A guide content generation method based on intention shunting and dual constraint enhancement

Abstract

The application discloses a kind of based on intention shunt and dual constraint enhancement's guide content generation method, comprising: constructing guide field knowledge graph and guide context information;Guide request voice signal is converted into text request;According to the control procedure of corresponding intention type, when it is judged that the text request of preset intention type is executed dual constraint type knowledge enhancement generation procedure: each component of guide context information is mapped as corresponding knowledge boundary constraint, selects the knowledge entity and relationship that satisfy knowledge boundary constraint from guide field knowledge graph, obtains knowledge subset;Guide sentence is generated based on the structure constraint constructed, parses the entity corresponding to guide sentence;The semantic similarity of knowledge entity of knowledge subset and the entity corresponding to guide sentence is calculated;Final entity is selected based on semantic similarity, and the guide sentence corresponding to final entity is output.The application realizes that guide content generation is significantly improved in accuracy, stability and scene adaptability.

Description

Technical Field

[0001] This invention relates to the field of navigation content generation technology, specifically to a navigation content generation method based on intent triage and dual constraint enhancement. Background Technology

[0002] With the development of intelligent tour guide technology, generative model-based tour guide content generation methods can dynamically generate personalized tour guide information according to user needs, and are widely used in intelligent tour guide equipment and platforms in scenic spots, museums, exhibition halls and other scenarios. However, existing generative model-based tour guide content generation methods have the following defects and shortcomings, making it difficult to meet the precise service needs of tour guide scenarios:

[0003] (1) There are problems with intent differentiation and computing power consumption. Intelligent guide devices need to respond to operation commands and guide questions and answers at the same time. There is no refined diversion mechanism for operation commands and question and answer commands. This lack of diversion control may lead to delayed execution of operation commands or misunderstanding, resulting in delayed triggering of mechanical actions. In extreme cases, emergency avoidance commands cannot be executed in time, which poses a safety hazard. At the same time, redundant reasoning of operation commands increases cloud computing power consumption and may also lead to equipment operation errors due to reasoning errors.

[0004] (2) Unclear boundaries of knowledge use: The guide generation model is prone to calling irrelevant or non-authoritative knowledge, resulting in problems such as factual errors, deviations in exhibit information, and outdated information in the generated content. It lacks an effective mechanism to limit the scope of knowledge use.

[0005] (3) Uncontrollable generated content: The structure and logic of the generated results of the tour guide generation model are loose, and it does not follow the narrative rules of the tour guide scene from shallow to deep and highlighting the key points. At the same time, there are inconsistencies in the factual statements, and the controllability is poor.

[0006] (4) Insufficient scene adaptability: The guide generation model does not fully integrate scene information such as the user's current location, tour progress, interests and preferences, and real-time environment. Redundancy (such as introducing attractions not yet visited) or missing information (such as omitting the core highlights of the current attraction) may occur in the generated guide model, resulting in a poor user experience.

[0007] (5) Low interactive response efficiency: Some technologies are deployed on the entire terminal or in the cloud, which does not achieve reasonable resource allocation, resulting in problems such as excessive energy consumption or response delay. Furthermore, they lack a flexible demand diversion and processing mechanism, and cannot balance the efficient response of dialogue and question answering with device control. Summary of the Invention

[0008] To overcome the shortcomings and deficiencies of existing technologies, this invention provides a guide content generation method based on intent triage and dual constraint enhancement. Before guide question-and-answer processing, this invention introduces intent type recognition and request triage mechanisms to determine the intent type of the text request and execute the corresponding control flow based on the intent type. This avoids the resource waste caused by all requests entering the generative language model for processing, improving overall processing efficiency and operational stability. Simultaneously, based on a dual constraint-based knowledge enhancement generation process, knowledge boundary constraints and generation behavior constraints are constructed to form a collaborative dual constraint control mechanism. This limits the scope of knowledge that can be invoked in the generated guide content and provides real-time control over the structural form and factual consistency of the generated guide content. This achieves full control over the generated content from knowledge input to output expression, effectively solving the problems of uncontrollable knowledge sources, susceptibility to factual errors, and difficulty in adapting to dynamic guide scenarios in existing guide content generation technologies.

