Guided question answering system and method based on implicit premise triple extraction

Abstract

The application relates to the technical field of natural language processing, and discloses a guided question and answer system and method based on implicit premise triple extraction, which comprises the following steps: receiving user question text and performing implicit premise triple extraction to construct a premise triple set; performing multi-layer conflict detection on the set, including internal mutual consistency verification and external domain knowledge collision detection, labeling internal mutual consistency states and external collision states for each premise triple, and generating a premise verification report by summarizing; generating a counter-question sentence according to the states by using a corresponding guided counter-question generation strategy, verifying a conflict premise correction state according to a user feedback reply text, executing a closed-loop correction according to the state, and outputting a final answer; and the application eliminates logical misdirection caused by a large language model blindly obeying input conditions through structured verification and closed-loop correction of a question premise, and ensures the correctness of a question and answer conclusion in a complex professional scenario.

Description

Technical Field

[0001] This invention relates to the field of natural language processing technology, and more specifically, to a guided question-answering system and method based on implicit premise triple extraction. Background Technology

[0002] With the widespread application of large-scale language modeling technology in professional question-answering systems, ensuring the accuracy and logical rigor of model outputs has become a key research focus in the field. Current question-answering generation schemes typically rely on external knowledge base retrieval enhancement or cue word engineering, attempting to improve response quality by supplementing background information or optimizing reasoning paths.

[0003] In existing patent literature, Chinese patent application CN121118873A discloses a multi-source knowledge conflict detection and repair scheme and system based on semantic graphs. It generates a structured conflict report by combining user queries with multi-source data and dynamically repairs the context using predefined repair operators, primarily addressing the problem of knowledge contradictions between external data from multiple sources. Chinese patent application CN118013051A discloses a question-answering generation method enhanced by a large language model. Its core lies in constructing entity links and knowledge subgraphs, textualizing extracted triples, and then inputting them into the model. The aim is to leverage the structured features of the knowledge graph to guide the model in generating more evidence-supported answers.

[0004] However, existing technical solutions mostly focus on conflict resolution from external knowledge sources or the assistance of knowledge graphs in the reasoning process, neglecting the implicit premise biases that may exist in the user's question text itself. Under the "conditional generation" paradigm of large language models, the model tends to generate the most probable output sequence under the constraints set by the input text, lacking a mechanism to verify the authenticity of the input premises. In professional domain question answering, users often embed incorrect basic rules or logical assumptions in their questions due to cognitive biases. For example, they might incorrectly describe the characteristics of series current in circuit analysis, or mistakenly believe that an incompletely configured service has taken effect in system maintenance. When these implicit erroneous premises are input into the model, the system performs "conditional obedience reasoning," generating a logically consistent but actually erroneous response. This phenomenon not only masks the user's initial cognitive errors but may also lead to subsequent decision-making errors or operational risks due to the misleading conclusions. Existing conflict detection mechanisms, unable to deconstruct and verify the implicit premises of the question from natural language, struggle to prevent such logical misguidance at its source. Summary of the Invention

[0005] To overcome the aforementioned shortcomings of existing technologies, this invention provides a guided question-answering system and method based on implicit premise triple extraction. Through implicit premise triple extraction and a multi-layer conflict detection mechanism, it achieves structured review and guided correction of the logical premises of user questions. The method can identify and intercept erroneous assumptions in the input text that contradict domain knowledge, and through a closed-loop correction process, it returns the question-answering logic to the correct premises, eliminating logical misdirection caused by large language models blindly following input conditions, and ensuring the accuracy of knowledge transfer in complex professional scenarios.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] Guided question answering methods based on implicit premise triple extraction include:

[0008] The system receives user question text, performs implicit premise triple extraction on the user question text to construct a premise triple set, performs multi-level conflict detection on the premise triple set, labels the internal consistency state and external collision state of each premise triple in the premise triple set, and summarizes and generates a premise verification report; the multi-level conflict detection includes internal consistency verification and external domain knowledge collision detection.

[0009] Based on the internal consistency state and external collision state of each premise triple, a corresponding guided question generation strategy is adopted to generate guided questions. User response text is received in response to the guided questions, and the conflict premise correction state of each conflicting premise triple in the premise verification report is verified. Closed-loop correction is performed based on the conflict premise correction state, and the final response based on the correct premise is output.

[0010] The method for extracting implicit premise triples includes:

[0011] The user's question text is embedded into a pre-defined premise extraction instruction template to construct a premise extraction prompt word. The premise extraction prompt word is sent to the large language model to perform implicit premise triple extraction, resulting in a set of premise triples that are bound to the user's question text and consist of a subject field, a relation field, and an object field.

[0012] The internal consistency state can be either internal consistency passed or internal consistency conflicted.

[0013] The method for performing the internal consistency verification includes:

[0014] The number of premise triples in the premise triple set is counted. When the number of premise triples is greater than or equal to 2, an internal consistency check is performed on the premise triple set to mark the internal consistency status of each premise triple in the premise triple set. When the number of premise triples is less than 2, the internal consistency status of the premise triples in the premise triple set is directly marked as internal consistency passed.

[0015] The method for performing the internal consistency verification also includes:

[0016] Each of the premise triples in the premise triple set is converted into an affirmative statement, and all the affirmative statements are concatenated to form a combined statement text. The combined statement text and a preset consistency judgment instruction are sent to the large language model to determine whether there is a logical contradiction between the affirmative statements in the combined statement text.

[0017] When a logical contradiction is determined, the contradictory premise triplet is identified as a contradictory premise triplet pair, and a contradiction explanation is generated. The internal consistency state of the premise triplets involved in the contradictory premise triplet pair is marked as internal consistency conflict. When the large language model determines that there is no logical contradiction, the internal consistency state of all premise triplets in the premise triplet set is marked as internal consistency passed.

[0018] The method for performing the external domain knowledge collision detection includes:

[0019] Determine the corresponding domain knowledge base, traverse each premise triplet in the premise triplet set, retrieve the nearest neighbor verified knowledge triplet that is semantically closest to the premise triplet from the domain knowledge base, perform external domain knowledge collision detection on the pair consisting of the premise triplet and the nearest neighbor verified knowledge triplet, and determine the external collision state.

[0020] The external collision state can be one of the following: external conflict premise, external consistency premise, conditional dependency premise, or no matching premise.

[0021] The retrieval method for the nearest neighbor verified knowledge triplet is as follows:

[0022] Retrieve the current premise triplet from the premise triplet set, and retrieve a subset of candidate verified knowledge triplets from the domain knowledge base using the subject field of the current premise triplet as the search key. Encode the semantic vector of each candidate verified knowledge triplet in the current premise triplet and the subset of candidate verified knowledge triplets, calculate the cosine similarity between the semantic vector of the current premise triplet and the semantic vector of each candidate verified knowledge triplet, and select the candidate verified knowledge triplet with the highest cosine similarity as the nearest neighbor verified knowledge triplet of the current premise triplet.

[0023] The method for performing external domain knowledge collision detection also includes:

[0024] The current premise triplet and the nearest neighbor verified knowledge triplet are transformed into affirmative statements, resulting in two affirmative statements corresponding to the current premise triplet and the nearest neighbor verified knowledge triplet, respectively. The two affirmative statements and a preset semantic relationship judgment instruction are sent to the large language model to determine the semantic relationship between the two affirmative statements. The external collision state of the current premise triplet is determined based on the semantic relationship judgment result. The semantic relationship judgment result is one of semantic consistency, semantic contradiction, and semantic irrelevance.

[0025] The method for determining the external collision state of the current premise triplet includes:

[0026] If the semantic relation judgment result is a semantic contradiction, then the current premise triplet is marked as an external conflicting premise and the conflict pair formed by the current premise triplet and the nearest neighbor verified knowledge triplet is recorded;

[0027] If the semantic relation is determined to be semantically consistent, then the current premise triplet is marked as an externally consistent premise.

[0028] The method for determining the external collision state of the current premise triplet further includes:

[0029] If the semantic relation is determined to be semantically irrelevant, then the verified knowledge triplet of conditional dependency is retrieved from the domain knowledge base. If found, the current premise triplet is marked as a conditional dependency premise and the conditional dependency pairing formed by the current premise triplet and the verified knowledge triplet of conditional dependency is recorded. If not found, the current premise triplet is marked as no matching premise.

[0030] The method for generating the prerequisite verification report includes:

[0031] The internal consistency states and external collision states of each premise triplet are merged to form a set of premise triplets labeled with conflict types.

[0032] Iterate through the internal consistency states and external collision states of all premise triples in the set of premise triples labeled with conflict types. When there is at least one premise triple with an internal consistency state that is internally conflicting, or an external collision state that is an externally conflicting premise, or an external collision state that is a conditional dependency premise, summarize the user question text, the set of premise triples labeled with conflict types, contradictory premise triple pairs and explanations of the contradiction, and conflict pairs and conditional dependency pairings into a premise verification report.

[0033] The method for generating the guided rhetorical question includes:

[0034] Extract all premise triples from the premise verification report whose internal consistency state is internally conflicting or whose external collision state is externally conflicting or conditionally dependent premises as conflicting premise triples. Based on the category of the internal consistency state or external collision state corresponding to each conflicting premise triple, adopt the corresponding guided question generation strategy to generate guided questions for each conflicting premise triple.

[0035] The value of the conflict premise correction state is either corrected or not corrected;

[0036] The method for correcting the conflicting premises of each conflicting premise triplet in the verification premise report includes: re-performing implicit premise triplet extraction, internal consistency verification, and external domain knowledge collision detection on the user's response text to obtain a set of premise triplets labeled with conflicting types in the user's response text; comparing each conflicting premise triplet in the premise verification report with the set of premise triplets labeled with conflicting types in the user's response text, and marking the conflicting premises correction status of each conflicting premise triplet in the premise verification report.

[0037] A guided question-answering system based on implicit premise triple extraction is provided to implement the aforementioned guided question-answering method based on implicit premise triple extraction. The system includes:

[0038] Premise detection module: This module receives user question text, performs implicit premise triple extraction on the user question text to construct a premise triple set, performs multi-level conflict detection on the premise triple set, labels the internal consistency state and external collision state of each premise triple in the premise triple set, and summarizes and generates a premise verification report; the multi-level conflict detection includes internal consistency verification and external domain knowledge collision detection.

[0039] The guidance and correction module generates guided rhetorical questions based on the internal consistency and external collision states of each premise triplet, using corresponding guided rhetorical question generation strategies. It receives user response texts for the guided rhetorical questions and verifies the conflict premise correction states of each conflicting premise triplet in the premise verification report. Based on the conflict premise correction states, it performs closed-loop correction and outputs the final response based on the correct premises.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] This invention establishes a mechanism for verifying the authenticity and logically reviewing input conditions by extracting and structurally representing implicit premises in user questions using triples. This solves the problem of conditional obedience reasoning in large language models when faced with questions containing erroneous premises. The introduction of a multi-layered conflict detection mechanism enables the system to perform dual verification of implicit assumptions from two dimensions: logical consistency and domain knowledge consistency, identifying potential deviations in the question text that contradict objective facts or domain rules. Combined with a guided questioning strategy and a closed-loop correction process, the system proactively guides users to confirm and correct erroneous perceptions in their questions, transforming question-and-answer logic from simple sequence generation to interactive reasoning based on fact verification. This ensures that the final reasoning path and response output are based on verified correct premises, avoiding conclusion deviations caused by false premises and guaranteeing the logical rigor and accuracy of information transmission when the knowledge question-answering system handles professional domain issues. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 A flowchart illustrating the guided question-answering method based on implicit premise triple extraction provided in this embodiment of the invention.

[0044] Figure 2 A schematic diagram of the premise triplet structure provided for an embodiment of the present invention;

[0045] Figure 3 This is a schematic diagram of internal consistency conflict detection provided in an embodiment of the present invention;

[0046] Figure 4 This is a schematic diagram of semantic vector similarity retrieval provided in an embodiment of the present invention;

[0047] Figure 5 This is a schematic diagram of three-way flow determination and final response generation provided in an embodiment of the present invention;

[0048] Figure 6 This is a schematic diagram illustrating the downward shift of knowledge levels provided in an embodiment of the present invention;

[0049] Figure 7 This is a functional block diagram of a guided question-answering system based on implicit premise triple extraction provided in an embodiment of the present invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] Example 1:

[0052] Please see Figure 1 As shown, this embodiment provides a guided question-answering method based on implicit premise triple extraction, including:

[0053] Step S10: Receive user question text, perform implicit premise triple extraction on the user question text to construct a premise triple set, perform multi-level conflict detection on the premise triple set, label the internal consistency state and external collision state of each premise triple in the premise triple set, and summarize to generate a premise verification report; the multi-level conflict detection includes internal consistency verification and external domain knowledge collision detection.

