Intelligent dialogue management method and system based on context understanding

Abstract

The invention belongs to the technical field of man-machine dialogue, and particularly relates to an intelligent dialogue management method and system based on context understanding, a reference relation confirmation mechanism is introduced after semantic analysis, comprehensive matching is carried out by utilizing vocabulary overlapping proportion and entity category consistency, and historical entities pointed by fuzzy expression are accurately recognized; and further judging whether the current request continues the existing intention or not by comparing the consistency of the operation action and the historical target demand and combining the semantic matching degree of weighted accumulation. According to the method, cross-round semantic accurate association and intention continuity judgment are realized, the problems of unknown anaphora and intention misjudgment caused by information splitting in multiple rounds of dialogues are effectively solved, and the context understanding ability and interaction naturalness of a system are remarkably improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of human-computer dialogue, and particularly relates to an intelligent dialogue management method and system based on context understanding. BACKGROUND

[0002] In the current human-computer dialogue system, a technical scheme based on single-round semantic recognition and intent classification is generally adopted. The system usually processes each sentence of the user independently, extracts keywords or intent labels through a natural language understanding module, and generates a reply in combination with a preset rule or model. Such a method relies on static analysis of the current input and lacks effective modeling of information continuity in multi-round dialogue. When the user uses pronouns, omits expressions or supplements and corrects the previous content, the system is difficult to accurately identify the association between sentences, resulting in context breakage, reference confusion or intent misjudgment.

[0003] For example, the user first asks for a Sichuan restaurant, and then supplements the position of the restaurant mentioned last night. Since the system fails to associate the restaurant with the Sichuan restaurant in the previous dialogue, it cannot determine the specific target and can only request the user to explain again. This information fragmentation problem seriously affects the coherence of the dialogue and the user experience, especially in complex task scenarios, where the user needs to repeat the information repeatedly, resulting in low interaction efficiency. SUMMARY

[0004] The purpose of the present application is to provide an intelligent dialogue management method and system based on context understanding, which realizes accurate association of cross-round semantics and intent continuity judgment, effectively solves the problem of unclear reference and intent misjudgment caused by information fragmentation in multi-round dialogue, and significantly improves the context understanding ability and interaction naturalness of the system.

[0005] To achieve the above purpose, the application adopts the following technical scheme: an intelligent dialogue management method based on context understanding, comprising the following steps: Obtain the complete sentence text of the current input of the user, and perform word segmentation and dependency syntax analysis to generate a syntax dependency structure; Recognize semantic roles based on the syntax dependency structure and generate semantic triples. Standardize and disambiguate the temporal expressions and entity references in the triples; Retrieve past semantic records in the historical dialogue memory library, locate the reference object in the historical records according to the disambiguated semantic triples, calculate the overlap ratio of indicative phrases and entity descriptions in the historical records and the consistency score of entity categories, and obtain a comprehensive matching value by weighting and summing the two, and confirm the reference relationship when the comprehensive matching value exceeds the preset judgment threshold; Replace the vague expressions in the current semantics after confirming the reference relationship with specific historical entity references; After the replacement, the operation action in the updated semantic triple is compared with the confirmed target demand in the historical dialogue memory library to check whether there is a constraint condition in violation, and a weighted cumulative value of semantic matching is calculated based on the decay weight of the historical semantic record of each round, and when the weighted cumulative value exceeds the set level and there is no violation constraint, it is determined that the current request continues the existing direction; The processed semantic triple is converted into a new node and connection in the dynamic semantic memory graph, and the activity and correlation strength of the related nodes in the graph are adjusted; According to the updated dynamic semantic memory graph, the complete intention of the user is determined, a natural language response containing historical associated information is constructed and output, and an external service call is started.

[0006] Preferably, the standardization and ambiguity resolution of the time expression and entity reference in the triple include: According to the current system time, the specific date and time point corresponding to the relative time expression are inferred; The inferred time is converted into a unified format time mark; Combined with the entity type information appearing in the previous dialogue, the specific direction of the action object in the current semantic is determined, and the action verb is normalized to a standard operation name.

[0007] Preferably, the positioning of the reference object includes: Identify the indicative phrase in the current semantic; Traverse the semantic records saved in each round of the historical dialogue memory library; Find the item that has been described as a specific type of entity, confirm that the entity has not been denied or replaced in the dialogue process and is still in a extendable state.

[0008] Preferably, the confirmation of the reference relationship includes: Replace the indicative phrase with the matching entity name in the historical record, add a role representing the target place in the semantic triple, and use the matching entity name as the value of the role, to generate an updated semantic triple containing the new role.

[0009] Preferably, the determination that the current request continues the existing direction includes: The operation action in the updated semantic triple is created as an action node, the unified format time mark is created as a time node, the target association connection between the action node and the entity node is established, and the time association connection between the action node and the time node is established.

[0010] Preferably, the adjustment of the activity of the related nodes in the graph includes: According to whether the action node and the entity node are referenced in this round, update the activity score thereof; Lower the activity of the early nodes which are not cited in this round by a fixed decay rate; When the node activity is lower than the set lower limit and is not activated for many rounds in succession, put it into the state of recycling.

[0011] Preferably, the determination of the complete intention of the user comprises: Scan the action nodes in the state of waiting for processing in the map, extract the entity nodes, time nodes and initiating subjects associated with the action nodes; Combine the extracted information into a request structure containing operations, targets, times and subjects.

[0012] Preferably, the construction of the natural language response containing historical association information comprises: Repeat the operation, target and time information in the request structure; Extract the previous expression content associated with the initiating subject in the map; Add the previous expression content as supplementary explanation to the response text.

[0013] Preferably, the starting of the external service call comprises: Encapsulate the request structure into a standard format and send it to the external service interface; Listen to and receive the execution result message returned by the external service, parse the key elements in it, create a credential node and establish a connection with the action node, and synchronously update the state of the action node.

