An AI artificial intelligence-based content generation and interactive dialogue method and system

Abstract

The application discloses a content generation and interactive dialogue method and system based on AI artificial intelligence, which models semantic label sequences in multiple rounds of conversation, identifies semantic derailment points in behavior evolution trajectory templates and triggers an intent clarification mechanism; constructs a semantic tension map based on user feedback, forms a tension closed cluster and generates multi-path response candidate segments; uses a misleading chain cross-analysis unit to identify high-risk response chains associated with historical negative feedback, and eliminates and shields content; and selects the optimal response segment through a context break span model and dynamically updates the behavior evolution model. The application can significantly improve the semantic stability, consistency and safety of the dialogue system, reduce semantic drift, misleading answers and illusion risks, and has high reliability and technical advancement. The system scheme can realize the modularization of the method functions and is suitable for application in intelligent customer service, general large models, decision support systems and other scenes.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information and communication interaction technology, and in particular to a content generation and interactive dialogue method and system based on AI artificial intelligence. BACKGROUND

[0002] With the rapid development of large language models (LLM) and natural language processing technology, intelligent dialogue systems based on deep learning have been widely applied in customer service systems, government consultation, online education, decision analysis, office assistants, etc. In such systems, user input often involves complex context, implicit intent, semantic jumps, multi-round dependencies, and context switching, so the system is required to accurately determine the user's intent and give a response content that is consistent with the context, has high continuity, and is controllable in risk.

[0003] However, the existing dialogue systems generally have the following shortcomings:

[0004] (1) Semantic drift is difficult to suppress. Traditional dialogue models mainly rely on context vectors or attention mechanisms to determine semantic associations, and do not have a clear semantic path structure. When the topic span increases or the context is inconsistent, the model is prone to off-topic or answering questions that are not asked.

[0005] (2) Lack of structured clarification mechanism when intent is uncertain. Existing systems can identify intent ambiguity through rules or classifiers, but usually only return template prompts, lack of structured analysis of semantic derailment reasons, and cannot provide user-friendly clarification options.

[0006] (3) Lack of structured filtering of historical negative feedback. Most systems only avoid dangerous answers based on blacklists, keyword matching, or RLHF training, but do not analyze the semantic chain of errors, the deviation of SVO structure, the risk level of semantic jump paths, etc. Therefore, it is easy to give wrong suggestions by changing the way of saying.

[0007] (4) Lack of semantic consistency structure measurement of response results. Existing systems usually let the model generate answers directly, without structured syntax dependency analysis of candidate fragments, and without measuring the degree of context break, resulting in incoherent logic before and after the answer.

[0008] (5) Insufficient dynamic learning ability. The dialogue tuning of existing LLMs relies on post-training or system-level memory mechanisms, and cannot update the dialogue trajectory model in real time according to user feedback.

[0009] In view of this, the present application provides a content generation and interactive dialogue method and system based on AI artificial intelligence to solve the above problems. SUMMARY

[0010] The application aims to provide an AI artificial intelligence-based content generation and interactive dialogue method and system, which solves the problems of the existing dialogue system, such as difficulty in recognizing the real intention of a user, easy to produce misleading replies, semantic jump unable to be monitored, and incoherent context response.

[0011] In order to achieve the above-mentioned purpose, the application provides the following technical scheme:

[0012] In a first aspect, the application provides an AI artificial intelligence-based content generation and interactive dialogue method, comprising the following steps:

[0013] S101: Collecting user input text, constructing a semantic label sequence according to at least three rounds of historical sentence pairs in a current session window; mapping the semantic label sequence to a behavior evolution trajectory template generated based on long-term interaction records of a user, and identifying a semantic derailment point and triggering an intention clarification request if a semantic node inconsistent with a historical trajectory appears in a mapping path;

[0014] S102: Obtaining a feedback sentence of the user to the intention clarification request, combining the feedback sentence with the historical sentence pairs, and inputting the combination to a semantic jump detection structure to construct a semantic tension spectrum for representing semantic coherence and evolution trend, and generating a candidate segment set of multi-path responses based on a tension closed cluster in the semantic tension spectrum;

[0015] S103: Inputting the candidate segment set of multi-path responses to a misleading chain cross-analysis unit, identifying a high-frequency overlapping response segment formed with a risk trigger main chain according to a semantic path coincidence degree of each candidate segment in a historical negative feedback context, and marking the high-frequency overlapping response segment as shielding content;

[0016] S104: In the candidate segments after removing the shielding content, selecting the smallest segment as a final response segment according to a context break span of a semantic tension closed cluster corresponding to the candidate segment, and feeding back user interaction behaviors caused by the final response segment to the behavior evolution trajectory template and the semantic tension spectrum respectively for dynamically updating an intention recognition structure and a candidate screening mechanism.

[0017] As a preferred technical scheme of the first aspect of the application, the determination of the semantic derailment point comprises:

[0018] In the semantic label sequence, determining a position of a last semantic label change, taking the position as a semantic jump base point, defining a semantic label at the semantic jump base point as a base point label, and selecting the first three label nodes before the semantic jump base point in the generation order of the semantic label sequence as presequence labels, so that the base point label and the first three presequence labels jointly constitute a bidirectional label segment group, wherein the presequence labels are used to represent the local context before the semantic jump, and the base point label is used to represent the semantic jump itself.

[0019] mapping the bidirectional label fragment group with the semantic path in the user historical interaction corpus, and counting the number of matched historical paths;

[0020] If the number of matches is lower than the preset threshold, it is determined that the bidirectional label fragment group constitutes a semantic derailment structure, and the last label in the bidirectional label fragment group is determined as a semantic derailment point;

[0021] Based on the semantic derailment point, its lexical semantic path and semantic adjacent node are extracted to construct a lexical and semantic divergence area;

[0022] Three branch nodes with the largest semantic level span are selected from the lexical and semantic divergence area as intent clarification request items.

[0023] As a preferred technical solution of the first aspect of the application, the lexical and semantic divergence area construction process comprises:

[0024] In the general knowledge graph, the direct hyponym and hypernym of the target label and its horizontal diffusion node are obtained;

[0025] The horizontal diffusion node with an appearance frequency greater than one in the user-confirmed semantic path is removed;

[0026] The remaining nodes are matched with the shortest path to filter the paths with semantic interruptions;

[0027] Based on the knowledge graph, the shortest semantic path between the remaining horizontal diffusion node and the semantic derailment point is calculated, and the position with semantic interruption is filtered;

[0028] The position with semantic interruption is taken as a divergence point, and the semantic path extending outward and not appearing before is extracted as a candidate branch path;

[0029] The candidate branch path is sorted based on the semantic jump level, and the top three terminal nodes are selected as intent clarification request items.

[0030] As a preferred technical solution of the first aspect of the application, the step of generating a multi-path response candidate fragment based on the semantic tension graph comprises:

[0031] In the semantic tension graph, a tension rising section on the current semantic path is identified, and a heterogeneous semantic fragment pair of the paths before and after the corresponding section is constructed;

[0032] The heterogeneous semantic fragment pair is executed in the historical corpus to perform structural migration matching, and a path group with cross mapping in the semantic space but not continuous in the path structure is obtained;

[0033] Taking the semantic jump position in the path group as the center, the three semantic nodes before and after it are extracted as a candidate fragment window.

[0034] determining whether the candidate segment window can be covered by two heterogeneous paths simultaneously, and if so, marking the semantic segment corresponding to the candidate segment window as a neutral tension segment;

[0035] outputting a candidate segment set of multi-path responses based on the closed feature of the neutral tension segment in the semantic tension map.