[0009] To achieve the above objectives, the present invention adopts the following technical solution:

[0010] This invention provides a method for generating navigation content based on intent triage and dual constraint enhancement, comprising the following steps:

[0011] Construct a knowledge graph for the navigation domain and build navigation context information;

[0012] Acquire the audio signal of the tour guide request and convert it into a text request;

[0013] Perform semantic analysis on text requests to determine the intent type of the text requests, and execute the corresponding control flow based on the intent type;

[0014] When a text request is determined to be of a preset intent type, a dual-constraint knowledge enhancement generation process is executed, specifically including:

[0015] Each component of the navigation context information is mapped to a corresponding knowledge boundary constraint. Knowledge entities and relationships that satisfy the knowledge boundary constraints are selected from the navigation domain knowledge graph to obtain a knowledge subset.

[0016] Construct structural constraints for the navigation content, generate navigation statements based on the structural constraints, and parse the entities corresponding to the navigation statements;

[0017] Knowledge entities are obtained based on the knowledge subset, and the semantic similarity between the knowledge entities and the entities corresponding to the navigation statements is calculated.

[0018] The final entity is selected based on the semantic similarity, and the corresponding navigation statement is output.

[0019] As a preferred technical solution, a knowledge graph for the tour guide domain is constructed, specifically including:

[0020] The knowledge graph of the tour guide domain contains multiple entities, including scenic spot entities, historical figure entities, and exhibit entities;

[0021] Establish entity relationships between various entities and configure attribute information mapping tables for each entity.

[0022] As a preferred technical solution, the navigation context information is constructed, specifically including:

[0023] Collect user location coordinates, determine the user's tour stage based on historical dialogues, obtain the tour object identifier, identify the user's identity, calculate the interaction time based on historical interaction information, and construct tour context information based on user location coordinates, tour object identifier, tour stage, user identity, and interaction time.

[0024] As a preferred technical solution, the corresponding control flow is executed according to the intent type, specifically including:

[0025] When a request is determined to be a navigation question-and-answer type request, a dual-constraint knowledge enhancement generation process is executed; when a request is determined to be an operation instruction type request, the corresponding operation process is executed.

[0026] As a preferred technical solution, each component of the navigation context information is mapped to a corresponding knowledge boundary constraint, specifically including:

[0027] Calculate the distance between each entity in the navigation domain knowledge graph and the user's current location, select entities that meet the preset spatial constraints from the navigation domain knowledge graph, and obtain a set of spatially constrained entities;

[0028] Thematic constraints are set based on the identification of the guided objects and the guided stages, and the thematic constraint entity set is selected from the spatial constraint entity set based on the thematic constraints.

[0029] Set knowledge depth constraints based on user identity and interaction time, and select a subset of knowledge from the set of subject-constrained entities based on the knowledge depth constraints.

[0030] As a preferred technical solution, the structural constraints for constructing the tour guide content specifically include:

[0031] The structural constraints of the guided tour content include the length of the generated content, language style, and narrative order.

[0032] Specifically, the length of the generated content is set according to the user's identity and the tour stage, the language style is set according to the user's identity, and the narrative order is set according to the tour object identifier and the tour stage.

[0033] As a preferred technical solution, selecting the final entity based on the semantic similarity specifically includes:

[0034] The cosine similarity is calculated by matching the embedding vectors of the knowledge entity with the entity corresponding to the navigation statement. The maximum cosine similarity is selected and summed. The average value is calculated based on the number of entities to obtain the knowledge alignment consistency score. The corresponding entities are then selected based on the consistency score threshold to obtain the final entity.

[0035] The present invention also provides a navigation content generation system based on intent triage and dual constraint enhancement, for implementing the above-mentioned navigation content generation method based on intent triage and dual constraint enhancement, including: a knowledge graph construction module, a navigation context information construction module, a text request acquisition module, an intent triage module, and a dual constraint generation module;

[0036] The knowledge graph construction module is used to construct a knowledge graph for the navigation domain;

[0037] The navigation context information construction module is used to construct navigation context information;

[0038] The text request acquisition module is used to convert the audio signal of the tour request into a text request.

[0039] The intent triage module is used to perform semantic analysis on text requests, determine the intent type of the text request, and execute the corresponding control flow according to the intent type.

[0040] The dual-constraint generation module is used to execute a dual-constraint knowledge enhancement generation process when a text request is determined to be of a preset intent type, specifically including:

[0041] Each component of the navigation context information is mapped to a corresponding knowledge boundary constraint. Knowledge entities and relationships that satisfy the knowledge boundary constraints are selected from the navigation domain knowledge graph to obtain a knowledge subset.

