Omnichannel intelligent customer service response interaction system and method based on artificial intelligence

Abstract

The invention discloses an omni-channel intelligent customer service response interaction system and method based on artificial intelligence, and belongs to the technical field of intelligent customer service. Multi-channel user session data is received and uniformly mapped into structured semantic data through a semantic coding rule, and cross-channel data is associated; constructing a dynamic session memory library based on the entity-feature-association three-dimensional data structure, and capturing session core information in real time; for anaphora and omission information in a dialogue, double-path parallel processing is carried out: a first path generates an entity candidate set through time sequence reverse retrieval, a second path generates a feature candidate set through semantic logic analysis, and a complete semantic expression is formed after cross validation; and respectively generating core response and potential demand guide content through double paths, and adjusting an output form in combination with channel adaptability to complete omni-channel intelligent interaction.

Description

Technical Field

[0001] This invention relates to the field of computer intelligent customer service technology, specifically to an omnichannel intelligent customer service response and interaction system and method based on artificial intelligence. Background Technology

[0002] Current AI-based omnichannel customer service systems generally rely on intent recognition models that independently analyze individual user queries for their core intelligent response capabilities. These models are typically built on pre-trained natural language processing techniques, aiming to extract key information from individual user statements and determine their intent. However, in their analysis process, there is a severe lack of effective perception and utilization of historical contextual information generated during continuous dialogue. When the user input is a self-contained complete question, the system can still respond effectively; but when the user's expression is based on information shared in previous dialogues, the system's understanding ability deteriorates significantly.

[0003] The essence of this technical limitation lies in the fact that the architecture of existing models does not include a memory mechanism that can dynamically maintain and update the dialogue state. The model cannot continuously track the core entities mentioned in the dialogue and their feature changes throughout the entire session lifecycle, making it difficult to parse the frequent referencing and omission phenomena in subsequent dialogues. This directly leads to the system's inability to accurately restore the user's complete semantics, and it can only trigger a preset clarification mechanism or give an incorrect response that is out of context. The ultimate consequence is that the dialogue process is manually interrupted, the efficiency of problem solving is reduced, and when transferring to a human agent, a coherent dialogue background cannot be provided, forcing the agent to re-question the user, which seriously affects the continuity of service and user experience. Summary of the Invention

[0004] The purpose of this invention is to provide an omnichannel intelligent customer service response and interaction system and method based on artificial intelligence to solve the problems mentioned in the background art.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: an omnichannel intelligent customer service response interaction method based on artificial intelligence, the method comprising: Step S100: Receive user session data from different channels, map all types of session data into structured semantic data according to preset standardized semantic encoding rules, and associate cross-channel session data through user unique identifiers; Step S200: Based on the entity-feature-association three-dimensional data structure, capture entity, feature information and semantic relationships in multi-turn dialogues in real time from structured semantic data to establish a dynamic conversation memory; and automatically extract new semantic data to supplement the dynamic conversation memory after each round of dialogue. Step S300: For the definite statements and omitted information in the dialogue, first locate the specific sentence position of the ambiguous statement through semantic features, and then perform parallel matching through two paths: the first path traverses the dynamic conversation memory in reverse order of interaction time to generate an entity candidate set, and selects the appropriate items by sorting by feature similarity; the second path identifies the missing feature information by parsing the semantic structure of the ambiguous statement and generates a feature candidate set; through cross-validation of the results of the two paths, the preceding information is associated and the omitted features are filled in to form a complete semantic statement; Step S400: Based on the complete semantic representation, the first path matches the knowledge base and business system data to generate the core response; the second path combines historical data from the memory bank to predict potential related interaction directions and generate supplementary guidance content; the results of the two paths are integrated, the optimal response combination is selected according to channel adaptability, and the interaction is completed by adapting the output format.

[0006] Furthermore, step S100 includes: Step S101: Receive various types of session data through a pre-defined standardized channel access interface, categorizing them into three types: text data, voice data, and speech-to-text data. For text data, filter non-semantic characters and meaningless expressions using pre-defined character filtering rules, and organize it into structured original text according to a fixed format of user expression-interaction timestamp-channel identifier. For voice data, analyze the spectral differences between the voice signal and environmental noise, separate and suppress noise components, and then set two corresponding thresholds based on the voice signal's intensity and density. When a voice segment simultaneously meets the requirement that both the intensity and density are not lower than the corresponding threshold, it is determined to be a valid voice segment; the rest are silent segments. After removing the silent parts, it is converted into the PCM standard format. For speech-to-text data, filter high-confidence text based on a pre-defined transcription confidence threshold, mark segments that do not meet the threshold as pending verification, and retain the association link with the original voice. Step S102: The preprocessed data is split into word groups with independent semantics according to a preset semantic dictionary and word segmentation rules; based on preset grammar rules, each word group is labeled with part of speech and sentence function role, and a unique grammar category code value is assigned; then, word frequency features, grammar features, and semantic association features are extracted respectively; the word frequency feature represents the frequency of each word group in the current session data and the frequency ratio, and the frequency ratio of the word group is calculated by counting the number of occurrences of each word group and combining it with the total number of word groups in the current session data; the grammar feature is composed of the grammar category code, part of speech, and sentence function role of each word group, and a unique numerical code is assigned to the part of speech and sentence function role respectively, and the grammar category code, part of speech numerical code, and sentence function role numerical code are concatenated in a preset order to form a fixed-length one-dimensional numerical sequence; the semantic association feature represents the association state between each word group, and based on the grammar annotation results, the logical association between word groups is analyzed, and a unique identifier value is assigned to different association types; Step S103: Preset three weight coefficients for the corresponding word frequency features, grammatical features, and semantic association features, with the sum of the weight coefficients being 1; first, standardize the word frequency features, grammatical features, and semantic association features to the 0-1 value range, and then calculate the fusion feature value of each word group through weighted fusion calculation; according to the preset fixed dimension, arrange the fusion feature values ​​of all word groups in semantic order, padding with 0 values ​​if the dimension is insufficient, and retaining the feature values ​​corresponding to the core word groups if the dimension exceeds the dimension, generating fixed-dimensional structured semantic data; at the same time, collect the user's unique identifier, sort the multi-channel conversation data of the same user in ascending order by interaction timestamp, bind the channel feature tags to generate a cross-channel conversation time sequence index, and form a standardized associated data structure of unique identifier, timestamp, channel tag, and semantic data.

