-enhanced semantic representation. Specifically, the generation of the context-enhanced semantic representation involves:

[0085] The current user interaction statement, historical user interaction statements, and corresponding customer service response statements are input into the context interaction enhancement encoding layer. The interaction round relationship corresponding to each interaction statement is identified by the statement position encoding. Cross-round semantic dependency features are extracted based on the semantic association relationship between the current user interaction statement and historical user interaction statements.

[0086] Feature fusion is performed on the question-and-answer feedback relationship between the current user interaction statement and the corresponding customer service response statement to extract customer service response feedback features. These customer service response feedback features are then associated with historical interaction semantic features to generate interaction round association features.

[0087] Based on the interaction round association feature analysis, the semantic continuity and semantic change relationships between user interaction statements in different rounds are analyzed. The topic retention information, entity change information and intent change information in consecutive interaction rounds are encoded to generate context semantic association features.

[0088] The interaction round association features, context semantic association features and semantic change features are concatenated and transformed into a unified semantic space through the semantic mapping mechanism in the context interaction enhancement coding layer, resulting in a context-enhanced semantic representation that integrates current interaction information and historical interaction information.

[0089] The context-enhanced semantic representation is input into the sentence encoding layer and the semantic projection layer. Feature mapping is then applied to the context-enhanced semantic representation to generate basic semantic embedding representations for each interaction round. A training sample set for the semantic transfer-aware contrastive learning layer is constructed based on the intent association relationships in the multi-round customer service interaction sentence sequences. Specifically, the feature mapping processing of the context-enhanced semantic representation involves:

[0090] The context-enhanced semantic representation is input into the sentence encoding layer. The sentence encoding layer encodes the interaction round association features, context semantic association features, and semantic change features in the context-enhanced semantic representation, and extracts the high-dimensional semantic feature representation corresponding to each interaction round.

[0091] The high-dimensional semantic feature representation output from the sentence encoding layer is input into the semantic projection layer. The semantic projection layer performs dimensional transformation on the high-dimensional semantic feature representation, mapping the semantic features corresponding to different interaction rounds to a unified semantic representation space, and generating the semantic projection representation corresponding to each interaction round.

[0092] Based on the interaction round relationship in the multi-round customer service interaction statement sequence, the semantic projection representation corresponding to the current user interaction statement is associated with the semantic projection representation corresponding to the historical user interaction statements, and the semantic preservation relationship, semantic change relationship and intent association relationship between different interaction rounds are analyzed.

[0093] Based on the intent association, the semantic projection representations are divided into samples. Semantic projection representations with the same business intent but different expression methods are identified as the same graph expression samples, and semantic projection representations with similar semantics but different business intents are identified as similar but different intent difficult negative samples. These samples are combined to generate the training sample set corresponding to the semantic transfer perception contrast learning layer.

[0094] The basic semantic embedding represents the input semantic transfer perception contrastive learning layer, based on the business meaning in the multi-round customer service interaction statement sequence. Figure 1 To generate consensus graph representation samples, generate similar but dissimilar intentional hard negative samples based on semantically similar but different business intents of interaction statements. Perform comparative optimization based on consensus graph representation samples and similar but dissimilar intentional hard negative samples to generate an intent-discriminating enhanced semantic representation, where:

[0095] Based on the business intent in the sequence of multiple customer service interaction statements Figure 1 The sample of the agreement graph representation for the relationship generation is as follows:

[0096] Obtain the business intent identifier corresponding to each interaction round in the multi-round customer service interaction statement sequence, group the basic semantic embedding representations based on the business intent identifiers, and divide the basic semantic embedding representations with the same business intent identifiers into the same intent set;

[0097] Within the same intent set, analyze the semantic differences in user interaction statements across different interaction rounds to extract business intent. Figure 1 Semantic embeddings with different sentence expression methods are paired to generate a consensus graph expression sample.

[0098] Based on the agreement diagram, express the business intent corresponding to the user interaction statements in the sample. Figure 1 To establish semantic representation associations, the basic semantic embedding representations corresponding to different expressions under the same business intent are made to form a positive contrast relationship, generating a set of consensus graph expression samples for optimization of the semantic transfer perception contrast learning layer.

[0099] Specifically, generating difficult negative samples with similar but different business intentions based on semantically similar but different interactive statements is as follows:

[0100] Obtain the basic semantic embedding representations corresponding to different business intentions in a multi-round customer service interaction statement sequence, calculate the semantic similarity relationship between different basic semantic embedding representations, and filter the interaction statements whose semantic similarity meets the preset similarity range and whose business intention identifiers are different.

[0101] Business intent difference analysis is performed on the filtered interactive statements to extract the differences in entity information, service requirements and interaction goals corresponding to different business intents. The basic semantic embedding representations with similar semantic expressions but different business intents are identified as the candidate set of difficult negative samples.

[0102] Based on the semantic similarity and business intent difference relationships in the candidate set of difficult negative samples, the basic semantic embedding representations with similar semantic expressions and high business intent discrimination are combined to generate a set of difficult negative samples with similar but different intents.

[0103] The set of consensus graph representation samples and the set of similar but different intentions difficult negative samples are input into the semantic transfer perception contrast learning layer. Based on the positive semantic association relationship and the negative intention distinction relationship, the distance relationship between the basic semantic embedding representations is adjusted so that the semantic representations corresponding to the same business intention are aggregated and the similar semantic representations corresponding to different business intentions are separated, thereby generating an intention-distinguishing enhanced semantic representation.

[0104] The intent-discriminating enhanced semantic representation is input into the historical semantic suppression gating layer. The correlation between the semantic representations of historical user interaction statements and the semantic representations of current user interaction statements is analyzed. Based on the degree of semantic change, the proportion of historical semantic information retained is determined, and the influence of historical semantic states on the current semantic state is adjusted to generate a dynamic semantic adjustment representation. Wherein:

[0105] The proportion of historical semantic information retained is determined based on the degree of semantic change, specifically as follows:

[0106] Obtain the semantic representations corresponding to the current user interaction statement and the historical user interaction statements in the intent-discriminative enhanced semantic representation, calculate the degree of semantic change between the current semantic representation and the historical semantic representation, and extract semantic shift features and intent transfer features;

[0107] Based on semantic offset features, the semantic change of the current user interaction statement relative to the historical user interaction statements is analyzed. Based on intent migration features, it is determined whether the business intent corresponding to the current user interaction statement has changed, and semantic change evaluation results are generated.