[0042] Construct structural constraints for the navigation content, generate navigation statements based on the structural constraints, and parse the entities corresponding to the navigation statements;

[0043] Knowledge entities are obtained based on the knowledge subset, and the semantic similarity between the knowledge entities and the entities corresponding to the navigation statements is calculated.

[0044] The final entity is selected based on the semantic similarity, and the corresponding navigation statement is output.

[0045] The present invention also provides a computer-readable storage medium storing a program that, when executed by a processor, implements the above-described method for generating navigation content based on intent triage and dual constraint enhancement.

[0046] The present invention also provides a computer device, including a processor and a memory for storing a processor-executable program, wherein when the processor executes the program stored in the memory, it implements the above-described method for generating navigation content based on intent triage and dual constraint enhancement.

[0047] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0048] (1) The present invention introduces an intent type recognition and request diversion mechanism before the navigation question and answer processing, determines the intent type of the text request, and executes the corresponding control flow according to the intent type. Specifically, navigation question and answer requests and operation instruction requests enter the appropriate processing path respectively, avoiding the resource waste caused by all requests entering the generative language model processing, improving the overall processing efficiency and operation stability, thereby solving the problem of response delay and insufficient security caused by mixed instruction processing in the prior art.

[0049] (2) The present invention constructs knowledge boundary constraints based on the navigation context information. Specifically, each component of the navigation context information is mapped to a corresponding executable knowledge boundary constraint, thereby limiting the scope of knowledge that can be called when generating navigation content. This avoids irrelevant knowledge and out-of-boundary knowledge from entering the generation process from the source, and significantly improves the authority and accuracy of navigation content.

[0050] Furthermore, by constructing behavioral constraints to control the structural form and factual consistency of the generated content in real time, a dual constraint control mechanism is formed in synergy with knowledge boundary constraints. This ensures that the generation of tour guide content is under control at both the knowledge input and content output levels, effectively avoiding the generalization bias of generative language models. This results in a significant improvement in the accuracy, stability, and scene adaptability of tour guide content generation, thereby effectively solving the problems of uncontrollable generated content, factual errors, and difficulty in timely correction in existing tour guide content generation technologies. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating the navigation content generation method based on intent triage and dual constraint enhancement of the present invention. Detailed Implementation

[0052] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0053] Example 1

[0054] like Figure 1 As shown, this invention provides a method for generating navigation content based on intent triage and dual constraint enhancement, comprising the following steps:

[0055] S1: Construct a knowledge graph for the navigation domain and build navigation context information;

[0056] In this embodiment, the knowledge graph of the tour guide domain has multiple entities, including scenic spot entities, historical figure entities, and exhibit entities. Entity relationships are constructed between each entity, including adjacent relationships, construction associations, etc., and attribute information mapping tables are configured for each entity, including attribute information mapping tables such as construction year, building size, and functional use.

[0057] In addition, the knowledge graph in the field of tour guides can integrate data released by the museum's official website, historical documents, scenic area spatial data provided by maps, and tourists' historical tour preferences data. It is dynamically updated by monitoring official data sources and conducting regular reviews to ensure the authority and timeliness of the knowledge.

[0058] In this embodiment, the current tour context information can be constructed using data collected by multimodal sensors, including devices such as GPS modules, UWB positioning modules, and cameras. These sensors collect user location coordinates, determine the user's tour stage based on historical dialogues, obtain the tour object identifier, identify the user's identity through user login information or historical behavior, calculate the interaction time based on historical interaction information, and construct the tour context information based on the user's location coordinates, tour object identifier, tour stage, user identity, and interaction time. Specifically, this is represented as follows:

[0059]

[0060] in, Indicates the user's location coordinates. Indicates the identifier of the guided object, such as the attraction ID. This indicates the guided tour phase, such as initial arrival, quick tour, in-depth explanation, Q&A, etc. Indicate user identity, such as ordinary tourist, expert, child, etc. Indicates the interaction time;

[0061] In this embodiment, a multi-source data preprocessing step is also included, in which feature extraction and normalization are performed on the multi-source data collected by the multi-modal sensors, and historical interaction information is fused using a time encoder (such as LSTM) to further construct the navigation context information.

[0062] S2: Obtain the audio signal of the tour guide request and convert it into a text request;

[0063] The guide robot continuously runs keyword wake-up detection to listen to the user's voice signal in real time. When the user speaks a preset keyword, it is triggered to wake up, obtain the user's guide request voice signal, and perform voice signal to text processing. Specifically, the collected voice signal is filtered to remove environmental noise and retain only valid voice segments. The filtered voice data is transmitted to the cloud via wireless network for speech recognition, completes speech to text processing, and generates the corresponding text request.