[0054] Step S10 aims to transform the implicit premises diffusely distributed in natural language form in the user's question text into premise triples with clear structural boundaries. A multi-level conflict detection mechanism is then used to verify the truth value of each premise triple, ultimately generating a premise verification report with conflict type annotations. Implicit premises refer to factual assumptions that are not explicitly stated by the user during the questioning process but are embedded as predetermined cognitions in the wording of the statement. These implicit premises constitute the input condition constraints when the large language model generates conditional probability maximization sequences. Premise triples are the basic units used in the knowledge representation domain to structurally describe single factual propositions. Each premise triple consists of three fields: the subject field indicates the object or concept described by the proposition; the relation field indicates the attribute type of the object or concept or the type of association between it and other concepts; and the object field indicates the specific value of the attribute type or the object it points to. Multi-layer conflict detection comprises two levels: internal consistency verification examines the logical consistency among the premise triples within the premise triple set, identifying contradictions among multiple implicit premises given by the user in the same user question text; external domain knowledge collision detection determines the semantic relationship between each premise triple and verified knowledge triples in the domain knowledge base, identifying whether the user's implicit premises contradict the actual knowledge of the corresponding professional domain in the domain knowledge base or whether necessary conditions are not met. The premise verification report is a structured summary of all the above detection results, including the user's original user question text, each premise triple in the premise triple set and its conflict type label, as well as the specific conflict pair or conditional dependency pair and explanation of the reason for each detected conflict. This premise verification report serves as input to the subsequent guided correction process in step S20, providing a structured data foundation for the generation of targeted rhetorical questions and the verification of user responses.

[0055] Further, step S10 includes:

[0056] Step S11: Receive user question text, embed the user question text into a preset premise extraction instruction template to construct premise extraction prompt words, send the premise extraction prompt words to the large language model to perform implicit premise triple extraction, and obtain a set of premise triples bound to the user question text and consisting of subject field, relation field and object field.

[0057] Step S11 involves constructing a premise extraction instruction template and calling the large language model to transform unstructured natural language text into a structured set of premise triples. When the intelligent question-answering system receives the user's input question text, it does not directly send the text to the large language model to generate a response. Instead, it first enters a premise extraction preprocessing stage. In this stage, the system reads a pre-set premise extraction instruction template stored in the system configuration file. This template consists of two parts: a placeholder area for embedding the user's question text and a task instruction requiring the large language model to identify and output all implicit factual premises in the question text. The system fills the placeholder area of ​​the premise extraction instruction template with the complete content of the user's question text, merging the previously separate task instruction with the text to be processed into a complete set of premise extraction prompts. The task instruction in the premise extraction instruction template imposes constraints on the output format of the large language model, requiring it to format each extracted implicit premise into a premise triple, see [link to relevant documentation]. Figure 2 This is a schematic diagram of the prerequisite triplet structure provided in the embodiments of this application. Figure 2 The document illustrates the standardized structure of premise triples, which consists of three core fields: a subject field, a relation field, and an object field, linked sequentially. An example structure is provided, using "circuit type" as the subject field, "belongs to" as the relation field, and "series circuit" as the object field. This visually demonstrates the complete composition of a single premise triple. Each premise triple consists of three fields: subject, relation, and object, separated by commas. Different premise triples are separated by semicolons, allowing the system to directly extract the content of each field through string parsing after receiving the output from the large language model. The system sends premise extraction prompts to the large language model via an application programming interface (API). Based on its semantic understanding capabilities learned from massive text corpora during pre-training, the large language model performs semantic parsing on the user's question text, identifying all implicit factual premises and outputting each implicit premise as a premise triple. The system receives the text returned by the large language model, extracts the subject, relation, and object fields of each premise triple through string parsing, and forms a set of premise triples bound to the current user's question text. This set of premise triples is stored in the system memory for subsequent steps to call.

[0058] For example, when a user asks the question "In a series circuit, the current through each resistor is different, so how should the total current be calculated?", the system embeds the user's question text into a premise extraction instruction template, forming premise extraction prompts, and sends these prompts to the large language model. The large language model identifies two implicit premises from the user's question text and outputs them as premise triples. The first premise triple is... Figure 2 The structural example shown corresponds perfectly. The subject field is "circuit type", the relation field is "belongs to", and the object field is "series circuit". This premise triple expresses the factual premise that the user implicitly assumes in the question that the circuit type being discussed is a series circuit. The second premise triple has the subject field "current of each resistor in the series circuit", the relation field is "attribute is", and the object field is "different". This premise triple expresses the factual premise that the user implicitly assumes in the question that the current values ​​through each resistor in the series circuit are different. For another example, when the user's question text is "My server has load balancing enabled, why is the single-node CPU utilization still 100%?", the large language model identifies two implicit premises from the user's question text and outputs them in the form of premise triples. The first premise triple has the subject field "server", the relation field is "configured", and the object field is "load balancing". This premise triple expresses the user's implicit assumption in the question that their server has completed the load balancing configuration. The second premise triple has "load balancing" as its subject, "status" as its relation, and "effectively running" as its object. This premise triple implicitly assumes that the load balancing function is currently effectively running. The system collects these premise triples to form a premise triple set, providing structured input for subsequent internal consistency checks and external domain knowledge collision detection.

[0059] Step S11 uses premise triples as a structured representation of implicit premises instead of directly identifying premises at the natural language level because premise triples have clear structural boundaries and definite semantic directions. In natural language, the same concept can have diverse wording and ambiguous references. For example, "the current is different" and "the current value is different at different places" express the same meaning but have significant literal differences; "load balancing is enabled" and "load balancing is configured" are semantically similar but have different syntactic structures. If semantic comparison is performed directly at the natural language level, the system needs to handle a large number of synonym substitutions, sentence transformations, and reference resolution issues, making it difficult to achieve accurate consistency determination. After transforming the diffusely distributed premise information in natural language into premise triples composed of three fields—subject, relation, and object—each premise becomes an independent proposition with a clear structure, which can be directly aligned and compared field-by-field with verified knowledge triples stored in the domain knowledge base using the same triple structure. Alignment of the subject field ensures that the comparison is of the same concept or object, alignment of the relation field ensures that the comparison is of the same attribute type or association type, and alignment of the object field focuses on whether the attribute values ​​or associations are consistent. This structured alignment method reduces the complex semantic comparison problem to a field-level matching problem, enabling subsequent external domain knowledge collision detection to achieve high-accuracy consistency determination with low computational overhead. Step S11 reuses the natural language semantic parsing capabilities already possessed by the large language model to complete the conversion from unstructured text to structured premise triples, without the need to train an additional dedicated premise extraction model, reducing the system's development cost and deployment complexity. At the same time, it utilizes the generalized understanding of different domain terms and expressions learned by the large language model during pre-training on massive corpora, making the premise extraction process more adaptable to user questions from different professional domains.

[0060] Step S12: Count the number of premise triples in the premise triple set. If the number of premise triples is greater than or equal to 2, perform an internal consistency check on the premise triple set to mark the internal consistency status of each premise triple in the premise triple set. If the number of premise triples is less than 2, directly mark the internal consistency status of the premise triples in the premise triple set as internal consistency passed.

[0061] The method for performing internal consistency verification includes: converting all premise triples in the premise triple set into affirmative statements one by one according to a preset statement transformation template; concatenating all affirmative statements to form a combined statement text; sending the combined statement text and a preset consistency judgment instruction to a large language model, which then determines whether there is a logical contradiction between the affirmative statements in the combined statement text; when the large language model determines that there is a logical contradiction, identifying the contradictory premise triples as contradictory premise triple pairs and generating a contradiction explanation; marking the internal consistency status of the premise triples involved in the contradictory premise triple pairs as internal consistency conflict; and recording the contradictory premise triple pairs and the contradiction explanation; when the large language model determines that there is no logical contradiction, marking the internal consistency status of all premise triples in the premise triple set as internal consistency passed; wherein the value of the internal consistency status is either internal consistency passed or internal consistency conflict.

[0062] Step S12 determines whether to perform an internal consistency check based on the number of premise triples. The system reads the set of premise triples obtained in step S11 and counts the number of premise triples contained in the set. The system compares this number of premise triples with the value 2 to determine if the number is greater than or equal to 2. If the number of premise triples is greater than or equal to 2, it indicates that there are multiple implicit premises in the user's question text. These premises may be logically contradictory but the user is unaware of them. The system needs to perform an internal consistency check to detect such contradictions. When the number of premise triples is less than two (i.e., the premise triple set contains only one premise triple or none), since internal consistency verification is essentially a logical consistency check within the premise triple set, its smallest operational unit is a contradiction determination between a pair of premise triples. When there is only one or zero premise triples, there is no second premise triple for comparison. Performing internal consistency verification would consume one inference call of the large language model without producing any valid detection results. Therefore, the system skips the internal consistency verification step, directly marking the internal consistency status of the unique premise triple as passed, and allowing the process to directly proceed to the subsequent external domain knowledge collision detection step. Using the quantity judgment condition as a pre-gating of internal consistency verification avoids unnecessary consumption of computational resources, ensuring that the system only initiates the logical consistency check process within the premise triple set when multiple premise triples exist.

[0063] The specific execution process of the internal consistency check is as follows. The system converts all premise triples in the premise triple set into affirmative statements one by one according to a preset statement conversion template. The statement conversion template is a pre-defined text template whose structure involves filling the subject field of the premise triple into the subject position, converting the relation field into a predicate verb or copula structure, and filling the object field into the object or predicate position, thus transforming the premise triples, originally stored as fields, into complete statements conforming to natural language grammar rules. For example, the premise triple (figure, belongs to, triangle) is converted into the affirmative statement "figure belongs to triangle" using the statement conversion template, and the premise triple (angle A, degree is, 30 degrees) is converted into the affirmative statement "angle A has a degree of 30 degrees". The system concatenates all the converted affirmative statements into a combined statement text according to the storage order of the premise triples in the premise triple set, separating each affirmative statement with a period. This combined statement text, along with the preset consistency judgment instruction, is sent to the large language model. The consistency judgment instruction is a pre-set task instruction text stored in the system configuration file. This instruction requires the large language model to read each affirmative statement in the combined statement text and determine whether these statements logically contradict each other. If a contradiction exists, it outputs the specific statement pair causing the contradiction and an explanation of the contradiction; otherwise, it outputs a consistency verification success flag. Based on its logical reasoning ability learned during the pre-training phase, the large language model analyzes the logical relationships between the affirmative statements in the combined statement text to determine whether there is a contradictory situation where the validity of one statement necessarily leads to the invalidity of another.

[0064] For example, when a user asks the question "In triangle ABC, angle A equals 30 degrees, angle B equals 60 degrees, and angle C equals 100 degrees, find the ratio of the side lengths", see [reference needed]. Figure 3 This is a schematic diagram of internal consistency conflict detection provided in an embodiment of this application. Figure 3The diagram illustrates the triangle ABC shape corresponding to this example, with its three interior angles labeled as 30°, 60°, and 100° respectively. It also fully presents the complete chain of angle value summation calculation process, triangle interior angle sum constraint conditions, internal consistency conflict judgment logic, and explanation of the contradiction reasons. The set of premise triples obtained in step S11 contains four premise triples: the first premise triple is (shape, belongs to, triangle), the second premise triple is (angle A, degree is, 30 degrees), the third premise triple is (angle B, degree is, 60 degrees), and the fourth premise triple is (angle C, degree is, 100 degrees). These four premise triples are converted into affirmative statements using the statement conversion template and then concatenated into a combined statement text: "The figure belongs to a triangle. Angle A measures 30 degrees. Angle B measures 60 degrees. Angle C measures 100 degrees." The system sends this combined statement text and the consistency judgment instruction to the large language model. During the analysis, the large language model recognizes that the first affirmative statement, "The figure belongs to a triangle," implicitly contains the geometric constraint of a triangle, namely, the sum of the interior angles of a triangle equals 180 degrees. Furthermore, the sum of the angle values ​​of 30 degrees, 60 degrees, and 100 degrees given in the subsequent three affirmative statements is 190 degrees. This calculation process is consistent with... Figure 3 The summation process of the angle values ​​30° + 60° + 100° = 190° shown in the image corresponds exactly, and the calculation result exceeds... Figure 3 The constraint that the sum of the interior angles of a triangle must equal 180° contradicts the geometric constraint that the sum of the interior angles of a triangle must equal 180 degrees. The large language model outputs contradictory premise triplet pairs as combinations of the first premise triplet and the second through fourth premise triplets, with the explanation that "the sum of the three angles is 190 degrees, exceeding the constraint that the sum of the interior angles of a triangle is 180 degrees." This explanation contradicts... Figure 3 The final output of the contradiction explanation is a perfect match. The system receives the output of the large language model, marks the internal consistency state of all premise triples involved in the contradictory premise triple pair as internal consistency conflict, and records the contradictory premise triple pair and contradiction explanation in the system memory for use in the subsequent step S14 when summarizing the premise verification report. When the large language model determines that there is no logical contradiction in the combined statement text, the system marks the internal consistency state of all premise triples in the premise triple set as internal consistency passed, and does not generate a record of internal consistency conflict.