[0108] The semantic change evaluation results are input into the historical semantic suppression gating layer. The retention ratio of the semantic representation corresponding to the historical user interaction statement is adjusted according to the degree of semantic change. When the current user interaction statement and the historical user interaction statement maintain the same business intent, the retention ratio of historical semantic information is increased. When a shift in business intent is detected, the retention ratio of historical semantic information is decreased.

[0109] Based on the adjusted retention ratio of historical semantic information, the semantic representation corresponding to historical user interaction statements is weighted and adjusted to suppress the influence of invalid historical semantic states on the current semantic state, while retaining the valid historical semantic information related to the current user interaction statement, and generating a dynamic semantic adjustment representation.

[0110] By integrating context-enhanced semantic representation, intent-discriminative enhanced semantic representation, and dynamic semantic adjustment representation, and arranging them according to the interaction rounds corresponding to the multi-round customer service interaction statement sequence, an enhanced semantic representation set containing semantic preservation features, semantic shift features, and intent transfer features is generated.

[0111] This invention addresses the problem that traditional SimCSE models, which primarily focus on single-sentence semantic representation learning, struggle to adapt to the continuous semantic changes in context during multi-round interactions in intelligent customer service. It structurally optimizes the SimCSE model by introducing a context-enhanced encoding layer. This layer jointly encodes the current user interaction statement, historical user interaction statements, and customer service response statements. Based on the interaction rounds, it extracts cross-round semantic association features and semantic change features, generating an enhanced semantic representation that includes contextual information. This effectively improves the SimCSE model's ability to model semantically continuous relationships in multi-round customer service interaction scenarios.

[0112] To address the problem that existing contrastive learning methods, which primarily rely on semantic similarity to distinguish samples, struggle to identify interactive statements that are semantically similar but have different business intentions, this invention introduces a semantic transfer-aware contrastive learning layer. This layer differentiates samples based on the business intentions expressed in multiple rounds of customer service interactions. Figure 1 The SimCSE model constructs a consensus graph to represent samples and generates difficult negative samples with similar but different business intentions based on semantically similar but different interaction statements. By optimizing the distance relationship between the same and different intentions in the semantic representation space, the SimCSE model enhances its ability to identify fine-grained intention differences and improves the accuracy of understanding user intentions in complex customer service interaction scenarios.

[0113] To further reduce the impact of historical semantic inertia on current intent judgment, this invention embeds a historical semantic suppression gating layer into the SimCSE model. By analyzing the correlation between historical semantic representation and current semantic representation, the proportion of historical semantic information retained is dynamically adjusted according to the degree of semantic change. This ensures that effective historical context information is preserved and suppresses the interference of expired historical semantic states, thereby improving the SimCSE model's adaptability to scenarios involving user implicit intent migration, topic switching, and changing needs. This enables dynamic optimization of the semantic understanding results of intelligent customer service interaction statements.

[0114] In this embodiment, constructing the semantic evolution trajectory of user interaction based on the direction of semantic change includes:

[0115] Obtain semantic preservation features, semantic offset features, and intent transfer features corresponding to each interaction round in the enhanced semantic representation set. Establish continuous semantic state nodes according to the interaction round order in the multi-round customer service interaction statement sequence, and associate each semantic state node with the corresponding interaction statement to generate a semantic state node set.

[0116] Calculate the semantic association degree between semantic state nodes corresponding to adjacent interaction rounds. Based on the feature matching relationship between the semantic representation of the current user interaction statement and the semantic representation of the historical interaction statements, determine the semantic continuity relationship between consecutive interaction rounds and generate a semantic association feature set, where:

[0117] Calculate the semantic association between semantic state nodes corresponding to adjacent interaction rounds, specifically as follows:

[0118] Obtain the preceding and following semantic state nodes corresponding to adjacent interaction rounds in the semantic state node set, read the semantic preservation features, semantic offset features, and intent transfer features corresponding to the two semantic state nodes, and align the corresponding features according to the same feature dimension.

[0119] Based on the aligned semantic preservation features, the feature similarity between the preceding and following semantic state nodes is calculated to obtain the semantic preservation matching degree; the semantic offset matching degree is obtained by matching based on the direction and magnitude of change of the semantic offset features; and the intent transfer matching degree is obtained by matching based on the consistent state and change trend of the intent transfer features.

[0120] The semantic preservation matching degree, semantic offset matching degree, and intent transfer matching degree are normalized respectively to map each matching degree to a unified value range, and the feature matching results of adjacent semantic state nodes are constructed according to the order of semantic preservation matching degree, semantic offset matching degree, and intent transfer matching degree.

[0121] The normalized semantic preservation matching degree, semantic offset matching degree, and intent migration matching degree are arithmetically averaged, and the average result is determined as the semantic association degree between semantic state nodes corresponding to adjacent interaction rounds. The semantic association degree corresponding to each adjacent semantic state node is recorded in the order of interaction rounds.

[0122] Based on the semantic association feature set analysis, the feature changes between adjacent semantic state nodes are analyzed, and the semantic preservation feature change, semantic offset feature change, and intent migration feature change are extracted. The semantic offset degree between the current user interaction statement and historical interaction statements is determined based on each feature change, generating a semantic offset feature set. Specifically, determining the semantic offset degree between the current user interaction statement and historical interaction statements based on each feature change is as follows:

[0123] Obtain the semantic preservation feature change, semantic offset feature change, and intent migration feature change corresponding to adjacent semantic state nodes, establish a feature change sequence according to the interaction round order, and normalize each feature change to map different types of feature changes to a unified feature space.

