Reply determination method based on large model decision, storage medium and electronic device

Abstract

The application discloses a reply determination method based on a large model decision, a storage medium and an electronic device, relates to the technical field of artificial intelligence, and the reply determination method based on the large model decision comprises the following steps: obtaining a query statement sent by a user end, querying in a question and answer decision tree to obtain an intention matching node matched with a query intention, and then matching the query statement with first node content of the intention matching node and second node content of a downstream node of the intention matching node respectively to obtain a matching result. In this way, the query statement can be matched with the first node content and the second node content respectively at one time, a reply node matched with the matching result is determined from the question and answer decision tree starting from the intention matching node, and the reply content of the query statement is determined according to node content of the reply node, so that the number of question and answer rounds can be reduced, and the reply determination efficiency is effectively improved.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and more specifically, to a method for determining responses based on large model decision-making, a storage medium, and an electronic device. Background Technology

[0002] With the rapid development of artificial intelligence and natural language processing technologies, intelligent customer service systems are being used more and more widely in smart homes, e-commerce, customer service and other scenarios.

[0003] Currently, intelligent question-answering systems are mainly based on entity relationship matching in structured knowledge graphs. Specifically, by constructing a knowledge graph, expert experience is abstracted into triple relationships of "phenomenon-possible cause-solution". Entity recognition and relationship extraction technologies are then used to match user questions with nodes in the knowledge graph, thereby obtaining the answer.

[0004] However, while such systems are efficient in handling clear and simple descriptions, they cannot cope with complex dialogue situations in real-world scenarios where user expressions are ambiguous, information is fragmented, and multiple factors are intertwined. In complex dialogue situations, it is necessary to ask questions layer by layer, resulting in redundant interactions and low efficiency.

[0005] Therefore, there is an urgent need for a new response determination method based on large model decision-making to overcome the shortcomings of conventional techniques and improve response determination efficiency. Summary of the Invention

[0006] This application provides a response determination method and apparatus, storage medium, and electronic device based on large model decision-making, to at least solve the problem of low response determination efficiency.

[0007] According to one aspect of the embodiments of this application, a response determination method based on large model decision-making is provided, comprising: acquiring a query statement sent by a user terminal; querying a pre-constructed question-and-answer decision tree based on the query intent of the query statement to obtain an intent matching node that matches the query intent; wherein, the question-and-answer decision tree includes: multiple nodes; each node includes: node content; matching the query statement with the first node content of the intent matching node and the second node content of the downstream node of the intent matching node respectively to obtain a matching result; the node content includes: the first node content and the second node content; starting from the intent matching node, determining a response node that matches the matching result from the question-and-answer decision tree, and determining the response content of the query statement based on the node content of the response node.

[0008] In an exemplary embodiment, the question-answering decision tree further includes: response options corresponding to the node content; the step of matching the query statement with the first node content of the intent-matching node and the second node content of the downstream node of the intent-matching node to obtain a matching result includes: matching the query statement with the first node content of the intent-matching node to obtain a first matching result, and matching the query statement with the second node content of the downstream node of the intent-matching node to obtain a second matching result; determining a first response content from the first response options of the intent-matching node based on the first matching result, and determining a second response content from the second response options of the downstream node based on the second matching result; and determining a matching result based on the first response content and the second response content.

[0009] In an exemplary embodiment, the step of matching the query statement with the first node content of the intent matching node to obtain a first matching result, and matching the query statement with the second node content of the downstream node of the intent matching node to obtain a second matching result, includes: inputting the query statement, the first node content, and the second node content into a pre-trained semantic judgment model; matching the query statement and the first node content according to preset semantic judgment instructions through the semantic judgment model to obtain a first matching result, and matching the query statement and the second node content according to preset semantic judgment instructions to obtain a second matching result.

[0010] In an exemplary embodiment, before inputting the query statement, the first node content, and the second node content into a pre-trained semantic judgment model, the method further includes: obtaining sample node content and obtaining sample query statements; the plurality of sample node contents include: first sample node content and second sample node content; inputting the first sample node content, the second sample node content, and the sample query statement into a large model according to a preset input format; using the large model, matching the first sample node content, the second sample node content, and the sample query statement according to preset semantic judgment instructions to obtain sample matching results; based on the sample matching results and the labeled matching results, adjusting the parameters of the large model and continuing training until training is completed to obtain a trained semantic judgment model.

[0011] In an exemplary embodiment, the matching result includes: a first response content and a second response content; any two nodes among the plurality of nodes are connected by directed edges; the step of determining the response node matching the matching result from the question-and-answer decision tree, starting from the intent matching node, includes: determining the content of a first directed edge connected to at least one directed edge of the intent matching node; performing semantic matching between the first directed edge content and the first response content to determine a first target directed edge of the intent matching node, and determining a target downstream node from the downstream nodes of the intent matching node based on the first target directed edge; determining the content of a second directed edge connected to at least one directed edge of the target downstream node; performing semantic matching based on the second response content and the second directed edge content to determine a second target directed edge of the target downstream node, continuing to determine the next target downstream node based on the second target directed edge, and returning to the step of determining the content of a second directed edge connected to at least one directed edge of the target downstream node to continue execution until a preset condition is met, thus determining the response node.

[0012] In an exemplary embodiment, before querying a pre-constructed question-and-answer decision tree based on the query intent of the query statement to obtain an intent-matching node that matches the query intent, the method further includes: determining M question statements from a pre-defined query library, and determining the question statement responses for the M question statements; the query library includes: question statements, and question statement responses corresponding to the question statements; M is an integer greater than 1; determining a node table based on the M question statements, and determining a directed edge table based on the question statement responses for the M question statements; and constructing a question-and-answer decision tree based on the node table and the directed edge table.

[0013] In an exemplary embodiment, the query intent based on the query statement is queried in a pre-constructed question-answering decision tree to obtain intent matching nodes that match the query intent. This includes: performing vectorization representation processing on the query intent to obtain an intent vector representation of the query intent, and performing vectorization representation processing on the node content of the plurality of nodes to obtain content vector representations of the plurality of nodes; performing similarity matching based on the intent vector representation and the content vector representations of the plurality of nodes to obtain similarity matching results of the plurality of nodes; and determining the nodes corresponding to the similarity matching results that meet preset matching conditions among the similarity matching results of the plurality of nodes as intent matching nodes.