[0036] As a preferred technical solution of the first aspect of the application, the generation logic of the shielding content comprises:

[0037] From the current candidate segment, all candidate SVO structures conforming to the subject-verb-object relationship are identified based on the dependency syntax analysis, and a set of misleading SVO structures marked as misleading, wrong or semantic deviation by the user is extracted from the historical negative feedback record;

[0038] Comparing the candidate SVO structure with the misleading structure set, calculating its longest co-occurrence semantic segment based on the semantic map, and determining whether the semantic tension direction of the longest co-occurrence semantic segment is consistent with the direction in the current semantic tension map;

[0039] If the semantic tension direction of the co-occurrence segment is consistent, further search whether the candidate SVO structure has ever triggered user error correction behavior in the historical dialogue;

[0040] If all the above conditions are met, the complete sentence of the candidate SVO structure is marked as a potential risk response unit;

[0041] Semantically aggregate all potential risk response units to form a shielding content set, and exclude the corresponding content from the candidate segment to prevent misleading segments from entering the final response generation process.

[0042] As a preferred technical solution of the first aspect of the application, the aggregation of the potential risk response unit comprises:

[0043] Extract the semantic jump path from the subject to the object from the SVO structure marked as risk, and calculate the semantic level change amount of the path based on the semantic map, and record its jump span;

[0044] Filtering jump paths with jump span greater than two levels and no intermediate semantic nodes in the path, and taking them as a set of fault path;

[0045] Performing path backtracking on the jump segment with the highest frequency in the fault path set, constructing its associated path influence chain along the semantic map, and representing the semantic diffusion range that the jump segment may trigger;

[0046] determining whether any fault jump segment in the path influence chain is contained in the candidate response segment as a basis for identifying potential misleading risks;

[0047] When the candidate segment contains the fault jump segment, the segment is marked as a strong risk response unit, and is added to the shielding content set for exclusion from the subsequent response generation process.

[0048] As a preferred technical solution of the first aspect of the present application, the calculation of the context break span includes:

[0049] The syntactic dependency structure of the context of each candidate segment is constructed, and a dependency edge connection density map between the context and the segment is generated based on the dependency relationship;

[0050] An interrupted area of the dependency edge between the context and the segment is identified in the connection density map;

[0051] Dependency path expansion is performed from the interrupted area to both sides of the context respectively until the dependency chain is reconnected or expanded to a set maximum boundary;

[0052] The shortest reconnection edge distance of each expanded path is recorded, and the distance is taken as the context break span of the corresponding segment;

[0053] The context break spans of all candidate segments are compared, and the segment with the smallest span is selected as the final output segment.

[0054] As a preferred technical solution of the first aspect of the present application, the break span comparison includes:

[0055] Dependency edge combinations of the context break path and the reconnection path are constructed, and the dependency edges are classified by type;

[0056] Dependency structure groups are constructed for each edge type, the dependency structure groups including subject-predicate structure groups, verb-object structure groups and bias-positive structure groups, different dependency structures having different frequencies of occurrence in normal semantic expression and different contributions to syntactic coherence;

[0057] It is determined whether the high-frequency backbone structure type is missing in the break area, and if so, a corresponding penalty value is added to the break path;

[0058] A break span total score model is constructed according to the break edge length and the penalty value, and the total score is standardized;

[0059] The one with the lowest break span total score is determined as the semantic break minimum segment.

[0060] As a preferred technical solution of the first aspect of the present application, the user interaction feedback update of the output segment includes:

[0061] Collecting user's first language input behavior in a dynamic short-time monitoring window after the output segment is presented, and extracting key word tags in the input;

[0062] Semantically pairing the key word tags with the end semantic tags of the output segment to determine whether the user input constitutes semantic continuation;

[0063] When determining semantic continuation, reversely mapping the interaction path corresponding to the user input to the semantic tension graph, and recording the semantic chain in the continuation direction;

[0064] Updating the weight value and path priority of the semantic chain to improve the recommendation probability of the semantic direction in subsequent response generation;

[0065] Synchronously writing the updated semantic tension graph state into the behavior evolution trajectory template to update the evolution sequence structure of the user's intention.

[0066] In a second aspect, the present application provides a content generation and interactive dialogue system based on AI artificial intelligence, based on the implementation of the first aspect, including a semantic tag construction module, a clarification feedback fusion module, a misleading chain analysis module and a response screening and updating module, each module transmits data through wired and / or wireless;

[0067] The semantic tag construction module collects user input text, and constructs a semantic tag sequence according to at least three rounds of historical statements in the current conversation window; map the semantic tag sequence to the behavior evolution trajectory template generated based on the user's long-term interaction record, if a semantic node inconsistent with the historical trajectory appears in the mapping path, identify it as a semantic derailment point and trigger an intention clarification request;

[0068] The clarification feedback fusion module obtains the feedback statement of the user to the intention clarification request, combines the feedback statement with the historical statement pair, and inputs it into the semantic transition detection structure to construct a semantic tension graph for representing semantic coherence and evolution trend, and generates a candidate segment set of multi-path responses based on the tension closed cluster in the semantic tension graph;

[0069] The misleading chain analysis module inputs the candidate segment set of multi-path responses into the misleading chain cross analysis unit, identifies the high-frequency overlapping response segment formed with the risk trigger main chain according to the semantic path coincidence degree of each candidate segment in the historical negative feedback context, and marks the high-frequency overlapping response segment as shielding content;

[0070] The response screening updating module selects the minimum segment as the final response segment according to the context break span of the semantic tension closed cluster corresponding to the candidate segment in the candidate segment after the shield content is removed, and feeds back the user interaction behavior triggered by the final response segment to the behavior evolution trajectory template and the semantic tension atlas respectively, so as to dynamically update the intention recognition structure and the candidate screening mechanism.

[0071] In the above technical solution, the present application provides technical effects and advantages:

[0072] The present application constructs the linkage mechanism among the semantic label sequence, the behavior evolution trajectory template and the semantic tension atlas, so that the system can identify the semantic derailment point in time when the user input appears semantic jump, and actively trigger the intention clarification, thereby avoiding the error response caused by the misjudged intention from the source.

[0073] The multiple-path response candidate segment is generated by the tension closed cluster in the semantic tension atlas, so that the system can maintain coherence in multiple potential semantic directions; the multi-dimensional risk screening is performed on the candidate segment by the misleading chain cross analysis unit, and the misleading structure overlapping with the high frequency historical negative feedback is removed, so as to effectively prevent the error understanding, fact deviation or semantic misleading information from entering the final response path.

[0074] The strategy of selecting the minimum segment based on the context break span makes the output content most consistent with the context semantic context in structure, significantly improves the naturalness and coherence of the system response, and realizes the adaptive learning of the user intention migration rule by the continuity analysis of the subsequent user interaction behavior and the dynamic writing of the action path to the semantic tension atlas and the behavior evolution trajectory template, so that the system can continuously improve the intention recognition accuracy and the response recommendation quality in the continuous use process.

[0075] The present application not only solves the problems of the existing dialogue system, such as difficult to capture real intention, easy to produce misleading content and context response break, but also significantly enhances the semantic understanding ability, risk control ability and long-term self-learning ability of the system, and improves the controllability, reliability and user experience of AI dialogue. BRIEF DESCRIPTION OF DRAWINGS

[0076] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced. Obviously, the drawings in the following description are only some embodiments described in the present application, and other drawings can also be obtained by those skilled in the art according to these drawings.

[0077] Figure 1 The flow chart of the content generation and interactive dialogue method based on AI artificial intelligence of the present application;

[0078] Figure 2A framework diagram of the AI artificial intelligence-based content generation and interactive dialogue system of the present application;

[0079] Figure 3 A framework diagram of the example judgment of semantic derailment points of the present application. DETAILED DESCRIPTION

[0080] Example implementations will now be described more fully with reference to the accompanying drawings. Example implementations may, however, be implemented in many different forms and should not be construed as limited to the examples set forth herein; rather, these example implementations are provided so that this disclosure will be thorough and complete, and will fully convey the concept of example implementations to those skilled in the art. Accompanying drawings are included to provide a further understanding of the present disclosure and are incorporated in and constitute a part of this specification, illustrate embodiments of the present disclosure and serve to explain the principles of the present disclosure. Features, structures or characteristics of the description can be combined in any suitable manner in one or more example implementations. In the following description, numerous specific details are provided to give a thorough understanding of example implementations of the present disclosure. One skilled in the relevant art will recognize, however, that the technology of the present disclosure can be practiced without one or more of the specific details, or with other methods, components, materials, and so forth. In other instances, well-known structures, materials, or actions are not shown or described in detail in order to avoid obscuring aspects of the present disclosure.