[0124] Based on the analysis of semantic preservation feature changes, the semantic continuity between the current user interaction statement and the historical interaction statement is determined. When the semantic preservation feature changes decrease, the semantic connection between the current user interaction statement and the historical interaction statement is weakened. Based on the analysis of semantic offset feature changes, the magnitude of changes in semantic expression content is determined to determine the degree of semantic deviation of the current user interaction statement relative to the historical interaction statement.

[0125] Based on the change of intent migration features, the relationship between the intent state corresponding to the current user interaction statement and the intent state corresponding to the historical interaction statement is analyzed to determine the direction of change of user business needs and extract the migration features corresponding to the intent change.

[0126] The semantic preservation feature change, semantic shift feature change, and intent transfer feature change are fused together. Based on the influence of each feature change on the semantic difference between the current user interaction statement and the historical interaction statement, the degree of semantic shift is calculated.

[0127] The consecutive interaction rounds are marked according to the numerical change range corresponding to the degree of semantic offset. The feature change information corresponding to the degree of semantic offset is associated with the semantic state node to generate a set of semantic offset features containing the direction of semantic offset, the magnitude of semantic offset, and the trend of intention change.

[0128] Based on the degree of semantic relevance and semantic offset, the semantic state nodes corresponding to each interaction round are connected. A positive association is established between consecutive semantic state nodes with the same direction of semantic change. Semantic state nodes whose direction of semantic change changes are marked as semantic change nodes, thus generating a user interaction semantic evolution trajectory. Specifically, generating the user interaction semantic evolution trajectory involves:

[0129] Obtain the semantic state nodes, semantic association degree, and semantic offset degree corresponding to each interaction round. Arrange the semantic state nodes according to the interaction round order in the multi-round customer service interaction statement sequence and establish the temporal connection relationship between adjacent semantic state nodes.

[0130] The semantic continuity of consecutive interaction rounds is determined based on the degree of semantic association between adjacent semantic state nodes. When the degree of semantic association between adjacent semantic state nodes meets the continuous association condition, a positive association relationship is established between the corresponding semantic state nodes to form a continuous semantic state transmission path.

[0131] Based on the semantic offset degree analysis, the semantic change direction between adjacent semantic state nodes is analyzed. When the semantic offset direction of the subsequent semantic state node is consistent with that of the preceding semantic state node, the corresponding nodes are connected to form a semantic evolution path in the same direction. When the semantic offset direction changes and the semantic offset degree exceeds the preset change range, the corresponding semantic state node is marked as a semantic change node.

[0132] The positive association, the same-direction semantic evolution path, and the semantic change nodes are combined according to the order of the interaction rounds. The connection relationship, change direction and change position between each semantic state node are recorded to generate the user interaction semantic evolution trajectory.

[0133] Based on the semantic evolution trajectory of user interaction, the semantic state change process corresponding to each interaction round is saved, forming a dynamic semantic evolution path that includes historical semantic state, current semantic state and semantic change nodes;

[0134] Based on the semantic evolution trajectory analysis of user interaction, the evolution path between each semantic state node is analyzed. The degree of intent change corresponding to different interaction rounds is determined by combining the change in intent transfer features. This generates a user interaction semantic evolution trajectory containing semantic state change paths and intent change information. Specifically, the degree of intent change corresponding to different interaction rounds is determined by combining the change in intent transfer features.

[0135] Obtain the change in intent migration features corresponding to each semantic state node in the semantic evolution trajectory of user interaction, establish a sequence of intent migration feature changes according to the order of interaction rounds, and associate the semantic offset features and semantic change direction of the corresponding semantic state nodes.

[0136] Based on the change in intent migration features between semantic state nodes corresponding to adjacent interaction rounds, we analyze the relationship between the user intent state corresponding to the previous interaction round and the user intent state corresponding to the subsequent interaction round, and extract intent retention features and intent migration features.

[0137] The degree of intent change between different interaction rounds is determined based on the magnitude of the change in intent migration features. When the change in intent migration features is within a continuous and stable range, it is determined that the corresponding interaction rounds maintain the same intent state. When the change in intent migration features changes significantly, it is determined that there is an intent change relationship between the corresponding interaction rounds.

[0138] By combining the changes in semantic offset features and the changes in intent migration features, the degree of intent change corresponding to each interaction round is corrected, reducing misjudgments caused solely by changes in semantic expression, and highlighting the changes in intent state corresponding to changes in business requirements.

[0139] Based on the corrected degree of intent change, the semantic state nodes in the semantic evolution trajectory of user interaction are marked, the interaction rounds in which intent changes occur are determined as intent change nodes, and the intent change direction, degree of change and associated semantic state information corresponding to the intent change nodes are written into the semantic evolution trajectory of user interaction.

[0140] In this embodiment, generating the intent migration detection result includes:

[0141] Obtain each semantic state node and its corresponding evolution path in the semantic evolution trajectory of user interaction, extract the historical semantic state information, current semantic state information, semantic preservation features, semantic offset features and intent migration features corresponding to each semantic state node, and construct a semantic state influence analysis set;

[0142] Based on the semantic state influence analysis set, the relationship between historical semantic states and current semantic states is analyzed. The influence relationship of historical semantic states is determined according to the degree of semantic association and semantic preservation of historical semantic states and the degree of semantic shift of current semantic states, and historical semantic influence features are generated.

[0143] Based on the historical semantic influence characteristics, change detection is performed on continuous semantic state nodes in the semantic evolution trajectory of user interaction, identifying semantic state nodes whose semantic deviation exceeds the preset change range, and combining the corresponding interaction round information to determine abnormal change nodes in the semantic evolution process, where the preset change range is 0.6 to 1.

[0144] Obtain the current and historical user interaction statements corresponding to the abnormally changed nodes, extract entity change information, intent offset information, and semantic conflict information between them, perform correlation analysis on the entity change information, intent offset information, and semantic conflict information, and generate implicit intent migration features. Specifically, the generation of implicit intent migration features includes:

[0145] Obtain the current user interaction statement and historical user interaction statement corresponding to the abnormal change node, extract entity information, business keyword information and semantic representation information from the two types of interaction statements, match the entity information corresponding to the current user interaction statement with the entity information corresponding to the historical user interaction statement, determine the entity addition information, entity replacement information and entity disappearance information, and generate entity change information.