[0014] According to another aspect of the embodiments of this application, a response determination device based on large model decision-making is also provided, comprising: an acquisition module, configured to acquire a query statement sent by a user terminal; a query module, configured to perform a query in a pre-constructed question-and-answer decision tree based on the query intent of the query statement, to obtain an intent matching node that matches the query intent; wherein, the question-and-answer decision tree includes: multiple nodes; each node includes: node content; a matching module, configured to match the query statement with the first node content of the intent matching node and the second node content of the downstream node of the intent matching node, respectively, to obtain a matching result; the node content includes: the first node content and the second node content; and a determination module, configured to determine, starting from the intent matching node, a response node that matches the matching result from the question-and-answer decision tree, and determine the response content of the query statement based on the node content of the response node.

[0015] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, wherein a computer program is stored in the computer program, and the computer program is configured to execute the above-described response determination method based on large model decision-making when running.

[0016] According to another aspect of the embodiments of this application, an electronic device is also provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the above-described response determination method based on large model decision through the computer program.

[0017] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program, which, when executed by a processor, provides the above-described response determination method based on large model decision-making.

[0018] This application obtains the query statement sent by the user, queries it in the question-and-answer decision tree to obtain the intent matching node that matches the query intent, and then matches the query statement with the first node content of the intent matching node and the second node content of the downstream node of the intent matching node to obtain the matching result. In this way, the query statement can be matched with the first node content and the second node content at the same time. Furthermore, starting from the intent matching node, the answer node that matches the matching result is determined from the question-and-answer decision tree, and the answer content of the query statement is determined based on the node content of the answer node. This can reduce the number of question-and-answer rounds and effectively improve the efficiency of answer determination. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a schematic diagram of the hardware environment for a response determination method based on large model decision-making according to an embodiment of this application;

[0022] Figure 2 This is one of the flowcharts of a response determination method based on large model decision-making according to an embodiment of this application;

[0023] Figure 3 This is a second flowchart of a response determination method based on large model decision-making according to an embodiment of this application;

[0024] Figure 4 This is a structural block diagram of a response determination device based on large model decision-making according to an embodiment of this application;

[0025] Figure 5 This is a schematic diagram of the structure of an optional electronic device according to an embodiment of this application. Detailed Implementation

[0026] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0027] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0028] According to one aspect of the embodiments of this application, a response determination method based on large model decision-making is provided. This response determination method based on large model decision-making is widely applicable to whole-house intelligent digital control application scenarios such as smart homes, smart home ecosystems, and intelligence house ecosystems. Optionally, in this embodiment, the above-mentioned response determination method based on large model decision-making can be applied to, for example... Figure 1 The hardware environment shown consists of terminal device 102 and server 104. For example... Figure 1 As shown, server 104 is connected to terminal device 102 via a network and can be used to provide services (such as application services) to the terminal or clients installed on the terminal. A database can be set up on the server or independently of the server to provide data storage services for server 104. Cloud computing and / or edge computing services can be configured on the server or independently of the server to provide data processing services for server 104. Specifically, terminal 102 can send a query statement to server 104. Server 104 can be equipped with an intelligent question-and-answer system. The intelligent question-and-answer system can obtain the query statement sent by terminal 102, and perform a query in a pre-constructed question-and-answer decision tree based on the query intent of the query statement to obtain an intent matching node that matches the query intent. The question-and-answer decision tree includes multiple nodes. Each node includes node content. The query statement is matched with the first node content of the intent matching node and the second node content of the downstream node of the intent matching node to obtain a matching result. The node content includes the first node content and the second node content. Starting from the intent matching node, the answer node that matches the matching result is determined from the question-and-answer decision tree, and the answer content of the query statement is determined based on the node content of the answer node.

[0029] The aforementioned network may include, but is not limited to, at least one of the following: wired network, wireless network. The aforementioned wired network may include, but is not limited to, at least one of the following: wide area network, metropolitan area network, local area network. The aforementioned wireless network may include, but is not limited to, at least one of the following: Wi-Fi (Wireless Fidelity), Bluetooth. The terminal device 102 may not be limited to PC, mobile phone, tablet computer, smart air conditioner, smart range hood, smart refrigerator, smart oven, smart stove, smart washing machine, smart water heater, smart washing equipment, smart dishwasher, smart projector, smart TV, smart clothes rack, smart curtains, smart audio-visual equipment, smart socket, smart speaker, smart speaker box, smart fresh air equipment, smart kitchen and bathroom equipment, smart bathroom equipment, smart robot vacuum cleaner, smart window cleaning robot, smart mopping robot, smart air purifier, smart steam oven, smart microwave oven, smart water heater, smart air purifier, smart water dispenser, smart door lock, etc.

[0030] To address the aforementioned issues, this embodiment provides a response determination method based on large model decision-making, which includes, but is not limited to, applications in whole-house intelligent digital control scenarios. Figure 2 This is a flowchart of a response determination method based on large model decision-making according to an embodiment of this application, the process including the following steps S202-S208:

[0031] Step S202: Obtain the query statement sent by the user;

[0032] The query is sent from the user's client to the intelligent question-answering system. Users can input their query on the user's client-side interface (such as a chatbot, voice assistant, or search box) when interacting with the intelligent question-answering system. The query can be entered directly as text or converted from the user's voice input.

[0033] For example, the query statement is a user's description of the fault phenomenon, equipment status, or specific request in everyday language.

[0034] Understandably, for text input, the intelligent question-answering system can receive it through a communication interface, including the user's query. For voice input, the intelligent question-answering system can first convert the voice signal into text using speech recognition technology, and then receive this text data.

[0035] In some embodiments, the server can receive query statements sent from user terminals such as mobile apps, smart voice terminals, WeChat mini-programs, and web pages. The server can cleanse, anonymize, and filter illegal characters from the raw input to ensure that the statement content conforms to security and semantic specifications. Subsequently, the input statement undergoes semantic validity testing, automatically blocking invalid, meaningless, or irrelevant requests. Finally, a standardized structured query statement is output, including the cleaned text, session context, device information, and input source, serving as the sole input source for subsequent intent recognition and matching.