[0014] On the other hand, the present application proposes an intelligent dialogue management system based on context understanding, comprising: an input analysis unit for obtaining the complete sentence text of the current input of the user, performing word segmentation and dependency syntax analysis, generating a syntax dependency structure, identifying semantic roles and generating semantic triples, and standardizing and disambiguating the temporal expressions and entity references in the triples; A memory retrieval unit for retrieving past semantic records in the historical dialogue memory library, locating the reference object in the historical records according to the disambiguated semantic triples, calculating the overlap ratio of the indicative phrases and the entity descriptions in the historical records and the consistency score of the entity categories, weighting and summing them to obtain a comprehensive matching value, and confirming the reference relationship when the comprehensive matching value exceeds the preset judgment threshold; A map construction unit for replacing the ambiguous expressions in the current semantics after confirming the reference relationship to be established with specific historical entity references, converting the replaced semantic triples into new nodes and connections in the dynamic semantic memory map, and adjusting the activity and association strength of the related nodes in the map; A response generation unit for determining the complete intention of the user according to the updated dynamic semantic memory map, constructing a natural language response containing historical association information and outputting, and simultaneously starting an external service call.

[0015] The technical effects and advantages of the present application: the intelligent dialogue management method and system based on context understanding has the following advantages compared with the prior art: The present application introduces a reference relationship confirmation mechanism after semantic analysis, uses the comprehensive matching of the proportion of lexical overlap and the consistency of entity categories to accurately identify the historical entity pointed by the ambiguous expression; further, by comparing the consistency of the operation action and the historical target demand, and combining the weighted cumulative semantic matching degree, it is judged whether the current request continues the existing intention. This method realizes the accurate association of cross-round semantics and the intention continuation judgment, effectively solves the problems of unclear reference and intention misjudgment caused by information fragmentation in multi-round dialogue, and significantly improves the context understanding ability and interaction naturalness of the system. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 The flowchart of the intelligent dialogue management method based on context understanding of the present application; Figure 2 The block diagram of the intelligent dialogue management system based on context understanding of the present application. DETAILED DESCRIPTION

[0017] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all the embodiments. The specific embodiments described herein are only used to explain the present application, and are not used to limit the present application. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0018] The present application provides an intelligent dialogue management method based on context understanding as shown in Figure 1 The present application provides an intelligent dialogue management method based on context understanding as shown in The system acquires the complete text of the user's current input, performs word segmentation and dependency parsing, and generates a syntactic dependency structure. Based on the syntactic dependency structure, it identifies semantic roles and generates semantic triples, standardizing and dissolving ambiguities in the time expressions and entity references within the triples. Specifically, this includes: inferring the specific date and time corresponding to the relative time expression based on the current system time; converting the inferred time into a unified time stamp; and combining entity type information from the preceding dialogue to determine the specific target of the action object in the current semantics, and unifying the action verb into a standard operation name.

[0019] By accurately parsing the grammatical structure of user statements and generating standardized semantic representations, the system effectively solves the problems of semantic understanding bias caused by ambiguous time expressions, unclear references, and varied verb forms in natural language, thereby improving the system's ability to accurately capture user intent.

[0020] Retrieve past semantic records from the historical dialogue memory bank, locate the referent in the historical records based on the resolved semantic triples, calculate the lexical overlap ratio and entity category consistency score between the indicative phrase and the entity description in the historical records, and obtain a comprehensive matching value by weighted summation of the two. When the comprehensive matching value exceeds the preset judgment threshold, the referential relationship is confirmed to be established. The process of locating the referent includes: identifying indicative phrases in the current semantics; traversing the semantic records saved in each round in the historical dialogue memory bank; searching for items that have been described as specific types of entities, and confirming that the entities have not been negated or replaced in the dialogue process and are still in a sustainable state.

[0021] By establishing a comprehensive matching mechanism based on lexical overlap and category consistency, combined with the validity judgment of historical entity states, the accuracy and robustness of referential resolution are significantly improved. This effectively solves the semantic ambiguity problem caused by pronouns or vague referentials in multi-turn dialogues and enhances the system's ability to coherently understand contextual information.

[0022] After confirming the establishment of the referential relationship, the vague expression in the current semantics is replaced with a specific historical entity reference; after confirming the establishment of the referential relationship, the indicative phrase is replaced with the entity name matched in the history record, a role representing the target location is added to the semantic triple, and the matched entity name is used as the value of the role, generating an updated semantic triple containing the new role.

[0023] It achieves accurate transformation of fuzzy descriptive expressions into specific semantic content, enhances the integrity and operability of semantic triples, provides an accurate and unambiguous input foundation for subsequent intent recognition and graph construction, and effectively supports the continuous evolution of dialogue states across rounds.

[0024] After the replacement is completed, the operation actions in the updated semantic triples are compared with the target requirements confirmed in the historical dialogue memory to see if they are consistent with the direction. It is checked whether there are any conflicting constraints. The weighted cumulative value of semantic matching is calculated based on the attenuation weight of the semantic records of each historical round. When the weighted cumulative value exceeds the set level and there are no constraints that are violated, it is determined that the current request continues the existing direction. After determining that the current request continues the existing direction, the operation actions in the updated semantic triples are created as action nodes, the unified format timestamps are created as time nodes, the target association connection between action nodes and entity nodes is established, and the time association connection between action nodes and time nodes is established.

[0025] By integrating the consistency judgment of operation direction with the weighted semantic accumulation mechanism, the system accurately identifies the continuity relationship between user requests and historical intents, effectively avoiding misjudgment of intent caused by zero reference or implicit correction. At the same time, the verified semantic structure is structured into graph nodes and associated connections, providing a traceable and scalable semantic foundation for the dynamic evolution of dialogue states, and enhancing the system's adaptability to complex interaction scenarios.

[0026] The processed semantic triples are transformed into new nodes and connections in a dynamic semantic memory graph, and the activity and association strength of related nodes in the graph are adjusted. The adjustment of the activity level of relevant nodes in the graph includes: updating the activity score of action nodes and entity nodes based on whether they are referenced in the current round; reducing the activity level of early nodes that are not referenced in the current round by a fixed decay rate; and placing a node into a pending recycling state when its activity level is below a set lower limit and it has not been activated for several consecutive rounds.

[0027] By dynamically maintaining node activity and association strength, the system continuously tracks important information in the dialogue context and gradually fades out low-relevance information, effectively ensuring the timeliness and accuracy of semantic memory and improving the system's ability to focus on core intentions and resource management efficiency in long-cycle, multi-round interactions.