[0007] Furthermore, step S200 includes: Step S201: Define the entity association degree variable R to quantify the association strength of each word group in the structured semantic data; based on the co-occurrence frequency and semantic association distance of each semantic feature in the structured semantic data, the formula is: R=α×(f / (f+1))+(1-α)×(1 / (d+1)); where f represents the co-occurrence frequency, referring to the number of times two word groups appear simultaneously in the same dialogue round; d represents the semantic association distance, referring to the association path length of two word groups in the semantic structure, obtained by counting the number of intermediate word groups that are directly or indirectly associated between two word groups; where the minimum value of d is 0, indicating that the two words are not related by any other word group. Each word group is directly associated; α is a preset balance coefficient used to adjust the weight ratio of co-occurrence frequency and semantic association distance in the calculation of entity association degree R; word groups with R ≥ preset threshold R0 are selected as core entities, and feature information corresponding to each core entity is extracted simultaneously to establish a mapping relationship between entities and features; then, based on semantic association features, the logical association between core entities is judged, the association type and association strength are labeled, and a three-dimensional structured data of core entities, feature information and logical association between entities is formed, thereby constructing a dynamic session memory bank; wherein the association strength is the final quantification value of the calculation result of the entity association degree variable R; Step S202: After each round of dialogue, define the new data matching degree variable M, using the formula M=β×S+γ×T+δ×U; where S is the degree of overlap between the new 3D structured data and existing entities in the dynamic session memory, obtained by counting the number of identical word groups between the new entity and each existing entity in the dynamic session memory, divided by the total number of word groups in the new entity; T is the degree of feature information matching between the new entity and existing entities, obtained by extracting the feature set of the new entity and the feature set of the corresponding candidate entities in the memory, counting the number of completely identical features in the two sets, divided by the total number of features in the new entity; U is the degree of logical relationship matching between the new entity and other entities, obtained by judging the logical relationship type and relationship path between the new entity and related entities based on semantic association features, comparing it with the logical relationship patterns of existing similar entities in the dynamic session memory, counting the number of items where the logical relationship type and relationship path are consistent, divided by the total number of logical relationship items in the new entity; β, γ, and δ are the preset weight coefficients for entity overlap degree, feature matching degree, and logical relationship matching degree, respectively, and their sum is 1; A preset threshold M0 is defined. If M ≥ the preset threshold M0, the newly added data is merged into the association set of the corresponding entity; if M < M0, a new independent entity entry is added to the dynamic session memory. At the same time, a memory weight decay factor λ is defined to weaken the retrieval priority of outdated historical information. A weight increment ΔW is defined to strengthen the information importance brought by the newly added association, ΔW = k × M × R; where k is a preset gain coefficient with a value range of 0.1-0.3. The retrieval weights of each entity, feature and relationship in the dynamic session memory are updated by the formula: W1 = W0 × λ + ΔW; where W1 is the new weight after the update and W0 is the historical weight before the update.

[0008] Furthermore, step S300 includes: Step S301: Define a fuzzy expression feature set, which includes two types of features: expression components without clear entity references and semantic structures lacking core features. Locate the specific statements of the fuzzy expression by identifying phrases in the current session data that match the feature set. The first path traverses the dynamic session memory in reverse order of the interaction timestamp to extract the semantic features of each entity; the semantic association strength between the entity features and the fuzzy expression interval is calculated by cosine similarity, a similarity threshold is set, entities that reach the threshold are filtered and arranged in descending order of similarity value to generate an entity candidate sequence; The second path identifies missing feature information by parsing the semantic structure of fuzzy expressions; it retrieves feature records of similar entities in the dynamic session memory, counts the frequency of occurrence and association strength of each feature, filters features that simultaneously meet the frequency threshold and association strength requirements, and constructs a feature candidate set. Step S302: Perform cross-validation on the dual paths. First, perform combined validation on the entity candidate sequence and feature candidate set, and delete invalid combinations where there is no direct relationship between the entity and the feature. Next, sort the validated combinations according to the most recent interaction time of the entities, and select the entity feature combination with the highest ranking as the parsing result. When multiple combinations with the same time sequence appear, the combination with the highest semantic similarity shall be used. Fill the missing positions of the fuzzy expression with the selected entities and features, adjust the expression structure in combination with the contextual semantic logic, eliminate referential ambiguity and information loss, and form a complete semantic expression.

[0009] Furthermore, step S400 includes: Step S401: The first path is based on the complete semantic expression formed in step S302, extracts core entities and feature information, retrieves data from the preset knowledge base and business system, filters suitable data according to the matching degree of core entities and feature information, integrates the filtered data according to semantic logic, and forms the core response content that directly responds to the current user. The second path retrieves historical interaction data of the same user from the dynamic session memory, statistically analyzes the interaction behaviors and frequencies corresponding to core entities and various feature combinations, and defines behaviors with frequencies greater than a preset interaction frequency threshold as high-frequency associated interaction behaviors; and combines this with current feature information to obtain potential interaction directions that have a logical relationship with the current interaction behavior, and generates targeted guidance content based on these directions; the preset knowledge base is a structured database that stores semantic rules, problem solutions, and entity feature associations related to customer service; the business system data is real-time business data corresponding to customer service. Step S402: First, integrate the core response and supplementary guidance content generated in step S401, delete guidance directions that users have explicitly rejected in historical interactions and content that has logical conflicts with the core response, and form a basic response set; the channel adaptability means adjusting the presentation form and expression style of the response content according to the interaction characteristics of different channels and user habits; the adapted response content is fed back to the user in real time according to the output format supported by the corresponding channel to complete the full-channel intelligent interactive response.