[0036] Step S204: Based on the query intent of the query statement, perform a query in the pre-constructed question-and-answer decision tree to obtain intent matching nodes that match the query intent; wherein, the question-and-answer decision tree includes: multiple nodes; each node includes: node content;

[0037] It should be noted that a question-and-answer decision tree can be a structured representation of a user's troubleshooting needs. A question-and-answer decision tree includes multiple nodes. Each node corresponds to a specific step in the troubleshooting process, and each node contains clearly defined content. Node content includes, but is not limited to: questions posed to the user by the intelligent question-and-answer system during the dialogue, statuses requiring user confirmation, or troubleshooting suggestions and solutions provided by the intelligent question-and-answer system. For example, one node's content might be "Is the refrigerator compressor running?", while another node's content might be "We suggest you check if the power plug is loose."

[0038] For example, the system receives a query statement input by the user. The query statement can specifically describe a fault phenomenon encountered by the user, a device model, or a specific problem description. Further semantic analysis and feature extraction are performed on the query statement, mapping it to a predefined set of standard intents. This transforms the query statement into standard intent labels; for example, "water on the refrigerator" is mapped to the standard intent "refrigerator leaking water," or "air conditioner not cooling" is mapped to the intent "air conditioner cooling failure." Next, a pre-built question-and-answer decision tree data structure is accessed. A query matching operation is performed. Using the standard query intent identified in the previous step as the index key, a search is conducted in the pre-built question-and-answer decision tree to determine the intent matching node that perfectly matches the query intent. For example, when the query intent is identified as "refrigerator exterior is dirty / scratched," the corresponding root node or first-level node is found in the question-and-answer decision tree. The node content of this node is the first round of questions for this fault type, such as "Is it dirt or scratches on the exterior?"

[0039] In an exemplary embodiment, the query intent based on the query statement is queried in a pre-constructed question-answering decision tree to obtain intent matching nodes that match the query intent. This includes: performing vectorization representation processing on the query intent to obtain an intent vector representation of the query intent, and performing vectorization representation processing on the node content of multiple nodes to obtain content vector representations of multiple nodes; performing similarity matching based on the intent vector representation and the content vector representations of multiple nodes to obtain similarity matching results of multiple nodes; and determining the nodes corresponding to the similarity matching results that meet the preset matching conditions among the similarity matching results of multiple nodes as intent matching nodes.

[0040] Understandably, query intent is vectorized to generate its corresponding intent vector representation. This process can employ a pre-trained language model to semantically encode query intents (such as "refrigerator exterior is dirty / scratched" or "gas water heater reports F4 code"), mapping natural language intents into a dense vector of fixed dimensions (e.g., 768-dimensional or 1024-dimensional). This vector can capture the deep semantic features of the intent, rather than relying solely on keyword matching.

[0041] Simultaneously, the node content of multiple nodes in the question-answering decision tree is vectorized to generate corresponding node content vector representations. The node content consists of expert-preset diagnostic questions or solution statements (such as "Is it dirt or scratches on the exterior?", "Please check if there is a protective film?", "We will arrange for after-sales service personnel to come to your door"). The vectorization process can use the same model architecture and parameters as the intent vectorization to ensure semantic space alignment.

[0042] Subsequently, the intent vector representation and the content vector representations of multiple nodes are matched using cosine similarity in a unified semantic vector space. The semantic similarity score between each node and the current query intent is calculated, forming a list of matching results containing multiple similarity values. Since each node in the question-answering decision tree is bound to a clear semantic context, this matching process can effectively identify nodes that are semantically similar but express different meanings. For example, both "scratches on the refrigerator door" and "abrasions on the refrigerator exterior" can be matched to the node corresponding to "dirt / scratches on the refrigerator exterior".

[0043] Optionally, preset matching conditions can be set, including but not limited to: a similarity threshold ≥ 0.85; retaining only the top-1 matching results, etc. Among the similarity results that meet the preset matching conditions, the node with the highest score and that meets the threshold is determined as the intent matching node, that is, the node in the question-answering decision tree that is most similar to the current query intent semantics. This node is the starting node of the subsequent dialogue process, and will be used to initiate the first round of questions or trigger the large language model to perform multi-condition semantic prediction.

[0044] In the above embodiments, the semantic vector similarity matching mechanism significantly improves the semantic generalization ability for users' non-standardized expressions. This mechanism not only achieves fine-grained semantic alignment between intents and nodes, but also realizes end-to-end integration of intent recognition and path localization based on the question-answering decision tree structure, providing a high-precision and robust starting point for subsequent multi-node conditional matching and intelligent dialogue flow.

[0045] Step S206: Match the query statement with the first node content of the intent matching node and the second node content of the downstream node of the intent matching node to obtain the matching results; the node content includes: the first node content and the second node content;

[0046] It's understandable that the first node content refers to the node content corresponding to the intent matching node. The intent matching node is the starting node or the first key branching node located in the question-and-answer decision tree after identifying the fault intent based on the user's query (e.g., "refrigerator exterior is dirty / scratched"). The first node content is typically the first round of core queries for the current fault type. For example, in the question-and-answer decision tree for "refrigerator exterior is dirty / scratched," the first node content of the intent matching node could be "Is it dirt or scratches on the exterior?"

[0047] The second node content refers to the node content corresponding to the downstream node of the intent-matching node. A downstream node is the next node logically the user should jump to based on the answer to the first node content (i.e., the edge / condition in the question-answering decision tree). Downstream nodes can include direct downstream nodes (directly connected to the intent-matching node) and indirect downstream nodes (indirectly connected to the intent-matching node). The second node content represents a second or more rounds of queries that further subdivide the problem. For example, if the answer to the first node content is "scratch," the user might jump to a downstream node, whose second node content could be, "Please check if there is a film on the outer surface of the refrigerator?" or "Please use a cleaning agent to wipe it; see if you can remove the scratch?" The second node content is used to predict whether the user's current query or historical dialogue has already implicitly answered these subsequent questions, thus achieving "jump-to-answer."

[0048] In an exemplary embodiment, the question-answering decision tree further includes: response options corresponding to node content; matching the query statement with the first node content of the intent-matching node and the second node content of the downstream node of the intent-matching node to obtain a matching result, including: matching the query statement with the first node content of the intent-matching node to obtain a first matching result, and matching the query statement with the second node content of the downstream node of the intent-matching node to obtain a second matching result; determining a first response content from the first response options of the intent-matching node based on the first matching result, and determining a second response content from the second response options of the downstream node based on the second matching result; and determining a matching result based on the first response content and the second response content.