[0146] Based on the intent migration features and target user intent representations corresponding to abnormal change nodes, the correlation between the current user interaction statement and the business requirements corresponding to the historical user interaction statements is analyzed, the direction of change between the business intent corresponding to the historical user interaction statements and the business intent corresponding to the current user interaction statements is determined, and intent offset information is generated.

[0147] Based on the differences between the semantic representations of current user interaction statements and those of historical user interaction statements, analyze the inconsistencies between the two in terms of semantic theme, business objectives, and interaction purpose, determine the location and degree of semantic conflict, and generate semantic conflict information.

[0148] By associating and fusing entity change information, intent shift information, and semantic conflict information, the correspondence between entity change and intent change is analyzed. Semantic conflict information is combined to correct semantic differences caused by non-intent changes and extract implicit intent migration features that can characterize the direction of changes in user needs.

[0149] The implicit intent migration features are associated with and marked with abnormal change nodes. The corresponding interaction rounds, historical intent states, current intent states and intent migration directions are recorded to generate a set of implicit intent migration features for identifying the implicit intent migration state of users.

[0150] Based on the implicit intent migration feature analysis, the direction of semantic state change before and after the abnormal change node is analyzed to determine whether the intent corresponding to the current user interaction statement is separated from the intent corresponding to the historical semantic state. The semantic state node where the intent changes is marked as intent migration node, and intent migration detection results containing intent migration node, intent migration direction and intent migration state are generated.

[0151] In this embodiment, generating the target user intent representation includes:

[0152] Obtain the intent migration node, intent migration direction and intent migration status from the intent migration detection results, read the historical semantic status and current semantic status associated with the corresponding intent migration node, extract the historical semantic representation, current semantic representation and semantic offset features, and construct a semantic state correction set;

[0153] Based on semantic state correction set analysis, the semantic association between historical semantic states and current semantic states is analyzed. The effective association range of historical semantic states is determined according to the intent transition state. Valid and invalid historical semantic information related to the current user interaction statement is identified, and historical semantic state correction information is generated. Specifically, determining the effective association range of historical semantic states based on the intent transition state involves:

[0154] Obtain the historical semantic state, current semantic state, and corresponding intent migration state from the semantic state correction set; read the evolution path information between the historical semantic state and the current semantic state in the user interaction semantic evolution trajectory; and determine the interaction round distance and semantic association relationship corresponding to each historical semantic state.

[0155] Based on the intent migration state analysis, the changes between the intent corresponding to the current user interaction statement and the intent corresponding to the historical user interaction statement are analyzed. When the intent corresponding to the historical semantic state is consistent with the intent corresponding to the current user interaction statement, the corresponding historical semantic state is determined as a valid candidate state for association. When the intent corresponding to the historical semantic state migrates from the intent corresponding to the current user interaction statement, the corresponding historical semantic state is determined as a state to be corrected.

[0156] Based on the semantic relevance of the effective candidate states, the distance between interaction rounds, and the meaning... Figure 1 The degree of semantic association determines the effective association range of historical semantic states, and historical semantic states that meet the semantic association degree threshold and are within the continuous path of intent are included in the effective association range.

[0157] Based on the migration direction and migration node position corresponding to the intent migration state, the effective association range is dynamically adjusted. When it is detected that the user intent has not migrated, the association range of the historical semantic state is expanded to retain continuous context information; when it is detected that the user intent has migrated, the association range of the historical semantic state is narrowed to reduce the impact of invalid historical semantic states.

[0158] The historical semantic states within the adjusted effective association range are determined as valid historical semantic information, and the historical semantic states outside the effective association range and inconsistent with the intent of the current user interaction statement are determined as invalid historical semantic information. Historical semantic state correction information containing the effective association range, valid historical semantic information and invalid historical semantic information is generated.

[0159] Based on historical semantic state correction information, the influence of failed historical semantic states on the current semantic state is adjusted, suppression processing is performed on failed historical semantic information, and contextual features corresponding to valid historical semantic information are retained to generate corrected historical semantic representations;

[0160] The modified historical semantic representation is fused with the current semantic representation. The entity change information, intent offset information and semantic transfer features in the current interaction statement are combined to supplement the intent features corresponding to the current user interaction statement and generate a fused semantic intent representation.

[0161] Based on the analysis of the target business intent corresponding to the current user interaction statement using the fused semantic intent representation, the fused semantic intent representation is verified according to the intent change path in the semantic evolution trajectory to determine the user's true intent corresponding to the current interaction stage, and a target user intent representation is generated. Specifically, the generation of the target user intent representation involves:

[0162] Obtain the current semantic features, corrected historical semantic features, and intent migration features from the fused semantic intent representation. Combine the intent change path in the semantic evolution trajectory of user interaction to read the semantic state node and associated intent state corresponding to the current interaction stage.

[0163] Based on the business requirement characteristics corresponding to the current semantic state node, the candidate intent information in the fused semantic intent representation is matched and analyzed to determine the degree of association between each candidate business intent and the current user interaction statement, and a candidate intent association set is generated.

[0164] Based on the intention change path in the semantic evolution trajectory, analyze the evolution relationship between candidate intentions, determine whether the current candidate intention conforms to the change direction of the user's historical intention to the current intention, correct candidate intentions that do not conform to the intention change path, and generate intention path verification results;

[0165] Based on the intent path verification results, the candidate intent association set is filtered, and the target intent features that are consistent with the semantic state of the current interaction stage and conform to the intent migration direction are retained. The entity information, business requirement information and effective context information in the current user interaction statement are integrated to generate target intent fusion features.

[0166] The target intent fusion features are encoded according to the semantic state corresponding to the current interaction stage, and the historical intent state, current intent state, and intent change path in the semantic evolution trajectory of user interaction are associated to form a target user intent representation that includes current business needs, intent migration direction, and semantic association information.