[0081] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more example implementations. In the following description, numerous specific details are provided to give a thorough understanding of example implementations of the present disclosure. One skilled in the relevant art will recognize, however, that the technology of the present disclosure can be practiced without one or more of the specific details, or with other methods, components, steps, etc. In other instances, well-known structures, materials, or actions are not shown or described in detail in order to avoid obscuring aspects of the present disclosure.

[0082] Example 1

[0083] As Figure 1 shown, the present application provides an AI artificial intelligence-based content generation and interactive dialogue method, comprising the following steps:

[0084] S101: Collecting user input text, constructing a semantic label sequence according to at least three rounds of historical statement pairs in the current session window; mapping the semantic label sequence to a behavior evolution trajectory template generated based on user long-term interaction records, and if a semantic node inconsistent with the historical trajectory appears in the mapping path, identifying it as a semantic derailment point and triggering an intent clarification request;

[0085] It should be noted that the semantic label sequence is generated by parsing the action, entity, emotion and intent elements in the user input and multiple rounds of context, combining syntax dependency and anaphora resolution, and concatenating structured semantic chains in chronological order to identify the trend of semantic changes. The behavior evolution trajectory template is constructed according to the label sequence of the user's long-term historical dialogue, and by statistically analyzing the typical semantic migration path and intent convergence rule, it is determined whether the current semantic sequence deviates from the usual mode, thereby realizing intent analysis and derailment detection.

[0086] The judgment of the semantic derailment point comprises:

[0087] In the semantic label sequence, the position of the last semantic label change is determined, which is taken as the semantic jump base point, and the semantic label at the semantic jump base point is defined as the base point label. The three previous label nodes in the generation order of the semantic jump base point along the semantic label sequence are selected as the pre-sequence labels, and the base point label and its three previous pre-sequence labels jointly constitute a bidirectional label segment group, wherein the pre-sequence labels are used to represent the local context before the semantic jump, and the base point label is used to represent the semantic jump itself.

[0088] It should be understood that the changes in label category and semantic direction are monitored in real time on the semantic label sequence. When it is detected that the semantic category, interaction role or behavior direction of the current label has a mutation relative to the previous label, the position of the label is recorded as the last semantic label change position, and it is taken as a candidate derailment point. On this basis, the candidate derailment point is taken as the base point, and three semantic label nodes are traced back along the time axis, which together with the candidate derailment point constitute a bidirectional label segment group before and after the label change. The bidirectional label segment group reflects the local context before the semantic jump and the jump position itself, which is the smallest analysis window for judging whether the jump is an abnormal deviation, and is used to construct a segment pair before and after the label change for subsequent derailment judgment. When updating the semantic label sequence in the user's current session window online, each semantic label is recorded in the order of the actual occurrence of the dialogue as ; wherein Each label corresponds to the semantic category, behavior type or emotional state and other structured labels obtained after semantic analysis in the user input sentence or system output sentence.

[0089] In the semantic derailment detection process, the system monitors the label category, semantic direction and interaction role attribute in real time on the semantic label sequence, and when it is detected that the current label has a mutation of a preset amplitude in any of the above attributes relative to the previous label, the position of the label is marked as the "last label change position". For example Figure 3As shown, assuming that label L4 is the last label change point, corresponding to the semantic jump base point, after determining the last label change position, the position is taken as the base point to backtrack along the generation order of the semantic label sequence, and the three label nodes L1, L2, and L3 immediately before the base point are sequentially selected, and together with the base point label L4, a local bidirectional label fragment group 〔L1, L2, L3, L4〕 is formed. Among them, L1-L3 are before L4, and are arranged in turn according to the dialog generation order, used to represent the local context continuous change before the semantic jump occurs; L4 is at the tail end of the fragment group, used to represent the position and semantic state of the semantic jump itself, and the dashed line is used to represent the corresponding mapping relationship between each label node in the semantic label sequence and the time generation order or position index.

[0090] It should be noted that the relationship between the labels in the bidirectional label fragment group 〔L1, L2, L3, L4〕 is a sequence position relationship, that is, arranged in turn from front to back according to the generation order of the semantic label in the dialog process, and does not depend on the absolute timestamp, but takes the label sequence index as the order reference. By taking the last label change position L4 as the center and backtracking three adjacent label nodes, a minimum analysis window is constructed, which can capture the evolution trend before the semantic jump and the semantic state at the jump moment, thereby providing a structured basis for subsequent position alignment and matching of the window with the semantic path in the user historical interaction corpus. If the occurrence frequency of the bidirectional label fragment group in the historical behavior evolution track is lower than a preset threshold, it is determined that the current semantic change belongs to an abnormal deviation structure, and the label L4 at the tail end of the fragment group is determined as the semantic derailment point.

[0091] The bidirectional label fragment group is positionally aligned and mapped with the semantic path in the user historical interaction corpus, and the number of matchable historical paths is counted.

[0092] It is further clarified that the user historical interaction corpus is a conversation path in the user real interaction log that has confirmed intent completion. The bidirectional label fragment group is positionally aligned and mapped with the semantic path with confirmed intent in the user historical interaction corpus. In the mapping process, a semantic sequence similarity calculation based on node index is adopted, and only when the label order in the bidirectional label fragment group is completely consistent with the order of a continuous label subsequence in the historical path, and the semantic direction matches, the historical path is counted as a matchable path. The number of matching paths that meet the conditions is counted, which is used to reflect the frequency of occurrence of the current semantic change structure in the historical behavior mode.

[0093] If the number of matches is lower than a preset threshold, it is determined that the bidirectional label fragment group constitutes a semantic derailment structure, and the last label in the bidirectional label fragment group is determined as the semantic derailment point.

[0094] It can be understood that when the number of matches is lower than the preset threshold (for example, 3), it means that the bidirectional label fragment group appears less frequently in the historical behavior evolution track and does not belong to the user's habitual semantic mode. At this time, the system determines that the bidirectional label fragment group is a semantic derailment structure, and confirms the last label in the fragment group (that is, the candidate derailment point) as the semantic derailment point. The derailment point identifies the position in the current semantic path that is most likely to cause misunderstanding or unclear intent, providing a basis for subsequent construction of the word sense divergence zone and triggering of the intent clarification request.

[0095] Based on the semantic derailment point, its lexical semantic path and semantic adjacent nodes are extracted to construct a word sense divergence zone.

[0096] It should be emphasized that after determining the semantic derailment point, the target label is taken as the target word, and its hypernyms, hyponyms and semantic adjacent nodes are extracted from the general knowledge graph to form an initial word sense structure network. By further screening the paths that have semantic interruption or semantic jump with the target label, and taking the breakpoints in these paths as semantic divergence points, the system can construct a word sense divergence zone that embodies multiple potential meaning directions.

[0097] The three farthest semantic branch nodes in the word sense divergence zone are taken as the option content of the intent clarification request and output.

[0098] It can be understood that the three terminal word sense nodes with the largest semantic level span and the highest path jump level between the target label are selected from the word sense divergence zone to output three optional intent items in natural language form, and an intent clarification request is initiated to the user, thereby realizing fine-grained explanation and correction of the semantic derailment point.