[0049] For example, the content of the first node and the content of the second node are organized into a list of questions to be judged. Each question is associated with answer options (such as "yes / no", "not sure", etc.; "dirty / scratched", "not sure", etc.). Based on the query "The refrigerator has scratches on its exterior", what is the user's answer to the corresponding question "Is it dirt or scratches on the exterior?"? The system identifies that "scratches" and "scratches" in the options are semantically consistent, obtaining the first matching result: outputting "scratches". Based on the fact that the user has confirmed "there are scratches", and whether the query contains other information, for the subsequent question "Please check if there is a film on the outer surface of the refrigerator?", has the user provided relevant information or an answer? "If a user's query is simply 'scratches on the refrigerator' without mentioning the film, the large model might infer that the user hasn't answered the question yet, or infer based on common sense that further confirmation is needed. If the query is 'scratches on the side of the refrigerator, still there after removing the film,' semantic reasoning identifies that 'still there after removing the film' implicitly answers 'whether there is a film' and 'the state after removing the film,' resulting in a second matching result: outputting 'yes' (regarding whether there is a film) or 'the film has been removed and it exists' (regarding the state after removing the film). Matching the query with the first node content of the intent matching node and the second node content of the downstream node of the intent matching node is essentially a multi-turn dialogue compression process based on the large model's semantic prediction. Through this matching mechanism, it can intelligently identify the information already provided by the user, avoid redundant questioning, and achieve a technological leap from passive question answering to proactive prediction and decision-making."

[0050] Step S208: Starting from the intent matching node, determine the response node that matches the matching result from the question-answering decision tree, and determine the response content of the query statement based on the node content of the response node.

[0051] It is understandable that, starting from the intent matching node, and combining the generated matching results, path search and node location are performed in the question-and-answer decision tree to determine the response node. The response node refers to the next node that needs to be interacted with or serves as a conclusion, as determined by the matching results.

[0052] In an exemplary embodiment, the matching result includes: first response content and second response content; any two nodes among multiple nodes are connected by directed edges; starting from the intent matching node, determining the response node that matches the matching result from the question-and-answer decision tree includes: starting from the intent matching node, determining the content of a first directed edge connected to at least one directed edge of the intent matching node; performing semantic matching between the first directed edge content and the first response content to determine the first target directed edge of the intent matching node, and determining the target downstream node from the downstream nodes of the intent matching node based on the first target directed edge; determining the content of a second directed edge connected to at least one directed edge of the target downstream node; performing semantic matching based on the second response content and the second directed edge content to determine the second target directed edge of the target downstream node, continuing to determine the next target downstream node based on the second target directed edge, and returning to the step of determining the content of the second directed edge of at least one directed edge connected to the target downstream node to continue execution until a preset condition is met, and the response node is determined.

[0053] It's important to note that, starting from the intent-matching node, all directed edges directly connected to it are identified. A directed edge represents a "condition" or "branch of the user's answer." For example, if the intent-matching node's question is "Is it dirty or scratched?", there are two outgoing edges: one labeled "dirty" and the other labeled "scratched." The label text of these two edges is extracted, which is the content of the first directed edge. The content of the first directed edge (i.e., all possible answer options, such as {"dirty", "scratched", "unsure"}) is semantically matched with the content of the first response, and the edge with the highest semantic consistency and confidence is found. This selected edge is the first target directed edge. For example, if the first response is "scratched", then the edge labeled "scratched" is determined as the target directed edge.

[0054] Following the direction pointed to by the first target directed edge, locate the downstream node of the target. Once the downstream node is determined, use it as the new center and find all directed edges connected to it. Extract the label text of these edges, i.e., the content of the second directed edge. For example, if the question of the downstream node is "Is there a coating?", the content of the second directed edge would be {"Yes", "No", "Unsure"}. Semantically match the content of the second directed edge with the content of the second response. Determine the second directed edge that is semantically consistent with the second response. The edge that matches successfully is determined as the second target directed edge. Jump along the second target directed edge to determine the next target downstream node. The process of obtaining edge content, semantic matching, determining the second target directed edge, and jumping to the node is a loop that stops when a preset condition is met.

[0055] In some embodiments, after reaching a new target downstream node, if the target downstream node still has outgoing edges, the content of the second directed edge connected to at least one directed edge of the target downstream node is determined. Semantic matching is performed based on the second response content and the second directed edge content, and the question-answering decision tree is continuously explored along the path determined by semantic matching. The loop stops when any of the following situations occur, and the node determined at this time is the final response node: 1. Reaching the end node: The current node is a leaf node of the fault tree, usually corresponding to "Solution", "No need to handle", or "Arrange after-sales service". At this time, there is no need to ask any more questions, and the content of this node is directly output. 2. No matching edge or insufficient information: It is impossible to infer the answer to the next round of questions from the user input, or the semantic matching confidence is lower than the threshold. At this time, the traversal stops, and the current node is the response node that needs to ask the user a question. 3. Reaching the maximum depth: To prevent infinite loops or excessively deep queries, a maximum number of traversal levels is set. If the level limit is reached, the loop stops at the current target downstream node. When the loop stops, the target downstream node is the response node.

[0056] In the above embodiments, the first response content drives the jump from the intent matching node to the first-level downstream node, and the second response content drives the jump from the first-level downstream node to the second-level downstream node. By continuously comparing the response content with the edge labels in the question-and-answer decision tree, it is possible to skip questions that have already been implicitly answered by the user and directly locate the logically deepest node. This ensures that the dialogue conforms to the rigorous logic of the question-and-answer decision tree while also possessing the flexibility of natural language understanding, thereby achieving efficient and accurate fault question answering.

[0057] In steps S202-S208 above, the query statement sent by the user is obtained. By querying in the question-and-answer decision tree, the intent matching node that matches the query intent is obtained. Then, the query statement is matched with the first node content of the intent matching node and the second node content of the downstream node of the intent matching node to obtain the matching result. In this way, the query statement can be matched with the first node content and the second node content at the same time. Furthermore, starting from the intent matching node, the answer node that matches the matching result is determined from the question-and-answer decision tree. The answer content of the query statement is determined according to the node content of the answer node. This can reduce the number of question-and-answer rounds and effectively improve the efficiency of answer determination.

[0058] In an exemplary embodiment, matching a query statement with the content of a first node of an intent-matching node to obtain a first matching result, and matching a query statement with the content of a second node of a downstream node of the intent-matching node to obtain a second matching result, includes: inputting the query statement, the content of the first node, and the content of the second node into a pre-trained semantic judgment model; matching the query statement and the content of the first node according to preset semantic judgment instructions through the semantic judgment model to obtain a first matching result, and matching the query statement and the content of the second node according to preset semantic judgment instructions to obtain a second matching result.