[0028] The system determines the user's complete intent based on the updated dynamic semantic memory graph, constructs and outputs a natural language response containing historical association information, and simultaneously initiates external service calls.

[0029] Determining the user's complete intent includes: scanning action nodes in the graph that are in a pending state, extracting the entity nodes, time nodes, and initiating entities associated with the action node; and combining the extracted information into a request structure that includes the operation, target, time, and subject.

[0030] Furthermore, a natural language response containing historical association information is constructed, including: restating the operation, target, and time information in the request structure; extracting the previous expressions associated with the initiating entity from the graph; and adding the previous expressions as supplementary explanations to the response text.

[0031] Specifically, initiating an external service call includes: encapsulating the request structure into a standard format and sending it to the external service interface; listening for and receiving the execution result message returned by the external service, parsing the key elements, creating a credential node and establishing a connection with the action node, and synchronously updating the status of the action node.

[0032] By integrating structured information from a dynamic semantic memory graph, a complete intent expression encompassing the current request and historical preferences is accurately generated, significantly improving the accuracy and contextual coherence of responses. Incorporating previously expressed content into natural language responses enhances the personalization and naturalness of the interaction. Simultaneously, the intent parsing results are transformed into standardized service requests and a closed-loop relationship for execution feedback is established, ensuring that system actions are traceable and states are synchronized, greatly improving the reliability of task execution and users' trust in the interaction process.

[0033] On the other hand, this invention proposes an intelligent dialogue management system based on context understanding, such as... Figure 2 As shown, it includes: The input parsing unit is used to obtain the complete text of the current user input, perform word segmentation and dependency parsing, generate syntactic dependency structures, identify semantic roles and generate semantic triples, and standardize and disambiguate the time expressions and entity references in the triples. The memory retrieval unit is used to retrieve past semantic records in the historical dialogue memory bank. Based on the resolved semantic triples, it locates the referent in the historical records. By calculating the lexical overlap ratio between the indicative phrase and the entity description in the historical records and the entity category consistency score, the two are weighted and summed to obtain a comprehensive matching value. When the comprehensive matching value exceeds the preset judgment threshold, the referential relationship is confirmed to be established. The graph construction unit is used to replace the vague expression in the current semantics after the establishment of the referential relationship with the specific historical entity reference, and to transform the replaced semantic triples into new nodes and connections in the dynamic semantic memory graph, and to adjust the activity and association strength of related nodes in the graph. The response generation unit is used to determine the user's complete intent based on the updated dynamic semantic memory graph, construct and output a natural language response containing historical association information, and simultaneously initiate external service calls.

[0034] In addition, each unit in the above system is also used to implement other steps of the aforementioned context-understanding-based intelligent dialogue management method, as follows: Step 1: Receive and parse user input statements, and extract basic semantic units. In multi-turn dialogue environments, each user input may carry new intentions or supplement and correct previous content. If the system cannot accurately capture the basic semantic components of the current input, all subsequent processing will lose its foundation. Therefore, the primary task is to perform structured parsing of the natural language text input by the user, separating language fragments with independent meaning as the basic units for subsequent context tracking and intent inference. This process involves not only lexical segmentation but also combining grammatical structures and common expression patterns to identify elements such as actions, objects, and modification relationships, forming a primary semantic representation that can be used for subsequent processing.

[0035] Sub-step 1.1: Obtain the complete text of the currently entered statement by the user: The system first listens for input from the user's terminal, and when it detects the arrival of text information, it immediately captures the complete sentence. ,in This indicates the time sequence number of the current conversation round. For example, if a user enters "I'd like to book a table for 7 PM tonight at that restaurant we talked about last time" in round 3, then... This statement serves as the raw material for this round of processing; its character sequence is recorded in its entirety and proceeds to the next stage. This step ensures that no surface information is omitted, providing the initial basis for subsequent in-depth analysis. Since all subsequent operations are based on this text, its completeness and accuracy must be guaranteed to avoid semantic deviations caused by truncation or encoding errors.

[0036] Sub-step 1.2: Perform tokenization and dependency parsing on the acquired statement text: In obtaining Then, the system initiates a language structure parsing program to perform fine-grained decomposition of the sentence. Specifically, it uses word segmentation technology to divide the continuous string into several word units and establishes grammatical relationships between the words using dependency parsing. For example, in "I want to book a seat at 7 PM tonight," the system identifies the subject "I," the predicate "want to book," the object "seat," and the time modifier "7 PM tonight." The dependency relationships between these components form a directed graph. , where nodes Representative words or phrases, edge This indicates the direction of grammatical dependencies (such as "modification," "dominance," etc.). This graph structure reveals the logical skeleton within a sentence, enabling the system to distinguish between core actions and auxiliary descriptions, thus laying the foundation for the next step of semantic role labeling. After this step, the original text is transformed into a structured linguistic expression, facilitating further machine understanding of its inherent meaning.

[0037] Sub-step 1.3: Identify semantic roles and generate semantic triples based on syntactic structure: After building Building upon this foundation, the system further analyzes the semantic roles of each component. For example, in the action of "booking a place," "I" is the agent, "place" is the patient, and "7 PM tonight" is an adverb of time. These roles are categorized into standard semantic classes (such as agent, patient, time, and place) and bound to corresponding language segments. Subsequently, the system organizes this information into a set of semantic triples. ,in Tag the character As a role taker, This refers to either a value or a description. Using the above statement as an example, a set of triples can be generated: ; This representation method strips away the surface-level form of natural language, highlighting the logical relationship between behavior and participating elements, enabling the same intent under different expressions to be uniformly identified. Furthermore, the triple structure has good scalability, facilitating comparison and integration with historical information.

[0038] Sub-step 1.4: Standardize and disambiguate the generated semantic triples: Although the triples have initially expressed semantic content, some expressions may still be ambiguous or subject to multiple interpretations. For example, "seven o'clock" could refer to 7 a.m. or 7 p.m., requiring judgment based on contextual common sense; "location" could refer to physical space, online ranking, or queue number, requiring determination of its specific meaning based on domain knowledge. Therefore, the system introduces external knowledge sources (such as a time common sense base and a domain ontology library) to... Each item in the system is calibrated. For the time expression "7 PM tonight", the system uses the current system time. Assuming it's January 1, 2025 at 14:13, we can deduce that "tonight" corresponds to January 1, 2025. Therefore, "7 o'clock" should be understood as... And standardize it into a unified time format. Regarding "location," considering the previously mentioned entity "restaurant," the system determines that "location" here should be "dining seat." After this step, the original triple is corrected into an unambiguous, standardized form: ; The standardized result serves as the final semantic representation of the input for this round and proceeds to the next stage of processing. This step ensures the clarity and consistency of semantic information, avoiding deviations in subsequent processing due to diversity or ambiguity in expression.