[0065] Internal consistency verification, by converting structured premise triples into natural language statements and then using a large language model to judge logical consistency, can detect self-contradictions among multiple implicit premises given by a user in the same user question text. Such self-contradictions often arise from the user's misunderstanding of the question's context or from the user confusing conditions in different scenarios, and are usually difficult for the user to detect. If the intelligent question-answering system does not detect such internal contradictions, the large language model, upon receiving a user question containing contradictory premises, may reason based on only one premise while ignoring the contradictory premises. This results in a response that, while seemingly reasonable in a local sense, is inconsistent with the other premises given by the user, causing confusion or errors when the user subsequently applies the response. Internal consistency verification works synergistically with premise triple extraction in step S11. The structured representation of premise triples allows internal consistency verification to accurately locate the specific premise triple pairs that cause contradictions and record the explanation of the contradiction, rather than simply providing a vague judgment of "contradiction exists." This precise positioning capability enables the subsequent step S21 to provide targeted guidance on the specific location and cause of the contradiction when generating guided rhetorical questions, guiding the user to re-examine whether the given conditions are consistent.

[0066] Step S13: Determine the corresponding domain knowledge base based on the domain involved in the user's question text; traverse each premise triplet in the premise triplet set; retrieve the nearest neighbor verified knowledge triplet that is semantically closest to the premise triplet from the domain knowledge base; perform external domain knowledge collision detection on the pair consisting of the premise triplet and the nearest neighbor verified knowledge triplet to determine the external collision state; merge the internal consistency state and the external collision state of each premise triplet to form a premise triplet set labeled with the conflict type; the value of the external collision state is one of the following: external conflict premise, external consistent premise, conditional dependency premise, and no matching premise.

[0067] The method for retrieving nearest neighbor verified knowledge triples is as follows: See [link / reference] Figure 4 This is a schematic diagram of semantic vector similarity retrieval provided in an embodiment of this application. Figure 4 The diagram illustrates the coordinate system structure of the semantic vector space, as well as the distribution relationship and retrieval matching logic of the premise triplet semantic vector, candidate verified knowledge triplets, and nearest neighbor verified knowledge triplets in the semantic vector space. The current premise triplet is retrieved from the premise triplet set, and a subset of candidate verified knowledge triplets is retrieved from the domain knowledge base using the subject field of the current premise triplet as the retrieval key. Semantic vector encoding is performed on each candidate verified knowledge triplet in the current premise triplet and the subset of candidate verified knowledge triplets. This encoding process corresponds to… Figure 4The process involves generating semantic vectors for premise triples and corresponding vectors for each candidate verified knowledge triple in the semantic vector space. It also involves calculating the cosine similarity between the semantic vector of the current premise triple and the semantic vector of each candidate verified knowledge triple, and selecting the candidate verified knowledge triple with the highest cosine similarity as the nearest neighbor verified knowledge triple of the current premise triple. This selection process is similar to... Figure 4 The logic of matching the nearest neighbor verified knowledge triplet from the candidate verified knowledge triplets to the one whose semantic vector direction is closest to that of the premise triplet is completely consistent.

[0068] The method for performing external domain knowledge collision detection is as follows: the current premise triple and the nearest neighbor verified knowledge triple are respectively converted into affirmative statements according to the statement transformation template defined in step S12, resulting in two affirmative statements corresponding to the current premise triple and the nearest neighbor verified knowledge triple respectively; the two affirmative statements and the preset semantic relationship judgment instruction are sent to the large language model, which judges the semantic relationship between the two affirmative statements, and determines the external collision state of the current premise triple based on the semantic relationship judgment result; the semantic relationship judgment result is one of semantic consistency, semantic contradiction, and semantic irrelevance;

[0069] The external collision state of the current premise triple is determined based on the semantic relation judgment result: if the semantic relation judgment result is semantic contradiction, the current premise triple is marked as an externally conflicting premise and the conflict pair formed by the current premise triple and the nearest neighbor verified knowledge triple is recorded; if the semantic relation judgment result is semantic consistency, the current premise triple is marked as an externally consistent premise; if the semantic relation judgment result is semantic irrelevance, the verified knowledge triple of conditional dependency is searched in the domain knowledge base. If found, the current premise triple is marked as a conditionally dependent premise and the conditional dependency pair formed by the current premise triple and the verified knowledge triple of conditional dependency is recorded; if not found, the current premise triple is marked as no matching premise.

[0070] Step S13 performs external domain knowledge collision detection on each premise triple in the premise triple set. This is done by semantically comparing the premise triple with verified knowledge triples in the domain knowledge base to determine whether the user's implicit premises contradict the actual knowledge of the professional domain involved in the user's question text. The system first determines a corresponding domain knowledge base based on the domain involved in the user's question text. The domain knowledge base is a pre-built structured knowledge storage unit stored in the system database. Each domain knowledge base corresponds to a specific professional domain and stores verified knowledge triples within that domain that have been verified by authoritative sources. Each verified knowledge triple consists of three fields: a subject field, a relation field, and an object field, using the same structural specifications as the premise triples. In the domain knowledge base, a hierarchical relationship is pre-established between verified knowledge triples. This hierarchical relationship describes the superior-inferior relationship between verified knowledge triples. The knowledge expressed by the superior verified knowledge triple is derived or supported by the more fundamental concepts or rules expressed by the inferior verified knowledge triple. The knowledge expressed by the inferior verified knowledge triple is at a more atomized level in the domain knowledge system. The hierarchical relationship is used in step S23 to perform a hierarchical shift operation on the knowledge anchor points of the guided rhetorical questions. The domain can be determined by keyword matching based on professional terms appearing in the user's question text, or by the user's conversational context in the intelligent question-answering system or by the user's pre-defined domain preferences. The system sequentially extracts each premise triple from the premise triple set. For the currently extracted premise triple, the subject field of the premise triple is used as the search key to search the domain knowledge base for all verified knowledge triples whose subject fields belong to the same conceptual category as the subject field of the premise triple, forming a subset of candidate verified knowledge triples. The determination of concept categories adopts a combination of lexical similarity matching and hypernym-hyponym relationship matching. For example, when the subject field of the current of the current proposal triple is "current of each resistor in a series circuit", all verified knowledge triples in the domain knowledge base with the subject field "current in a series circuit" or "current in a series circuit" are included in the subset of candidate verified knowledge triples, because "current of each resistor in a series circuit" and "current in a series circuit" belong to the same concept category, and the former is a specific expression of the latter.

[0071] Each candidate verified knowledge triplet from the currently extracted premise triplet and the subset of candidate verified knowledge triplets is input into a semantic encoder. The semantic encoder is a pre-trained text embedding model whose function is to transform the input triplet text into a fixed-dimensional semantic vector. The position of this semantic vector in a high-dimensional vector space reflects the features of the semantic content expressed by the triplet, and... Figure 4The design logic of representing semantic features by the spatial position of each vector in the semantic vector space is consistent. The subject, relation, and object fields of each triple are concatenated sequentially into a text segment, which is then input into the semantic encoder to obtain the semantic vector corresponding to that triple. The cosine similarity between the semantic vector of the currently retrieved premise triple and the semantic vector of each candidate verified knowledge triple is calculated. Cosine similarity is a measure of how close two vectors are in direction. It is calculated by dividing the dot product of the two vectors by the product of their magnitudes. The value ranges from -1 to 1; a value closer to 1 indicates that the two vectors are closer in direction, a value closer to -1 indicates that the two vectors are more opposite in direction, and a value close to 0 indicates that the two vectors have no significant directional correlation. The candidate verified knowledge triple with the highest cosine similarity value is selected as the comparison object with the current premise triple. This comparison object is recorded as the nearest neighbor verified knowledge triple corresponding to the premise triple. This result is compared with... Figure 4 The nearest neighbor in the final location has been verified to be a perfect match of the knowledge triple.

[0072] The use of semantic vector cosine similarity for nearest-neighbor verified knowledge triplet retrieval is because users' erroneous premises are usually not completely random fabrications, but rather partial misunderstandings or directional reversals of correct knowledge. For example, the erroneous premise "the current in a series circuit is different at each stage" and the correct knowledge "the current in a series circuit is equal everywhere" both involve the core concepts of "series circuit" and "current." These two statements have a high cosine similarity in the semantic vector space, a characteristic that... Figure 4 In the semantic vector space, the semantic vector directions of erroneous premises and their corresponding correct knowledge are highly similar because their subject fields and the physical quantities involved are the same; the only difference is the opposite direction of the statement on the current characteristics. Through semantic vector cosine similarity retrieval, the system can find the most relevant verified knowledge triples to the erroneous premise semantics from the domain knowledge base with high recall, avoiding the omission of correct knowledge directly related to the user's erroneous premise. If the semantic vector retrieval stage is skipped and each premise triple is directly compared with all verified knowledge triples in the domain knowledge base, the number of large language model inference calls required will be proportional to the size of the domain knowledge base. When the domain knowledge base contains hundreds of thousands of verified knowledge triples, the computational cost becomes unacceptable, and the system's response latency will severely impact the user experience. Semantic vector retrieval, as an efficient approximate nearest neighbor search method, has a time complexity that is logarithmically related to the size of the domain knowledge base rather than nonlinearly, allowing external domain knowledge collision detection to be completed within a controllable time overhead.

[0073] The system performs semantic relation judgment on pairs consisting of the current premise triple and its corresponding nearest neighbor verified knowledge triple. The current premise triple and its corresponding nearest neighbor verified knowledge triple are each converted into two affirmative statements using the same statement transformation template from step S12. These two affirmative statements, along with a preset semantic relation judgment instruction, are sent to the large language model. The semantic relation judgment instruction is a pre-set task instruction text stored in the system configuration file. The instruction requires the large language model to determine whether the semantic relation between the two affirmative statements falls into one of three categories: semantic consistency, semantic contradiction, or semantic irrelevance. Semantic consistency means that the factual content expressed by the two affirmative statements is the same in direction or mutually implied; if one statement is true, the other statement is also true or can be derived from it. Semantic contradiction means that the factual content expressed by the two affirmative statements is opposite in direction; if one statement is true, the other statement is necessarily false. Semantic irrelevance means that the factual content expressed by the two affirmative statements is not directly related semantically; the truth or falsehood of one statement has no effect on the truth or falsehood of the other statement. Based on its ability to understand semantic relationships learned from massive text corpora during the pre-training stage, the large language model analyzes two affirmative statements and outputs the judgment result of semantic relationship.

[0074] The external collision state of the current premise triplet is determined based on the semantic relationship judgment result. If the large language model determines that the semantic relationship between two affirmative statements is semantically contradictory, the system marks the external collision state of the current premise triplet as an external conflict premise, and records the pair consisting of the current premise triplet and its corresponding nearest neighbor verified knowledge triplet as a conflict pair. For example, the current premise triplet is (the current of each resistor in a series circuit, the attribute is, not the same), and the nearest neighbor verified knowledge triplet retrieved from the domain knowledge base is (the current in a series circuit, the attribute is, equal everywhere). The large language model performs semantic relationship judgment on the affirmative statements "the current of each resistor in a series circuit is not the same" and "the current in a series circuit is equal everywhere" transformed from these two premise triplets, determines that the two are semantically contradictory, marks the premise triplet as an external conflict premise, and records the conflict pair. If the large language model determines that the semantic relationship between two affirmative statements is semantically consistent, the system marks the external collision state of the current premise triple as an externally consistent premise, indicating that the factual statement expressed by the premise triple is consistent with the verified knowledge triple in the domain knowledge base and there is no factual error.

[0075] If the large language model determines that the semantic relationship between two affirmative statements is semantically unrelated, the system enters the conditional dependency detection branch. The semantic unrelatedness determination indicates that there is no direct semantic implication or contradiction between the content stated by the current premise triple and the content stated by the nearest neighbor verified knowledge triple. However, this does not mean that the current premise triple is necessarily correct; there may be cases where necessary conditions are not met. The system uses the concatenated text of the subject and object fields of the current premise triple as the search key to search the domain knowledge base for a verified knowledge triple whose subject field is the same as or semantically equivalent to the concatenated text of the subject and object fields of the current premise triple, and whose relation field expresses conditional dependency semantics. Conditional dependency semantics means that the verified knowledge triple describes the necessary conditions for a certain state or attribute to be true. Typical values ​​for its relation field include "the necessary condition for effectiveness is", "the premise for success is", and "must be satisfied". If such a verified knowledge triple is retrieved, the system records the verified knowledge triple as a conditionally dependent verified knowledge triple, marks the external collision state of the current premise triple as a conditionally dependent premise, and records the pairing of the current premise triple and the conditionally dependent verified knowledge triple as a conditionally dependent pairing. For example, if the current premise triple is (load balancing, state is, effectively running), and the domain knowledge base contains a conditionally dependent verified knowledge triple (load balancing effectively running, necessary condition is, number of backend nodes is greater than or equal to 2), the system marks this premise triple as a conditionally dependent premise and records the conditionally dependent pairing. A conditionally dependent premise means that the fact stated by the user can be true when a specific condition is met, but the user has not declared whether the condition has been met. The system cannot definitively determine its correctness or error based solely on the user's question text; it needs to further determine this by asking the user in subsequent step S21 whether the necessary condition is met. If no verified knowledge triples of this conditional dependency are found, the system marks the external collision state of the current premise triple as no matching premise, indicating that there are no verified knowledge triples directly related to the premise triple in the domain knowledge base. The premise triple can neither be confirmed nor denied, and the system does not mark it as conflicting.