[0167] In this embodiment, the output of dynamic semantic understanding results includes:

[0168] Obtain the current user intent features, intent change path and semantic evolution trajectory information from the target user intent representation, extract the business semantic features, business entity features and business association features corresponding to each business intent in the intelligent customer service business knowledge base, and construct a business intent matching set;

[0169] The semantic matching relationship between the target user intent representation and each business intent information is calculated based on the business intent matching set. A candidate business intent set is determined according to the current user intent features, business semantic features, and business entity features. The candidate business intent set is then sorted according to the degree of matching. Specifically, the semantic matching relationship between the target user intent representation and each business intent information is calculated based on the business intent matching set as follows:

[0170] Obtain the business semantic features, business entity features, and business association features corresponding to each business intent information in the business intent matching set, and read the current user intent features, entity association features, and semantic change features in the target user intent representation, and perform feature dimension alignment processing on the target user intent representation and each business intent information;

[0171] Matching analysis is performed based on the semantic relationship between the current user intent features and the semantic features of each business to calculate the semantic consistency between the target user intent representation and the business intent information, and to determine the semantic relationship between the business requirements and business intent information corresponding to the current user interaction statement.

[0172] Entity matching is performed based on the entity association features and business entity features in the target user intent representation. The correspondence between the target entity, business object and business attribute in the current user interaction statement and the corresponding entity information of each business intent information is analyzed to generate entity matching features.

[0173] Based on semantic consistency and entity matching features, combined with business association features, we analyze the comprehensive matching relationship between the target user's intent representation and various business intent information, and determine the semantic matching degree corresponding to different business intent information.

[0174] The business intent information in the business intent matching set is associated and marked according to the semantic matching degree of the business intent information. The business intent information that meets the matching conditions is determined as the candidate business intent set, and sorted according to the semantic matching degree to generate the business intent matching result.

[0175] The candidate business intent set is correlated with the semantic evolution trajectory of user interaction. The intent migration nodes, intent migration directions and intent change paths in the semantic evolution trajectory are read. The consistency relationship between the candidate business intent and the current user intent representation is verified, and the business intent verification result is generated.

[0176] Based on the business intent verification results, target business intents corresponding to the current user's actual needs are filtered, and customer service processing flow information, service response information and interaction status information corresponding to the target business intents are associated to generate dynamic semantic understanding information.

[0177] The dynamic semantic understanding information is organized according to the current interaction round, intent migration state, and business processing type to generate dynamic semantic understanding results that include the user's current intent, semantic change process, and business processing direction.

[0178] refer to Figure 4 A semantic understanding system for intelligent customer service interaction statements, comprising:

[0179] The interactive data acquisition module is used to collect user interaction statements, customer service response statements, and interaction information, and generate a multi-round customer service interaction statement sequence after preprocessing.

[0180] The semantic enhancement representation module is used to generate an enhanced semantic representation set by fusing contextual information, intent transfer information and historical semantic information based on the improved SimCSE model;

[0181] The semantic trajectory construction module is used to analyze the semantic change relationship between rounds based on the enhanced semantic representation set, calculate semantic association, offset and intent change features, and construct the semantic evolution trajectory of user interaction.

[0182] The intent migration analysis module is used to analyze semantic state changes based on the semantic evolution trajectory of user interaction, locate abnormal change nodes and identify implicit intent migration states, and generate intent migration detection results.

[0183] The intent state correction module is used to adjust the influence of historical semantic states based on intent migration detection results and integrate current interaction semantic information to generate a target user intent representation.

[0184] The semantic result generation module is used to match the target user's intent representation with business intent information, combine semantic evolution trajectory to determine the user's real needs, and output dynamic semantic understanding results.

[0185] In this embodiment, the computer-readable storage medium stores computer program instructions, which, when executed by a processor, perform a method for semantic understanding of intelligent customer service interaction statements.

[0186] Example 1: In an intelligent customer service after-sales interaction scenario, the system receives a batch of anonymized multi-round customer service conversation data, containing a total of 6200 conversation samples, including 34860 ​​user interaction statements and 33620 customer service response statements. Each conversation contains 3–8 rounds of interaction, with an average of 5.6 rounds. Business types include refund progress, logistics inquiry, exchange application, repair consultation, coupon usage, price compensation, and invoice processing. The samples show instances where users change their needs midway through the conversation; 940 samples show explicit intent switching, and 1260 samples show implicit intent migration, accounting for 20.3% of all conversations. A typical problem in the implicit intent migration samples is that the user's first three rounds revolve around refund inquiries, followed by the input "How long will it take to arrive if I buy another one?", which traditional methods easily interpret as a refund arrival inquiry.

[0187] After the data enters the processing flow, the system collects user interaction statements, customer service response statements, interaction round information, and time sequence information. Taking a set of conversations as an example, in the first round, the user inputs "Has the refund been approved?", and the customer service responds "The refund application has been submitted and is being processed." In the second round, the user inputs "When will the money be returned?", and the customer service responds "The refund will be returned according to the original payment method." In the third round, the user inputs "If I buy another one, how long will it take to deliver?" The system performs preprocessing on the above statements, removing duplicate punctuation and invalid interjections, standardizing "money returned" as "refund credited", marking "buy another one" as a purchase behavior entity, and marking "delivered" as a delivery target entity. Before preprocessing, this conversation contains 46 Chinese word units; after preprocessing, 38 valid word units are retained. Entity recognition identifies three types of entities: refund item, purchase behavior, and delivery target, and generates a multi-round customer service interaction statement sequence according to the interaction rounds.

[0188] The multi-round customer service interaction sequence is fed into the improved SimCSE model. This model embeds a context-enhanced coding layer at the input of the original SimCSE model, a historical semantic suppression gating layer between the semantic projection layer and the contrastive learning layer, and replaces the original contrastive learning layer with a semantic transfer-aware contrastive learning layer. The system pairs each round of user interaction statements with the corresponding customer service response statements, adding round position markers for rounds 1, 2, and 3. After processing by the context-enhanced coding layer, the semantic representation of the refund topic response value in round 1 is 0.84, and the logistics delivery response value is 0.12; in round 2, the refund topic response value is 0.88, and the logistics delivery response value is 0.09; in round 3, the refund topic response value decreases to 0.31, and the logistics delivery response value increases to 0.79. The resulting context-enhanced semantic representation has a dimension of 768, forming 3–8 round semantic vectors for each session.