[0039] Step 2: Establish a mapping between the current semantics and the historical dialogue state. After semantically parsing the current input, the system needs to compare and correlate it with information accumulated from previous rounds of dialogue to determine whether there are referential, continuation, modification, or supplementary relationships. Ignoring historical context and processing only the current statement in isolation can easily lead to information gaps. For example, if a user says "How is it?", without tracing back to the previously mentioned "that restaurant," it's impossible to clearly identify what "it" refers to. Therefore, it's essential to establish a connection path between the current semantic unit and past dialogue content, identifying potential referents, unfulfilled intentions, or implicit premises, thereby providing a basis for contextual continuity.

[0040] Sub-step 2.1: Retrieve past semantic records stored in the history dialogue memory bank: Before entering this stage, the system has maintained a chronologically ordered sequence of dialogue memories. Each of them Indicates the first The semantic state saved after a round of dialogue ends contains the set of normalized triples parsed from that round. And its contextual tags. For example, when a user says "find a Sichuan restaurant" in the first round, the system records... In the second round, users added "I want a place close to my home," and the system recorded this. At this point, when processing the third round of input, the system reads from the memory bank. and , obtain and As a reference benchmark, this step is significant because it provides a searchable historical context framework, so that the current input is no longer isolated, but is examined within the entire dialogue context.

[0041] Sub-step 2.2: Locate the potential referent related to the current input in the historical semantic record: based on The system handles vague or pronoun expressions (such as "it," "that," "what I said last time") that appear in the text. Search for candidate entities that are semantically similar or structurally matched. For example, in The phrase "the restaurant we talked about last time" appears in the text. The system recognizes "that restaurant" as an indicator phrase, triggering the reference parsing mechanism. The system iterates through... Find entities that have been described as "restaurant" or "dining place" in all rounds of triples. The system identified "Sichuan cuisine restaurant" as a restaurant type, and since it wasn't subsequently negated or replaced, it was marked as a potential referential target. Further, the system checked whether the entity possessed persistent attributes (such as whether it had been named or was active), confirming that it remained a valid topic in the current conversation. This process relied on semantic similarity calculation, defining two entity descriptions... and The degree of matching is: ; The first item is the word overlap rate, and the second item is the type consistency score (e.g., "Sichuan cuisine restaurant" and "restaurant" belong to the same category). This is the weighting coefficient (taken as 0.6). If (If the threshold is set to 0.7), then a referential relationship is determined to exist. In this example, "Sichuan cuisine restaurant" and "restaurant" are of the same type, with some overlap in vocabulary, and the overall score is 0.8, which exceeds the threshold, so the referential relationship is confirmed.

[0042] Sub-step 2.3: Replace pronouns or vague expressions in the current semantics with specific historical entity references: Once the referential relationship is confirmed, the system performs a semantic substitution operation, replacing the ambiguous component in the current triple with the actual entity name or identifier it points to. For example, in In the text, "that restaurant" was replaced with "Sichuan restaurant" (from...). The core entity), thus obtaining the updated triplet: ; A new "Target Location" role has been added to clearly define the service recipient for booking actions. This replacement not only eliminates linguistic ambiguity but also enables the transmission of semantic information across rounds, seamlessly connecting the current intent with historical topics. More importantly, this operation integrates information that was originally scattered across different rounds into a complete request structure, providing clear preconditions for subsequent intent confirmation and execution.

[0043] Sub-step 2.4: Evaluate the consistency and conflicts between the current semantics and historical records: After completing the referential parsing and semantic completion, the system needs to perform further checks. and The system checks for contradictions or duplications in existing information. For example, if a user previously stated "no spicy food" but now requests "find a Sichuan restaurant," there is a preference conflict; if a user repeatedly requests to "make a reservation," it may indicate urgency or lack of response. The system makes this judgment by comparing the changing trends of actions, constraints, and target entities. A consistency metric is defined. for: ; in The decay weight is determined by the historical rounds (the more recent the round, the higher the weight, e.g., ...). This is a semantic consistency function, taking values ​​of 1 (consistent), 0 (irrelevant), or -1 (conflicting). If the sum... If a significant conflict is found, a clarification process must be initiated; if Furthermore, if the action is repeated, it is considered user emphasis. In this example, "Reserving a table" and The search for "Sichuan restaurants" aligns with this direction and contains no negative constraints, therefore... This indicates that the current request continues the existing intent without conflict. This assessment result serves as an important basis for the next stage of state updates, ensuring that the system behavior conforms to the evolution path expected by the user.

[0044] Step 3: Construct a dynamic semantic memory graph and update the dialogue state After completing the correlation analysis between the current input and historical information, the system needs to integrate the newly acquired semantic content into the overall dialogue state, forming a memory structure that evolves over time. Traditional flat recording methods are difficult to express complex relationships, while graph-based structures can naturally present entities, attributes, actions, and their interrelationships. By transforming the semantic triples of each round into graph nodes and edges, and maintaining their lifecycle and activity, the system can track long-term dependencies and retain key information, preventing the loss of important details in multiple rounds of interaction.

[0045] Sub-step 3.1: Initialize the structural framework of the dynamic semantic memory graph: The system maintains a global semantic graph. ,in It is a set of nodes, representing various entities that appear in the dialogue (such as people, places, events, attributes, etc.). This is the set of edges connecting nodes, representing semantic relationships between entities (such as "belongs to", "occurs at", "proposed by", etc.). The graph is initialized empty and gradually expands as the dialogue progresses. Each node... Carrying attribute set This is used to track its state. Indicates from node arrive Semantic connections, with accompanying relation tags This graph, serving as the core carrier of dialogue states, supports efficient query, update, and reasoning operations. In this example, as... Processed, the system creates a node. (Type: Restaurant, Value: Sichuan Cuisine Restaurant, Source: Round 1), and add it to the list. .