[0076] After performing the aforementioned external domain knowledge collision detection on each of the premise triplet sets in the system, the system merges the internal consistency verification result obtained in step S12 with the external domain knowledge collision detection result obtained in this step to generate a premise triplet set labeled with conflict types. Each premise triplet in this set of premise triplet sets labeled with conflict types carries two types of annotation information. The first type is the internal consistency status, which takes the values ​​of internal consistency passed or internal consistency conflict. If it is an internal consistency conflict, it also carries the corresponding contradictory premise triplet pair and an explanation of the contradiction reason. The second type is the external collision status, which takes the values ​​of external conflicting premise, external consistent premise, conditional dependency premise, or no matching premise. If it is an external conflicting premise, it also carries the corresponding conflict pair. If it is a conditional dependency premise, it also carries the corresponding conditional dependency pair.

[0077] External domain knowledge collision detection distinguishes between two different types of conflict: externally conflicting premises and conditionally dependent premises. This is because the nature of these two types of premise errors is fundamentally different. Externally conflicting premises mean that the facts stated by the user directly contradict the verified knowledge triples, and can be definitively identified as incorrect premises. In subsequent step S21, the system can directly generate guiding questions to guide the user to correct these incorrect premises. Conditionally dependent premises mean that the facts stated by the user are valid only when the necessary conditions they depend on are met. The system cannot definitively determine whether these premises are correct or incorrect and needs to obtain more information by asking the user whether the necessary conditions are met before making a judgment. If both types of premise errors are treated as definitive errors, the system might directly give a misjudgment of "your premise is incorrect" for conditionally dependent premises. When the user has actually met the necessary conditions, this misjudgment will reduce the user's trust in the system. The two conflict types are distinguished and marked so that the subsequent step S21 can adopt different guided question generation strategies for different conflict types. For external conflict premises, a question is generated to guide the user to examine the erroneous premises from the underlying principles. For conditional dependency premises, a question is generated to guide the user to confirm whether the necessary conditions are met.

[0078] Step S13 and step S11's premise triple extraction work synergistically. The structured representation of premise triples enables external domain knowledge collision detection to efficiently locate the most relevant verified knowledge triples to the user's premises using a two-stage retrieval strategy of subject field retrieval plus semantic vector cosine similarity calculation, avoiding the computational overhead of traversing the entire database. The three-field structure of premise triples also allows the system to retrieve conditionally dependent verified knowledge triples using the concatenated text of the subject and object fields as the search key, achieving accurate identification of conditionally dependent premises. Steps S13 and S12's internal consistency verification complement each other. Internal consistency verification detects logical contradictions between premise triples within the premise triple set, while external domain knowledge collision detection detects contradictions or conditional dependencies between each premise triple and verified knowledge in the domain knowledge base. Both cover two sources of potential errors in the user's implicit premises: the former addresses the inherent inconsistency in the user's own statements, and the latter addresses the deviation between the user's statements and objective knowledge. The combination of two detection mechanisms enables the system to perform comprehensive correctness verification of user implicit premises, reducing the probability of missing incorrect premises.

[0079] Step S14: Traverse the internal consistency states and external collision states of all premise triples in the set of premise triples labeled with conflict types. When there is at least one premise triple with an internal consistency state that is internally conflicting, or an external collision state that is an externally conflicting premise, or an external collision state that is a conditionally dependent premise, summarize the user query text, the set of premise triples labeled with conflict types, contradictory premise triple pairs and explanations of contradictions, conflict pairs and conditionally dependent pairings into a premise verification report and proceed to the guided correction process in step S20. When the internal consistency states of all premise triples are internally consistent and all external collision states are externally consistent premises or no matching premises, send the user query text directly to the large language model to generate a response.

[0080] Step S14 involves traversing the conflict type annotations of all premise triples in the set of premise triples labeled with conflict types to determine if there are any conflict situations requiring the initiation of a guided correction process. The system reads the set of premise triples labeled with conflict types obtained in step S13 and sequentially checks the internal consistency state and external collision state of each premise triple in the set. The system determines whether at least one premise triple satisfies one of the following three conditions: internal consistency state is internally contradictory, external collision state is an externally contradictory premise, or external collision state is a conditionally dependent premise. These three conditions correspond to three types of problems with the user's implicit premises: internal consistency conflict indicates that the multiple premises given by the user are logically self-contradictory; externally contradictory premises indicate that a premise given by the user directly contradicts verified knowledge in the domain; and conditionally dependent premises indicate that the validity of a premise given by the user depends on necessary conditions that the user has not declared are satisfied.

[0081] If at least one premise triple carries any of the aforementioned conflict type markers, the process proceeds to the premise verification report summary stage. The user's original user question text, the set of premise triples marked with conflict types, the contradictory premise triple pairs and their explanations recorded in step S12, and the conflict pairs and conditional dependency pairs recorded in step S13 are summarized into a premise verification report. The premise verification report is a structured data object, and its content organization facilitates direct reading and processing by the guided correction process in step S20: the user's original user question text serves as contextual information for reference when generating guided rhetorical questions; the set of premise triples marked with conflict types provides all extracted implicit premises and their conflict type markers; the contradictory premise triple pairs and their explanations are used when generating guided rhetorical questions for internal consistency conflicts; the conflict pairs are used when generating guided rhetorical questions for external conflict premises; and the conditional dependency pairs are used when generating guided rhetorical questions that inquire whether necessary conditions are met for conditional dependency premises. The system stores this premise verification report in system memory and jumps the process to the guided correction process entry point in step S20.

[0082] If all premise triples in the set of premise triples labeled with conflict types do not exhibit any of the three conflict scenarios mentioned above—that is, all premise triples have passed internal consistency and all have either consistent external premises or no matching premises—it is determined that the implicit premises in the user's query text do not contain any contradictions or conditional dependencies with the domain knowledge base, and the user's implicit premises are correct within the currently verifiable range. At this point, the user's query text is directly sent to the large language model, which generates a response based on the query text and returns the response to the user through the user interface. The process ends normally here, without proceeding to the guided correction stage.

[0083] Step S14 sets a conflict existence check as a gate condition for entering the guided correction process because not all user questions contain incorrect premises. If guided questions are applied to all questions regardless of their premises, it will have two negative impacts. First, for questions with correct premises, users expect a direct answer rather than being questioned. If the system still asks the user "Are you sure your premise is correct?" even when the premise is correct, the user experience will significantly deteriorate, and the user may feel that the system is deliberately making things difficult or that the system is incapable of answering directly. Second, each guided question requires waiting for the user's response before starting a new round of premise extraction and conflict detection, increasing the number of interaction rounds and waiting time required for the user to obtain a final answer. Applying this process to questions with correct premises would incur unnecessary time overhead. By setting a conflict existence check at the end of step S10, the system only initiates the guided correction process when a premise conflict is detected. When the premise is correct, it directly returns a response through a fast response path, achieving a balance between correction capability and response efficiency.

[0084] Step S10 serves as a pre-detection step for the entire guided question answering method based on implicit premise triple extraction. It transforms the implicit premises that are originally scattered in the natural language question text into a set of premise triples with clear structural boundaries and definite semantic orientation. Through a two-level detection mechanism of internal consistency verification and external domain knowledge collision detection, each premise triple is labeled with conflict type. Finally, a premise verification report is generated as the input for the subsequent guided correction process. The premise verification report output in step S10 includes each conflicting premise triplet and its corresponding conflict pair or conditional dependency pair, as well as an explanation of the reason for the contradiction. This structured information enables step S21 to adopt different guided question generation strategies based on different conflict types, rather than using a general question template. The premise triplet structure specification used in step S10 is consistent with the structure specification of verified knowledge triplets stored in the domain knowledge base, allowing step S22 to reuse the extraction and detection process of steps S11 to S13 when verifying the premise correction status in the user's response text, without the need to develop another set of verification logic. The gating judgment in step S14 that distinguishes between conflicting and non-conflicting situations enables the system to respond quickly to questions with correct premises and to perform targeted correction for questions with problematic premises, avoiding efficiency losses or correction omissions caused by a one-size-fits-all approach.

[0085] Step S10 expands the processing flow of the large language model after receiving a user question from a single response generation to a three-stage process of "premise extraction, multi-layer conflict detection, and triage decision-making." In existing intelligent question-answering systems, the large language model directly generates a sequence that maximizes conditional probability after receiving a user question, treating all information in the user question (including implicit premises) as predetermined input conditions without questioning their correctness. This processing mode can efficiently generate accurate responses when the user premises are correct, but when the user premises contain factual errors, the large language model will reason based on erroneous premises, producing misleading responses that appear logically consistent under erroneous premises but whose conclusions contradict real-world domain knowledge. Step S10 inserts a premise verification layer between the user question and response generation. This premise verification layer does not change the large language model's own generation mechanism, but checks whether there are conflicts in the implicit premises in the user question before calling the large language model to generate the response. If a conflict is detected, the process jumps to a guided correction branch instead of directly generating a response, thus avoiding the situation where the large language model reasones based on incorrect premises. If no conflict is detected, the process enters the direct response branch, where the large language model generates a response based on the validated user question. This "validate before generating" process design enables the intelligent question-answering system to distinguish between "questions with correct premises" and "questions with problematic premises" and adopt different processing strategies accordingly. This makes the final response presented to the user more probabilistically inclined to be a conclusion derived from correct premises, reducing the risk of misleading responses being delivered to the user and causing cognitive bias or operational errors.

[0086] Step S20: Based on the internal consistency state and external collision state of each premise triplet, a corresponding guided question generation strategy is used to generate guided questions. The user's response text to the guided questions is received, and the conflict premise correction state of each conflicting premise triplet in the premise verification report is verified. Closed-loop correction is performed based on the conflict premise correction state, and the final response based on the correct premise is output.

[0087] Specifically, step S20 transforms the detected conflict information into user-oriented guided interactive behavior. By constructing guided rhetorical questions, it prompts users to independently examine and correct their erroneous premises. After the user responds, the premise correction status is verified to determine the subsequent processing path. Guided rhetorical questions are a special form of question-and-answer interaction. Their core characteristic is that they do not directly tell the user the answer or directly negate the user's erroneous premise. Instead, they guide the user to discover logical flaws or factual errors in their premises by posing a question related to the user's erroneous premise. This interactive form draws on the basic concept of the Socratic method of questioning in education, which uses continuous questioning to prompt learners to discover contradictions in their own cognition during independent thinking, thereby actively correcting erroneous cognitions. Compared to the explicit correction method of directly telling the user "your premise is wrong," guided rhetorical questions are more easily accepted by users in human-computer interaction scenarios because conclusions reached through independent reasoning are more likely to form deep memories than those passively accepted external corrections, while avoiding the resistance that may be caused by directly negating the user's cognition. The conflict premise correction status is a determination of whether the user's original incorrect premise has been corrected after responding to the guided rhetorical question. This determination requires re-executing the implicit premise triple extraction and multi-layer conflict detection process (similar to step S10) on the user's response text to verify in a structured manner whether the user has corrected the incorrect premise in the original question in their response. Closed-loop correction refers to a closed-loop process from conflict detection, generating a rhetorical question, receiving a response, verifying the correction status, to the final response generation. This ensures that the system has fully verified the correctness of the user's premises before outputting the final response, preventing the generation of misleading responses based on incorrect premises.

[0088] Further, step S20 includes:

[0089] Step S21: Read the premise verification report, extract all premise triples in the premise verification report whose internal consistency state is internal consistency conflict or whose external collision state is external conflict premise or conditional dependency premise as conflict premise triples, and adopt the corresponding guided question generation strategy according to the category of internal consistency state or external collision state corresponding to each conflict premise triple to generate guided question for each conflict premise triple.

[0090] For a conflict premise triplet where the external collision state is the external conflict premise, extract its corresponding conflict pair, construct a question generation instruction with a set of guided question constraints as input, send the conflict pair and the question generation instruction to the large language model to generate guided questions, and establish the underlying principle on which the large language model generates guided questions as a knowledge anchor.

[0091] For conflict premise triples where the external collision state is a conditional dependency premise, extract the verified knowledge triples of conditional dependency and necessary condition information from the corresponding conditional dependency pair, construct a conditional dependency rhetorical question generation instruction, and send the conditional dependency rhetorical question generation instruction to the large language model to generate a guided rhetorical question that guides the user to confirm whether the necessary condition is satisfied.