[0010] An AI-based omnichannel intelligent customer service response and interaction system, comprising a data preprocessing module, a context management module, a semantic parsing module, and an intelligent response module; The data preprocessing module receives user session data from different channels, maps all types of session data into structured semantic data according to preset standardized semantic coding rules, and associates cross-channel session data through user unique identifiers; The context management module is based on a three-dimensional data structure of entity-feature-relationship. It captures the core entities, feature information and semantic relationships in multi-turn dialogues in real time from structured semantic data and establishes a dynamic conversation memory. After each round of dialogue, it automatically extracts new semantic data to supplement the dynamic conversation memory. The semantic parsing module targets the ambiguity of expressions and omitted information in the dialogue. First, it locates the position of the ambiguous expression through semantic features, and then performs parallel matching through two paths: the first path traverses the dynamic conversation memory in reverse order of interaction time to generate an entity candidate set, and selects suitable items by sorting by feature similarity; the second path deduces the missing core feature information of the ambiguous expression based on semantic logic and generates a feature candidate set; through cross-validation of the results of the two paths, it associates the core information in the preceding text and completes the omitted features to form a complete semantic expression. The intelligent response module is based on complete semantic expression. The first path matches the knowledge base and business system data to generate the core response; the second path combines historical data from the memory bank to predict potential related needs and generate supplementary guidance content; the results of the two paths are integrated, the optimal response combination is selected according to channel adaptability, and the interaction is completed by adapting the output format.

[0011] Furthermore, the data preprocessing module includes a multimodal data receiving unit, a semantic vectorization unit, and a cross-channel association unit. The multimodal data receiving unit receives user session data from different channels and processes it according to preset standardized semantic encoding rules. The semantic vectorization unit extracts and weights the semantic features of the preprocessed data to generate structured semantic data. The cross-channel association unit collects unique user identifiers, sorts multi-channel session data by interaction timestamps, binds channel feature tags to generate a time-series index, and establishes cross-channel data association. Furthermore, the context management module includes an entity relationship building unit and a dynamic memory maintenance unit: the entity relationship building unit calculates the word association degree, filters core entities, extracts corresponding features and marks the logical associations between entities to form three-dimensional structured data of entity-feature-association; the dynamic memory maintenance unit determines whether data should be merged or added by calculating the data matching degree, and updates the retrieval weight by combining the weight decay and incremental mechanism.

[0012] Furthermore, the semantic parsing module includes a fuzzy expression recognition unit, a dual-path parsing unit, and a semantic completion unit. The fuzzy expression recognition unit identifies referential and omitted information in the dialogue based on a preset fuzzy expression feature set, and locates the specific interval of the fuzzy expression. The dual-path parsing unit retrieves entity candidate sets from the memory in reverse order of interaction time, and generates feature candidate sets through semantic logic deduction to complete the parallel matching of dual paths. The semantic completion unit performs cross-validation and logical filling on the dual-path results to form a complete semantic expression.

[0013] Furthermore, the intelligent response module includes a dual-path response generation unit, an intelligent response fusion unit, and a multi-channel output adaptation unit. The dual-path response generation unit retrieves data from the knowledge base and business system to generate core responses, and combines historical data from the memory bank to deduce potential related needs and generate targeted guidance content. The intelligent response fusion unit integrates the results of dual-path responses, eliminates logical conflicts and guidance directions that users explicitly refuse, and optimizes them into a unified response combination. The multi-channel output adaptation unit performs content format conversion and presentation optimization according to the characteristics of different channels.

[0014] Compared with the prior art, the beneficial effects achieved by the present invention are: This invention establishes a context-aware capability that continuously tracks dialogue states by constructing a unified semantic space across channels and a dynamically evolving conversation memory. By employing a dual-path parallel processing mechanism, it considers both temporal features and semantic logic in the referential resolution stage, significantly improving the accuracy of intent recognition in complex dialogue scenarios. Furthermore, through a dual-path response generation and intelligent fusion strategy, it ensures accurate responses to core needs while proactively serving potential needs. Ultimately, it effectively solves technical challenges in multi-turn dialogues, such as incomplete semantic understanding and disconnect between response content and context caused by referential omission. This significantly improves the initial response accuracy, demand coverage, and cross-channel service consistency of the intelligent customer service system, providing users with a coherent and accurate interactive experience. Attached Figure Description

[0015] 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: Figure 1 This is a flowchart of an AI-based omnichannel intelligent customer service response and interaction method. Figure 2 This is a system module diagram of an AI-based omnichannel intelligent customer service response and interaction system. Detailed Implementation

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

[0017] Please see Figures 1-2 This invention provides a technical solution: an omnichannel intelligent customer service response and interaction method based on artificial intelligence, comprising: Step S100: Receive user session data from different channels, map all types of session data into structured semantic data according to preset standardized semantic encoding rules, and associate cross-channel session data through user unique identifiers; Step S200: Based on the entity-feature-association three-dimensional data structure, capture entity, feature information and semantic relationships in multi-turn dialogues in real time from structured semantic data to establish a dynamic conversation memory; and automatically extract new semantic data to supplement the dynamic conversation memory after each round of dialogue. Step S300: For the definite statements and omitted information in the dialogue, first locate the specific sentence position of the ambiguous statement through semantic features, and then perform parallel matching through two paths: the first path traverses the dynamic conversation memory in reverse order of interaction time to generate an entity candidate set, and selects the appropriate items by sorting by feature similarity; the second path identifies the missing feature information by parsing the semantic structure of the ambiguous statement and generates a feature candidate set; through cross-validation of the results of the two paths, the preceding information is associated and the omitted features are filled in to form a complete semantic statement; Step S400: Based on the complete semantic representation, the first path matches the knowledge base and business system data to generate the core response; the second path combines historical data from the memory bank to predict potential related interaction directions and generate supplementary guidance content; the results of the two paths are integrated, the optimal response combination is selected according to channel adaptability, and the interaction is completed by adapting the output format.