[0059] Understandably, the query statement, the content of the first node, and the content of the second node are input into a pre-trained semantic judgment model. The semantic judgment model is a semantic reasoning engine fine-tuned based on a large language model (such as Qwen2-7B-Instruct), which is configured to execute preset semantic judgment instructions for parallel semantic alignment and answer prediction between the user query statement and multiple question-answer pairs.

[0060] Optionally, the semantic judgment instruction can be structured as follows: You are a home appliance troubleshooting expert. Based on the user's original description, please determine the standard answer options for each of the following multiple questions. User description: {query statement} Question 1: {first node content}, optional answer: {first response option}; Question 2: {second node content}, optional answer: {second response option}. Please return the best answer options for each question strictly in JSON array format, in the same order as the input, without adding any explanations or descriptions.

[0061] During execution, the semantic judgment model performs multiple semantic matching operations in parallel according to the above instructions. For example, the first matching: based on the query statement and the content of the first node (e.g., "Is it dirty or scratched?") and its corresponding first answer options (e.g., ["dirty", "scratched", "no_answer"]), the most matching candidate answer is inferred, and the first matching result (e.g., "scratched") is obtained. The second matching: based on the query statement and the content of the second node (e.g., "Please check if there is a film on the outer surface of the refrigerator?") and its corresponding second answer options (e.g., ["yes", "no", "no_answer"]), the most matching candidate answer is inferred, and the second matching result (e.g., "no") is obtained. Subsequently, based on the first matching result, the corresponding first answer content is extracted from the first answer option; similarly, based on the second matching result, the corresponding second answer content is extracted from the second answer option.

[0062] In the above embodiments, the mechanism of matching node content through semantic judgment model enables a single model to complete semantic reasoning of multiple dialogue nodes at once without relying on round-by-round interaction. This greatly reduces user waiting time and customer service system load, and significantly improves the response efficiency and automation level of intelligent question answering.

[0063] In an exemplary embodiment, before inputting the query statement, the content of the first node, and the content of the second node into the pre-trained semantic judgment model, the method further includes: obtaining sample node content and obtaining sample query statements; the multiple sample node contents include: the content of the first sample node and the content of the second sample node; inputting the content of the first sample node, the content of the second sample node, and the sample query statement into the large model according to a preset input format; using the large model, matching the content of the first sample node, the content of the second sample node, and the sample query statement according to preset semantic judgment instructions to obtain sample matching results; based on the sample matching results and the labeled matching results, adjusting the parameters of the large model and continuing training until training is completed to obtain a trained semantic judgment model.

[0064] Specifically, sample node content is collected and constructed from historical after-sales dialogue logs, expert knowledge bases, and fault tree specification documents. Sample node content includes, but is not limited to, first sample node content and second sample node content. Each sample node content consists of a "question text" and a "set of preset standard answer options," for example: First sample node content: {"question":"Is it dirt or scratches on the exterior?","options":["dirt","scratches","no_answer"]}; Second sample node content: {"question":"Please check if there is a film on the outer surface of the refrigerator?","options":["yes","no","no_answer"]}.

[0065] Simultaneously, sample query statements corresponding to the above sample node content were collected, namely the original descriptions entered by real users during the consultation process, such as: "There are scratches on the refrigerator door that I can't wipe off no matter what I do" or "I see scratches on the outside, but there's no dirt." These statements are characterized by linguistic diversity, unstructured language, and colloquialism, truly reflecting users' expression habits.

[0066] Next, the content of the first sample node, the content of the second sample node, and the corresponding sample query statement are combined into a structured prompt according to a unified preset input format and input into the large language model (e.g., Qwen2-7B-Instruct). The input format is as follows: You are a home appliance troubleshooting expert. Based on the user's original description, please select the best matching option from the preset answer choices for the following questions.

[0067] User description: {Sample query statement}; Question 1: {Question containing the content of the first sample node}, Optional answers: {Set of options for the content of the first sample node}; Question 2: {Question containing the content of the second sample node}, Optional answers: {Set of options for the content of the second sample node}. Please return only the best answer options for each question as a JSON array, in the same order as the input, without adding any explanatory text. During the model inference phase, obtain the sample matching results output by the model, i.e., the sequence of standard answer options predicted by the model for each question.

[0068] Subsequently, the sample matching result is compared with the manually labeled true standard answer (i.e., the labeled matching result) to construct a training loss function. The labeled matching result is determined manually by after-sales experts, customer service managers, or the rule engine based on the fault tree standard path to ensure the authority and consistency of the labels.

[0069] Based on the comparison results, supervised learning methods (such as cross-entropy loss) are used to fine-tune the parameters of the large model to be trained. Combined with efficient parameter update strategies such as LoRA (Low-Rank Adaptation), only some low-rank matrices are optimized to reduce training costs and maintain the stability of pre-trained knowledge.

[0070] In the above embodiments, a customized training mechanism was implemented for an end-to-end semantic judgment model oriented towards multi-turn question-answering decision trees. Compared with traditional models that only train intent classification models or single-turn question-answering models, this application significantly improves the model's contextual understanding ability and answer prediction accuracy in complex multi-hop diagnostic scenarios, laying a reliable and reproducible model foundation for realizing an intelligent diagnostic system that achieves one-time input, multi-point hits, and direct solution access.

[0071] In an exemplary embodiment, before the query intent based on the query statement is queried in a pre-constructed question-and-answer decision tree to obtain an intent-matching node that matches the query intent, the response determination method based on the large model decision further includes: determining M question statements from a pre-set query library, and determining the question statement responses for the M question statements; the query library includes: question statements, and question statement responses corresponding to the question statements; M is an integer greater than 1; determining a node table based on the M question statements, and determining a directed edge table based on the question statement responses for the M question statements; and constructing a question-and-answer decision tree based on the node table and the directed edge table.

[0072] Understandably, the system extracts M question statements (where M is a positive integer greater than 1) from a pre-defined query database and simultaneously retrieves the corresponding responses. The query database can be a structured knowledge base compiled from appliance after-sales service experts, technical documents, and historical work orders. Each record contains two core fields: Question Statement: Standardized queries posed to the user during fault diagnosis, such as: "Is it dirt or scratches?"; "Please check if there is a film on the outer surface of the refrigerator?"; "Is the angle valve open?". Question Statement Response: Standardized response options that the user might give to the above questions, such as: "Dirty", "Scratched", "no_answer"; "Yes", "No", "no_answer"; "On", "Off", "no_answer".