[0092] For conflicting premise triples whose internal consistency state is internally contradictory, extract their corresponding contradictory premise triple pairs and explanations of the contradiction reasons, construct a consistency rhetorical question generation instruction, and send the consistency rhetorical question generation instruction to the large language model to generate guided rhetorical questions.

[0093] Specifically, step S21 transforms the structured conflict information detected in step S10 into user-oriented guided rhetorical questions in natural language. It reads the set of premise triples labeled with conflict types from the premise verification report output in step S14, iterates through each premise triple in this set, and filters out all premise triples whose internal consistency state is internally consistent conflict, or whose external collision state is externally conflicting premises or conditionally dependent premises. These premise triples are collectively recorded as conflicting premise triples. Conflicting premise triples are structured representations of the implicit premises of the user's question that contain problems. Each conflicting premise triple carries its corresponding conflict type label and pairing information related to that conflict. Based on the conflict type label carried by each conflicting premise triple, the system divides all conflicting premise triples into three categories: externally conflicting premises, conditionally dependent premises, and internally consistent conflicting premises. These three categories correspond to three different types of premise errors, requiring different guided rhetorical question generation strategies for processing.

[0094] For conflicting premise triples of the external conflicting premise category, the conflict pair recorded in step S13 is retrieved. The conflict pair consists of two parts: the user's erroneous premise triple and the nearest neighbor verified knowledge triple retrieved from the domain knowledge base that semantically contradicts the erroneous premise triple. The system uses this conflict pair as input to construct a rhetorical question generation instruction. This instruction is a prompt sent to the large language model, and its structure includes two components: the first part is the complete information of the conflict pair, including the erroneous premise triple and its corresponding nearest neighbor verified knowledge triple; the second part is the constraint condition on the output content of the large language model. A guided rhetorical question constraint condition set is applied to the rhetorical question generation instruction, which contains three constraints. The first constraint condition requires that the guided rhetorical question generated by the large language model must involve the underlying principle or basic constraint condition corresponding to the erroneous premise triple, rather than merely restating the erroneous premise itself. The underlying principles refer to concepts, laws, or rules at a more fundamental level within a knowledge system. The correctness of these concepts or laws is often a prerequisite for the validity of higher-level knowledge. For example, the erroneous premise that "the currents in series circuits are all different" is supported by underlying principles including the "law of conservation of charge" and the "definition of current." These more fundamental physical principles are prerequisites for understanding the characteristics of currents in series circuits. The second constraint requires that the guided rhetorical questions generated by the large language model guide users to re-examine the validity of the premise based on these underlying principles. In other words, the wording of the guided rhetorical questions should prompt users to use their existing basic knowledge to verify the correctness of the premise, rather than passively waiting for the system to provide the correct answer. The third constraint requires that the guided rhetorical questions generated by the large language model should not directly provide the specific content of the nearest neighbor verified knowledge triplet, i.e., it should not directly tell the user what the correct answer is, leaving room for the user's independent thinking and verification. The command to generate rhetorical questions containing conflict pairs and guided rhetorical question constraint sets is sent to the large language model. Based on its language generation capabilities and domain knowledge learned during the pre-training phase, the large language model generates guided rhetorical questions that satisfy the above three constraints. The system receives guided rhetorical questions returned by the large language model and simultaneously establishes the underlying principles upon which the large language model generated these guided rhetorical questions as knowledge anchors, storing them in system memory. These knowledge anchors are the core knowledge points upon which the guided rhetorical questions are based. Each knowledge anchor corresponds to a verified knowledge triple describing the underlying principle in the domain knowledge base. Its hierarchical position within the domain knowledge system will be used in the subsequent correction and upgrade process in step S23 to generate more fundamental guided rhetorical questions.

[0095] For example, when the erroneous premise triplet in a conflicting pair is (the current in each resistor in a series circuit is different), and the nearest neighbor verified knowledge triplet is (the current in a series circuit is equal everywhere), the conflicting pair, along with a set of guiding rhetorical question constraints, is sent to the large language model. The large language model analyzes the conflicting pair and identifies a directional contradiction between the erroneous premise triplet's statement that "the current in a series circuit is different" and the nearest neighbor verified knowledge triplet's statement that "the current in a series circuit is equal everywhere." Based on the first constraint, the large language model traces down the physics knowledge system to the underlying principles related to "the characteristics of current in a series circuit," identifying the "law of conservation of charge" as the knowledge anchor. Based on the second constraint, the large language model constructs a rhetorical question guiding the user to examine its erroneous premise from the perspective of charge conservation. Based on the third constraint, the large language model avoids directly including the correct knowledge statement "the current in a series circuit is equal everywhere" in the wording of the rhetorical question. The large language model generates the following guiding question: "In a series circuit, there is only one current path. If the current decreases after passing through the first resistor, where does the reduced charge go?" This guiding question uses "charge conservation" as its anchor point, guiding users to consider that in a single current path of a series circuit, if the current decreases at a certain point, it means that the charge disappears into thin air at that point, which contradicts the law of charge conservation. If users realize that charge cannot disappear into thin air while answering this question, they can naturally deduce the correct conclusion that the current should be equal at all points in a series circuit, thus autonomously correcting their original erroneous premise.

[0096] For conflicting premise triples of the conditional dependency premise category, the conditional dependency pairing recorded in step S13 is retrieved. The conditional dependency pairing consists of two parts: the user's premise triples and the verified conditional dependency knowledge triples retrieved from the domain knowledge base. The verified conditional dependency knowledge triples describe the necessary conditional information required for the state or attribute stated in the user's premise triples to be true. Necessary conditional information is extracted from the verified conditional dependency knowledge triples and stored in the object field of the verified conditional dependency knowledge triples. A conditional dependency rhetorical question generation instruction is constructed. This instruction is a prompt sent to the large language model and has two components: the first part is the complete information of the conditional dependency pairing, including the user's premise triples and the verified conditional dependency knowledge triples; the second part is the task instruction. This instruction requires the large language model to generate a guiding rhetorical question to guide the user to confirm whether the necessary condition is met. The conditional dependency question generation instruction differs from the question generation instruction for conflicting external premises in its constraints: the conditional dependency question generation instruction does not require the guided question to involve underlying principles, but directly asks the user whether the necessary conditions are met. The technical consideration behind this differentiated design lies in the fact that the nature of conditional dependency premises differs from that of conflicting external premises—conditional dependency premises do not directly contradict the verified knowledge triples, but their validity depends on whether the necessary conditions not declared by the user are met. The system cannot definitively determine whether the user's premise is correct or incorrect and needs to obtain more information through follow-up questions before making a judgment. The conditional dependency question generation instruction is sent to the large language model, which generates the guided question and then receives it. The system establishes the necessary condition information in the verified knowledge triples as conditional dependency anchors and stores them in system memory for use in subsequent step S23 when generating more specific follow-up guided questions for conditional dependency premises that have not yet been corrected.

[0097] For example, when the user premise triplet in the conditional dependency pair is (load balancing, status is, effectively running) and the verified knowledge triplet of the conditional dependency is (load balancing is effectively running, necessary condition is, number of backend nodes is greater than or equal to 2), the necessary condition information "number of backend nodes is greater than or equal to 2" is extracted from the object field of the verified knowledge triplet of the conditional dependency, and a conditional dependency rhetorical question generation instruction is constructed. The guided rhetorical question generated by the large language model is: "You can check the list of backend nodes of the load balancer. How many nodes are currently actually mounted?" This guided rhetorical question guides the user to check whether the necessary condition for the effective operation of the load balancer is met. If the user finds that the number of backend nodes is 0 or only 1 after checking, the user can realize that their premise of "load balancing is effectively running" is actually not true, thereby correcting their understanding and re-describing the problem. If the user confirms after checking that the number of backend nodes is indeed greater than or equal to 2, the conditional dependency premise is confirmed, and the system can treat it as a verified correct premise in subsequent processing.

[0098] For conflicting premise triples of the internal consistency conflict category, the contradictory premise triple pairs and the explanation of the contradiction reason recorded in step S12 are extracted. A contradictory premise triple pair refers to a combination of two or more premise triple pairs that contain logical contradictions among multiple implicit premises given by the user in the same user question text. The explanation of the contradiction reason is a textual explanation of why there is a logical contradiction between these premise triple pairs. The system constructs a consistency-based rhetorical question generation instruction, which is a prompt word sent to the large language model. Its structure includes two parts: the first part is the complete information of the contradictory premise triple pairs and the explanation of the contradiction reason; the second part is the task instruction. This consistency-based rhetorical question generation instruction requires the large language model to generate a guided rhetorical question that guides the user to verify the consistency between multiple given premises. The core requirement of the consistency-based rhetorical question generation instruction is to guide the user to jointly verify the multiple given premises, enabling the user to discover the logical conflicts between these premises during the verification process. The system sends the consistency question generation instruction to the large language model, which generates a guiding question and then receives it. The system establishes the specific numerical values, constraints, or logical relationships involved in the explanation of the contradiction as consistency verification anchor points and stores them in the system memory for use in subsequent step S23 when generating more specific verification guiding questions for internal consistency conflicts that have not yet been corrected.

[0099] For example, when a user asks the question "In triangle ABC, angle A equals 30 degrees, angle B equals 60 degrees, and angle C equals 100 degrees, find the ratio of the side lengths," step S12 detects a contradictory premise triplet pair: (figure, belongs to, triangle) and (the sum of angle A (30 degrees), angle B (60 degrees), and angle C (100 degrees) equals 190 degrees). The explanation for the contradiction is "The sum of the three angles is 190 degrees, exceeding the constraint that the sum of the interior angles of a triangle is 180 degrees." The system uses this contradictory premise triplet pair and the explanation for the contradiction as input to construct a consistent rhetorical question generation instruction. The guided rhetorical question generated by the large language model is: "You can add up the three angles and see what condition the sum of the interior angles of a triangle should satisfy?" This guided rhetorical question guides the user to discover, through a simple arithmetic operation, that the sum of the given angle values ​​(190 degrees) is inconsistent with the geometric constraint of the triangle (the sum of the interior angles equals 180 degrees). Users can realize that the angle value they provided is incorrect after performing the addition operation, thus correcting the conditions set in their original question.

[0100] Step S21 employs different guided question generation strategies based on different conflict type markers. This differentiated approach allows guided questions to accurately match premise errors of different natures. If the same general questioning strategy is used for three different types of premise errors, the targeted nature of the guided questions will decrease—asking about necessary conditions for external conflict premises will fail to address the user's knowledge errors; citing underlying principles for conditionally dependent premises will render the questioning ineffective because the user's premises may be logically correct; and correcting a single premise for internally consistent conflicts will fail to guide the user to discover logical contradictions between multiple premises. The differentiated guidance strategy ensures that each type of conflict premise receives the most suitable guidance method, thereby maximizing the probability of successful guidance. When generating guided questions for external conflict premises, the constraints of "must involve underlying principles" and "cannot directly provide correct knowledge" are imposed because users' mastery of professional domain knowledge typically exhibits a hierarchical distribution—users may not understand specific knowledge at a higher level, but they usually have a grasp of more fundamental concepts or principles. For example, a user might not be aware of the circuit characteristic that the current is the same everywhere in a series circuit, but they are usually familiar with the more fundamental physical law of charge conservation. By guiding the user to use their existing correct underlying knowledge to examine their erroneous upper-level premises, the user experiences cognitive conflict and corrects their premises during the process of autonomous reasoning. Compared to passively accepting external correction, this process of users discovering and correcting errors autonomously results in higher user acceptance of the corrections and deeper long-term memory, while avoiding the resistance that might arise from directly denying the user's cognition. The recording of knowledge anchors provides a hierarchical positioning basis for the correction and escalation process in subsequent step S23, enabling the system to move down the knowledge hierarchy after the user's initial guidance fails, generating new guided rhetorical questions based on more fundamental principles.

[0101] Step S22: Send a guiding rhetorical question to the user and receive the user's response text in response to the guiding rhetorical question. Re-execute the implicit premise triple extraction from step S11, the internal consistency verification from step S12, and the external domain knowledge collision detection from step S13 on the user's response text to obtain a set of premise triples labeled with conflict types in the user's response text. Compare each conflicting premise triple in the premise verification report with the set of premise triples labeled with conflict types in the user's response text, and label the conflicting premise correction status of each conflicting premise triple in the premise verification report. The value of the conflicting premise correction status is either "corrected" or "not corrected".

[0102] For each conflicting premise triplet in the premise verification report, search the set of premise triplets in the user's response text for premise triplets whose subject and relation fields are the same or semantically equivalent to the conflicting premise triplet. If a premise triplet with the same or semantically equivalent subject and relation fields is found, and the internal consistency state of the found premise triplet is no longer internally conflicting, and the external collision state is no longer an external conflicting premise or a conditional dependency premise, then mark the conflicting premise triplet as having a conflicting premise correction status. If a conflicting premise triplet with the same or semantically equivalent subject and relation fields is found, but the internal consistency of the found premise triplet is still internally conflicting or the external collision state is still an external conflicting premise or a conditional dependency premise, then the conflicting premise triplet is marked as uncorrected. If no conflicting premise triplet with the same or semantically equivalent subject and relation fields is found, then the conflicting premise triplet is marked as uncorrected.