[0018] Step S100 includes: Step S101: Receive various types of session data through a pre-defined standardized channel access interface, categorizing them into three types: text data, voice data, and speech-to-text data. For text data, filter non-semantic characters and meaningless expressions using pre-defined character filtering rules, and organize it into structured original text according to a fixed format of user expression-interaction timestamp-channel identifier. For voice data, analyze the spectral differences between the voice signal and environmental noise, separate and suppress noise components, and then set two corresponding thresholds based on the voice signal's intensity and density. When a voice segment simultaneously meets the requirement that both the intensity and density are not lower than the corresponding threshold, it is determined to be a valid voice segment; the rest are silent segments. After removing the silent parts, it is converted into the PCM standard format. For speech-to-text data, filter high-confidence text based on a pre-defined transcription confidence threshold, mark segments that do not meet the threshold as pending verification, and retain the association link with the original voice. Step S102: The preprocessed data is split into word groups with independent semantics according to a preset semantic dictionary and word segmentation rules; based on preset grammar rules, each word group is labeled with its part of speech and sentence function role, and a unique grammar category code value is assigned; then, word frequency features, grammar features, and semantic association features are extracted respectively; the word frequency feature represents the frequency and frequency ratio of each word group in the current session data. The frequency ratio of each word group is calculated by counting the number of occurrences of each word group and combining it with the total number of word groups in the current session data; the grammar feature is composed of the grammar category code, part of speech, and sentence function role of each word group. A unique numerical code is assigned to the part of speech and sentence function role respectively, and the grammar category is... The part-of-speech (POS) encoding, part-of-speech (POS) numerical encoding, and statement function role (SFL) numerical encoding are concatenated in a preset order to form a fixed-length one-dimensional numerical sequence. The semantic association features represent the association status between each word group. Based on the grammatical annotation results, the logical association between word groups is analyzed. Semantic associations such as "core word group - modifying word group" and "action word group - object word group" are used as edges. Based on the preset association type rule base (including subject-predicate association, verb-object association, and modifier-head association type definitions), the part-of-speech combinations (such as verb + noun, matching verb-object association) and statement function role combinations (such as subject + predicate, matching subject-predicate association) between nodes are matched to identify and determine the specific logical association type between word groups. Finally, a unique identifier value is assigned to different logical association types.

[0019] The preset weight matrix contains weight coefficients corresponding to three features, namely ω1, ω2, and ω3, with a total weight coefficient of 1. Based on historical customer service conversation data, the semantic representation contribution of the three types of features is analyzed: word frequency feature contribution = (keyword recognition accuracy with word frequency feature - keyword recognition accuracy without word frequency feature), grammatical feature contribution = (sentence structure parsing accuracy with grammatical feature - sentence structure parsing accuracy without grammatical feature), and semantic association feature contribution = (logical association recognition accuracy with semantic association feature - logical association recognition accuracy without this feature). The contribution of each type of feature is divided by the sum of the contributions of the three types of features to obtain the contribution ratio of each feature, and this ratio is used as the initial weight coefficient. Then, the weighted average of semantic encoding accuracy, entity recognition accuracy, and semantic parsing accuracy is used as the comprehensive evaluation index. The weight coefficients are fine-tuned at a fixed step size, and the comprehensive evaluation index is recalculated after each adjustment until the index reaches the preset optimal threshold. Finally, the weight coefficients that balance the effects of the three types of features are determined. First, the word frequency features, grammatical features, and semantic association features are standardized to the 0-1 value range to eliminate the magnitude differences between different features. Then, the fused feature value of each word group is calculated using the formula F=ω1×F1+ω2×F2+ω3×F3, where F is the total feature value after fusion, F1 is the standardized word frequency feature value, F2 is the standardized grammatical feature value, and F3 is the standardized semantic association feature value. Following a preset fixed dimension (e.g., 512 dimensions), the fused feature values ​​of all word groups are arranged in semantic order (first sorted according to the grammatical structure of the sentence, prioritizing the subject and predicate). The system first extracts phrases corresponding to core semantic roles such as objects, then arranges phrases corresponding to auxiliary semantic roles such as attributives, adverbs, and complements (phrases with the same role are arranged according to their order of appearance in the original sentence). If the dimension is insufficient, it is padded with 0 values; if the dimension exceeds the limit, the feature values ​​corresponding to the core phrases are retained, generating fixed-dimensional structured semantic data. At the same time, the system collects unique user identifiers, sorts multi-channel conversation data of the same user in ascending order by interaction timestamps, binds channel feature tags to generate cross-channel conversation time-series indexes, and forms a standardized associated data structure of unique identifiers, timestamps, channel tags, and semantic vectors.

[0020] Example 1: User A first sends "View my orders for the past 3 months" via text message through an e-commerce app, and then inquires about "its payment status" via customer service phone. The system first receives two types of data through a standardized interface: text data is filtered to remove meaningless interjections such as "oh" and "ne," and formatted according to "User Statement-202405201015-APP"; voice data is processed by separating ambient noise through spectral differences, setting a sound intensity threshold of 60dB and a density threshold of 0.8, filtering effective speech segments, removing silences, and converting to PCM format. After transcription, the system filters for highly reliable text "Query its payment status" based on a confidence threshold of 0.9. Fragments that do not reach the threshold are marked as "to be verified" and associated with the original speech; after preprocessing, word groups are split into "order", "payment status" and "query", and labeled with part of speech and functional role, and word frequency / grammar / semantic association features are extracted; initial weights are calculated by contribution (ω1=0.3, ω2=0.4, ω3=0.3), and fine-tuned to the optimal level by semantic encoding accuracy; after feature standardization, weighted fusion is used to generate 512-dimensional semantic data, sorted by "subject-verb-object", and padded with 0 if necessary; two data are associated by user mobile phone number (unique identifier), and a cross-channel time-series index is generated in ascending order of timestamp to complete the structured mapping.