[0064] In this embodiment, the guide robot device continuously runs keyword wake-up detection. When a new wake-up word is detected, the current output is immediately interrupted and the state is reset in order to prioritize responding to new guide requests.

[0065] S3: After obtaining the text request, the cloud performs semantic analysis on the text request to determine the intent type of the text request, and executes the corresponding control flow according to the intent type;

[0066] Specifically, it determines whether the text request is a navigation question and answer type request or an operation instruction type request. If it is determined to be a navigation question and answer type request, it enters the dual-constraint knowledge enhancement generation process; if it is determined to be an operation instruction type request, it directly triggers the corresponding control module and executes the corresponding operation process.

[0067] In this embodiment, a text request is made. As input, it is processed through a pre-defined intent discrimination function. Output Intent Category The discriminant relation is expressed as:

[0068] ;

[0069] in, This indicates a request for an operation instruction. This indicates a navigation question-and-answer type request. The intent discrimination function is implemented based on semantic analysis and introduces model reasoning to ensure high determinism in the intent discrimination process. Through this pre-discrimination mechanism, the processing path selection is completed before entering the generative language model, structurally avoiding competition for resources between operation instructions and question-and-answer instructions in the same processing channel.

[0070] Specifically, when the intent determination result is an operation instruction type request. At this time, a fast instruction processing strategy is adopted, which directly sends control signals according to the preset instruction-execution mapping relationship, bypassing the generative language model and semantic generation process. Before executing the instruction, the legality and security of the control signal are verified, including the constraint judgment on motion parameters, motor load and execution timing. The instruction is only allowed to be executed when the verification passes. Through this processing method, the response path of operation control instructions is minimized, and the potential impact of generative language models on equipment safety is reduced through structural isolation.

[0071] When the intent determination result is a navigation question and answer type request When entering the guided tour scenario, a dual-constraint knowledge enhancement generation process is adopted. Based on the knowledge graph of the guided tour domain and the guided tour context information, the guided tour content generation process is constrained to avoid cross-regional knowledge interference. This can effectively avoid the generalization bias of the generative language model and improve the professionalism and accuracy of the guided tour content.

[0072] This embodiment introduces an intent type recognition and request diversion mechanism before the navigation question and answer processing. Different processing paths are selected according to the user request type, which avoids the waste of resources caused by all requests entering the generative language model for processing, and improves the overall processing efficiency and running stability.

[0073] S4: Construct knowledge boundary constraints as the first level of constraints, map each component of the navigation context information to a specific knowledge constraint, select knowledge entities and relationships that satisfy the knowledge boundary constraints from the navigation domain knowledge graph, and obtain a knowledge subset;

[0074] In this embodiment, knowledge boundary constraints are used to limit the spatial scope, topic scope, and knowledge depth of knowledge. Constraint matching is performed in the knowledge graph of the tour guide domain, retaining only knowledge entities and relationships that simultaneously satisfy all constraint rules, thereby forming a unique subset of knowledge corresponding to the current tour guide scenario, represented as:

[0075] ;

[0076] in, This represents a knowledge graph representing the navigation domain. The set representing knowledge boundary constraints. This indicates the number of knowledge entities selected. This formula represents the selection of all knowledge entities in the navigation domain knowledge graph that satisfy the knowledge boundary constraints. To form a subset of knowledge This subset of knowledge serves as the only knowledge input space that the generative language model is allowed to access in this interaction, thus pre-limiting the knowledge sources of the generative language model.

[0077] In this embodiment, the set of knowledge boundary constraints is: ,

[0078] The specific meanings of each rule are as follows:

[0079] Calculate the distance between each entity in the navigation domain knowledge graph and the user's current location, select entities that meet the preset spatial constraints from the navigation domain knowledge graph, and obtain a set of spatially constrained entities;

[0080] Specifically, the spatial constraints are expressed as follows:

[0081] ;

[0082] This spatial constraint is derived from the knowledge graph of the tour guide domain. Select all entities that meet the spatial conditions, as long as the entity's location is within the specified range. With the user's current location distance Not exceeding the maximum allowable distance The entity is then stored in the spatial constraint entity set to limit the spatial scope of knowledge, ultimately resulting in a spatial constraint entity set that conforms to the spatial constraint rules. ;

[0083] Thematic constraints are set based on the identification of the guided objects and the guided stages, and the thematic constraint entity set is selected from the spatial constraint entity set based on the thematic constraints.