[0103] Specifically, step S22 aims to complete the sending of the guided rhetorical question, the receiving of the user's response, and the verification of the premise correction status. The guided rhetorical question generated in step S21 is sent to the user interface as a system response, and the user interface presents the guided rhetorical question to the user. After reading the guided rhetorical question, the user, based on their thinking about the issues involved in the question, inputs a user response text as a response to the guided rhetorical question. The user response text input by the user is received through the user interface and stored in the system memory for subsequent processing. The user response text is the user's natural language response to the guided rhetorical question. Its content may include the user's correction of the original incorrect premise, or it may maintain the original incorrect premise unchanged, or it may even introduce a new incorrect premise. The literal content of the user response text alone cannot reliably determine whether the user has truly corrected the incorrect premise—the user may use different wording in the response than the original incorrect premise, but the premise expressed is still essentially the same or equivalent to the original incorrect premise. Therefore, it is necessary to re-execute the structured premise verification process on the user response text, using the same standards and methods as the original detection to verify whether the premises in the user response still conflict.

[0104] The implicit premise triple extraction process described in step S11 is re-executed on the user's reply text. The user's reply text is embedded in the same premise extraction instruction template used in step S11 to form premise extraction prompts for the user's reply text, which are then sent to the large language model via the application programming interface. The large language model performs semantic parsing on the user's reply text, identifies all implicit factual premise assumptions in the user's reply text, and outputs each implicit premise as a premise triple. All premise triples returned by the large language model are collected to form the premise triple set of the user's reply text. The internal consistency check described in step S12 is re-executed on the premise triple set of the user's reply text—if the number of premise triples in the premise triple set is greater than or equal to 2, all premise triples are converted into affirmative statements according to the statement conversion template and then concatenated into a combined statement text, which is sent to the large language model for logical consistency judgment; if the number is less than 2, the internal consistency status of the premise triples is directly marked as internal consistency passed. For the set of premise triples in the user's reply text, re-execute the external domain knowledge collision detection described in step S13—for each premise triple, search for the nearest neighbor verified knowledge triple, perform directional consistency judgment, and label the external collision state according to the semantic relationship judgment result. Combine the internal consistency verification result with the external domain knowledge collision detection result to obtain the set of premise triples in the user's reply text labeled with conflict types.

[0105] Then, the premise triples marked as conflicting in the premise verification report recorded in step S14 are compared with the set of premise triples marked with conflicting types in the user's reply text to determine whether the original conflicting premises still have the same conflict after the user's reply. The specific comparison method is as follows: For each conflicting premise triple in the premise verification report, a search is conducted in the set of premise triples in the user's reply text to see if there is a premise triple whose subject field and relation field are the same as or semantically equivalent to the conflicting premise triple. The semantic equivalence determination of the subject field and relation field adopts the same semantic vector encoding and cosine similarity calculation method as in step S13. If the cosine similarity of the semantic vectors of the two fields exceeds the preset equivalence determination threshold, they are determined to be semantically equivalent. The method for determining the equivalence determination threshold is as follows: Select several sets of field pairs known to be semantically equivalent and field pairs known to be semantically inequivalent as calibration samples, calculate the cosine similarity of each set of field pairs, and take the cosine similarity boundary point that can correctly distinguish between equivalence and inequivalence as the equivalence determination threshold. For example, the equivalence determination threshold can be set to a value between 0.85 and 0.95.

[0106] If a premise triplet with the same or semantically equivalent subject and relation fields as the original conflicting premise triplet is found in the set of premise triplets in the user's reply text, the conflict type label of the retrieved premise triplet in the user's reply text is further examined. If the internal consistency state of the retrieved premise triplet is no longer internally consistent and conflicting, and the external collision state is no longer an externally conflicting premise or a conditionally dependent premise—that is, the conflict type label of the premise triplet in the user's reply text is internally consistent and the external collision state is an externally consistent premise or no matching premise—it indicates that the user has corrected the premise in the reply to be consistent with the domain knowledge base or at least no longer contradicts the domain knowledge base, and the conflict premise correction status of the original conflicting premise triplet is marked as corrected. If the retrieved premise triplet's internal consistency state is still internally conflicting, or its external collision state is still an externally conflicting premise or a conditionally dependent premise, it indicates that although the user mentioned content on the same topic as the original conflicting premise in their response, they still maintained the incorrect premise or introduced a new conflict. The conflicting premise correction state of the original conflicting premise triplet is marked as uncorrected. If no premise triplet with the same or semantically equivalent subject and relation fields as the original conflicting premise triplet is found in the user's response text's premise triplet set, it indicates that the user did not involve content related to the original conflicting premise in their response. The original conflicting premise was neither explicitly corrected nor explicitly maintained. The conflicting premise correction state of the original conflicting premise triplet is marked as uncorrected—this conservative marking strategy ensures that a correction is only determined when the user explicitly expresses the correct premise, avoiding misjudgment as corrected due to the user's response not involving relevant content.

[0107] For example, Table 1 shows the correspondence between the original conflicting premise triples and the premise correction status after the user's response.

[0108] Table 1. Correspondence between the original conflicting premise triples and the premise correction status after the user's response.

[0109]

[0110] After performing the above comparison on each conflicting premise triplet in the premise verification report, a complete set of conflicting premise correction states is obtained. This set indicates whether each conflicting premise triplet in the premise verification report is in a corrected or uncorrected state after the user's response. This set of conflicting premise correction states serves as the input for the flow determination in step S23.

[0111] Step S22 re-executes the same implicit premise triple extraction and multi-layer conflict detection process as steps S11 to S13 on the user's reply text. This design allows the determination of the premise correction status to be based on structured semantic analysis, rather than on literal matching of natural language. Users may use different wording in their replies to express the same or similar meanings as the original erroneous premises. For example, a user may rephrase "the currents are not the same" as "the currents are different at each location" or "the currents are different." These different wordings are obviously different in natural language, but semantically they express the same erroneous premise. If the system judges whether the premise has been corrected based solely on literal matching, it may misjudge that it has been corrected because the user has changed the wording. By re-executing structured premise triple extraction and conflict detection on the user's reply text, the system can analyze the premises in the user's reply with the same standards and methods as the original detection, so that premises with different wording but the same semantics are extracted into the same or equivalent premise triples, thereby avoiding misjudgments that may occur based on literal matching. Step S22 reuses the processing flow of steps S11 to S13 instead of developing a dedicated user response analysis flow. This reuse design ensures that user questions and user responses use the same premise extraction and conflict detection standards, thus ensuring the consistency and comparability of premise verification throughout the guided correction process.

[0112] Step S22 works in tandem with step S21. The guiding question generated in step S21 prompts the user to independently discover the premise error by thinking about the underlying principles or checking necessary conditions. Step S22 then verifies whether the user has indeed corrected the premise through structured analysis of the user's response. Together, these two steps form a single-round correction loop of "guidance-verification." Sending the guiding question initiates the user's thinking process, while analyzing the user's response verifies the effectiveness of that process. Without the verification step in S22, it would be impossible to confirm whether the user has corrected the erroneous premise in their response, and subsequent processing would lack a basis for routing—the system would not know whether it should enter the corrected path to generate the final response, nor whether it should enter the uncorrected path to continue guiding or switch to explicit correction.

[0113] Step S23, see Figure 5 Based on the conflict premise, the state is modified to perform a three-way flow determination;

[0114] When all conflicting premise triples in the premise verification report are in the "corrected" state, the corrected premise triples are extracted from the set of premise triples in the user's response text, and the final response is generated.

[0115] When at least one conflicting premise triplet in the premise verification report has an uncorrected conflicting premise status, the correction attempt counter is incremented by 1, and it is determined whether the value of the incremented correction attempt counter is less than the maximum number of attempts threshold: the correction attempt counter is initialized to 0 when the current session first enters step S20;

[0116] If the incremented value of the correction attempt counter is less than the maximum number of attempts threshold, generate new guided rhetorical questions for all conflicting premise triples that are still in the uncorrected state, send the new guided rhetorical questions to the user, and return to step S22.

[0117] If the incremented value of the correction attempt counter is greater than or equal to the maximum number of attempts threshold, the explicit correction mode is triggered to replace the incorrect premise of the user's question text and generate the final response.

[0118] Specifically, step S23 performs a three-way splitting judgment based on the conflict premise correction state obtained in step S22, and adopts different subsequent processing strategies according to different judgment results. The system maintains a correction attempt counter bound to the current session, which records the number of times the system has sent guided rhetorical questions to the user in the current session. The correction attempt counter is initialized to 0 when the current session first enters step S20, and increments by 1 each time the system sends a round of guided rhetorical questions to the user. At the same time, a maximum attempt threshold is preset, which specifies the maximum number of rounds the system can attempt to correct premises through guided rhetorical questions in a single session. The maximum attempt threshold is determined by comprehensively considering the expected success rate of guided correction, user interaction patience, and the reasonable duration of a single session, selecting a value that can achieve a balance between correction effect and user experience. For example, the maximum attempt threshold can be set to an integer between 2 and 5, indicating that the system will attempt to correct the user's erroneous premises through a maximum of 2 to 5 rounds of guided rhetorical questions in a single session.

[0119] The first step of the traffic splitting judgment process addresses the corrected path. If the conflict premise correction status obtained in step S22 shows that all conflicting premise triples in the premise verification report are in a corrected state in the user's response text, it indicates that the user has independently corrected all original erroneous premises in their response. The system then performs the following operations: Construct a confirmation feedback generation instruction, and send the corrected correct premise triples along with the confirmation feedback generation instruction to the large language model. The large language model generates a positive confirmation feedback text, which confirms to the user that their correction is correct, enhancing the user's trust in the system and reinforcing the user's memory of the correct knowledge. Extract the corrected correct premise triples from the set of premise triples in the user's response text. These correct premise triples are the premises expressed by the user in their response that have been confirmed to be consistent with the domain knowledge base through external domain knowledge collision detection. Combine the user's original user question text with the corrected correct premise triples into a premise-corrected question text. The merging method is as follows: replace the wording related to the erroneous premises in the original user question text with the content expressed by the correct premise triples, while retaining the core of the question and other non-conflicting content in the original user question text. The revised question text, with corrected premises, is sent to the large language model, which generates a response based on the correct premises. Since the input conditions received by the large language model no longer contain the original incorrect premises, its generated response will reason based on the correct premises, and the response conclusion will be consistent with real-world domain knowledge. This response, along with positive confirmation feedback text, is then returned to the user, and the guided error correction process for the current session concludes normally.

[0120] For example, when a user's original question is "In a series circuit, the current through each resistor is different, so how should the total current be calculated?", and the user replies after the guiding question, "I remember now, a series circuit has only one path, so the current should be the same everywhere," step S22 compares the premise triplet in the user's reply (the current through each resistor in the series circuit is the same, with the attribute "the same") with the original conflicting premise triplet (the current through each resistor in the series circuit is different, with the attribute "the current through each resistor in the series circuit is not the same"), and determines that the conflicting premise correction status is corrected. The system enters the corrected path, outputs the positive confirmation feedback text "You are absolutely right, the current in a series circuit is indeed the same everywhere," and replaces the incorrect premise in the original question with the correct premise to form the premise-corrected question text "In a series circuit, the current through each resistor is the same, so how should the total current be calculated?", and sends it to the large language model to generate a response. The large language model reasones based on the correct premise "the current in a series circuit is the same everywhere," and generates the response "Since the current in a series circuit is the same everywhere, the total current is the current through any one resistor, and no further calculation is needed." The response was derived based on correct premises and is consistent with real knowledge in the field of physics.