[0021] Step S200 includes: Step S201: Define the entity association degree variable R to quantify the association strength of each word group in the structured semantic data; based on the co-occurrence frequency and semantic association distance of each semantic feature in the structured semantic data, the formula is: R=α×(f / (f+1))+(1-α)×(1 / (d+1)); where f represents the co-occurrence frequency, which refers to the number of times two word groups appear simultaneously in the same dialogue round; d represents the semantic association distance, which refers to the length of the association path between two word groups in the semantic structure, obtained by counting the number of intermediate word groups that are directly or indirectly related between two word groups; where the minimum value of d is 0, indicating that the two word groups are directly related; α is a preset balance coefficient used to adjust the weight ratio of co-occurrence frequency and semantic association distance in the calculation of entity association degree R. Based on the historical customer service conversation data with labeled core entities and association relationships, 0.5 is used as the initial value (the default value for the two types of features in determining entity association). (With comparable basic impact), a fixed adjustment step size is set, using the core entity recognition accuracy (core entity recognition accuracy = the number of core entities selected based on the current α value that matches the real core entities labeled in historical data divided by the total number of real core entities labeled in historical data) and entity association matching accuracy (entity association matching accuracy = the number of core entity associations determined based on the current α value that matches the real associations labeled in historical data divided by the total number of real associations labeled in historical data) as core evaluation indicators, and iteratively adjusting the value of α: if the two indicators corresponding to the current α do not reach the preset optimal threshold, and the indicators increase after increasing α, then α is increased by the step size; if the indicators decrease or remain unchanged after increasing α, then α is decreased by the step size; after each adjustment, the indicators are recalculated, and the iteration is repeated until both indicators reach the preset optimal threshold, which is then determined as the final balance coefficient; Phrases with R ≥ a preset threshold R0 are selected as core entities. Feature information corresponding to each core entity is extracted simultaneously to establish a mapping relationship between entities and features. Then, based on semantic association features, the logical association between core entities is judged, and the association type and association strength are labeled to form three-dimensional structured data of core entities, feature information and logical association between entities, thereby constructing a dynamic conversation memory. The association strength is calculated using the entity association degree variable R as the final quantification value. Step S202: After each round of dialogue, define the new data matching degree variable M, using the formula M=β×S+γ×T+δ×U; where S is the degree of overlap between the new 3D structured data and existing entities in the dynamic session memory, obtained by counting the number of identical word phrases between the new entity and each existing entity in the dynamic session memory, divided by the total number of word phrases in the new entity; T is the degree of feature information matching between the new entity and existing entities, obtained by extracting the feature set of the new entity and the feature set of the corresponding candidate entities in the memory, counting the number of completely identical features in the two sets, divided by the total number of features in the new entity; U is the degree of logical relationship matching between the new entity and other entities, determined based on semantic association features, judging the logical association type and association path between the new entity and related entities, comparing it with the logical association patterns of existing similar entities in the dynamic session memory, counting the number of items where the logical association type and association path are consistent, divided by the total number of logical association items in the new entity; β, γ, and δ are the preset weight coefficients for entity overlap degree, feature matching degree, and logical relationship matching degree, respectively, with a sum of 1, as follows: First, the actual processing results of the newly added data that has been labeled are selected, that is, the historical customer service dialogue data that should be updated to existing entities or stored as new entities in the memory bank, as the basis for weight calibration; the degree of entity overlap S is the basic basis for judging whether they are the same entity, and the initial weight β is set to 0.4-0.5; the degree of feature matching T is the key dimension for verifying entity consistency, and the initial weight γ is set to 0.3-0.4; the degree of relational logic matching U is an auxiliary verification dimension, and the initial weight δ is set to 0.1-0.2, ensuring that the sum of the three is 1; Set a fixed adjustment step size, with data update accuracy and memory redundancy as the core evaluation indicators (data update accuracy = the number of results of "should be updated / should be added" determined based on the M value calculated based on the current weight that are consistent with the actual processing results marked in the historical data, divided by the total number of historical data samples; memory redundancy = the number of entities repeatedly stored in the memory after being determined based on the current weight, divided by the total number of entities in the memory), and iteratively adjust the weights: If the data update accuracy corresponding to the current weight does not reach the preset optimal threshold, or the memory redundancy rate is higher than the preset acceptable threshold, and the accuracy increases and the redundancy rate decreases after increasing a certain coefficient, then the coefficient is increased by a step size, and the other two types of coefficients are decreased simultaneously (keeping the sum to 1); if the accuracy decreases or the redundancy rate increases after adjustment, then the adjustment is reversed by a step size; the process is iterated until the data update accuracy reaches the preset optimal threshold and the memory redundancy rate is lower than the preset acceptable threshold, and finally the values ​​of β, γ, and δ are determined. A preset threshold M0 is defined. If M ≥ the preset threshold M0, the newly added data is merged into the association set of the corresponding entity; if M < M0, a new independent entity entry is added to the dynamic session memory. At the same time, a memory weight decay factor λ is defined to weaken the retrieval priority of outdated historical information. A weight increment ΔW is defined to strengthen the information importance brought by the newly added association, ΔW = k × M × R; where k is a preset gain coefficient with a value range of 0.1-0.3. The retrieval weights of each entity, feature and relationship in the dynamic session memory are updated by the formula: W1 = W0 × λ + ΔW; where W1 is the new weight after the update and W0 is the historical weight before the update.

[0022] Example 2: Entity association degree is calculated from semantic vectors. The co-occurrence frequency of "order," "query," and "payment status" is f=2, the semantic association distance is d=0, and the initial value of α is 0.5. After iterative adjustment based on the core entity recognition accuracy, it is determined to be 0.55. Substituting into the formula, we get R≈0.85≥R0=0.6. "Order" is identified as the core entity. The features of "last 3 months" and "payment status" are extracted, and the subject-verb association of "query-order" is labeled (association strength R=0.85). This forms three-dimensional structured data and is stored in the memory bank. After adding new data to the conversation, calculate the M value: S=0.9 (entity overlap), T=1 (feature matching degree), U=0.95 (relationship matching degree), β=0.45, γ=0.35, δ=0.2, resulting in M≈0.93≥M0=0.7, which is then merged into the "order" association set; set λ=0.9, k=0.2, calculate ΔW=0.2×0.93×0.85≈0.158, and update the weights according to the formula W1=W0×0.9+0.158 to strengthen the retrieval priority of "payment status"; Step S300 includes: Step S301: Define a fuzzy expression feature set, which includes two types of features: expression components without clear entity references and semantic structures lacking core features. By identifying phrases in the current session data that match the feature set, the specific statements of the fuzzy expression are located. The first path iterates through the dynamic session memory in reverse order of the interaction timestamps to extract the semantic features of each entity; the semantic association strength between the entity features and the fuzzy representation interval is calculated by cosine similarity, a similarity threshold is set, entities that reach the threshold are filtered and sorted in descending order of similarity value to generate an entity candidate sequence. The second approach identifies missing core features by parsing the semantic structure of fuzzy expressions, including unique entity identifiers (such as order numbers, account IDs, business application numbers, etc.), entity characteristics (such as product specifications, account types, business processing entities, etc.), current entity status (such as payment status, review progress, service validity period, etc.), operation methods (such as modifying fields, query dimensions, processing methods, etc.), time range (such as processing time, query period, validity period range, etc.), geographical range (such as service area, processing outlet affiliation, delivery area, etc.), and business scenarios (such as e-commerce orders, operator packages, etc.). It then retrieves feature records of similar entities from the memory, calculates the frequency and association strength of each feature, and filters features that simultaneously meet the frequency threshold and association strength requirements to construct a feature candidate set. Step S302: Perform cross-validation on the dual paths. First, perform combined validation on the entity candidate sequence and feature candidate set, and delete invalid combinations where there is no direct relationship between the entity and the feature. Next, sort the combinations that pass the consistency check according to the most recent interaction time of the entities, and prioritize the entity feature combination with the most recent interaction. Select the entity feature combination with the highest ranking as the parsing result. When multiple combinations with the same time sequence appear, the combination with higher semantic similarity shall prevail. Accurately fill the missing positions of the ambiguous expression with the selected entities and features, adjust the expression structure in combination with the contextual semantic logic, eliminate referential ambiguity and information loss, and form a complete semantic expression.