[0046] Sub-step 3.2: Transform the currently parsed semantic triples into new nodes and connections in the graph: exist Once confirmed, the system maps its content to graph operations. First, a new node is created for the "Reservation" action. (Type: Operation, Value: Seat Reservation), Create a time node for "2025-01-01 19:00" (Type: Time, Value: 19:00). Next, establish an edge connection: from arrive Add edge The relationship is "target location"; from arrive Add edge The relationship is "occurrence time"; from the user node (which exists by default) to... Add edge The relationship is "initiator". Simultaneously, update the node. The last active time of a term is the current moment, reflecting its re-mention in this round. This process transforms discrete triples into a structured knowledge network, making semantic relationships visible and easy to retrieve later. For example, by traversing... It can quickly identify "who initiated what action against which restaurant, when, and what action".

[0047] Sub-step 3.3: Adjust the activity and association strength of existing nodes in the graph: Not all historical information is equally important; the system needs to dynamically adjust the prominence of each node based on usage frequency and time elapsed. Therefore, an activity score is introduced. and correlation strength Two metrics. Activity level is updated as follows: ; in As the attenuation factor, This is an indicator function (1 if the node is referenced in this round, 0 otherwise). For (Sichuan restaurant), because in Nodes explicitly mentioned in the context of the topic saw increased activity; however, nodes mentioned earlier but not reappearing (such as "recommendation rating" mentioned in round 1) experienced decreasing activity with each subsequent round. (Association strength) Then, based on the increasing number of co-occurrences: ; Whenever reasoning or querying is successful through this edge, This mechanism makes frequently used paths more likely to be activated, forming "semantic inertia," which helps predict what the user might focus on next. For example, if the relationship "restaurant—service—cuisine" is used frequently, the system is more likely to prioritize this path when generating a response.

[0048] Sub-step 3.4: Remove inactive nodes to maintain graph simplicity and response efficiency: To prevent the graph from expanding indefinitely and causing processing delays, the system periodically performs lightweight cleanup operations. An activity threshold is set. If a node is not referenced for three consecutive rounds and If a node is not actively retrieved, it is marked as awaiting recycling. However, it is not immediately deleted but instead enters a "dormant zone," remaining for a certain period (e.g., 5 rounds) for possible backtracking. If it is not awakened during this period, it is completely removed. For example, if a user mentioned "the waiter was nice" but did not discuss it further, the node's activity gradually decreases to 0.15, meeting the removal criteria and being removed from the main graph. This step ensures that the graph focuses on the current main dialogue, reduces noise interference, and retains short-term backtracking capability, balancing memory capacity and processing efficiency.

[0049] Step 4: Generate context-aware response content and confirm the execution path. After updating the semantic graph, the system now possesses a complete understanding of the current intent and historical context. The next task is to generate natural, coherent responses that meet user expectations based on this state, and to determine whether further action is needed (such as calling external services or asking follow-up questions). Response generation must not only express semantic content but also demonstrate an understanding of the context, such as using appropriate pronouns, echoing previous details, and avoiding repetition of known information, thereby improving the naturalness of the dialogue and user satisfaction.

[0050] Sub-step 4.1: Determine the complete intent expression of the current user based on the updated semantic graph: System Scan The most recently added or updated nodes and edges are identified, and the most significant incomplete action is determined. In this example, the node... The "Reservation" action is the most recent and has not yet been marked as "Completed," therefore the system determines that the user's current intention is to complete a reservation for this Sichuan restaurant. Simultaneously, the system extracts relevant constraints: time (19:00), location (Sichuan restaurant), and user identity (current session holder). This information together constitutes a complete request packet. The key to this step lies in extracting executable instruction structures from the graph, rather than simply restating user statements. By relying on the structured representation of the graph, the system can accurately identify "booking" as the core action, rather than "inquiry" or "cancellation".

[0051] Sub-step 4.2: Construct a context-appropriate natural language response draft: based on The system begins generating a response text. First, the response type is determined: because the request is clear and the information is complete, no further questions are needed, and the system can proceed directly to the confirmation phase. The response structure is designed with three parts: confirmation action, restatement of key parameters, and prompts for the next step. The initial draft is: "Okay, reserving a table for you at a Sichuan restaurant for 7 PM tonight." Here, "Sichuan restaurant" comes from... The value, "7 PM tonight" is determined by The timeline was generated using reverse formatting. This draft ensures that all key information is included and the language is fluent. To enhance context awareness, the system avoids using phrases like "that restaurant" or "the one mentioned before" when referring to "Sichuan cuisine restaurant," as its uniqueness has been confirmed in the map, and direct naming is clearer.

[0052] Sub-step 4.3: Embed historical context information to enhance the coherence and personalization of responses: To demonstrate an understanding of the dialogue process, the system includes a supplementary explanation related to the history in its responses. (Review) The user emphasized "close to my home," indicating that location was a key consideration. Therefore, the draft was updated with: "You previously mentioned wanting stores close to home, so I've prioritized nearby locations." This sentence, while not strictly necessary, significantly improves the user experience, demonstrating that the system remembers the user's preferences. This information originates from the implicit path of "user—preference—location proximity" in the graph, assessed through the strength of the association. They are activated at a higher level. This approach ensures that the response not only completes the task communication but also demonstrates memory and a proactive service mindset.

[0053] Sub-step 4.4: Output the final response and prepare to execute the external call: The final response text, after grammatical checks and fluency optimization, reads: "Okay, we are reserving a table for you at a Sichuan restaurant tonight at 7 PM. Since you mentioned wanting to stay close to home, we have prioritized restaurants in the nearby area." This text is sent to the user interface via the output channel. Simultaneously, the system will... The request is packaged into a standard request format and submitted to the backend booking service interface to initiate the actual booking process. While waiting for the result, the dialog status is marked as "Waiting for external response," and relevant nodes (such as...) Set a temporary lock to prevent duplicate submissions. This step completes the loop from understanding to action, ensuring that user intent is not only heard but also implemented.