[0073] Next, a node table is constructed using the statement identifiers of the M question statements as node identifiers. Each record corresponds to a diagnostic node, containing a unique node ID, node content (i.e., the question statement), the fault type ID, and a flag indicating whether it is a starting node. Correspondingly, a directed edge table is constructed based on the response representations of the answers to the M question statements. Finally, based on the node table and the directed edge table, a complete question-answering decision tree is constructed using a graph database or adjacency list structure. The question-answering decision tree has the following structural characteristics: each node is a diagnostic question, corresponding to a question statement; each directed edge is a semantic condition, uniquely identified by the question statement response; it converges to a terminal node (out-degree 0), the content of which is the preset final solution; the graph structure is a directed acyclic graph (DAG), ensuring that the diagnostic path is non-cyclic and terminateable.

[0074] In the above embodiments, the constructed question-answering decision tree fully retains the diagnostic logic, and its topology can support efficient navigation and automated reasoning based on user intent.

[0075] Obviously, the embodiments described above are only some embodiments of this application, and not all embodiments. To better understand the above method, the following description, in conjunction with embodiments, illustrates the process, but is not intended to limit the technical solutions of the embodiments of this application:

[0076] This embodiment uses home appliance malfunctions as an example. The home appliance fault tree involves different appliances and different fault types. These fault trees are compiled from the valuable experience accumulated by experts, customer service personnel, and after-sales repair personnel. Utilizing this fault tree information effectively is key to intelligent question answering and rapid, efficient problem resolution. Therefore, intelligent question answering based on home appliance fault trees has become an important task and a challenge. Refer to Table 1 for the fault tree data format:

[0077] Table 1

[0078]

[0079]

[0080]

[0081] From a data perspective, following the dialogue logic, we would need to find the node corresponding to the user's answer in the fault tree data format table, then trace back to the parent node's question based on that node, and then ask the user that same question again. This process is overly cumbersome. To make it easier to operate and understand, we convert the above data into a directed acyclic graph, as shown in the reference... Figure 3 The diagram shown is a schematic representation of the structure of a directed acyclic graph. For ease of understanding, node identifiers are used instead of node content. Figure 3 In the code, all nodes are stored in the form of (nodes inside circles (e.g., 20781875), directed edges (e.g., 21276106)). The following describes... Figure 3 The meanings of each node and directed edge are explained below: 20781875: Dirt / scratches on the refrigerator exterior; 21276105: Is it dirt or scratches on the exterior?; 21276107: Scratches; 21276109: Please check if there is a protective film on the outer surface of the refrigerator; 21276110: Yes; 21276112: Please check if the scratches remain after removing the protective film; 21276118: The refrigerator comes with a protective film to prevent scratches. You can use it with confidence after removing it; 21276117: Yes; 21276120: Please use a cleaning agent to wipe it and see if you can remove the scratches; 21276122: Cannot be removed; 21276125: We will arrange for an after-sales technician to provide on-site service; 21276113: None; 21276119: Please use a cleaning agent to wipe it and see if you can remove the scratches; 21276121: Cannot be removed; 21276126: We will arrange for an after-sales technician to provide on-site service; 21276123: Can be removed; 21276124: Can be removed; 21276127: End point; 21276106: Dirt; 21276108: Please use a cleaning agent to wipe it and see if you can remove the dirt; 21276111: Cannot be removed; 21276116: We will arrange for an after-sales technician to provide on-site service; 21276115: Can be removed. Examples are as follows: (1) Intent classification: When a user says "the refrigerator is dirty", "the refrigerator is scratched", or "the refrigerator is dirty / scratched", the intent classification model indicates that the intent is "the refrigerator is dirty / scratched"; (2) Confirmation node: At this time, based on the intent obtained in the first step, "21276105" is found to be the conversation start node, i.e., the intent matching node; (3) Further construct the question-answer pair based on the intent matching node; (4) Input the question-answer pair into the large language model.

[0082] The prompt description of the large language model is as follows: You are now a question language expert. You need to fully understand the given text. The following questions may be further questions to be asked. Please deduce which questions may be asked next. Combine the given text and questions and find one of the multiple answers given: For example, if the air conditioner is in heating mode, then it must be powered on and in heating mode.

[0083] Here are the basic requirements: 1. If no answer is found, give "no_answer"; 2. If the text does not mention anything related to the question, give "no_answer"; 3. Carefully analyze the given answers and text, find the relevant text content, and deduce the answer that is closest in meaning; 4. The format must be JSON, and the outermost layer must contain []; 5. Return all given question-answer pairs, regardless of the answer, without omission.

[0084] Here is an example: Text: The water from my under-sink water purifier has a taste. It's a brand new machine, and the kettle is also working properly. What should I do? Questions: [(question: Is this a newly installed machine? Answer: [Yes, No, no_answer]), (question: Was the filter cartridge recently replaced? Answer: [Yes, No, no_answer]), (question: Does the kettle have an odor or is it a new kettle? Answer: [Yes, No, no_answer]), (question: Does the water container have an odor or is it a new water container (plastic, paper cup, etc.)? Answer: [Yes, No, no_answer])]. Returns: [(question: Is this a newly installed machine? Answer: Yes), (question: Was the filter cartridge just replaced? Answer: No_answer), (question: Does the water heater have an odor or is it a new water heater? Answer: No), (question: Does the water container have an odor or is it a new water container (plastic, paper cup, etc.)? Answer: No_answer).

[0085] Because user queries and standard answers (such as "yes" or "no") may differ significantly, this difference can alter the flow of the decision graph. Therefore, it is crucial to accurately determine the standard answer based on the user's actual response to ensure the correct flow of the decision graph. From the perspective of the decision graph's structure (nodes, directed edges), the condition values ​​for the corresponding question need to be determined based on the user's query to find the appropriate ending node. To this end, a combination of fine-tuning and Prompt techniques is employed to improve the accuracy of answer determination.

[0086] The qwen2-7B-Instruct model was used as the base model, and LoRa tuning was employed. The data format is as follows:

[0087] {"instruction": "Question: Press the mute button on the remote control to observe if the noise still exists; Standard answer: Answer 1: The sound disappears; Answer 2: The sound still exists. Current answer: There is noise","input":"","output":"The sound still exists"}.