[0084] Specifically, the subject constraints are:

[0085] ;

[0086] Topic constraints are used to limit the topic categories of knowledge, and the navigation object identifiers of entities. and guided tour stage Guide object identifier based on guide context information and guided tour stage Determine, from the set of spatially constrained entities Select the set of subject constraint entities that meet the subject constraint conditions. ;

[0087] Set knowledge depth constraints based on user identity and interaction time, and select a subset of knowledge from the set of topic-constrained entities based on the knowledge depth constraints;

[0088] Specifically, the knowledge depth constraint is as follows:

[0089] ;

[0090] In the knowledge depth constraint, the user identity of the entity and interaction time Based on the user's identity in the navigation context information and interaction time Determine, from the subject-constrained entity set Selecting a subset of knowledge Preferably, the different levels in the knowledge depth constraint conditions correspond to different granularities of knowledge attributes, specifically used to limit the level of detail of knowledge. For example, values ​​of 1, 2, and 3 correspond to basic, standard, and detailed, respectively.

[0091] This embodiment constructs knowledge boundary constraints based on the tour context information, transforming the tour context into executable knowledge constraint rules to limit the scope of knowledge that the generative language model can use. For example, it limits the generation process to only allowing the use of authoritative knowledge directly related to the attraction entity, including its construction date, building structure, and functional use. At the same time, it explicitly prohibits the use of other scenic area knowledge related to attractions that have not yet been reached. This avoids irrelevant and out-of-boundary knowledge from entering the generation process from the source, significantly improving the authority and accuracy of the tour content.

[0092] S5: After completing the knowledge boundary constraints, construct the generative behavior constraints as the second layer of constraints to regulate the consistency between the expression of the generated content and the facts, represented as:

[0093] ;

[0094] in, This indicates the generation of behavioral constraints. This represents structural constraints used to limit the length, language style, and narrative order of the generated content, in order to adapt to different guided tour scenarios and the understanding habits of different types of users. This represents a knowledge alignment constraint, used to ensure that the factual representations of the generated content are consistent with the authoritative knowledge in the knowledge subset;

[0095] Structural constraints Specifically, it is expressed as follows:

[0096] ;

[0097] in, Indicates the length of the generated content. Indicates language style, Indicates the narrative order;

[0098] Preferably, the length of the generated content can be set according to the user's identity and the stage of the tour. For example, if the user is a child, the length of the generated content can be set accordingly. The length of the generated content will be smaller if the user's identity is an expert. The size will be larger, but if the guided tour is a fast-track guided tour, it will be smaller.

[0099] Preferably, the generative language model can adjust the length of the generated content. By changing the length of the navigation statement, the generated words in the navigation statement can be calculated. The final probability:

[0100] ;

[0101] in, This indicates generated words in the navigation statements. The final probability, This represents the generation probability of the generative language model. Indicates the bias adjustment term. For hyperparameters, Indicates the length of the output statement of the generative language model;

[0102] Preferably, the language style can be set according to the user's identity. The generative language model can change the word style of the navigation statements according to the language style and calculate the style matching score, which is expressed as:

[0103] ;

[0104] in, Indicator Style matching score corresponding to the language style Indicating in language style The probability of style matching for the given word;

[0105] Preferably, the narrative order can be set according to the guide object identifier and the guide stage. If the guide object identifier is an already explained attraction ID, or the guide stage is a quick guide stage, the narrative order will not cover all the knowledge that needs to be introduced, that is, it will filter out what has been explained and what has not been explained.

[0106] In this embodiment, navigation statements are generated based on the above structural constraints, and the entities corresponding to the navigation statements are parsed through named entity recognition and relation extraction. , to obtain the entity set ;

[0107] Knowledge Alignment Constraints The process involves identifying each generated entity within the generated navigation statement, finding the closest entity in the knowledge subset, and then comparing the average similarity of all entities within the entire statement.

[0108] In this embodiment, a knowledge subset is defined. The set of fact triples contained therein is Each triple represents the head entity. With tail entity There is a relationship Obtaining knowledge entities based on knowledge subsets Computational knowledge entities Entities corresponding to navigation statements Based on semantic similarity, the final entity is selected and the corresponding navigation statement is output.