[0099] To further ensure that the options generated by the intent clarification request have semantic distinction and representativeness, after triggering the semantic derailment point, the following word sense divergence zone construction process is continued, which includes:

[0100] All direct hypernyms and hyponyms of the target label are found in the general knowledge graph, and their horizontal diffusion nodes in the knowledge classification path are recorded.

[0101] It should be noted that the target label is a label that is determined to have derailment risk as a semantic derailment point in the semantic label sequence. The embedded general knowledge graph (such as the concept hierarchy network or the entity class atlas) is called to find all direct hypernym semantic classes (i.e. higher abstraction layers into which it is classified) and direct hyponym semantic classes (i.e. specific subclasses it contains) to which the label belongs. Based on these hypernyms / hyponyms as center points in the knowledge classification path, their parallel classifications (i.e. other entities at the same level) are obtained horizontally, marked as horizontal diffusion nodes, and a semantic neighborhood set around the target label is formed.

[0102] All transverse diffusion nodes appearing more than once in the user's confirmed semantic path are excluded, and only the unappeared label paths are retained.

[0103] It should be noted that, in order to prevent the user from repeatedly displaying known or confirmed semantic content, the transverse diffusion nodes are matched with the nodes appearing in all historical confirmed semantic paths of the user before the options are constructed, and all nodes appearing more than once are excluded.

[0104] This operation is based on the user behavior reaction library, and ensures that the retained label paths are paths that have not been explicitly identified or confirmed by the user, thereby improving the exploratory and ambiguity deconstruction ability of the intent clarification request.

[0105] The remaining node set is matched with the shortest path to filter the paths having semantic interruption positions between the target label.

[0106] Further, in the knowledge graph structure, the shortest path between the above-mentioned retained node set and the target label is calculated, and it is checked whether there is an interruption point in the path, i.e., the necessary intermediate edges between nodes are missing or there is a semantic jump (for example, directly connecting "emotion" from "object").

[0107] If there is an interruption, the path is regarded as a candidate path having "semantic divergence", which is used to generate "high-jump semantic options".

[0108] The nodes before the path interruption are taken as divergence points, and all unappeared paths extending from each divergence point are taken as branch paths.

[0109] After determining the semantic interruption occurrence node, the node is taken as a semantic divergence point, and further expansion is performed on other unappeared semantic paths of the node to form a set of non-traditional extended paths.

[0110] These paths have never been used in the user's past dialogues, and represent potential understanding directions that have not been explored in the current context, which can be used as divergent options for intent clarification.

[0111] Three end-point word sense nodes having the largest semantic path jump level from the target label are extracted and output as user intent options.

[0112] Finally, from all the constructed paths, three end-point nodes having the largest path jump level from the target label in the graph structure (i.e., having the largest semantic span and the largest direction deviation) are selected, and are output in the form of natural language as three alternative intent items for the user to select in the intent clarification request.

[0113] These jump-maximum but legally existing semantic directions ensure that the generated options have clear semantic boundaries and cover the main directions of the actual intent that the user may not have expressed clearly.

[0114] To avoid repeated attention to the semantic direction that the user has explicitly identified, the system compares the above-mentioned lateral diffusion nodes with the user's historical confirmed semantic path, eliminates nodes that appear more than once in the historical path, and only retains label paths that are not widely used to improve the exploratory nature of the clarification options.

[0115] Subsequently, the shortest semantic path between the remaining node set and the semantic derailment point is calculated, the positions of the interrupt points or cross-domain jumps in the path are identified, and these positions are regarded as semantic divergence points.

[0116] Starting from each divergence point, further expanding along its unused semantic edges, a set of branch paths that have not appeared in the user's historical dialogue is obtained. For these candidate branch paths, based on indicators such as semantic jump level, path depth, and semantic category span, the three terminal word sense nodes with the largest path jump level are selected as intent options representing different potential understanding directions and output to the user.

[0117] The word sense divergence area constructed through the above process can effectively cover the main semantic directions to which the user's real intent belongs and ensure that different options have clear semantic boundaries.

[0118] S102: Obtain the feedback statement of the user in the intent clarification request, and input the feedback statement and the historical statement pair into a semantic transition detection structure to construct a semantic tension map representing semantic coherence and evolution trend; the map is used to identify tension closed clusters in the current semantic path and generate a candidate fragment set of multi-path responses accordingly;

[0119] It should be noted that based on the clarification content of the user feedback, the potential multi-response trend is identified in combination with the current semantic evolution background, and the semantic coherence and trend are represented in a structured map to avoid generating misleading or broken answers; the semantic transition detection structure is used for semantic stage change analysis of the input semantic sequence, by detecting the changes of semantic categories, semantic directions or semantic roles between adjacent nodes, the semantic transition features are extracted; and according to the continuity and offset degree of each transition feature, the coherence and evolution trend of the current semantic sequence are judged.

[0120] The tension relationship between complex semantic paths is modeled through the modeling of the semantic tension map, and the candidate fragment set of multi-path responses covering different semantic trends is generated through the extraction logic of the tension closed clusters in the map. Its internal structure has significant nonlinear characteristics, avoiding the monotony of linear sentence concatenation.

[0121] Specifically, the generation of the candidate fragment set of multi-path responses includes:

[0122] Selecting each tension rising segment on the current path in the semantic tension map, constructing the longest non-overlapping segment in the longest path before and after it, forming a heterogeneous path pair.

[0123] It should be understood that the semantic tension map is a directed graph structure constructed based on the change trend of context semantic connection tension. The tension rising segment refers to a paragraph in the map where the semantic consistency between multiple nodes is reduced but the logical close connection exists, which usually represents a potential semantic divergence point.

[0124] In the semantic tension map, when comparing the two paths before and after the tension rising segment, the longest segment in the continuous path segment where the semantic nodes do not correspond to each other in the two paths is the longest non-overlapping segment, which represents the maximum divergence area of the path in semantic evolution and is used to construct a heterogeneous path pair. The heterogeneous path pair is used for subsequent semantic migration and candidate segment generation modeling.

[0125] Mapping the heterogeneous path pair to the structure migration matrix in the corpus, identifying all path groups that exist in history with cross mapping but discontinuous structure.

[0126] It can be understood that in order to determine whether the above heterogeneous path has ever indirectly co-occurred in different forms in user history interaction, it is mapped to the pre-constructed structure migration matrix. The structure migration matrix records the label evolution and migration path (including jump connection, fault connection, bridging node, etc.) between any two structures appearing in user interaction.

[0127] If a group of heterogeneous paths exist in the history migration matrix with cross mapping (i.e. two segments have co-occurred with other paths but not directly connected), it is considered that the path group has migration rationality and is included in the candidate path group for constructing the response bridge between structure cracks.

[0128] Taking the three nodes before and after the jump label position in the above path group as a segment candidate window, extracting its text segment to construct a path segment pool.

[0129] For each identified path group, further detect its semantic jump point, i.e. the position where the label evolution appears mutation or fault, and extract the text content corresponding to the three label nodes before and after it to form a segment candidate window.

[0130] Subsequently, the text segments extracted in all candidate windows are collected to construct a path segment pool, which is used for subsequent structure screening and tension evaluation.

[0131] For each path segment, determine whether it can be covered by two heterogeneous paths, if it can be covered, mark it as a neutral tension segment.

[0132] For each text segment in the path segment pool, calculate whether it has a label structure alignment relationship with both paths in the above heterogeneous path pair (i.e. the segment can be interpreted by both paths or contains).

[0133] If a segment is in the neutral or low tension section of the tension map, and the text semantics do not constitute a structural conflict under the two paths, it is marked as a neutral tension segment, i.e. the segment has high multi-directional adaptability in the current semantic direction.

[0134] All neutral tension segments are classified according to the tension closure characteristics, and the output is a candidate segment set for multi-path response.