[0121] The second step in the flow control process addresses paths that are not corrected and have not reached the maximum attempt threshold. If the conflict premise correction status obtained in step S22 shows that at least one conflicting premise triplet in the premise verification report remains uncorrected in the user's response text, it indicates that the user failed to completely correct all erroneous premises after the first round of guided questioning, and the correction attempt counter is incremented by 1. Then, it is determined whether the incremented correction attempt counter value is less than the maximum attempt threshold. If the correction attempt counter value is less than the maximum attempt threshold, it indicates that the system still has the opportunity to continue attempting correction through guided questioning, and the system performs the following operations: New guided question sentences are generated for all conflicting premise triplets that are still uncorrected. Compared to the first guided question sentence generated in step S21, the new guided question sentences are moved one level down in the knowledge anchor level. See [link to relevant documentation]. Figure 6 This is a schematic diagram illustrating the downward shift of knowledge levels provided in an embodiment of this application. Figure 6 The diagram illustrates the distribution structure of the knowledge system from high to low levels, the hierarchical progression between higher-level knowledge and basic concepts, and the complete link of the gradual downward shift of knowledge anchors. It also highlights the core design logic of gradually shifting down to the user's cognitive reach, making more basic concepts easier for users to understand and apply. The specific implementation of the hierarchical shift of knowledge anchors is as follows: starting with the verified knowledge triplet corresponding to the knowledge anchor established in step S21 and stored in the system memory, the diagram retrieves verified knowledge triplets of more basic concepts or rules related to the knowledge anchor from the pre-established knowledge hierarchy in the domain knowledge base. The knowledge hierarchy is the hierarchical relationship between verified knowledge triplets in the domain knowledge base defined in step S13, and... Figure 6 The hierarchical correspondence between high-level overarching knowledge and low-level foundational concepts is perfectly consistent. Overarching verified knowledge triples are more complex knowledge derived from or supported by lower-level verified knowledge triples, which in turn are more fundamental and atomized concepts or rules. Using the retrieved, more fundamental verified knowledge triples as new knowledge anchors is a process similar to... Figure 6 The operation logic of moving the knowledge anchor point down one level from the initial guidance level is completely matched. Taking the new knowledge anchor point and the conflict pair that is still in an uncorrected state as input, the question generation instruction is reconstructed according to the question generation instruction and the guidance question constraint group described in step S21, and sent to the large language model to obtain a new guidance question sentence.

[0122] For example, if the initial guiding question in step S21 is constructed based on the "law of conservation of charge" as a knowledge anchor, this knowledge anchor is related to... Figure 6The initial guiding knowledge anchor point marked in the code corresponds perfectly, but the user still maintains the erroneous premise that "the currents in series circuits are all different" in their reply. Therefore, the system traces down from the "law of conservation of charge" in the domain knowledge base to the more fundamental level of "definition of current." This tracing process is consistent with... Figure 6 The process of shifting the knowledge anchor from the law of conservation of charge to the definition of electric current is completely consistent with the hierarchical movement of the definition of electric current. Figure 6 The knowledge anchor point corresponds to the shift of the marker. The definition of "current" refers to the amount of charge passing through a cross-section of a conductor per unit time. This definition is a more fundamental concept for understanding the application of the law of conservation of charge in circuits. A new guiding question is generated using "the definition of current" as the new knowledge anchor point. For example, the new guiding question is: "In a series circuit, all charges must pass through each resistor sequentially. Assuming the amount of charge passing through the first resistor cross-section per second is fixed, how much charge will pass through the second resistor cross-section per second?" This new guiding question is constructed based on the more fundamental "definition of current," guiding the user to think from the perspective of the definition of current: in a single path of a series circuit, the amount of charge passing through any two cross-sections per unit time should be the same. The newly generated guiding question is sent to the user as a system response. The process returns to step S22, waiting for the user's next response and re-performing implicit premise triplet extraction and multi-level conflict detection.

[0123] For conflict premise triples that are still in an uncorrected state and whose external collision state is a conditional dependency premise, the system extracts the verified knowledge triples and necessary condition information of the conditional dependency from their corresponding conditional dependency pairs, and reads the conditional dependency anchors established and stored in the system memory in step S21. Based on the conditional dependency question generation instruction used for the first time in step S21, it adds specific guidance on the necessary conditions to construct an upgraded conditional dependency question generation instruction. This upgraded conditional dependency question generation instruction requires the large language model to generate a guiding question that guides the user to provide specific evidence or operational verification results regarding whether the necessary conditions are met. The upgraded conditional dependency question generation instruction is then sent to the large language model to generate a new guiding question. For example, if the initial guiding question in step S21 is "You can check the list of backend nodes of the load balancer. How many nodes are currently actually mounted?" and the user does not give a clear reply, the new guiding question generated in the retry phase can be "Please execute the command to check the status of the backend nodes in the server management console, confirm whether the number of backend nodes in the current healthy state is greater than or equal to 2, and report the execution result to me."

[0124] For conflicting premise triples whose internal consistency is conflicting and are still in an uncorrected state, the system retrieves the corresponding contradictory premise triple pair and the explanation of the contradiction, and reads the consistency verification anchor points established and stored in the system memory in step S21. Based on the consistency rhetorical question generation instruction used for the first time in step S21, it adds specific guidance on the focus of the contradiction and constructs an upgraded consistency rhetorical question generation instruction. This upgraded consistency rhetorical question generation instruction requires the large language model to generate a guiding rhetorical question that guides the user to clearly point out which of its multiple premises needs to be modified. The upgraded consistency rhetorical question generation instruction is then sent to the large language model to generate a new guiding rhetorical question. For example, if the initial guiding question in step S21 is "You can add up the degrees of the three angles and see what condition the sum of the interior angles of a triangle should satisfy?" and the user still does not correct the angle values, then the new guiding question generated in the retry phase can be "The sum of the interior angles of a triangle must be equal to 180 degrees. You gave angle A as 30 degrees, angle B as 60 degrees, and angle C as 100 degrees, and the sum of the three is 190 degrees. Which angle's degree do you want to change?"

[0125] The technical consideration behind the design of progressively lowering the knowledge anchor level in guided rhetorical questions lies in the fact that if a user fails to correct their erroneous premise in the first round of guided rhetorical questions, it may indicate that the knowledge level upon which the initial guided rhetorical question was based exceeds the user's current cognitive reach. Although the user received the guided rhetorical question, they were unable to independently complete the reasoning and verification using that level of knowledge. By progressively lowering the knowledge anchor level to more basic concepts or rules, the guidance gradually approaches the user's comprehensible boundaries—that is, the level of basic knowledge that the user has truly mastered and can apply. Only when the knowledge anchor level is lowered to a level that the user can understand and apply can they independently complete the reasoning based on that level of knowledge and discover the contradiction between their erroneous premise and the basic knowledge, thus correcting the premise. This progressively lowering guidance strategy has a higher probability of success than guidance methods that repeatedly use the same level of knowledge anchor.

[0126] The third step in the flow judgment process addresses paths that are uncorrected and have reached the maximum number of attempts threshold. If the correction attempt counter is greater than or equal to the maximum number of attempts threshold, it indicates that the user has not corrected all erroneous premises after the number of guided rhetorical questions specified by the maximum number of attempts threshold. The guided correction strategy has failed to achieve the expected effect for this user, and the system switches to explicit correction mode. Explicit correction mode is a correction method that directly informs the user of their incorrect premises and provides the correct knowledge. Compared to the indirect guidance of guided rhetorical questions, explicit correction is direct and clear, but may trigger a cognitive defense response from the user. In explicit correction mode, the system performs the following operations: It reads all conflicting premise triples and their corresponding conflicting pairs, conditional dependency pairs, or contradictory premise triples that are still in an uncorrected state from the premise verification report output in step S14. It generates explicit correction text according to the conflict type of the conflicting premise triples. For conflicting premise triples whose external collision state is an external conflict premise, the corresponding conflict pair is extracted, and an explicit correction text is generated. This explicit correction text contains three parts: The first part indicates which specific premise in the user's question contradicts the domain knowledge base. The system converts the erroneous premise triple into a natural language statement using a declarative sentence conversion template and clearly marks it as a "premise that needs correction," so that the user clearly understands which part of their question is problematic. The second part provides the content of the nearest neighbor verified knowledge triple corresponding to the erroneous premise triple. The system converts the nearest neighbor verified knowledge triple into a natural language statement using a declarative sentence conversion template and adds a brief explanation, so that the user understands what the correct knowledge in the domain is. The third part explains the specific reasons for the contradiction between the erroneous premise and the correct knowledge, so that the user understands why their original premise is wrong.

[0127] For conflicting premise triples whose external collision state is a conditional dependency premise, the corresponding conditional dependency pair is extracted, and an explicit correction text is generated. This explicit correction text contains three parts: The first part indicates which premise in the user's question depends on a necessary condition that the user has not confirmed. The system then converts the conditional dependency premise triple into a natural language statement using a declarative sentence conversion template. The second part provides the necessary condition described by the verified knowledge triple of this conditional dependency, so that the user understands the conditions that need to be met for the premise to be valid. The third part explains that since the user has not confirmed whether the necessary condition is met in multiple rounds of interaction, the system will not treat this premise as a confirmed correct premise in subsequent response generation and suggests that the user re-ask the question after verifying the necessary condition.

[0128] For conflicting premise triples whose internal consistency is conflicting, the system extracts the corresponding contradictory premise triple pairs and explanations of the contradictions, generating an explicit correction text. This explicit correction text contains three parts: the first part identifies which premises in the user's question are logically contradictory; the system converts each premise triple in the contradictory premise triple pair into a natural language statement using a declarative sentence conversion template; the second part explains the reasons for the contradictions, allowing the user to understand why these premises cannot be logically true simultaneously; the third part requires the user to explicitly indicate which premise needs modification, and the system will replace the premise according to the user's specification. The generation method of the above explicit correction text is as follows: the system sends the erroneous premise triples, nearest neighbor verified knowledge triples or conditional dependency verified knowledge triples or contradictory premise triple pairs, and explanations of the contradictions, along with the preset explicit correction generation instructions, to the large language model. The large language model generates natural language text conforming to the above three-part structure based on the input information. The system then summarizes all the explicit correction texts and sends them to the user interface.

[0129] For example, if a user maintains the erroneous premise that "the current in a series circuit is not the same in each resistor" even after multiple rounds of guided questioning, the system generates an explicit corrective text: "In your question, you mentioned 'the current through each resistor is different,' which needs to be corrected. According to the basic principles of circuit theory, the current in a series circuit is the same everywhere, that is, the current value through each resistor is the same. The contradiction lies in the fact that a series circuit has only one current path, and according to the law of conservation of charge, the amount of charge passing through any cross-section of the circuit per unit time is the same, therefore the current remains equal throughout the series circuit." This explicit corrective text clearly points out the user's erroneous premise, correct knowledge, and the reason for the contradiction between the two, allowing the user to clearly understand their cognitive bias after reading it.

[0130] After outputting the explicit correction text, the incorrect premises in the user's original question text are replaced with the correct content of the corresponding nearest-neighbor verified knowledge triples, forming a premise-corrected question text. This premise-corrected question text is sent to the large language model, which generates a response based on the correct premises. This response is returned to the user along with the explicit correction text, ending the guided correction process for the current session. Although the explicit correction mode does not allow users to independently discover and correct incorrect premises, it ensures that the final response received by the user is a conclusion derived from correct premises, and the user also receives a clear explanation of their incorrect premises, enabling them to correct their understanding accordingly.

[0131] Setting a maximum attempt threshold provides a clear termination condition for guided questions. While guided questions are advantageous in promoting deeper user understanding, they are not effective for all users—some users may lack sufficient domain knowledge, and even if the system lowers the knowledge anchor of the guided questions to the most basic conceptual level in the domain, they may still be unable to independently discover and correct their erroneous premises. If the system repeatedly sends guided questions indefinitely without providing explicit correction, the user experience will deteriorate—users may feel frustrated or question the system's capabilities due to repeatedly receiving guided questions but never receiving the answers they need. By setting a maximum attempt threshold, when guided questions fail to achieve the correction goal within a limited number of attempts, the system switches to explicit correction mode to directly point out the error and provide the correct knowledge to the user, ensuring that the erroneous premise is eventually corrected, while ensuring that the user receives the required response within a reasonable number of interaction rounds.

[0132] The three-way routing design in step S23 ensures that regardless of the user's response, the system will ultimately generate and present a response to the user based on verified correct premises. Through the corrected path of routing judgment one, the user independently corrects the incorrect premise, and the system generates a response based on the user-confirmed correct premise. Through the re-guidance path of routing judgment two, the user corrects the incorrect premise after multiple rounds of guidance, and the process ultimately merges into the corrected path of routing judgment one. Through the explicit correction path of routing judgment three, the system directly points out the error to the user, replaces the incorrect premise with correct knowledge, and generates a response. The endpoint of all three paths is generating a response based on correct premises, fundamentally preventing the occurrence of "large language models generating misleading responses that contradict true knowledge through conditional obedience reasoning based on user's incorrect premises."

[0133] Step S20, as the latter half of the guided question-answering method based on implicit premise triple extraction, forms a complete collaborative closed loop with step S10. Step S10 completes the structured extraction of implicit premises and multi-layer conflict detection, outputting a premise verification report. Step S20 reads the premise verification report and generates guided rhetorical questions based on the conflict type. After receiving the user's reply, it verifies the premise correction status and performs triage processing based on the correction status until the final response is generated. The two steps together constitute a complete processing chain of "extraction-detection-rhetorical question-verification-re-detection-confirmation-response," with the output of each step directly driving the execution of the next step, and there are no information breaks or logical jumps between steps. Step S22 reuses the processing flow of steps S11 to S13 for the user's reply text, ensuring that the premise verification standard remains consistent between the user's question and the user's reply, and ensuring that the premise verification of the entire process is comparable and traceable. The correction attempt counter mechanism in step S23 sets a limited number of attempts for guided correction, which not only gives guided correction sufficient opportunities to try and leverage its advantage of promoting deeper user understanding, but also ensures that the erroneous premise is eventually corrected through the explicit correction mode as a fallback strategy, thus achieving a balance between correction effect and user experience.