[0189] In the semantic transfer-aware contrastive learning layer, the system constructs agreement graph expression samples and difficult negative samples with similar but different intentions from the training samples. A total of 18,600 pairs of agreement graph expression samples are formed in the training set. For example, "How long will it take for the refund to arrive?" and "When will the money be returned?" are labeled as agreement graph expression samples, both with the business intention of refund progress. Simultaneously, 9,200 pairs of difficult negative samples with similar but different intentions are constructed. For example, "How long will it take for the refund to arrive?" and "How long will it take to deliver a new one?" both contain "how long" in the text, but the former corresponds to refund progress, while the latter corresponds to delivery time. Before training, the average semantic cosine similarity of the above difficult negative samples is 0.74, and the model tends to group them close together. After training with the semantic transfer-aware contrastive learning layer, the average similarity of the agreement graph expression samples increases from 0.71 to 0.89, while the average similarity of the difficult negative samples with similar but different intentions decreases from 0.74 to 0.43, clearly separating semantically similar but different business intention interactive statements in the representation space.

[0190] In the historical semantic suppression gating layer, the system calculates the degree of semantic change between the current user interaction statement and historical user interaction statements. Taking the third round of the above conversation as an example, the semantic correlation between the current statement and the statement in the first round is 0.39, the semantic correlation with the statement in the second round is 0.42, the semantic offset is 0.76, and the intent migration feature response value is 0.81. Since the semantic offset exceeds the preset variation range of 0.6–1.0, the historical semantic suppression gating layer reduces the retention ratio of refund-related historical semantic information from 0.78 in the second round to 0.26 in the third round, while retaining the effective contextual relationship of "repurchase after refund" at a retention ratio of 0.52. After processing, the system generates an enhanced semantic representation set, in which the delivery timeliness intent response value of the statement in the third round increases from 0.79 before gating to 0.86, and the refund progress intent response value decreases from 0.31 to 0.18.

[0191] The system extracts semantic change relationships between consecutive interaction rounds based on an enhanced semantic representation set and constructs a semantic evolution trajectory of user interaction. Between rounds 1 and 2, the semantic correlation is 0.87, the semantic offset is 0.16, and the intent change is 0.11. The system establishes a positive correlation, determining it as a continuation of the refund intent. Between rounds 2 and 3, the semantic correlation decreases to 0.42, the semantic offset increases to 0.76, and the intent change increases to 0.83. The system marks the semantic state node of round 3 as a semantic change node and records the intent change path as a migration from refund progress to delivery timeliness in the user interaction semantic evolution trajectory.

[0192] During the intent migration detection phase, the system reads the current user interaction statement corresponding to the abnormal change node in the third round and the historical user interaction statements from the previous two rounds, extracting entity change information, intent offset information, and semantic conflict information. Entity change information shows that the core entity in the historical statement was a refund request, while the current statement adds a purchase action and a delivery target. Intent offset information shows that the business requirement has shifted from "when will the money be returned?" to "how long after repurchasing will it be delivered?". Semantic conflict information shows that historical customer service responses revolved around the refund processing flow, while the current user statement requests delivery timeliness; the two business objectives are inconsistent. The system identifies this abnormal change node as an intent migration node and outputs the intent migration detection result, where the intent migration status is "migrated," the intent migration direction is from refund progress to delivery timeliness, and the migration confidence is 0.88.

[0193] During the target user intent representation generation phase, the system correlates and corrects the historical and current semantic states based on intent migration detection results. The first-round refund semantic state is determined to be invalid historical semantic information, and its influence weight decreases from 0.41 to 0.12; the second-round refund arrival semantic state is determined to be partially valid historical semantic information, and its influence weight decreases from 0.45 to 0.21; the purchase behavior and delivery goal in the third-round current semantic state are determined to be valid semantic information, and their combined weight increases to 0.86. After integrating current semantic features, corrected historical semantic features, and intent migration features, the system generates a target user intent representation, with a delivery timeliness score of 0.91, a refund progress score of 0.19, and an exchange request score of 0.08.

[0194] During the business matching phase, the system matches the target user's intent representation with business intent information in the intelligent customer service business knowledge base. The business knowledge base contains 142 business intent nodes and 386 business processing rules. For the above conversation, the semantic matching degree between the target user's intent representation and the delivery timeliness business intent is 0.92, the semantic matching degree with the refund progress business intent is 0.21, and the semantic matching degree with the order cancellation business intent is 0.17. Based on this, the system outputs a dynamic semantic understanding result: the current user's actual need is for inquiries about the delivery timeliness after repurchasing, the business processing type is delivery timeliness query, and the recommended customer service response is to explain the estimated delivery period based on the product inventory status and delivery rules.

[0195] To verify the effectiveness, the system compared the proposed method with the traditional method on the same test set. The test set contained 1600 conversations, including 420 implicit intent transfer samples and 510 similar but different intent samples. The traditional method uses a common sentence encoding model plus a static intent classifier, concatenating only the most recent three rounds of historical statements; the proposed method uses an improved SimCSE model, user interaction semantic evolution trajectory, and a historical semantic suppression gating layer. The test results show that, in terms of overall intent recognition accuracy, the traditional method is 86.2%, while the proposed method is 94.1%; in terms of implicit intent transfer recognition accuracy, the traditional method is 69.8%, while the proposed method is 90.5%; in terms of similar but different intent sample recognition accuracy, the traditional method is 67.4%, while the proposed method is 88.9%; and in terms of semantic conflict sample recognition accuracy, the traditional method is 71.6%, while the proposed method is 87.8%. Regarding the error matching rate of business processes, the traditional method is 11.3%, while the method of this invention is 4.5%; regarding the rate of continuous follow-up inquiries from users, the traditional method is 17.9%, while the method of this invention is 9.1%; regarding the average processing time per round, the traditional method is 82 milliseconds, while the method of this invention is 97 milliseconds, both still meeting the requirements for real-time interaction of intelligent customer service.