[0135] Finally, all neutral tension segments are classified according to their tension closure cluster in the semantic tension map, which refers to the low tension subgraph area where all tension paths converge in the map, representing the semantic convergence direction. Extract neutral segments from each tension closure cluster to construct a response candidate set with multiple potential semantic directions, and use it as the basic material set for subsequent misleading content screening and final response generation.

[0136] S103: Input the candidate segment set for multi-path response to the misleading chain cross analysis unit, identify high-frequency overlapping response segments with the risk trigger main chain according to the semantic path coincidence degree of each candidate segment in the historical negative feedback context, and mark them as shielding content;

[0137] It should be noted that the so-called misleading chain cross analysis unit refers to a comparative analysis module based on SVO (subject-verb-object) structure, which constructs a dynamic and evolving misleading structure detection strategy by integrating semantic tension map directional information and user negative feedback data.

[0138] Specifically, the generation logic of the shielding content includes:

[0139] Extract all subject-verb-object structures in the current response candidate segment, and extract the misleading SVO structure set from the historical negative feedback.

[0140] It can be understood that for each candidate segment, perform syntax structure analysis, extract all semantic units that meet the subject-predicate-object structure based on the dependency syntax analyzer, and mark them as candidate SVO structures. At the same time, access the feedback logs in the user behavior database that have been explicitly marked by users as "misleading", "wrong understanding" or "ambiguous meaning", extract the SVO structures therein, and constitute the misleading SVO structure set.

[0141] The semantic path of the candidate SVO structure and the misleading structure set is compared, the longest co-occurrence semantic segment is calculated based on the semantic atlas, and it is judged whether the semantic tension direction of the segment is consistent with the direction in the current semantic tension atlas.

[0142] Further, the semantic path of the candidate SVO structure and the structure in the misleading SVO structure set is compared, the longest alignable sub-path is identified through semantic atlas node sequence comparison, and if it exists, its path direction is recorded and compared with the directionality in the current semantic tension atlas. Only when the path direction is consistent, it is considered that there is potential semantic misleading.

[0143] If the semantic tension directions of the co-occurrence segments are the same, further analysis is performed on whether the structure triggers user correction behavior in history.

[0144] The behavior path record of the candidate SVO structure in historical interaction is retrieved, and it is detected whether a correction response has been triggered, such as the user sending “you misunderstood” or “not this meaning” keyword expressions. If such interaction events are strongly related to the structure, the risk level is adjusted, and it is considered as a misleading structure with user verification support.

[0145] When the above conditions are met, the complete sentence corresponding to the candidate SVO structure is marked as a potential risk response unit.

[0146] At this time, the complete sentence containing the candidate SVO structure is recorded as suspicious response content, and is combined with other risk sentence segments to form a potential risk response unit, which is used for jump path deep analysis in the next stage.

[0147] All potential risk response units are aggregated into a shielding content set and removed from the response candidate set.

[0148] The potential risk response units identified above are subjected to semantic consistency clustering processing, and after removing non-related structures caused by context differences, a shielding content set is finally formed. All contents in the set do not participate in the subsequent final response candidate generation.

[0149] Further, to enhance the technical disclosure of the shielding content construction logic, the aggregation of the potential risk response units includes:

[0150] The semantic jump path between the subject and the object is extracted from each SVO structure marked as risk, and the jump span in the semantic relationship graph is recorded.

[0151] The relationship graph of each structure node in the semantic atlas is called, the path is expanded from the subject node to the object node, and the number of all non-continuous semantic level change nodes in the path is recorded as the jump span.

[0152] Screen out the jump path with a jump span greater than two levels and without the participation of intermediate nodes as the fault path set.

[0153] If there is a label level transition greater than two levels in the jump path, and no neutral node appears in the path (i.e. not through the conventional semantic transition node), it is considered as a semantic fault, and such paths constitute the fault path set.

[0154] The jump segment with the highest frequency in all fault path sets is connected back to generate the path impact chain.

[0155] Further frequency statistics are performed on all fault paths to identify the jump path segment with the highest number of repetitions, and the path structure is unfolded in reverse to construct the path impact chain associated with the semantic graph.

[0156] Determine whether there is a response structure in the response candidate set that contains any fault jump segment in the path impact chain.

[0157] Using the path impact chain as an index, all structures in the response candidate set are quickly indexed and matched. If a candidate response contains any jump segment of the impact chain, it is marked as a high-repetition-risk response.

[0158] If so, mark the corresponding segment of the response structure as a strong-risk-response unit and add it to the shielded content set.

[0159] Such structures are identified as high-risk, user-misleading responses with a high probability of user misdirection. The complete segment will be excluded and added to the final shielded content set, which will not enter the output process.

[0160] S104: In the segment set after removing the shielded content, select the segment with the smallest context break span in the corresponding semantic tension closed cluster as the final response output to the user terminal; and the user interaction behavior triggered by the output segment is fed back to the behavior evolution trajectory template and the semantic tension graph, respectively, for dynamic adjustment of the intent recognition and candidate selection structure.

[0161] It should be noted that the current step undertakes the dual tasks of final response segment determination and dynamic feedback mechanism construction in the present application. The design goal is to optimize the segment that is least likely to cause context break in multiple path response segments, and to use the real user behavior triggered by the segment to continuously optimize the intent recognition and candidate generation model structure.

[0162] It can be understood that this step first performs dependency structure analysis on all unshielded response candidate segments to measure the coherence of the context semantic tension transition, then selects the segment with the smallest context break based on the tension closed cluster characteristics, and implements structured feedback mapping based on user response after output.

[0163] Exemplarily, the calculation of the context break span includes:

[0164] The context corresponding syntax dependency structure is constructed for each candidate segment respectively, and a connection density map of dependency edges is generated.

[0165] The context and the candidate response segment are subjected to syntax analysis, all dependency edges are extracted based on the dependency syntax tree, and a two-dimensional connection density map is generated according to the syntax structure direction, the part-of-speech relationship and the edge weight, which depicts the density of the inter-word grammar coherence and is used to identify the strong and weak areas of the semantic structure connection.

[0166] The areas where the connection edges between the context and the segment are discontinuous are identified in the connection density map as initial context break areas.

[0167] The dependency edge density between the context words and the response segment words in the connection density map is scanned to identify the density drop or edge missing section. If the dependency edge length of a certain area exceeds a set interval threshold and the edge type is missing the main structure (such as subject-predicate or verb-object), the area is marked as an initial context break area.

[0168] The dependency path is expanded from each context break area to both sides of the context until it is reconnected to the known context dependency chain or reaches the maximum boundary.

[0169] From the break area, the syntax path is expanded to the left (previous context) and right (segment interior) along the dependency edge respectively. If the expansion path is connected to the known high-density area in the context or forms a complete subgraph, the path is recorded as a reconnection path, otherwise it is terminated at the boundary.

[0170] The shortest reconnection edge distance in the expansion path is recorded as the context break span of the segment.

[0171] In each expansion path, the shortest dependency edge number from the break point to the successful reconnection point, i.e. the reconnection edge distance, is calculated to quantify the break span between the response segment and the context.

[0172] The context break spans of all candidate segments are compared, and the one with the smallest break span is selected as the final output segment.

[0173] The break spans recorded for all candidate segments are compared, and the segment with the shortest span is selected as the response with the strongest semantic coherence in the current context, which is output to the user terminal to complete the segment selection operation.

[0174] Further, to make the evaluation of the break span more accurate and scene-adaptive, the break span comparison includes:

[0175] The dependency edge combination of the context breaking path and the reconnection path is constructed, and the dependency edges are classified according to types.

[0176] Specifically, all edges in the breaking path and the reconnection path are labeled and classified according to grammatical types, such as subject-verb, verb-object, modifier-noun, etc., to form a complete structure vector group. By combining the missing dependency edges in the breaking path and the complementary parts in the reconnection path, a set of dependency edges for structure comparison is formed, and is divided into different structure categories such as subject-verb, verb-object, and modifier-noun according to edge types.