[0054] Step 5: Receive external service responses and synchronously update the dialogue semantic graph. After the system submits a request to an external service (such as a restaurant reservation platform or a schedule management interface), it needs to wait for the response. This response may be in various states, such as successful confirmation, insufficient resources, time conflict, or restricted permissions. If the external response is treated as an independent event without integrating its semantic content with the current dialogue state, information fragmentation may occur. For example, when a user asks "Have you made your reservation?", the system may not be able to accurately provide details. Therefore, the response from the external service must be considered as part of the dialogue semantic flow, its key information parsed, and mapped back into a dynamic semantic graph to achieve a unified view of human-computer interaction and system operation.

[0055] Sub-step 5.1: Listen for and retrieve the execution result message returned by the external service: After the system initiates a reservation request it starts an asynchronous listening mechanism to continuously detect the response channel from the external server. When a return data packet is received, its complete content is immediately captured. For example, a certain reservation platform returns: { "status":"confirmed", "reservation_id":"R20250101J900CN7", "venue":"Chuanxiangge", "time":"2025-01-01T19:00:00", "seats":2, "note":"Window seat has been reserved" } This message serves as a credential for the system's interaction with the outside world, containing the operation result, entity identifier, time and location, and additional instructions. The core function of this step is to ensure that the system can promptly perceive the feedback from the external world and avoid status stagnation caused by timeouts or packet losses. All returned fields are completely recorded, providing the original basis for subsequent semantic parsing to prevent the loss of potentially useful information due to truncation or omission of secondary fields.

[0056] Sub-step 5.2: Structured parsing and key element extraction of the service response message: After obtaining the system performs field traversal and semantic annotation on it. First, it identifies the operation status field (`status`), and maps its value "confirmed" to an internal status code ; then it extracts the unique identifier `reservation_id` assigned by the server as the credential number for this operation; then it identifies the value of `venue` "Chuanxiangge" as the actual execution location, noting that there is a name difference from the user's original mention of "Sichuan restaurant", but the type is the same; the `time` field matches the time in the request, indicating that the time has not been adjusted; `seats` represents the number of seats, defaulting to 2 people; the "window seat" in `note` is a newly added constraint satisfaction item. These information are organized into a set of server semantic triples: ; This representation converts machine-readable JSON data into an expression form isomorphic to the user's semantics, facilitating subsequent integration with the knowledge graph.

[0057] Sub-step 5.3: Map the server semantic information to the dynamic semantic graph and establish a two-way link: The system will Each item in the graph is transformed into a graph update operation. First, a new node is created. (Type: Voucher, Value: R202501011900CN7), and establish a slave operation node. (Reserving a seat) to edge The relationship is "generate voucher". Secondly, update the original "Sichuan Restaurant" node. Set its "Official Name" to "Chuanxiangge" to avoid future confusion. Then create a new node. (Type: Special Requirements, Value: Window Position), and connect to The relationship is "satisfying the condition". Furthermore, in The status attribute of the superscript is set to "Executed," replacing the original "Pending." More importantly, backlinks are established: from... Pointing to the user node indicates the "owner user," ensuring that the context can be queried through credentials in the future. This process not only records the result but also maintains the causal chain, making the path of "who—initiating—what operation—generating—what credentials—involving—which entities" clearly traceable.

[0058] Sub-step 5.4: Evaluate the consistency between the service response and the original request and mark the deviations: Although the service returned "confirmed," it is still necessary to check whether its content fully meets the user's initial expectations. System comparison. and Key fields: consistent time, default matching of number of people, and similar but geographically close locations despite different names (verified via external map API that "Chuanxiangge" is 1.2 kilometers from the user's address, meeting the "close" criterion). The only new item is "window seat," which is a positive addition and does not violate the user's wishes. Define consistency deviation. for: ; in For comparison of field sets (time, location, number of people, status). For the measure of difference (0 indicates consistency, 1 indicates conflict). This assigns a field weight (e.g., time has the highest weight of 1.0, and the number of people has a weight of 0.8). In this example, all... Therefore If the values ​​are completely consistent, then the time will be determined. If there is any inconsistency (e.g., the time was changed to 20:00), then... The system will initiate a clarification process. Since there are currently no deviations, the operation status has been officially updated to "Completed," and the system is preparing to notify the user of the result.

[0059] Step Six: Generate a results notification and response and maintain the continuity of the dialogue. After integrating external service responses, the system needs to provide feedback to the user in natural language. Simply repeating "booking successful" mechanically sounds impersonal and lacks sufficient information; proactively presenting details based on context, echoing prior preferences, and anticipating potential follow-up questions can significantly improve the quality of interaction. Therefore, response generation should be based on the updated semantic graph, selecting the most relevant nodes and relationships to construct information-rich, natural-sounding, and potentially continuation-oriented statements, ensuring the dialogue doesn't abruptly end upon task completion but smoothly transitions to the next stage.

[0060] Sub-step 6.1: Extract the set of key information to be communicated to users from the updated semantic graph: The system scans the graph for highly active nodes related to the current operation, filtering out content the user might be interested in. This includes: operation result (confirmed), voucher number (R202501011900CN7), actual venue name (Chuanxiangge), time (19:00), number of people (2), and special arrangements (window seat). Simultaneously, it reviews historical preference nodes, confirming that the "close to home" preference has been met (distance 1.2 km). This information constitutes the core set of the notification. This step avoids information overload by selecting only content that is directly relevant to the user and not obvious, such as avoiding repeating known premises like "you want to reserve a seat".

[0061] Sub-step 6.2: Construct a response statement framework containing a multi-layered information structure: based on The logical hierarchy of the system design response is as follows: the first sentence declares the result, the second sentence lists the key parameters, the third sentence emphasizes the personalized satisfaction items, and the last sentence opens up a continuation channel. The preliminary framework is as follows: 1. Reservation completed; 2. The venue is Chuanxiang Pavilion, the time is 7 PM tonight, and there are two people in total; 3. A window seat has been arranged according to your preference; 4. Do I need to send an electronic voucher to my email address? This structure follows a "general-specific-supplement-activation" pattern, clearly conveying the facts while demonstrating a proactive service mindset. In particular, the phrase "according to your preferences" in the third sentence directly links to the historical node "user-preference-location proximity," activating it through semantic paths and achieving cross-round responsiveness.