[0196] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for semantic understanding of intelligent customer service interaction statements, characterized in that, include: Collect user interaction statements, customer service response statements, interaction round information and time sequence information during the intelligent customer service interaction process, perform preprocessing on user interaction statements and customer service response statements, and generate multi-round customer service interaction statement sequences according to the interaction round arrangement; The sequence of multi-round customer service interaction statements is input into the improved SimCSE model, which includes a context interaction enhancement coding layer, a semantic transfer awareness contrastive learning layer, and a historical semantic inhibition gating layer to generate an enhanced semantic representation set. Based on the enhanced semantic representation set, the semantic change relationship between consecutive interaction rounds is extracted. The semantic association degree, semantic offset degree and intention change degree between the current user interaction statement and the historical interaction statement are calculated. The semantic evolution trajectory of user interaction is constructed according to the direction of semantic change. Based on the semantic evolution trajectory of user interaction, the influence relationship between historical semantic states and current semantic states is analyzed, abnormal change nodes in the semantic evolution process are located, and the implicit intent migration state of users is identified by combining entity change information, intent offset information and semantic conflict information, and intent migration detection results are generated. Based on the intent migration detection results, the historical semantic state and the current semantic state are correlated and corrected to reduce the impact of invalid historical semantic states on the current intent judgment, and the effective semantic information in the current interaction statement is integrated to generate the target user intent representation; The system matches the target user's intent representation with the business intent information in the intelligent customer service business knowledge base, combines the semantic evolution trajectory of user interaction to determine the current user's real needs, and outputs dynamic semantic understanding results.

2. The method for semantic understanding of intelligent customer service interaction statements according to claim 1, characterized in that, The preprocessing of user interaction statements and customer service response statements includes text cleaning, entity extraction, and statement segmentation.

3. The method for semantic understanding of intelligent customer service interaction statements according to claim 1, characterized in that, The step of generating a multi-round customer service interaction statement sequence based on the interaction rounds includes: Obtain the interaction time information, user identification information, and interaction round identification information corresponding to the user interaction statements. Sort the user interaction statements corresponding to the same user identification according to the order of interaction time to determine the relationship of consecutive interaction rounds. Based on the interaction round identifier information, the user interaction statement is divided into multiple interaction statement fragments, and associated with the customer service response statement of the corresponding round, forming an interaction statement pair containing user input information and customer service feedback information; Based on the sequential relationships between pairs of interactive statements, semantic continuity, topic continuation, and topic change relationships between adjacent rounds of interaction are extracted to construct a multi-round interactive context chain. The interaction statement fragments and their corresponding context chains are combined according to the interaction rounds to generate a multi-round customer service interaction statement sequence that includes historical interaction information and current interaction information.

4. The method for semantic understanding of intelligent customer service interaction statements according to claim 1, characterized in that, The generated enhanced semantic representation set includes: An improved SimCSE model is constructed based on the sentence encoding layer, semantic projection layer, and contrast learning layer of the SimCSE model. A context interaction enhancement encoding layer is embedded at the input of the SimCSE model, and a historical semantic suppression gating layer is embedded between the semantic projection layer and the contrast learning layer. The contrast learning layer is replaced with a semantic transfer-aware contrast learning layer, forming an improved SimCSE model for multi-turn customer service interaction statements. Obtain the current user interaction statement, historical user interaction statement, and corresponding customer service response statement from a multi-round customer service interaction statement sequence. Associate and arrange the current user interaction statement, historical user interaction statement, and customer service response statement according to the interaction round, and input them into the context interaction enhancement coding layer to extract interaction round association features, context semantic association features, and semantic change features to generate a context-enhanced semantic representation. The context-enhanced semantic representation is input into the sentence encoding layer and the semantic projection layer. Feature mapping is performed on the context-enhanced semantic representation to generate the basic semantic embedding representation corresponding to each round of interaction. The training sample set of the semantic transfer perception contrast learning layer is constructed based on the intent association relationship in the sequence of customer service interaction statements in multiple rounds. The basic semantics are embedded into the semantic transfer perception contrastive learning layer. Based on the consistency of business intent in the multi-round customer service interaction statement sequence, a consensus graph expression sample is generated. Based on the semantically similar but different business intents of the interaction statements, similar but different intents hard negative samples are generated. Based on the consensus graph expression sample and the similar but different intents hard negative samples, a contrast optimization is performed to generate an intent-discriminating enhanced semantic representation. The intent-discriminative semantic representation is input into the historical semantic suppression gating layer. The relationship between the semantic representations corresponding to historical user interaction statements and the semantic representations corresponding to current user interaction statements is analyzed. The proportion of historical semantic information retention is determined based on the degree of semantic change. The influence of historical semantic state on current semantic state is adjusted to generate dynamic semantic adjustment representation. By integrating context-enhanced semantic representation, intent-discriminative enhanced semantic representation, and dynamic semantic adjustment representation, and arranging them according to the interaction rounds corresponding to the multi-round customer service interaction statement sequence, an enhanced semantic representation set containing semantic preservation features, semantic shift features, and intent transfer features is generated.

5. The method for semantic understanding of intelligent customer service interaction statements according to claim 1, characterized in that, The construction of the user interaction semantic evolution trajectory based on the direction of semantic change includes: Obtain semantic preservation features, semantic offset features, and intent transfer features corresponding to each interaction round in the enhanced semantic representation set. Establish continuous semantic state nodes according to the interaction round order in the multi-round customer service interaction statement sequence, and associate each semantic state node with the corresponding interaction statement to generate a semantic state node set. Calculate the semantic association degree between semantic state nodes corresponding to adjacent interaction rounds, determine the semantic continuity relationship between consecutive interaction rounds based on the feature matching relationship between the semantic representation corresponding to the current user interaction statement and the semantic representation corresponding to the historical interaction statement, and generate a semantic association feature set. Based on the semantic association feature set, the feature changes between adjacent semantic state nodes are analyzed, and the semantic preservation feature change, semantic offset feature change, and intent migration feature change are extracted. The semantic offset degree between the current user interaction statement and the historical interaction statement is determined according to the feature change, and a semantic offset feature set is generated. Based on the degree of semantic association and semantic offset, the semantic state nodes corresponding to each interaction round are connected. A positive association relationship is established between continuous semantic state nodes with the same direction of semantic change. Semantic state nodes whose direction of semantic change changes are marked as semantic change nodes, thus generating the semantic evolution trajectory of user interaction. Based on the semantic evolution trajectory analysis of user interaction, the evolution path between each semantic state node is analyzed, and the degree of intent change corresponding to different interaction rounds is determined by combining the change amount of intent migration features. This generates a semantic evolution trajectory of user interaction that includes semantic state change path and intent change information.