[0177] Dependency structure groups are constructed for each edge type, including subject-verb structure groups, verb-object structure groups, and modifier-noun structure groups. The frequency of occurrence of different dependency structures in normal semantic expression and their contribution to syntactic coherence are counted according to the combination mode to generate a connection mode group of three types of main structure, which is used to analyze the missing or abnormality of structure connection.

[0178] It is identified whether the breaking area is missing high-frequency main structure types. If missing, the breaking path penalty value is increased. If the subject-verb, verb-object, and other high-frequency main dependency structures are missing in the breaking area, it indicates that there is a more obvious semantic jump in the context of the segment, and the breaking span is expanded by increasing the penalty value.

[0179] A breaking span total score model is constructed according to the breaking edge length and the penalty value, and the total score is standardized;

[0180] It can be understood that the breaking edge length is used to reflect the basic separation degree of the segment and the context at the structure level, and the penalty value is used to supplement the coherence loss caused by the missing of the main structure. The two together constitute the total score of the breaking span. After standardization, the total score can be compared between different candidate segments. The breaking length + penalty value of each segment is summed and normalized to construct a breaking span score model of a unified scale to compare the coherence of all segments.

[0181] The segment with the lowest breaking span total score is selected and marked as the semantic breaking minimum candidate segment and output.

[0182] The lowest breaking span indicates that it is most closely related to the original context structure and requires the least compensation for semantic reconnection. The semantic breaking minimum segment is considered to be the most suitable for the current context structure in all dimensions (structure, length, frequency, and missing compensation), and it will be output to the user terminal as the final response.

[0183] In addition, in step S104, for the final output segment, a dynamic adjustment mechanism of the following user interaction feedback path is also included, which is used to improve the semantic adaptation ability and behavior prediction ability for future response generation. The logic of this mechanism is mainly used to update the intent recognition path and response candidate screening logic, to ensure continuous self-adjustment and optimization according to user habits.

[0184] The user interaction feedback update of the output segment includes:

[0185] The first language input behavior of the user within 5 seconds after receiving the output segment is collected, and the key word label in the behavior is recorded.

[0186] It can be understood that a short monitoring window is started after the user is presented with the response segment, and a 5-second monitoring window is set. If the user terminal generates a language input behavior (such as a voice instruction or text input) within the window, the key terms in the behavior are immediately extracted and subjected to semantic labeling processing, such as through named entity recognition (NER) and context intent decoding structure, to label the label attributes of each key term. The key terms are extracted from the input content and subjected to semantic labeling, which is used to represent the user's immediate semantic focus.

[0187] The key word label is semantically paired with the label at the end of the current output segment to determine whether it constitutes semantic continuation.

[0188] The connection relationship between the labels in the semantic tension map is called to calculate the correlation degree between the key word label and the semantic label at the end of the output segment. When there is a direct semantic connection or they are in a closable short-range jump path, it is determined that the user input and the current segment constitute semantic continuation.

[0189] If it constitutes semantic continuation, the user interaction behavior path is reversely mapped to the semantic tension map, and the corresponding semantic continuation chain is recorded.

[0190] For the identified continuation path, it is reversely mapped with the known path in the semantic tension map, the mapping path segment of the interaction behavior in the semantic map is identified, and its tension extension structure (i.e. the semantic inertia direction triggered by the user) in the map is extracted, forming a continuation chain consistent with the semantic direction of the current response segment, which is used to describe the natural development trend of the user's semantic behavior.

[0191] The weight value and path priority of the mapped semantic continuation chain are updated, which is used to improve the subsequent response recommendation probability.

[0192] According to the semantic chain triggered by the user feedback behavior, the edge weight and node priority corresponding to the continuation chain in the semantic tension graph are adaptively updated, so that the semantic direction obtains a higher recommendation probability in the subsequent response candidate screening process, thereby enhancing the tracking ability of the system to the user intention evolution.

[0193] The updated semantic tension graph state is written back to the behavior evolution trajectory template for updating the intention evolution sequence structure.

[0194] The weight-adjusted semantic tension graph is synchronously written into the user's behavior evolution trajectory template to update the intention evolution sequence. The sequence is used to describe the semantic migration pattern presented by the user in multiple rounds of interaction, thereby supporting subsequent individualized intention prediction and response generation.

[0195] Embodiment 2

[0196] This embodiment is based on embodiment 1, and takes the real consultation process of a user on the implicit charge issue in an online intelligent customer service platform as the background. First, the user inputs in the intelligent customer service platform: “Does your platform automatically charge after the trial of the membership? I am worried about being charged money.”

[0197] The input is subjected to semantic analysis to identify the core semantic units therein, including “membership trial”, “automatic charge” and “fear of charge”. According to the semantic category division, it is divided into “service type label”, “charge behavior label” and “user risk emotion label”, etc., and a semantic label sequence is generated according to the input order: service_trial → billing_auto → risk_fear_charge; the sequence presents a cross-semantic field change from service consultation to risk concern.

[0198] The above label sequence is mapped to the behavior evolution trajectory template based on the historical interaction. The template usually shows a regular fee consultation path: service_trial → fee_standard → billing_method;

[0199] And in the user's current input, “billing_auto” and “risk_fear_charge” appear directly after “service_trial”, which is not in the typical user path of the structure, and is marked as a semantic derailment point.

[0200] After triggering the derailment point judgment, the bidirectional label segment group before and after the label change is constructed by backtracking from the jump node, and the matching degree is calculated in the historical trajectory. Since in the million-level customer service conversation history, the continuous path from trial to automatic charge concern has a very low occurrence rate (only 0.3%), which is much lower than the set threshold, it is determined that the jump has a semantic deviation feature and needs to enter the intention clarification stage.

[0201] To accurately identify the user's true intention, based on the core label of automatic renewal, extract its hypernyms, hyponyms, lateral diffusion nodes and semantic adjacency nodes from the knowledge graph. The obtained semantic structure includes multiple concepts such as "charging mechanism", "deduction policy", "trial clause", "automatic renewal rule" and "deduction reminder mechanism", forming a word meaning divergence area.

[0202] According to the user's historical behavior characteristics and the word meaning jump level, the semantic space is comprehensively divided into three directions that the user is most likely to pay attention to:

[0203] Confirm whether there is an automatic renewal behavior;

[0204] Understand the specific charging standard of the member;

[0205] Understand how to avoid mistaken deduction or involuntary deduction.

[0206] Based on the above directions, generate an intention clarification request: "In order to accurately understand your question, do you pay more attention to:

[0207] Will there be automatic deduction after the trial period ends?

[0208] The charging standard of the member?

[0209] How to avoid mistaken deduction?"

[0210] Assuming the user chooses: "I am worried about automatic deduction." After obtaining the clear intention, enter the next stage.

[0211] Take "automatic deduction worry" as the semantic core node, introduce the semantic tension graph model to construct the semantic development trend of the current dialogue. The tension graph automatically identifies that there are two significant directions in the current context:

[0212] 1. Tension closed cluster (safe direction): automatic renewal → trial period rule → user confirmation mechanism;

[0213] 2. Tension rising direction (risk direction): automatic renewal → complaint path → strong emotion or platform accusation.

[0214] When generating candidate response fragments, only select semantic stable content from the tension closed cluster. For example: "After the trial period ends, you will not be automatically deducted." "All charges must be confirmed by the user before they take effect." And "There will be a clear reminder before charging." Instead of choosing fragments that may lead to emotional escalation or complaint direction.

[0215] In the candidate fragments, it is found that some fragments may contain historically misunderstood sentence patterns, such as: "As long as you manage your subscription, you will not be deducted."

[0216] SVO structure analysis of the sentence extracts "You - manage - subscription". In the historical negative feedback corpus, this structure has been repeatedly identified by users as a way to shift responsibility and lack of security, which is a high-risk sentence pattern. Identifying that the content belongs to the misleading trend coincides with the direction of the complaint chain and risk chain, so it is excluded from the candidate set, and the final retained fragments are all from the low-risk semantic stable area.