[0023] Example 3: First, through the fuzzy expression feature set, identify "it" as an expression without a clear entity reference, and locate the specific fuzzy statement: its payment status; the first path traverses the memory in reverse order, extracts the semantic features of the "order" entity, calculates the cosine similarity with the fuzzy interval S_sim=0.8≥S0=0.65, and generates the entity candidate sequence [order]; the second path parses the semantic structure, identifies the missing core feature of "unique entity identifier", retrieves the same "order" entity feature records, and includes the "user-associated order number" with a frequency of occurrence of 3≥F0=3 and an association strength R=0.85≥0.6 in the feature candidate set; during cross-validation, invalid combinations without association are deleted, and after verification, the "order-user-associated order number" combination is selected according to the most recent interaction time sequence, and the fuzzy expression is filled and the structure is adjusted to form the complete semantics: "Query the payment status of my own associated orders in the last 3 months".

[0024] Step S400 includes: Step S401: The first path, based on the complete semantic representation formed in step S302, extracts core entities and feature information, retrieves data from a preset knowledge base and business system, filters suitable data according to the matching degree of core entities and feature information, and integrates the filtered data according to semantic logic to form core response content that directly responds to the current user; the second path retrieves historical interaction data of the same user from the dynamic session memory, statistically analyzes the interaction behaviors and corresponding frequencies corresponding to core entities and each feature combination, defines behaviors with frequencies greater than a preset interaction frequency threshold as high-frequency associated interaction behaviors; and, combined with the current feature information, deduces potential interaction directions that have a logical relationship with the current interaction behavior, and generates targeted responses based on these directions. The guidance content is as follows: This derivation process is based on historical interaction records in a dynamic session memory bank. Statistical analysis methods are used to identify high-frequency associated interaction behavior sequences related to the current core entity and feature combination. Combined with the currently identified feature information, logically coherent potential interaction direction options not mentioned in the current dialogue are selected from the historical sequences. Then, based on the frequency of interaction behavior and business relevance (determined by business rules based on a preset knowledge base), priority is given to the direction ranked first as the core guidance direction. The preset knowledge base is a structured database storing semantic rules, problem solutions, and entity feature relationships related to customer service. The business system data is the real-time business data corresponding to customer service. Step S402: First, integrate the core responses and supplementary guidance content generated in step S401, delete guidance directions that users have explicitly rejected in historical interactions and content that logically conflicts with the core responses, forming a basic response set; the channel adaptability means adjusting the presentation format and expression style of the response content according to the interaction characteristics of different channels and user habits, so as to make information transmission efficient and adaptable; adaptation is achieved based on the channel feature tags in step S100: text channels adopt a point-based structured format to distinguish between core responses and guidance content; voice channels adopt short sentences and colloquial expressions, simplify complex sentence structures and supplement natural tone connection components; speech-to-text channels retain the integrity of text information and mark and emphasize core entities, features and key business nodes; the adapted response content is fed back to the user in real time according to the output format supported by the corresponding channel to complete the intelligent interactive response across all channels.

[0025] Example 4: The first path extracts the core entity "orders," features "last 3 months," and "payment status," searches the knowledge base and e-commerce business system, filters data based on a core entity matching degree of 0.95 and a feature matching degree of 0.9, and integrates to obtain the core response "Your related orders in the last 3 months have been paid for." The second path retrieves historical data, statistically analyzes the high-frequency association of "order-payment status" with "logistics query" (frequency = 12 ≥ threshold 8), sorts by business relevance, deduces potential interaction directions, and generates guiding content "Do you need to query the corresponding logistics progress?" After integration, conflict-free content is deleted, and based on telephone channel features, a conversational short sentence is used: "Your related orders in the last 3 months have been paid for. If you want to know the logistics progress, I can help you query it quickly," supplemented with natural connecting words, and real-time feedback is provided in the voice stream format to complete the omnichannel interaction.

[0026] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An omnichannel intelligent customer service response and interaction method based on artificial intelligence, characterized by: The method includes: Step S100: Receive user session data from different channels, map all types of session data into structured semantic data according to preset standardized semantic encoding rules, and associate cross-channel session data through user unique identifiers; Step S200: Based on the entity-feature-association three-dimensional data structure, capture entity, feature information and semantic associations in multi-turn dialogues in real time from structured semantic data to establish a dynamic conversation memory; and automatically extract new semantic data to supplement the dynamic conversation memory after each round of dialogue. Step S300: For the definite statements and omitted information in the dialogue, first locate the position of the ambiguous statement through semantic features, and then perform parallel matching through two paths: the first path traverses the dynamic conversation memory in reverse order of interaction time to generate an entity candidate set, and selects the appropriate items by sorting by feature similarity; the second path identifies the missing feature information by parsing the semantic structure of the ambiguous statement and generates a feature candidate set; through cross-validation of the results of the two paths, the preceding information is associated and the omitted features are filled in to form a complete semantic statement; Step S400: Based on the complete semantic representation, the first path matches the knowledge base and business system data to generate the core response; the second path combines historical data from the memory bank to predict potential related interaction directions and generate supplementary guidance content; the results of the two paths are integrated, the optimal response combination is selected according to channel adaptability, and the interaction is completed by adapting the output format.