[0134] Compared to direct explicit correction, the guided correction process in step S20 has unique value in the following aspects: In terms of user acceptance, guided rhetorical questions guide users to discover errors independently by asking questions rather than denying them, avoiding the resistance that might arise from directly negating user cognition. The cognitive conflict generated during the user's independent reasoning process makes them have a higher intrinsic sense of acceptance of the corrected result. In terms of knowledge consolidation, correct conclusions reached by users through independent reasoning have a higher retention rate in long-term memory compared to passively accepted external corrections, allowing users to more effectively recall and apply this knowledge when encountering similar situations in the future. In terms of educational application, the guided correction process aligns with the educational philosophies of Socratic questioning and discovery learning, making it particularly suitable for application scenarios such as digital teaching platforms in universities where knowledge acquisition is the core objective. It can correct users' misconceptions while promoting a deeper understanding of relevant knowledge. At the system reliability level, the three-way split design ensures that the system can output responses based on correct premises under various user response scenarios. This elevates the reliability of the intelligent question-answering system from relying on the judgment ability of the large language model itself to relying on structured premise verification and multi-way guarantee mechanisms, reducing the probability of misleading responses being delivered to users. The design of progressively shifting the knowledge anchor level in step S23 allows guided rhetorical questions to gradually approach the user's cognitive reach. This adaptive knowledge level adjustment capability enables the system to provide differentiated guidance depths for users with different levels of knowledge, expanding the user applicability of guided correction.

[0135] Example 2:

[0136] This embodiment, based on Embodiment 1, provides a guided question-answering system based on implicit premise triple extraction, such as... Figure 7 As shown, it includes:

[0137] Premise detection module: This module receives user question text, performs implicit premise triple extraction on the user question text to construct a premise triple set, performs multi-level conflict detection on the premise triple set, labels the internal consistency state and external collision state of each premise triple in the premise triple set, and summarizes and generates a premise verification report; the multi-level conflict detection includes internal consistency verification and external domain knowledge collision detection.

[0138] The guidance and correction module generates guided rhetorical questions based on the internal consistency and external collision states of each premise triplet, using corresponding guided rhetorical question generation strategies. It receives user response texts for the guided rhetorical questions and verifies the conflict premise correction states of each conflicting premise triplet in the premise verification report. Based on the conflict premise correction states, it performs closed-loop correction and outputs the final response based on the correct premises.

[0139] Furthermore, in the premise detection module, the execution method for extracting the implicit premise triplet includes:

[0140] The user's question text is embedded into a pre-defined premise extraction instruction template to construct a premise extraction prompt word. The premise extraction prompt word is sent to the large language model to perform implicit premise triple extraction, resulting in a set of premise triples that are bound to the user's question text and consist of a subject field, a relation field, and an object field.

[0141] The internal consistency state can be either internal consistency passed or internal consistency conflicted.

[0142] The method for performing the internal consistency verification includes:

[0143] The number of premise triples in the premise triple set is counted. When the number of premise triples is greater than or equal to 2, an internal consistency check is performed on the premise triple set to mark the internal consistency status of each premise triple in the premise triple set. When the number of premise triples is less than 2, the internal consistency status of the premise triples in the premise triple set is directly marked as internal consistency passed.

[0144] The method for performing the internal consistency verification also includes:

[0145] Each of the premise triples in the premise triple set is converted into an affirmative statement, and all the affirmative statements are concatenated to form a combined statement text. The combined statement text and a preset consistency judgment instruction are sent to the large language model to determine whether there is a logical contradiction between the affirmative statements in the combined statement text.

[0146] When a logical contradiction is determined, the contradictory premise triplet is identified as a contradictory premise triplet pair and a contradiction explanation is generated. The internal consistency state of the premise triplets involved in the contradictory premise triplet pair is marked as internal consistency conflict. When the large language model determines that there is no logical contradiction, the internal consistency state of all premise triplets in the premise triplet set is marked as internal consistency passed.

[0147] The method for performing the external domain knowledge collision detection includes:

[0148] Determine the corresponding domain knowledge base, traverse each premise triplet in the premise triplet set, retrieve the nearest neighbor verified knowledge triplet that is semantically closest to the premise triplet from the domain knowledge base, perform external domain knowledge collision detection on the pair consisting of the premise triplet and the nearest neighbor verified knowledge triplet, and determine the external collision state.

[0149] The external collision state can be one of the following: external conflict premise, external consistency premise, conditional dependency premise, or no matching premise.

[0150] The retrieval method for the nearest neighbor verified knowledge triplet is as follows:

[0151] Retrieve the current premise triplet from the premise triplet set, and retrieve a subset of candidate verified knowledge triplets from the domain knowledge base using the subject field of the current premise triplet as the search key. Encode the semantic vector of each candidate verified knowledge triplet in the current premise triplet and the subset of candidate verified knowledge triplets, calculate the cosine similarity between the semantic vector of the current premise triplet and the semantic vector of each candidate verified knowledge triplet, and select the candidate verified knowledge triplet with the highest cosine similarity as the nearest neighbor verified knowledge triplet of the current premise triplet.

[0152] The methods and systems of this application may be implemented in many ways. For example, they may be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The above-described order of steps for the method is for illustrative purposes only, and the steps of the method of this application are not limited to the order specifically described above, unless otherwise specifically stated.

[0153] In addition, the parts of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of the corresponding technical solutions in the prior art have not been described in detail, so as to avoid excessive elaboration.

[0154] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A guided question answering method based on implicit premise triple extraction, characterized in that, The method includes: The system receives user query text, performs implicit premise triple extraction on the user query text to construct a premise triple set, performs multi-level conflict detection on the premise triple set, labels each premise triple in the premise triple set with internal consistency status and external collision status, and generates a premise verification report. The internal consistency status takes the value of internal consistency pass or internal consistency conflict, and the multi-level conflict detection includes internal consistency verification and external domain knowledge collision detection. The method for performing the internal consistency check includes: counting the number of premise triples in the premise triple set; when the number of premise triples is greater than or equal to 2, performing an internal consistency check on the premise triple set to mark the internal consistency status of each premise triple in the premise triple set; when the number of premise triples is less than 2, directly marking the internal consistency status of the premise triples in the premise triple set as internal consistency passed. The method for performing internal consistency verification on the premise triple set is as follows: Each premise triple in the premise triple set is converted into an affirmative statement, and all affirmative statements are concatenated to form a combined statement text. The combined statement text, along with a preset consistency judgment instruction, is sent to the large language model to determine whether there is a logical contradiction between the affirmative statements in the combined statement text. When a logical contradiction is determined, the contradictory premise triples are identified as contradictory premise triple pairs, and a contradiction explanation is generated. The internal consistency status of the premise triples involved in the contradictory premise triple pairs is marked as internal consistency conflict. When the large language model determines that there is no logical contradiction, the internal consistency status of all premise triples in the premise triple set is marked as internal consistency passed. The method for performing external domain knowledge collision detection includes: determining the corresponding domain knowledge base, traversing each premise triplet in the premise triplet set, retrieving the nearest neighbor verified knowledge triplet that is semantically closest to the premise triplet from the domain knowledge base, performing external domain knowledge collision detection on the pair consisting of the premise triplet and the nearest neighbor verified knowledge triplet, and determining the external collision state; the value of the external collision state is one of external conflict premise, external consistency premise, conditional dependency premise, and no matching premise; The current premise triplet and the nearest neighbor verified knowledge triplet are transformed into affirmative statements, resulting in two affirmative statements corresponding to the current premise triplet and the nearest neighbor verified knowledge triplet, respectively. The two affirmative statements and a preset semantic relationship judgment instruction are sent to the large language model to determine the semantic relationship between the two affirmative statements. The external collision state of the current premise triplet is determined based on the semantic relationship judgment result. The semantic relationship judgment result is one of semantic consistency, semantic contradiction, and semantic irrelevance. Based on the internal consistency state and external collision state of each premise triple, a corresponding guided question generation strategy is adopted to generate guided questions. User response text is received in response to the guided questions, and the conflict premise correction state of each conflicting premise triple in the premise verification report is verified. Closed-loop correction is performed based on the conflict premise correction state, and the final response based on the correct premise is output.

2. The guided question-answering method based on implicit premise triple extraction according to claim 1, characterized in that, The method for extracting implicit premise triples includes: The user's question text is embedded into a pre-defined premise extraction instruction template to construct a premise extraction prompt word. The premise extraction prompt word is sent to the large language model to perform implicit premise triple extraction, resulting in a set of premise triples that are bound to the user's question text and consist of a subject field, a relation field, and an object field.

3. The guided question-answering method based on implicit premise triple extraction according to claim 2, characterized in that, The retrieval method for the nearest neighbor verified knowledge triplet is as follows: Retrieve the current premise triple from the premise triple set, and use the subject field of the current premise triple as the search key to retrieve a subset of candidate verified knowledge triples from the domain knowledge base; Semantic vector encoding is performed on each candidate verified knowledge triple in the current premise triple and candidate verified knowledge triple subset. The cosine similarity between the semantic vector of the current premise triple and the semantic vector of each candidate verified knowledge triple is calculated. The candidate verified knowledge triple with the highest cosine similarity is selected as the nearest neighbor verified knowledge triple of the current premise triple.

4. The guided question-answering method based on implicit premise triple extraction according to claim 3, characterized in that, The method for determining the external collision state of the current premise triplet includes: If the semantic relation judgment result is a semantic contradiction, then the current premise triplet is marked as an external conflicting premise and the conflict pair formed by the current premise triplet and the nearest neighbor verified knowledge triplet is recorded; If the semantic relation is determined to be semantically consistent, then the current premise triplet is marked as an externally consistent premise.

5. The guided question-answering method based on implicit premise triple extraction according to claim 4, characterized in that, The method for determining the external collision state of the current premise triplet further includes: If the semantic relation is determined to be semantically irrelevant, then the verified knowledge triplet of conditional dependency is retrieved from the domain knowledge base. If found, the current premise triplet is marked as a conditional dependency premise and the conditional dependency pairing formed by the current premise triplet and the verified knowledge triplet of conditional dependency is recorded. If not found, the current premise triplet is marked as no matching premise.

6. The guided question-answering method based on implicit premise triple extraction according to claim 5, characterized in that, The method for generating the prerequisite verification report includes: The internal consistency states and external collision states of each premise triplet are merged to form a set of premise triplets labeled with conflict types. Iterate through the internal consistency states and external collision states of all premise triples in the set of premise triples labeled with conflict types. When there is at least one premise triple with an internal consistency state that is internally conflicting, or an external collision state that is an externally conflicting premise, or an external collision state that is a conditional dependency premise, summarize the user question text, the set of premise triples labeled with conflict types, contradictory premise triple pairs and explanations of the contradiction, and conflict pairs and conditional dependency pairings into a premise verification report.

7. The guided question-answering method based on implicit premise triple extraction according to claim 6, characterized in that, The method for generating the guided rhetorical question includes: Extract all premise triples from the premise verification report whose internal consistency state is internally conflicting or whose external collision state is externally conflicting or conditionally dependent premises as conflicting premise triples. Based on the category of the internal consistency state or external collision state corresponding to each conflicting premise triple, adopt the corresponding guided question generation strategy to generate guided questions for each conflicting premise triple.

8. The guided question-answering method based on implicit premise triple extraction according to claim 7, characterized in that, The value of the conflict premise correction state is either corrected or not corrected; The method for correcting the conflicting premises of each conflicting premise triplet in the verification premise report includes: re-performing implicit premise triplet extraction, internal consistency verification, and external domain knowledge collision detection on the user's response text to obtain a set of premise triplets labeled with conflicting types in the user's response text; comparing each conflicting premise triplet in the premise verification report with the set of premise triplets labeled with conflicting types in the user's response text, and marking the conflicting premises correction status of each conflicting premise triplet in the premise verification report.

9. A guided question-answering system based on implicit premise triple extraction, used to implement the guided question-answering method based on implicit premise triple extraction as described in any one of claims 1-8, characterized in that, The system includes: Premise detection module: This module receives user question text, performs implicit premise triple extraction on the user question text to construct a premise triple set, performs multi-level conflict detection on the premise triple set, labels the internal consistency state and external collision state of each premise triple in the premise triple set, and summarizes and generates a premise verification report; the multi-level conflict detection includes internal consistency verification and external domain knowledge collision detection. The guidance and correction module generates guided rhetorical questions based on the internal consistency and external collision states of each premise triplet, using corresponding guided rhetorical question generation strategies. It receives user response texts for the guided rhetorical questions and verifies the conflict premise correction states of each conflicting premise triplet in the premise verification report. Based on the conflict premise correction states, it performs closed-loop correction and outputs the final response based on the correct premises.

Images