[0217] Context break span analysis is performed on the remaining candidate fragments. This mechanism calculates the semantic coherence and information burden by constructing the syntactic dependency structure of the candidate fragments and the current dialogue context. In this embodiment, the break span of the following fragment is the smallest: "Our service will not automatically charge after the trial ends, all charges will be manually confirmed by you before they take effect." The semantic structure completely corresponds to the user's core concern about the risk of automatic charges, and the syntactic dependency relationship has no abrupt jumps, meeting the judgment conditions for selecting the optimal fragment, which is output as the final response.

[0218] After receiving the reply, the user inputs: "OK, thank you." This statement is identified as a "understanding confirmation" class label and is mapped to the behavior evolution trajectory template along with the current dialogue path, updating the user's intent weight in the "charge risk category" to enable faster identification of their focus in similar scenarios in the future, improving response efficiency and stability.

[0219] Embodiment 3

[0220] As shown in Figure 2 Embodiments 1-2, the present embodiment provides an AI-based content generation and interactive dialogue system. The system can be deployed on a cloud server, enterprise local server or edge computing node, and includes at least one processor, memory and communication interface. The communication interface interacts with the user terminal through wired and / or wireless networks. The memory pre-stores a computer program that can run on the processor. When the program is executed by the processor, the following functional modules are formed:

[0221] The semantic label construction module is used to interact with the user terminal for session data, collecting the user's current input text and at least three rounds of historical sentence pairs (including user input statements and system output statements) in the session window.

[0222] Specifically, the semantic label construction module has a semantic element analysis unit and a behavior trajectory mapping unit:

[0223] The semantic element analysis unit performs morphological analysis, syntactic dependency analysis and anaphora resolution on the input text and the historical statement, identifies action words, entity object words, emotional attitude words and intention association words, maps the semantic elements to preset semantic category labels, and concatenates the semantic label sequence in chronological order;

[0224] The behavior trajectory mapping unit aligns and matches the semantic label sequence with the behavior evolution trajectory template generated based on the long-term interaction record of the user, marks a label node as a semantic derailment point when it is found that the node is obviously inconsistent with the corresponding position in the template in terms of semantic category, semantic direction or semantic role attribute, sends an intention clarification request to the clarification feedback fusion module, and outputs the current semantic label sequence and the candidate derailment segment.

[0225] The clarification feedback fusion module is configured to receive the semantic derailment point information and output a clarification question interface to the user terminal, and collect the feedback statement of the user to the intention clarification request.

[0226] Specifically, the clarification feedback fusion module includes a feedback fusion unit and a semantic transition detection unit:

[0227] The feedback fusion unit combines the user feedback statement with the historical statement pair when the clarification is triggered to form an extended statement pair set;

[0228] The semantic transition detection unit inputs the extended statement pair set into a semantic transition detection structure, constructs a semantic tension spectrum for representing semantic coherence and evolution trend based on label change trend and path directionality, identifies tension rising sections and tension closed clusters in the spectrum, and generates a candidate segment set of multi-path responses according to the multiple reachable paths of the tension closed clusters in the spectrum and sends it to the misleading chain analysis module.

[0229] The misleading chain analysis module is configured to perform risk screening on the candidate segment set from the clarification feedback fusion module.

[0230] Further, the misleading chain analysis module is internally provided with a misleading chain cross-analysis unit and a risk aggregation unit:

[0231] The misleading chain cross-analysis unit extracts a subject-verb-object (SVO) structure from each candidate segment, accesses a misleading SVO structure set stored in a historical negative feedback library, calculates the longest co-occurrence semantic segment of the candidate SVO structure and the misleading SVO structure in the semantic spectrum, and judges whether they are high-frequency coincidences in the same tension direction in combination with the directional information in the semantic tension spectrum; when the co-occurrence segment direction is consistent and the corresponding structure has triggered user correction behavior in history, the candidate segment is marked as a potential risk response unit;

[0232] A risk aggregation unit analyzes fault paths for all potential risk response units, identifies fault paths with large jump spans and missing backbone dependency structures, constructs path influence chains, marks candidate segments containing any fault jump segment of the path influence chains as strong risk response units, and aggregates all risk response units into a shielding content set and outputs the shielding content set to a response screening update module.

[0233] The response screening update module is used to complete final response selection after shielding risk responses and drive system adaptive updates.

[0234] Specifically, the response screening update module includes a response screening subunit and a user feedback update subunit.

[0235] The response screening subunit, after receiving the candidate segment set and the corresponding shielding content set, first removes all segments marked as shielding content from the candidate set; then, constructs a context dependency structure according to the semantic tension closed cluster in which the remaining candidate segments are located, calculates the context break span of each segment, and if a segment has a short reconnection path in the dependency edge connection density graph and the lowest penalty value for missing the backbone structure, the total score of the break span of the segment is recorded as the minimum, and the segment is determined as the final response segment of the current round of conversation, which is sent to the output management module and returned to the user terminal for display.

[0236] The user feedback update subunit starts a short time window after the final response segment is output, collects the first language input behavior of the user within a preset time (for example, 5 seconds), extracts the keyword label therein, and performs semantic pairing judgment with the semantic label at the end of the final response segment to determine whether it constitutes semantic continuation; if it constitutes semantic continuation, the interaction behavior path is reversely mapped to the semantic tension map, the weight value and path priority of the corresponding semantic chain are updated, and the updated map state is written back to the behavior evolution trajectory template, so as to dynamically adjust the intention evolution sequence of the user and realize adaptive optimization of the intention recognition structure and candidate screening mechanism.

[0237] The above only describes certain exemplary embodiments of the present application by way of illustration, and it is needless to say that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present application. Therefore, the above figures and descriptions are illustrative in nature and should not be understood as limiting the scope of protection of the claims of the present application.