[0062] Sub-step 6.3: Refine and optimize the delivery of the response statements to make them more natural. To avoid a mechanical feel, the system adjusted the language style of the frame text. "The venue is Chuanxiang Pavilion" was changed to "At Chuanxiang Pavilion," and "Two people in total" was changed to "Two seats have been reserved for you," enhancing the conversational tone. Regarding the referent: "According to your preference" clearly refers to the user's previous comment about "being close to my home." Although "window seat" wasn't directly mentioned, the system inferred from external knowledge that "window seat" is often considered a premium location and falls under the same category of experience optimization as "close," thus reasonably classifying it as a preference response. The final polished sentence is: "Okay, your seat reservation has been successfully completed. Two seats have been reserved for you at Chuanxiang Pavilion tonight at 7 PM, and they are window seats. Would you like me to send you the e-voucher?" This version uses a friendly tone, provides complete information, and consistently employs the formal "you" to enhance the sense of individuality.

[0063] Sub-step 6.4: Output the results and set the dialog continuation guidance status: The response is delivered to the user via voice or text. Simultaneously, the system internally marks the dialogue status as "awaiting user feedback" and activates potential paths related to "credential delivery," "booking modification," and "cancellation." (This is related to the graph.) The "Sending Method" node linked to the (credential) is preloaded, supporting rapid responses to subsequent requests. This step not only completes the information transmission but also proactively paves the way for the next stage of the dialogue, allowing the interaction chain to extend naturally rather than abruptly end.

[0064] Step 7: Analyze user feedback on the results and determine the direction of the conversation. After receiving the result notification, users may indicate acceptance, request modifications, ask for cancellation, or inquire about details. If the system cannot accurately identify the intent behind the feedback, it may miss service opportunities or cause misunderstandings. Therefore, it is essential to analyze the user's response within its complete context, combining the current task status, historical preferences, and active paths in the semantic graph to determine their true needs, and then decide whether to proceed to a new task, adjust the existing arrangements, or end the conversation.

[0065] Sub-step 7.1: Receive subsequent input statements from the user regarding the result notification: Suppose a user replies, "Let's change it to three people, as an extra friend is joining us." The system captures this statement. The system identifies this as a change request for a confirmed booking. Unlike the initial input, this statement occurs after the task has been executed, falling under the "post-event adjustment" scenario. The system immediately suspends other pending processes to prioritize this feedback, as it involves modifying already generated vouchers and therefore has a higher priority. This statement becomes the core input for this round of processing, and its content will directly affect the graph status and subsequent actions.

[0066] Sub-step 7.2: Identify the change intent and new constraints in the statements: right Perform semantic parsing to extract the action "change to", the object "three people", and the reason "an extra friend came temporarily". Generate triples: ; The system recognizes "change to" as a modification command and triggers an operation on the existing code. The adjustment process involves three people. Simultaneously, the new headcount constraint for these three individuals must cover the original value of "2". The key to this step is distinguishing between "new requirements" and "revise existing arrangements." The former initiates a new task, while the latter enters the change process. Here, because the phrase "changed to" is mentioned, and the context clearly refers to already booked seats, it is determined to be a revision.

[0067] Sub-step 7.3: Locate the nodes to be modified in the semantic graph and assess the feasibility of the changes: The system locates and in the map (Reservation Operation) The associated "Number of People" node originally had a value of 2. Checking its status: Since the voucher has already been generated, direct modification might affect external systems. The system queries the API to see if the reservation supports changes in the number of people, returning "Adjustable to a maximum of 4 people." Therefore, the change is feasible. The "Number of People" node value in the graph is updated from 2 to 3 and marked as "Modified," recording the modification round as the 4th round. Simultaneously, a new node is created. (Type: Change Record, Value: +1 person, Reason: Increased Visitors), and connect to This creates an audit trail. This operation ensures that all changes are traceable and supports future retrospection.

[0068] Sub-step 7.4: Generate a change confirmation request and submit it to the external service: The system constructs a new service request. This includes the original voucher number, the number of new participants (3), and the unchanged time and location. The system submits this information to the booking platform, requesting an update. The external system returns updated confirmation information, including the same voucher number and the number of new participants. The system then updates the map again to synchronize the status. This step dynamically adjusts existing operations, demonstrating the system's adaptability to real-world changes, rather than rigidly executing initial instructions.

[0069] Step 8: End the dialogue or guide the player to a new task loop. After processing user feedback, the system needs to determine whether the current conversation should end or continue. If all intentions have been satisfied and there are no new requests, it should end naturally; if there are unresolved issues or the user shows a willingness to continue, it should remain open. By analyzing the convergence trend of active nodes in the semantic graph and the ending characteristics of the user's latest statement, the system can intelligently decide on the next stage of the conversation lifecycle.

[0070] Sub-step 8.1: Evaluate the completion status of all currently active intents: The system traverses all nodes in the graph marked as "operation" and checks their state attributes. It discovers... The (seat reservation) status is "Updated," and the external service has confirmed its success with no pending exceptions. Other nodes, such as "User Preferences" and "Credential Information," have also been responded to or stored. (Definition completion rate) ;when If no new input is triggered, the main dialogue is considered complete. In this example... The convergence condition is met.

[0071] Sub-step 8.2: Detect whether the user's latest statement implicitly contains the starting point for a new task. The user's last input, "Let's change it to three people, one more friend is joining us temporarily," ends with a period, contains no interrogative words or imperative structures, and only states a change request without introducing any new demands (such as "order a car while we're at it"). There is no new action initiator among the semantic roles. Therefore, it is determined that there is no new task initiation signal. This step prevents the closing statement from being mistaken for the start of a new instruction.

[0072] Sub-step 8.3: Generate a final response to confirm the closure of the dialogue: The system constructs a closing statement: "I have updated the number of diners to three for you; all other arrangements remain unchanged. Please let me know if you require anything else." This response is characterized by: confirming the change, demonstrating stability, and allowing for future interaction. It avoids using absolute terms like "goodbye" or "end," maintaining a lightweight standby state. After this response is sent, the system reduces the overall activity decay rate of the graph and enters a low-power listening mode.