[0109] Specifically, the cosine similarity is calculated using the embedding vectors of the knowledge entities and the entities corresponding to the navigation statements. The maximum cosine similarity values ​​are then summed, and the average value is calculated based on the number of entities to obtain the knowledge alignment consistency score, which is expressed as:

[0110] ;

[0111] in, The semantic similarity between two fact triples can be obtained by calculating cosine similarity. This represents the knowledge alignment consistency score. This represents the set of entities corresponding to the navigation statements. This represents the entity corresponding to the navigation statement. Knowledge entities that represent subsets of knowledge;

[0112] If the knowledge alignment consistency score is higher than the consistency score threshold, the corresponding entity is retained.

[0113] In this embodiment, the text request, the knowledge subset obtained by the constrained retrieval, and the generation behavior constraints are input into the generative large language model. During the generation process, the generative large language model generates content based only on the knowledge subset under the constraint of knowledge boundary, and adjusts the generation path under the real-time control of the generation behavior constraints, thereby generating guide text content that meets the expression requirements of the guide scenario and is consistent with authoritative knowledge. In this way, the knowledge boundary constraints limit the knowledge input space of the generative large language model, and the generation behavior constraints control the output path of the generative large language model. The two form a dual constraint collaborative mechanism that is connected and mutually reinforcing.

[0114] In this embodiment, the context information of the tour is continuously updated during multiple rounds of guided tour interaction, and the knowledge boundary constraints and generation behavior constraints are dynamically adjusted according to the changes in the context. This ensures that the dual constraint mechanism remains effective in real time throughout the entire tour process. Through the implementation of the above complete technical process, the present invention can achieve full-process control over the generated content from knowledge input to output expression in the tour scenario, significantly improving the accuracy, authority and scene adaptability of the generated tour content.

[0115] To verify the application effect of the method of the present invention in actual tour guide scenarios, a test dataset oriented towards tour guide interaction was constructed. The test dataset includes two parts: operation instruction requests and tour guide question and answer requests. The operation instruction requests include common control commands such as stop navigation and stop dialogue, while the tour guide question and answer requests include attraction introductions, exhibit backgrounds, historical event descriptions, route inquiries, multi-round follow-up questions, and personalized questions for different user identities. At the same time, corresponding independent knowledge bases were established around different floors, different exhibition halls, and different areas, and a tour guide domain knowledge graph was constructed by combining attraction entities, exhibit entities, and their attribute mapping relationships.

[0116] During the experiment, by pre-setting contextual information such as the user's current location, tour stage, and user identity, the interactive request input in a real tour scenario was simulated. For each test request, both the processing scheme using only the large model and the scheme of this invention were used for processing. The intent recognition results, response time, large model call status, and final generated content were recorded. Then, combined with standard factual information in the knowledge base, the facts accuracy, out-of-bounds knowledge citation rate, length compliance rate, and current attraction matching rate were statistically analyzed, as shown in Table 1 below. The ablation experiment results of this invention are as follows:

[0117] Table 1 shows the ablation experiment results of this invention.

[0118]

[0119] Among them, the intent recognition accuracy is obtained by statistically analyzing the proportion of requests correctly identified in the test set out of the total number of requests; the average response latency and high-quantile latency of operation instructions are calculated by recording the time from the generation of speech recognition text to the start of execution of control instructions on the device; the P95 latency of operation instructions represents the 95th percentile response latency, that is, the response latency of 95% of requests is lower than this value, used to characterize the worst response performance of the system; the question-and-answer first response latency is obtained by recording the time from the input of the question to the output of the first segment of guide text or the first sentence of the broadcast by the system; the large model call ratio is obtained by statistically analyzing the proportion of requests entering the generative model processing link out of the total number of requests. For the generation quality-related indicators, the factual accuracy is obtained by comparing the entities, attributes, and factual relationships in the generated content with the standard information in the knowledge base; the out-of-bounds knowledge citation rate is obtained by statistically analyzing the proportion of information in the generated content that exceeds the knowledge range allowed in the current scenario; the length compliance rate is obtained by comparing whether the length of the generated text falls within the target range set by the corresponding guide stage and user identity; and the current attraction matching rate is statistically analyzed by judging whether the generated content revolves around the current attraction or exhibition area.