6. The method for semantic understanding of intelligent customer service interaction statements according to claim 1, characterized in that, The generated intent migration detection results include: Obtain each semantic state node and its corresponding evolution path in the semantic evolution trajectory of user interaction, extract the historical semantic state information, current semantic state information, semantic preservation features, semantic offset features and intent migration features corresponding to each semantic state node, and construct a semantic state influence analysis set; Based on the semantic state influence analysis set, the relationship between historical semantic states and current semantic states is analyzed. The influence relationship of historical semantic states is determined according to the degree of semantic association and semantic preservation of historical semantic states and the degree of semantic shift of current semantic states, and historical semantic influence features are generated. Based on the historical semantic influence characteristics, change detection is performed on continuous semantic state nodes in the semantic evolution trajectory of user interaction, identifying semantic state nodes whose semantic deviation exceeds the preset change range, and combining the corresponding interaction round information to determine abnormal change nodes in the semantic evolution process. Obtain the current user interaction statement and historical user interaction statement corresponding to the abnormal change node, extract the entity change information, intent offset information and semantic conflict information between the two, perform correlation analysis on the entity change information, intent offset information and semantic conflict information, and generate implicit intent migration features. Based on the implicit intent migration feature analysis, the direction of semantic state change before and after the abnormal change node is analyzed to determine whether the intent corresponding to the current user interaction statement is separated from the intent corresponding to the historical semantic state. The semantic state node where the intent changes is marked as intent migration node, and intent migration detection results containing intent migration node, intent migration direction and intent migration state are generated.

7. The method for semantic understanding of intelligent customer service interaction statements according to claim 1, characterized in that, The generation of the target user intent representation includes: Obtain the intent migration node, intent migration direction and intent migration status from the intent migration detection results, read the historical semantic status and current semantic status associated with the corresponding intent migration node, extract the historical semantic representation, current semantic representation and semantic offset features, and construct a semantic state correction set; Based on the semantic state correction set analysis, the semantic association between historical semantic state and current semantic state is analyzed. The effective association range of historical semantic state is determined according to the intent migration state. The effective historical semantic information and invalid historical semantic information related to the current user interaction statement are identified, and historical semantic state correction information is generated. Based on historical semantic state correction information, the influence of failed historical semantic states on the current semantic state is adjusted, suppression processing is performed on failed historical semantic information, and contextual features corresponding to valid historical semantic information are retained to generate corrected historical semantic representations; The modified historical semantic representation is fused with the current semantic representation. The entity change information, intent offset information and semantic transfer features in the current interaction statement are combined to supplement the intent features corresponding to the current user interaction statement and generate a fused semantic intent representation. Based on the analysis of the target business intent corresponding to the current user interaction statement, the fused semantic intent representation is verified according to the intent change path in the semantic evolution trajectory to determine the user's true intent corresponding to the current interaction stage and generate the target user intent representation.

8. The method for semantic understanding of intelligent customer service interaction statements according to claim 1, characterized in that, The output dynamic semantic understanding result includes: Obtain the current user intent features, intent change path and semantic evolution trajectory information from the target user intent representation, extract the business semantic features, business entity features and business association features corresponding to each business intent in the intelligent customer service business knowledge base, and construct a business intent matching set; Based on the business intent matching set, the semantic matching relationship between the target user intent representation and each business intent information is calculated. Based on the current user intent features, business semantic features and business entity features, the candidate business intent set is determined and sorted according to the degree of matching. The candidate business intent set is correlated with the semantic evolution trajectory of user interaction. The intent migration nodes, intent migration directions and intent change paths in the semantic evolution trajectory are read. The consistency relationship between the candidate business intent and the current user intent representation is verified, and the business intent verification result is generated. Based on the business intent verification results, target business intents corresponding to the current user's actual needs are filtered, and customer service processing flow information, service response information and interaction status information corresponding to the target business intents are associated to generate dynamic semantic understanding information. The dynamic semantic understanding information is organized according to the current interaction round, intent migration state, and business processing type to generate dynamic semantic understanding results that include the user's current intent, semantic change process, and business processing direction.

9. A semantic understanding system for intelligent customer service interaction statements, comprising executing the semantic understanding method for intelligent customer service interaction statements as described in any one of claims 1 to 8, characterized in that, include: The interactive data acquisition module is used to collect user interaction statements, customer service response statements, and interaction information, and generate a multi-round customer service interaction statement sequence after preprocessing. The semantic enhancement representation module is used to generate an enhanced semantic representation set by fusing contextual information, intent transfer information and historical semantic information based on the improved SimCSE model; The semantic trajectory construction module is used to analyze the semantic change relationship between rounds based on the enhanced semantic representation set, calculate semantic association, offset and intent change features, and construct the semantic evolution trajectory of user interaction. The intent migration analysis module is used to analyze semantic state changes based on the semantic evolution trajectory of user interaction, locate abnormal change nodes and identify implicit intent migration states, and generate intent migration detection results. The intent state correction module is used to adjust the influence of historical semantic states based on intent migration detection results and integrate current interaction semantic information to generate a target user intent representation. The semantic result generation module is used to match the target user's intent representation with business intent information, combine semantic evolution trajectory to determine the user's real needs, and output dynamic semantic understanding results.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer program instructions, which, when executed by a processor, perform a semantic understanding method for intelligent customer service interaction statements as described in any one of claims 1-8.

Images