[0073] Sub-step 8.4: Freeze the current dialogue semantic graph and set backtracking anchors: Current map Serialization storage, generating conversation snapshots The system establishes index anchors containing keywords such as "Chuanxiangge," "2025-01-01," and "Reservation Voucher R202501011900CN7." If a user mentions "the meal they ordered last time" in the future, the system can quickly recover this memory segment through keyword matching without having to parse it from scratch. This mechanism enables efficient retrieval of long-term memory, bridging the gap between single conversations and ongoing relationships.

[0074] In summary, by introducing a referential relationship confirmation mechanism after semantic parsing, and using a comprehensive matching based on the proportion of word overlap and the consistency of entity category, the historical entity pointed to by the ambiguous expression can be accurately identified. Furthermore, by comparing the consistency between the operation action and the historical target requirement, and combining the weighted cumulative semantic matching degree, it can be determined whether the current request continues the existing intent.

[0075] This method achieves accurate semantic association and intent continuity judgment across multiple rounds, effectively solving the problems of ambiguous referencing and misjudgment of intent caused by information fragmentation in multi-turn dialogues, and significantly improving the system's contextual understanding ability and interaction naturalness. Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method of intelligent dialogue management based on contextual understanding, characterized in that, The method comprises the following steps: Obtaining the complete sentence text input by the user at present, and performing word segmentation and dependency syntax analysis to generate a syntax dependency structure; Recognizing semantic roles based on the syntax dependency structure and generating semantic triples, and standardizing and disambiguating the time expressions and entity references in the triples; Accessing the past semantic records in the historical dialogue memory library, locating the reference object in the historical records according to the disambiguated semantic triples, calculating the overlap ratio of the indicative phrases and the entity descriptions in the historical records and the consistency score of the entity categories, weighting and summing to obtain a comprehensive matching value, and confirming that the reference relationship is established when the comprehensive matching value exceeds a preset judgment threshold; Replacing the ambiguous expressions in the current semantics with specific historical entity references after confirming that the reference relationship is established; After the replacement is completed, comparing the operation actions in the updated semantic triples with the target requirements confirmed in the historical dialogue memory library to check whether there are conflicting constraint conditions, calculating a weighted cumulative value of semantic matching based on the decay weights of the semantic records in each round, and determining that the current request continues the existing direction when the weighted cumulative value exceeds a set level and there is no conflicting constraint; Converting the processed semantic triples into new nodes and connections in the dynamic semantic memory graph, and adjusting the activity and correlation strength of the related nodes in the graph; Determining the complete intention of the user according to the updated dynamic semantic memory graph, constructing a natural language response containing historical associated information and outputting the same, and starting an external service call. 2.The context understanding based intelligent dialog management method according to claim 1, characterized in that: The standardization and disambiguation of the time expressions and entity references in the triples comprises: Inferencing the specific date and time point corresponding to the relative time expression according to the current system time; Converting the inferred time into a unified format time mark; Combining the entity type information appearing in the previous dialogue to determine the specific direction of the action object in the current semantics, and normalizing the action verbs into standard operation names. 3.The context understanding based intelligent dialog management method according to claim 1, characterized in that: The locating of the reference object comprises: Recognizing the indicative phrases in the current semantics; Traversing the semantic records saved in each round of the historical dialogue memory library; Finding the items that have been described as specific type entities, and confirming that the entities have not been denied or replaced and are still in a extendable state in the dialogue process. 4.The context understanding based intelligent dialog management method according to claim 1, characterized in that: The confirmation of the reference relationship comprises: Replacing the indicative phrases with the matching entity names in the historical records, adding a new role representing the target place in the semantic triples, taking the matching entity name as the value of the role, and generating an updated semantic triple containing the new role. 5.The context understanding based intelligent dialog management method according to claim 1, characterized in that: The determination of the current request to continue the existing direction comprises: Creating the operation action in the updated semantic triple as an action node, creating the unified format time mark as a time node, establishing a target association connection between the action node and the entity node, and establishing a time association connection between the action node and the time node.

6. The context understanding based intelligent dialog management method according to claim 5, characterized in that: The adjustment of the activity of the related nodes in the graph comprises: Updating the activity score of the action node and the entity node according to whether they are referenced in the current round; Reducing the activity of the early nodes that are not referenced in the current round by a fixed decay rate; When the node activity is lower than a set lower limit and has not been activated for a plurality of consecutive rounds, the node is put into a state to be recycled.

7. The context understanding based intelligent dialog management method according to claim 5, characterized in that: The determining the complete intention of the user comprises: Scanning the action node with a state of being processed in the map, extracting the entity node, time node and initiating principal associated with the action node; Combining the extracted information into a request structure containing operation, target, time and principal. 8.The context understanding based intelligent dialog management method according to claim 7, characterized in that: The constructing the natural language response containing historical association information comprises: Repeating the operation, target and time information in the request structure; Extracting the previous expression content associated with the initiating principal in the map; Adding the previous expression content as supplementary explanation into the response text. 9.The context understanding based intelligent dialog management method according to claim 8, characterized in that: The starting the external service call comprises: Encapsulating the request structure into a standard format and sending it to the external service interface; Listening and receiving the execution result message returned by the external service, parsing the key elements in the message, creating a credential node and establishing a connection with the action node, and synchronously updating the state of the action node.

10. A context understanding based intelligent dialog management system for implementing the method according to any one of claims 1 to 9, characterized by: Comprise: An input analysis unit for obtaining the complete sentence text of the current input of the user, performing word segmentation and dependency syntax analysis, generating syntax dependency structure, identifying semantic roles and generating semantic triples, and standardizing and disambiguating the time expression and entity reference in the triples; A memory retrieval unit for retrieving past semantic records in the historical dialogue memory library, locating the reference object in the historical records according to the disambiguated semantic triples, calculating the overlap ratio of indicative phrases and entity descriptions in the historical records and the consistency score of entity categories, weighting and summing them to obtain a comprehensive matching value, and confirming the reference relationship when the comprehensive matching value exceeds the preset judgment threshold; A map construction unit for replacing the ambiguous expressions in the current semantics after confirming the reference relationship to be established with specific historical entity references, converting the replaced semantic triples into new nodes and connections in the dynamic semantic memory map, and adjusting the activity and association strength of related nodes in the map; A response generation unit for determining the complete intention of the user according to the updated dynamic semantic memory map, constructing a natural language response containing historical association information and outputting, and starting an external service call.

Images