[0120] As shown in Table 1 above, compared with processing schemes that only use large models, this invention demonstrates significant advantages in several key performance indicators by introducing an intent-based traffic splitting mechanism and a dual-constraint knowledge-enhanced generation process. Firstly, this invention can perform pre-processing intent recognition and traffic splitting for operation command requests, eliminating the need for local operation commands to enter the large model generation processing chain. Secondly, compared to the baseline scheme, the average response latency of operation commands is reduced by 75.0%, and P95 is lower. The high-resolution response latency was reduced by 46.9%, the initial response latency for question answering was reduced by 43.5%, and the overall proportion of large model calls decreased by 40.0%, significantly improving the overall real-time response and reducing the consumption of large model computing resources in the cloud. On the other hand, the invention constructs a dual constraint mechanism based on knowledge boundary constraints and generation behavior constraints in the context of the tour guide, effectively limiting the scope of model knowledge calls and standardizing the generation output process. Compared with the baseline solution, the factual accuracy rate was improved by 15.5%, the out-of-bounds knowledge citation rate was reduced by 73.5%, the length compliance rate was improved by 36.6%, and the current attraction matching rate was improved by 13.1%. The factual accuracy, knowledge compliance, length controllability, and scene adaptability of the tour guide content were all comprehensively improved.

[0121] The experimental results above show that the solution of the present invention can significantly improve real-time response speed and reduce resource consumption, while simultaneously achieving comprehensive optimization of the accuracy, controllability, and scene adaptability of guide content generation.

[0122] Example 2

[0123] This embodiment provides a navigation content generation system based on intent triage and dual constraint enhancement, used to implement the navigation content generation method based on intent triage and dual constraint enhancement in Embodiment 1 above, including: a knowledge graph construction module, a navigation context information construction module, a text request acquisition module, an intent triage module, and a dual constraint generation module;

[0124] In this embodiment, the knowledge graph construction module is used to construct a knowledge graph for the navigation domain;

[0125] In this embodiment, the navigation context information construction module is used to construct navigation context information;

[0126] In this embodiment, the text request acquisition module is used to convert the audio signal of the tour guide request into a text request;

[0127] In this embodiment, the intent triage module is used to perform semantic analysis on text requests, determine the intent type of the text request, and execute the corresponding control flow according to the intent type.

[0128] In this embodiment, the dual-constraint generation module is used to execute a dual-constraint knowledge enhancement generation process when a text request is determined to be of a preset intent type, specifically including:

[0129] Each component of the navigation context information is mapped to a corresponding knowledge boundary constraint. Knowledge entities and relationships that satisfy the knowledge boundary constraints are selected from the navigation domain knowledge graph to obtain a knowledge subset.

[0130] Construct structural constraints for the navigation content, generate navigation statements based on the structural constraints, and parse the entities corresponding to the navigation statements;

[0131] Knowledge entities are obtained based on knowledge subsets, and the semantic similarity between the knowledge entities and the entities corresponding to the navigation statements is calculated.

[0132] The final entity is selected based on semantic similarity, and the corresponding navigation text is output.

[0133] Example 3

[0134] This embodiment provides a storage medium, which may be a ROM, RAM, disk, optical disk, or other storage medium. The storage medium stores one or more programs. When the programs are executed by the processor, they implement the navigation content generation method based on intent triage and dual constraint enhancement of Embodiment 1.

[0135] Example 4

[0136] This embodiment provides a computing device, which may be a desktop computer, laptop computer, smartphone, PDA handheld terminal, tablet computer or other terminal device with display function. The computing device includes a processor and a memory. The memory stores one or more programs. When the processor executes the program stored in the memory, it implements the navigation content generation method based on intent triage and dual constraint enhancement of embodiment 1.

[0137] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for generating navigation content based on intent triage and dual constraint enhancement, characterized in that, Includes the following steps: Construct a knowledge graph for the navigation domain and build navigation context information; Acquire the audio signal of the tour guide request and convert it into a text request; Perform semantic analysis on text requests to determine the intent type of the text requests, and execute the corresponding control flow based on the intent type; When a text request is determined to be of a preset intent type, a dual-constraint knowledge enhancement generation process is executed, specifically including: Each component of the navigation context information is mapped to a corresponding knowledge boundary constraint. Knowledge entities and relationships that satisfy the knowledge boundary constraints are selected from the navigation domain knowledge graph to obtain a knowledge subset. Construct structural constraints for the navigation content, generate navigation statements based on the structural constraints, and parse the entities corresponding to the navigation statements; Knowledge entities are obtained based on the knowledge subset, and the semantic similarity between the knowledge entities and the entities corresponding to the navigation statements is calculated. The final entity is selected based on the semantic similarity, and the corresponding navigation statement is output.