2. The omnichannel intelligent customer service response interaction method based on artificial intelligence according to claim 1, characterized in that: Step S100 includes: Step S101: Receive various types of session data through a pre-defined standardized channel access interface, categorizing them into three types: text data, voice data, and speech-to-text data. For text data, filter non-semantic characters and meaningless expressions using pre-defined character filtering rules, and organize it into structured original text according to a fixed format of user expression-interaction timestamp-channel identifier. For voice data, analyze the spectral differences between the voice signal and environmental noise, separate and suppress noise components, and then set two corresponding thresholds based on the voice signal's intensity and density. When a voice segment simultaneously meets the requirement that both the intensity and density are not lower than the corresponding threshold, it is determined to be a valid voice segment; the rest are silent segments. After removing the silent parts, it is converted into the PCM standard format. For speech-to-text data, filter high-confidence text based on a pre-defined transcription confidence threshold, mark segments that do not meet the threshold as pending verification, and retain the association link with the original voice. Step S102: The preprocessed data is split into word groups with independent semantics according to a preset semantic dictionary and word segmentation rules; based on preset grammar rules, each word group is labeled with part of speech and sentence function role, and a unique grammar category code value is assigned; then, word frequency features, grammar features, and semantic association features are extracted respectively; the word frequency feature represents the frequency of each word group in the current session data and the frequency ratio, and the frequency ratio of the word group is calculated by counting the number of occurrences of each word group and combining it with the total number of word groups in the current session data; the grammar feature is composed of the grammar category code, part of speech, and sentence function role of each word group, and a unique numerical code is assigned to the part of speech and sentence function role respectively, and the grammar category code, part of speech numerical code, and sentence function role numerical code are concatenated in a preset order to form a fixed-length one-dimensional numerical sequence; the semantic association feature represents the association state between each word group, and based on the grammar annotation results, the logical association between word groups is analyzed, and a unique identifier value is assigned to different association types; Step S103: Preset three weight coefficients for the corresponding word frequency features, grammatical features, and semantic association features, with the sum of the weight coefficients being 1; first, standardize the word frequency features, grammatical features, and semantic association features to the 0-1 value range, and then calculate the fusion feature value of each word group through weighted fusion calculation; according to the preset fixed dimension, arrange the fusion feature values ​​of all word groups in semantic order, padding with 0 values ​​if the dimension is insufficient, and retaining the feature values ​​corresponding to the core word groups if the dimension exceeds the dimension, generating fixed-dimensional structured semantic data; at the same time, collect the user's unique identifier, sort the multi-channel conversation data of the same user in ascending order by interaction timestamp, bind the channel feature tags to generate a cross-channel conversation time sequence index, and form a standardized associated data structure of unique identifier, timestamp, channel tag, and semantic data.

3. The omnichannel intelligent customer service response interaction method based on artificial intelligence according to claim 2, characterized in that: Step S200 includes: Step S201: Define the entity association degree variable R to quantify the association strength of each word group in the structured semantic data; based on the co-occurrence frequency and semantic association distance of each semantic feature in the structured semantic data, the formula is: R=α×(f / (f+1))+(1-α)×(1 / (d+1)); where f represents the co-occurrence frequency, referring to the number of times two word groups appear simultaneously in the same dialogue round; d represents the semantic association distance, referring to the association path length of two word groups in the semantic structure, obtained by counting the number of intermediate word groups that are directly or indirectly associated between two word groups; where the minimum value of d is 0, indicating that the two words are not related by any other word group. Each word group is directly associated; α is a preset balance coefficient used to adjust the weight ratio of co-occurrence frequency and semantic association distance in the calculation of entity association degree R; word groups with R ≥ preset threshold R0 are selected as core entities, and feature information corresponding to each core entity is extracted simultaneously to establish a mapping relationship between entities and features; then, based on semantic association features, the logical association between core entities is judged, the association type and association strength are labeled, and a three-dimensional structured data of core entities, feature information and logical association between entities is formed, thereby constructing a dynamic session memory bank; wherein the association strength is the final quantification value of the calculation result of the entity association degree variable R; Step S202: After each round of dialogue, define the new data matching degree variable M, using the formula M=β×S+γ×T+δ×U; where S is the degree of overlap between the new 3D structured data and existing entities in the dynamic session memory, obtained by counting the number of identical word groups between the new entity and each existing entity in the dynamic session memory, divided by the total number of word groups in the new entity; T is the degree of feature information matching between the new entity and existing entities, obtained by extracting the feature set of the new entity and the feature set of the corresponding candidate entities in the memory, counting the number of completely identical features in the two sets, divided by the total number of features in the new entity; U is the degree of logical relationship matching between the new entity and other entities, obtained by judging the logical relationship type and relationship path between the new entity and related entities based on semantic association features, comparing it with the logical relationship patterns of existing similar entities in the dynamic session memory, counting the number of items where the logical relationship type and relationship path are consistent, divided by the total number of logical relationship items in the new entity; β, γ, and δ are the preset weight coefficients for entity overlap degree, feature matching degree, and logical relationship matching degree, respectively, and their sum is 1; A preset threshold M0 is defined. If M ≥ the preset threshold M0, the newly added data is merged into the association set of the corresponding entity; if M < M0, a new independent entity entry is added to the dynamic session memory. A memory weight decay factor λ is defined to weaken the retrieval priority of outdated historical information. At the same time, a weight increment ΔW is defined to strengthen the information importance brought by the newly added association, ΔW = k × M × R; where k is a preset gain coefficient with a value range of 0.1-0.

3. The retrieval weights of each entity, feature, and relationship in the dynamic session memory are updated using the formula: W1 = W0 × λ + ΔW; where W1 is the new weight after the update, and W0 is the historical weight before the update.