Claims

1. An AI artificial intelligence based content generation and interactive conversation method, characterized in that, The method comprises the following steps: S101: Collecting user input text, constructing a semantic label sequence according to at least three rounds of historical sentence pairs in the current session window; mapping the semantic label sequence to an action evolution track template generated based on long-term user interaction records, and identifying a semantic derailment point and triggering an intent clarification request if an inconsistent semantic node appears in the mapping path with the historical track; S102: Obtaining a feedback sentence of the user to the intent clarification request, combining the feedback sentence with the historical sentence pairs, inputting the combined result into a semantic transition detection structure, constructing a semantic tension spectrum for representing semantic coherence and evolution trend, and generating a candidate segment set of multi-path responses based on the tension closed cluster in the semantic tension spectrum; S103: Inputting the candidate segment set of multi-path responses into a misleading chain cross-analysis unit, identifying a high-frequency overlapping response segment formed with a risk trigger main chain according to the semantic path coincidence degree of each candidate segment in the historical negative feedback context, and marking the high-frequency overlapping response segment as shielded content; S104: Selecting the smallest segment as the final response segment from the candidate segments after removing the shielded content according to the context break span of the corresponding semantic tension closed cluster of the candidate segment, and feeding back the user interaction behavior caused by the final response segment to the action evolution track template and the semantic tension spectrum for dynamically updating the intent recognition structure and the candidate screening mechanism; The calculation of the context break span comprises: constructing the syntactic dependency structure of the context for each candidate segment, and generating a dependency edge connection density map between the context and the segment based on the dependency relationship; identifying the dependency edge discontinuity area between the context and the segment in the connection density map; performing dependency path expansion from the discontinuity area to both sides of the context until the dependency chain is reconnected or expanded to the set maximum boundary; recording the shortest reconnection edge distance of each expanded path as the context break span of the corresponding segment; comparing the context break spans of all candidate segments, and selecting the smallest one as the final output segment. 2.The AI artificial intelligence based content generation and interactive dialog method according to claim 1, characterized in that, The determination of the semantic derailment point comprises: determining the position of the last semantic label change in the semantic label sequence as a semantic jump base point, defining the semantic label at the semantic jump base point as a base point label, and sequentially selecting the previous three label nodes as pre-sequence labels in the order of generation of the semantic jump base point along the semantic label sequence, so that the base point label and the previous three pre-sequence labels together constitute a bidirectional label segment group, wherein: the pre-sequence labels are used to represent the local context before semantic jump, and the base point label is used to represent the semantic jump itself; mapping the bidirectional label segment group to the semantic path in the user historical interaction corpus by position alignment, and counting the number of matchable historical paths; if the number of matches is lower than a preset threshold, determining that the bidirectional label segment group constitutes a semantic derailment structure, and determining the last label in the bidirectional label segment group as a semantic derailment point; extracting the lexical semantic path and semantic adjacent node based on the semantic derailment point, and constructing a lexical divergence area; Select three branch nodes with the largest semantic hierarchy span from the word sense divergence area as the intent clarification request items. 3.The AI artificial intelligence based content generation and interactive dialog method according to claim 2, characterized in that, The word sense divergence area construction process comprises: Obtain the direct hypernym, hyponym and lateral diffusion nodes of the target label in the general knowledge graph; Remove the lateral diffusion nodes appearing more than once in the user-confirmed semantic path; Perform shortest path matching on the remaining nodes to filter the paths with semantic discontinuity; Calculate the shortest semantic path between the remaining lateral diffusion nodes and the semantic derailment points based on the knowledge graph, and filter the positions with semantic discontinuity; Take the semantic discontinuity positions as divergence points, and extract the semantic paths extending outward and not appearing before as candidate branch paths; Sort the candidate branch paths based on the semantic jump level, and select the top three terminal nodes as the intent clarification request items. 4.The AI artificial intelligence based content generation and interactive dialog method according to claim 1, characterized in that, The steps of generating multi-path response candidate segments based on the semantic tension graph comprise: Identify the tension rising section on the current semantic path in the semantic tension graph, and construct the heterogeneous semantic segment pair of the paths before and after the corresponding section; Perform structural migration matching on the heterogeneous semantic segment pair in the historical corpus to obtain a path group that has cross mapping in the semantic space but is not continuous in the path structure; Take the semantic jump position in the path group as the center, and extract the previous and next three semantic nodes as a candidate segment window; Determine whether the candidate segment window can be covered by two heterogeneous paths simultaneously, and if so, mark the semantic segment corresponding to the candidate segment window as a neutral tension segment; Based on the closed feature of the neutral tension segment in the semantic tension graph, output the candidate segment set of the multi-path response. 5.The AI artificial intelligence based content generation and interactive dialog method according to claim 1, characterized in that, The generation logic of the shielding content comprises: Identify all candidate SVO structures conforming to the subject-verb-object relationship based on dependency syntax analysis from the current candidate segment, and extract a set of misleading SVO structures marked as misleading, wrong or semantically biased by the user from the historical negative feedback records; Compare the semantic paths of the candidate SVO structure and the misleading structure set, calculate their longest co-occurrence semantic segment based on the semantic graph, and determine whether the semantic tension direction of the longest co-occurrence semantic segment is consistent with the direction in the current semantic tension graph; If the semantic tension direction of the co-occurrence segment is consistent, further retrieve whether the candidate SVO structure has ever triggered user error correction behavior in the historical dialogue; If all the above conditions are met, mark the complete sentence of the candidate SVO structure as a potential risk response unit; Perform semantic aggregation on all potential risk response units to form a shielding content set, and remove the corresponding content from the candidate segment to prevent misleading segments from entering the final response generation process. 6.The AI artificial intelligence based content generation and interactive dialog method according to claim 5, characterized in that, The aggregation of the potential risk response unit comprises: Extract the semantic jump path from subject to object from the SVO structure marked as risk, and calculate the semantic level change amount spanned in the path based on the semantic graph, and record the jump span; Filter the jump paths with a jump span greater than two levels and without any intermediate semantic nodes in the path, and take them as a set of fault paths; performing path backtracking on the jump segment with the highest frequency of occurrence in the fault path set, constructing a path influence chain associated therewith along the semantic graph, for representing the semantic diffusion range possibly triggered by the jump segment; judging whether any fault jump segment in the path influence chain is contained in the candidate response segment, as the identification basis of the potential misleading risk; when the candidate segment contains the fault jump segment, marking the segment as a strong risk response unit and adding it to the shielding content set, for eliminating it from the subsequent response generation process.

7. The AI artificial intelligence-based content generation and interactive dialogue method according to claim 1, wherein the fracture span comparison comprises: constructing a combination of dependency edges of context fracture paths and reconnection paths, and classifying the dependency edges by type; constructing dependency structure groups for each edge type, the dependency structure groups comprising subject-predicate structure groups, verb-object structure groups and bias-positive structure groups, the occurrence frequency of different dependency structures in normal semantic expression and the contribution degree thereof to syntactic coherence; judging whether the fracture region is missing a high-frequency main structure type, and if so, adding a corresponding penalty value to the fracture path; constructing a fracture span total score model according to the fracture edge length and the penalty value, and performing standardization processing on the total score; determining the fracture span with the lowest total score as the smallest semantic fracture segment. 8.The AI artificial intelligence based content generation and interactive dialog method according to claim 7, characterized in that, The user interaction feedback update of the output segment comprises: collecting user input within a short-time monitoring window after the presentation of the output segment, extracting key word labels in the first language input behavior within the short-time monitoring window, and performing semantic pairing between the key word labels and the semantic labels at the end of the output segment to determine whether the user input constitutes semantic continuation; when determining that the user input constitutes semantic continuation, reversely mapping the interaction path corresponding to the user input to the semantic tension graph, and recording the semantic chain in the continuation direction; updating the weight value and path priority of the semantic chain to improve the recommendation probability of the semantic direction in subsequent response generation; synchronously writing the updated semantic tension graph state into the behavior evolution trajectory template to update the evolution sequence structure of the user intent. The system comprises a semantic label construction module, a clarification feedback fusion module, a misleading chain analysis module and a response screening update module, and data transmission is performed between the modules through wired and / or wireless transmission; 9. An AI artificial intelligence based content generation and interactive dialogue system, based on the implementation of the AI artificial intelligence based content generation and interactive dialogue method of any one of claims 1-8, characterized in that, The semantic label construction module collects user input text, constructs a semantic label sequence according to at least three rounds of historical sentences in the current conversation window, maps the semantic label sequence to a behavior evolution trajectory template generated based on long-term user interaction records, and if a semantic node inconsistent with the historical trajectory appears in the mapping path, identifies it as a semantic derailment point and triggers an intent clarification request. The clarification feedback fusion module obtains the feedback sentence of the user to the intent clarification request, combines the feedback sentence with the historical sentence pair, inputs them into a semantic transition detection structure, constructs a semantic tension graph for representing semantic coherence and evolution trend, and generates a candidate segment set of multi-path responses based on the tension closed clusters in the semantic tension graph. ​ The misleading chain analysis module inputs the candidate fragment set of the multi-path response into the misleading chain cross-analysis unit, identifies high-frequency overlapping response fragments formed with the risk trigger main chain according to the semantic path coincidence degree of each candidate fragment in the historical negative feedback context, and marks the high-frequency overlapping response fragments as shielding content; The response screening update module selects the smallest fragment as the final response fragment output from the candidate fragments after removing the shielding content according to the context break span of the corresponding semantic tension closed cluster of the candidate fragment, and feeds back the user interaction behavior triggered by the final response fragment to the behavior evolution trajectory template and the semantic tension atlas respectively, for dynamically updating the intent recognition structure and the candidate screening mechanism.

Images