[0088] The fine-tuned large model is accessed using the prompt technique to improve the accuracy of model inference. The prompt is described as follows: You are now a question-answering expert, given a question and standard answers. There may be multiple standard answers, but only one existing answer. The task is to determine which standard answer the existing answer matches semantically, identifying the existing answer that best matches the standard answer; then, simply provide the standard answer. Consider the following example: Question: Is this a newly installed machine? Standard Answer: Answer 1: No; Answer 2: Yes; Existing Answer: No; Return: No.

[0089] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods of the various embodiments of this application.

[0090] This embodiment also provides a response determination device based on large model decision-making, which is used to implement the above embodiments and preferred embodiments, and will not be repeated as already described. As used below, the term "module" can be a combination of software and / or hardware that implements a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.

[0091] Figure 4 This is a structural block diagram of a response determination device based on a large model decision according to an embodiment of this application. The device includes:

[0092] The acquisition module 42 is used to acquire the query statement sent by the user.

[0093] The query module 44 is used to perform a query in a pre-constructed question-and-answer decision tree based on the query intent of the query statement, and obtain intent matching nodes that match the query intent; wherein, the question-and-answer decision tree includes: multiple nodes; each node includes: node content;

[0094] The matching module 46 is used to match the query statement with the first node content of the intent matching node and the second node content of the downstream node of the intent matching node to obtain a matching result; the node content includes: the first node content and the second node content;

[0095] The determination module 48 is used to determine the response node that matches the matching result from the question-and-answer decision tree, starting from the intent matching node, and to determine the response content of the query statement based on the node content of the response node.

[0096] The aforementioned device acquires the query statement sent by the user terminal, queries the question-and-answer decision tree to obtain the intent matching node that matches the query intent, and then matches the query statement with the first node content of the intent matching node and the second node content of the downstream node of the intent matching node to obtain the matching result. In this way, the query statement can be matched with the first node content and the second node content at the same time. Furthermore, starting from the intent matching node, the device determines the response node that matches the matching result from the question-and-answer decision tree, and determines the response content of the query statement based on the node content of the response node. This can reduce the number of question-and-answer rounds and effectively improve the response determination efficiency.

[0097] In an exemplary embodiment, the question-answering decision tree further includes: response options corresponding to the node content; a matching module 46, further configured to match the query statement with the first node content of the intent matching node to obtain a first matching result, and match the query statement with the second node content of the downstream node of the intent matching node to obtain a second matching result; determine a first response content from the first response options of the intent matching node based on the first matching result, and determine a second response content from the second response options of the downstream node based on the second matching result; and determine a matching result based on the first response content and the second response content.

[0098] In an exemplary embodiment, the matching module 46 is further configured to input the query statement, the first node content, and the second node content into a pre-trained semantic judgment model; and through the semantic judgment model, match the query statement and the first node content according to a preset semantic judgment instruction to obtain a first matching result, and match the query statement and the second node content according to a preset semantic judgment instruction to obtain a second matching result.

[0099] In an exemplary embodiment, the apparatus further includes: a training module, configured to acquire sample node content and sample query statements; the sample node content includes: first sample node content and second sample node content; the first sample node content, the second sample node content, and the sample query statements are input into a large model according to a preset input format; the large model is used to match the first sample node content, the second sample node content, and the sample query statements according to preset semantic judgment instructions to obtain sample matching results; based on the sample matching results and the labeled matching results, the parameters of the large model are adjusted and training continues until the training ends, thereby obtaining a trained semantic judgment model.

[0100] In an exemplary embodiment, the matching result includes: first response content and second response content; any two nodes among the plurality of nodes are connected by directed edges. The determining module 48 is further configured to, starting from the intent matching node, determine the first directed edge content of at least one directed edge connected to the intent matching node; perform semantic matching between the first directed edge content and the first response content to determine the first target directed edge of the intent matching node, and determine a target downstream node from the downstream nodes of the intent matching node based on the first target directed edge; determine the second directed edge content of at least one directed edge connected to the target downstream node; perform semantic matching based on the second response content and the second directed edge content to determine the second target directed edge of the target downstream node, continue to determine the next target downstream node based on the second target directed edge, and return to the step of determining the second directed edge content of at least one directed edge connected to the target downstream node to continue execution until a preset condition is met, and then stop to determine the response node.

[0101] In an exemplary embodiment, the query module 44 is further configured to perform vectorization representation processing on the query intent to obtain an intent vector representation of the query intent, and perform vectorization representation processing on the node content of the N nodes to obtain a content vector representation of the N nodes; perform similarity matching based on the intent vector representation and the content vector representation of the N nodes to obtain a similarity matching result of the N nodes; and determine the nodes corresponding to the similarity matching results that meet the preset matching conditions in the similarity matching results of the N nodes as intent matching nodes.

[0102] In an exemplary embodiment, the query module 44 is further configured to perform vectorization representation processing on the query intent to obtain an intent vector representation of the query intent, and perform vectorization representation processing on the node content of the plurality of nodes to obtain a content vector representation of the plurality of nodes; perform similarity matching based on the intent vector representation and the content vector representation of the plurality of nodes to obtain a similarity matching result of the plurality of nodes; and determine the node corresponding to the similarity matching result that meets the preset matching conditions in the similarity matching results of the plurality of nodes as the intent matching node.

[0103] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to perform the steps in any of the above method embodiments when it is run.

[0104] Optionally, in this embodiment, the storage medium may be configured to store a computer program for performing the following steps:

[0105] S1, retrieve the query statement sent by the user;

[0106] S2, based on the query intent of the query statement, a query is performed in a pre-constructed question-and-answer decision tree to obtain intent matching nodes that match the query intent; wherein, the question-and-answer decision tree includes: multiple nodes; each node includes: node content;

[0107] S3, the query statement is matched with the first node content of the intent matching node and the second node content of the downstream node of the intent matching node to obtain the matching result; the node content includes: the first node content and the second node content;

[0108] S4, starting from the intent matching node, determine the response node that matches the matching result from the question-answering decision tree, and determine the response content of the query statement based on the node content of the response node.

[0109] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0110] Specific examples in this embodiment can be found in the examples described in the above embodiments and exemplary implementations, and will not be repeated here.

[0111] Embodiments of this application also provide an electronic device, such as... Figure 5 As shown, the electronic device includes a memory 502 and a processor 504. The memory 502 stores a computer program, and the processor 504 is configured to execute the steps in any of the above method embodiments via the computer program.