4. The omnichannel intelligent customer service response interaction method based on artificial intelligence according to claim 3, characterized in that: Step S300 includes: Step S301: Define a fuzzy expression feature set, which includes two types of features: expression components without clear entity references and semantic structures lacking core features. Locate the specific statements of the fuzzy expression by identifying phrases in the current session data that match the feature set. The first path traverses the dynamic session memory in reverse order of the interaction timestamp to extract the semantic features of each entity; the semantic association strength between the entity features and the fuzzy expression interval is calculated by cosine similarity, a similarity threshold is set, entities that reach the threshold are filtered and arranged in descending order of similarity value to generate an entity candidate sequence; The second path identifies missing feature information by parsing the semantic structure of fuzzy expressions; it retrieves feature records of similar entities in the dynamic session memory, counts the frequency of occurrence and association strength of each feature, filters features that simultaneously meet the frequency threshold and association strength requirements, and constructs a feature candidate set. Step S302: Perform cross-validation on the dual paths. First, perform combined validation on the entity candidate sequence and feature candidate set, and delete invalid combinations where there is no direct relationship between the entity and the feature. Next, sort the validated combinations according to the most recent interaction time of the entities, and select the entity feature combination with the highest ranking as the parsing result. When multiple combinations with the same time sequence appear, the combination with the highest semantic similarity shall be used. Fill the missing positions of the fuzzy expression with the selected entities and features, adjust the expression structure in combination with the contextual semantic logic, eliminate referential ambiguity and information loss, and form a complete semantic expression.

5. The omnichannel intelligent customer service response interaction method based on artificial intelligence according to claim 4, characterized in that: Step S400 includes: Step S401: The first path is based on the complete semantic expression formed in step S302, extracts core entities and feature information, retrieves data from the preset knowledge base and business system, filters suitable data according to the matching degree of core entities and feature information, integrates the filtered data according to semantic logic, and forms the core response content that directly responds to the current user. The second path retrieves historical interaction data of the same user from the dynamic session memory, statistically analyzes the interaction behaviors and frequencies corresponding to core entities and various feature combinations, and defines behaviors with frequencies greater than a preset interaction frequency threshold as high-frequency associated interaction behaviors. Combined with current feature information, it obtains potential interaction directions that have a logical relationship with the current interaction behavior, and generates targeted guidance content based on these directions. The preset knowledge base is a structured database storing semantic rules, problem solutions, and entity feature associations related to customer service. The business system data is real-time business data corresponding to customer service. Step S402: First, integrate the core response and supplementary guidance content generated in step S401, delete guidance directions that users have explicitly rejected in historical interactions and content that has logical conflicts with the core response, and form a basic response set; the channel adaptability means adjusting the presentation form and expression style of the response content according to the interaction characteristics of different channels and user habits; the adapted response content is fed back to the user in real time according to the output format supported by the corresponding channel to complete the full-channel intelligent interactive response.

6. An omnichannel intelligent customer service response and interaction system based on artificial intelligence, characterized in that: The system includes a data preprocessing module, a context management module, a semantic parsing module, and an intelligent response module; The data preprocessing module receives user session data from different channels, maps various types of session data into structured semantic data according to preset standardized semantic coding rules, and associates cross-channel session data through user unique identifiers; The context management module is based on the entity-feature-association three-dimensional data structure. It captures the core entities, feature information and semantic relationships in multi-turn dialogues in real time from structured semantic data to establish a dynamic conversation memory. After each round of dialogue, it automatically extracts new semantic data to supplement the dynamic conversation memory. The semantic parsing module targets the ambiguity of expressions and omitted information in the dialogue. First, it locates the position of the ambiguous expression through semantic features, and then performs parallel matching through two paths: the first path traverses the dynamic conversation memory in reverse order of interaction time to generate an entity candidate set, and filters the matching items by sorting by feature similarity; the second path deduces the core feature missing information of the ambiguous expression based on semantic logic to generate a feature candidate set. By cross-validating the results from two paths, we can associate the core information from the preceding text and complete the omitted features to form a complete semantic representation. The intelligent response module is based on complete semantic expression. The first path matches the knowledge base and business system data to generate the core response; the second path combines historical data in the memory bank to predict potential related needs and generate supplementary guidance content. The system integrates the results from both paths, filters the optimal response combination based on channel compatibility, and completes the interaction by adapting the output format.

7. The omnichannel intelligent customer service response and interaction system based on artificial intelligence according to claim 6, characterized in that: The data preprocessing module includes a multimodal data receiving unit, a semantic vectorization unit, and a cross-channel association unit; The multimodal data receiving unit receives user session data from different channels and encodes it according to preset standardized semantic encoding rules. The semantic vectorization unit extracts and weights the semantic features of the preprocessed data to generate structured semantic data. The cross-channel association unit collects the user's unique identifier, sorts multi-channel session data by interaction timestamp, binds channel feature tags to generate a time-series index, and establishes cross-channel data association.

8. The omnichannel intelligent customer service response and interaction system based on artificial intelligence according to claim 6, characterized in that: The context management module includes an entity relationship construction unit and a dynamic memory maintenance unit; The entity relationship construction unit calculates the word association degree, filters core entities, extracts corresponding features and marks the logical associations between entities, forming three-dimensional structured data of entity-feature-association; The dynamic memory maintenance unit determines whether data should be merged or added by calculating the data matching degree, and updates the retrieval weight by combining the weight decay and incremental mechanism.

9. The omnichannel intelligent customer service response and interaction system based on artificial intelligence according to claim 6, characterized in that: The semantic parsing module includes a fuzzy representation recognition unit, a dual-path parsing unit, and a semantic completion unit; The fuzzy expression recognition unit identifies the descriptive and omitted information in the dialogue based on a preset fuzzy expression feature set, and locates the position of the fuzzy expression in the sentence. The dual-path parsing unit retrieves entity candidate sets from the memory in reverse order of interaction time, and generates feature candidate sets through semantic logic deduction to complete dual-path parallel matching; The semantic completion unit performs cross-validation and logical filling on the dual-path results to form a complete semantic representation.

10. The omnichannel intelligent customer service response and interaction system based on artificial intelligence according to claim 6, characterized in that: The intelligent response module includes a dual-path response generation unit, an intelligent response fusion unit, and a multi-channel output adaptation unit. The dual-path response generation unit retrieves data from the knowledge base and business system to generate core responses, and combines historical data from the memory bank to deduce potential related needs and generate targeted guidance content. The intelligent response fusion unit integrates the dual-path response results, eliminates logical conflicts and the guidance direction explicitly rejected by the user, and optimizes to form a unified response combination; The multi-channel output adaptation unit performs content format conversion and presentation optimization based on the characteristics of different channels.

Images