2. The navigation content generation method based on intent triage and dual constraint enhancement according to claim 1, characterized in that, Constructing a knowledge graph for the tour guide domain, specifically including: The knowledge graph of the tour guide domain contains multiple entities, including scenic spot entities, historical figure entities, and exhibit entities; Establish entity relationships between various entities and configure attribute information mapping tables for each entity.

3. The navigation content generation method based on intent triage and dual constraint enhancement according to claim 1, characterized in that, Constructing navigation context information specifically includes: Collect user location coordinates, determine the user's tour stage based on historical dialogues, obtain the tour object identifier, identify the user's identity, calculate the interaction time based on historical interaction information, and construct tour context information based on user location coordinates, tour object identifier, tour stage, user identity, and interaction time.

4. The guide content generation method based on intent triage and dual constraint enhancement according to claim 1, characterized in that, The corresponding control flow is executed based on the intent type, specifically including: When a request is determined to be a navigation question-and-answer type request, a dual-constraint knowledge enhancement generation process is executed; when a request is determined to be an operation instruction type request, the corresponding operation process is executed.

5. The navigation content generation method based on intent triage and dual constraint enhancement according to claim 3, characterized in that, Mapping each component of the navigation context information to a corresponding knowledge boundary constraint, specifically including: Calculate the distance between each entity in the navigation domain knowledge graph and the user's current location, select entities that meet the preset spatial constraints from the navigation domain knowledge graph, and obtain a set of spatially constrained entities; Thematic constraints are set based on the identification of the guided objects and the guided stages, and the thematic constraint entity set is selected from the spatial constraint entity set based on the thematic constraints. Set knowledge depth constraints based on user identity and interaction time, and select a subset of knowledge from the set of subject-constrained entities based on the knowledge depth constraints.

6. The navigation content generation method based on intent triage and dual constraint enhancement according to claim 3, characterized in that, The structural constraints for constructing the tour guide content include: The structural constraints of the guided tour content include the length of the generated content, language style, and narrative order. Specifically, the length of the generated content is set according to the user's identity and the tour stage, the language style is set according to the user's identity, and the narrative order is set according to the tour object identifier and the tour stage.

7. The navigation content generation method based on intent triage and dual constraint enhancement according to claim 1, characterized in that, The final entity is selected based on the semantic similarity, specifically including: The cosine similarity is calculated by matching the embedding vectors of the knowledge entity with the entity corresponding to the navigation statement. The maximum cosine similarity is selected and summed. The average value is calculated based on the number of entities to obtain the knowledge alignment consistency score. The corresponding entities are then selected based on the consistency score threshold to obtain the final entity.

8. A navigation content generation system based on intent triage and dual constraint enhancement, characterized in that, The method for generating navigation content based on intent triage and dual constraint enhancement as described in any one of claims 1-7 includes: a knowledge graph construction module, a navigation context information construction module, a text request acquisition module, an intent triage module, and a dual constraint generation module; The knowledge graph construction module is used to construct a knowledge graph for the navigation domain; The navigation context information construction module is used to construct navigation context information; The text request acquisition module is used to convert the audio signal of the tour request into a text request. The intent triage module is used to perform semantic analysis on text requests, determine the intent type of the text request, and execute the corresponding control flow according to the intent type. The dual-constraint generation module is used to execute a dual-constraint knowledge enhancement generation process when a text request is determined to be of a preset intent type, specifically including: Each component of the navigation context information is mapped to a corresponding knowledge boundary constraint. Knowledge entities and relationships that satisfy the knowledge boundary constraints are selected from the navigation domain knowledge graph to obtain a knowledge subset. Construct structural constraints for the navigation content, generate navigation statements based on the structural constraints, and parse the entities corresponding to the navigation statements; Knowledge entities are obtained based on the knowledge subset, and the semantic similarity between the knowledge entities and the entities corresponding to the navigation statements is calculated. The final entity is selected based on the semantic similarity, and the corresponding navigation statement is output.

9. A computer-readable storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the navigation content generation method based on intent triage and dual constraint enhancement as described in any one of claims 1-7.

10. A computer device comprising a processor and a memory for storing a processor-executable program, characterized in that, When the processor executes the program stored in the memory, it implements the navigation content generation method based on intent triage and dual constraint enhancement as described in any one of claims 1-7.

Images