[0112] Optionally, in this embodiment, the processor 504 can be configured to perform the following steps via a computer program:

[0113] S1, retrieve the query statement sent by the user;

[0114] S2, based on the query intent of the query statement, a query is performed in a pre-constructed question-and-answer decision tree to obtain intent matching nodes that match the query intent; wherein, the question-and-answer decision tree includes: multiple nodes; each node includes: node content;

[0115] S3, the query statement is matched with the first node content of the intent matching node and the second node content of the downstream node of the intent matching node to obtain the matching result; the node content includes: the first node content and the second node content;

[0116] S4, starting from the intent matching node, determine the response node that matches the matching result from the question-answering decision tree, and determine the response content of the query statement based on the node content of the response node.

[0117] Specific examples in this embodiment can be found in the examples described in the above embodiments and exemplary implementations, and will not be repeated here.

[0118] Alternatively, as those skilled in the art will understand, Figure 5 The structure shown is for illustrative purposes only. Figure 5 This does not limit the structure of the aforementioned electronic device. For example, the electronic device may also include components that are more... Figure 5 The more or fewer components shown (such as network interfaces, etc.), or having the same Figure 5 The different configurations shown.

[0119] The memory 502 can be used to store software programs and modules, such as the program instructions / modules corresponding to the response determination method and apparatus based on large model decision-making in this embodiment. The processor 504 executes various functional applications and data processing by running the software programs and modules stored in the memory 502, thereby realizing the aforementioned response determination method based on large model decision-making. The memory 502 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 502 may further include memory remotely located relative to the processor 504, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof. Specifically, the memory 502 may be used, but is not limited to, to store information such as system configuration files. As an example, such as... Figure 5 As shown, the memory 502 may include, but is not limited to, the acquisition module 42, query module 44, matching module 46, and determination module 48 of the response determination device based on large model decision-making described above. Furthermore, it may include, but is not limited to, other module units (such as the first determination module and the second determination module) of the response determination device based on large model decision-making described above, which will not be elaborated further in this example.

[0120] Optionally, the transmission device 506 described above is used to receive or send data via a network. Specific examples of the network described above may include wired networks and wireless networks. In one example, the transmission device 506 includes a Network Interface Controller (NIC), which can be connected to other network devices and a router via a network cable to communicate with the Internet or a local area network. In another example, the transmission device 506 is a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0121] In addition, the above-mentioned electronic device also includes: a display 508; and a connection bus 510 for connecting the various module components in the above-mentioned electronic device.

[0122] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0123] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in any of the above method embodiments.

[0124] The embodiments described herein also provide a computer program that includes computer instructions stored in a computer-readable storage medium; a processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the steps in any of the above method embodiments.

[0125] Specific examples in this embodiment can be found in the examples described in the above embodiments and exemplary implementations, and will not be repeated here.

[0126] Obviously, those skilled in the art should understand that the modules or steps of this application described above can be implemented using general-purpose computing devices. They can be centralized on a single computing device or distributed across a network of multiple computing devices. They can be implemented using computer-executable program code, and thus can be stored in a storage device for execution by a computing device. In some cases, the steps shown or described can be performed in a different order than those presented here, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, this application is not limited to any particular combination of hardware and software.

[0127] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A response determination method based on large model decision-making, characterized in that, include: Get the query statement sent by the user; Based on the query intent of the query statement, a query is performed in a pre-constructed question-and-answer decision tree to obtain intent matching nodes that match the query intent; wherein, the question-and-answer decision tree includes: multiple nodes; each node includes: node content; The query statement is matched with the first node content of the intent matching node and the second node content of the downstream node of the intent matching node to obtain the matching result; the node content includes: the first node content and the second node content; Starting from the intent matching node, the response node that matches the matching result is determined from the question-and-answer decision tree, and the response content of the query statement is determined based on the node content of the response node.

2. The method according to claim 1, characterized in that, The question-and-answer decision tree further includes: response options corresponding to the node content; the matching of the query statement with the first node content of the intent-matching node and the second node content of the downstream node of the intent-matching node to obtain the matching result includes: The query statement is matched with the content of the first node of the intent matching node to obtain a first matching result, and the query statement is matched with the content of the second node of the downstream node of the intent matching node to obtain a second matching result; The first response content is determined from the first response options of the intent matching node based on the first matching result, and the second response content is determined from the second response options of the downstream node based on the second matching result; The matching result is determined based on the content of the first response and the content of the second response.

3. The method according to claim 2, characterized in that, The step of matching the query statement with the content of the first node of the intent matching node to obtain a first matching result, and matching the query statement with the content of the second node of the downstream node of the intent matching node to obtain a second matching result, includes: The query statement, the content of the first node, and the content of the second node are input into a pre-trained semantic judgment model; The semantic judgment model is used to match the query statement and the content of the first node according to the preset semantic judgment instructions to obtain a first matching result, and the query statement and the content of the second node are matched according to the preset semantic judgment instructions to obtain a second matching result.

4. The method according to claim 3, characterized in that, Before inputting the query statement, the content of the first node, and the content of the second node into the pre-trained semantic judgment model, the method further includes: Obtain sample node content and sample query statement; the multiple sample node contents include: first sample node content and second sample node content; Input the content of the first sample node, the content of the second sample node, and the sample query statement into the large model according to a preset input format; Using the large model, the content of the first sample node, the content of the second sample node, and the sample query statement are matched according to preset semantic judgment instructions to obtain sample matching results; Based on the sample matching results and the annotation matching results, the parameters of the large model are adjusted and training continues until the training is completed, thus obtaining a semantic judgment model that has been trained.

5. The method according to claim 1, characterized in that, The matching result includes: the content of a first response and the content of a second response; any two nodes among the plurality of nodes are connected by a directed edge; the step of determining the response node that matches the matching result from the question-and-answer decision tree, starting from the intent matching node, includes: Starting from the intent matching node, determine the content of the first directed edge of at least one directed edge connected to the intent matching node; Semantic matching is performed between the content of the first directed edge and the content of the first response to determine the first target directed edge of the intent matching node, and the target downstream node is determined from the downstream nodes of the intent matching node based on the first target directed edge. Determine the content of the second directed edge that connects at least one directed edge to the downstream node of the target; Based on the second response content and the second directed edge content, semantic matching is performed to determine the second target directed edge of the target downstream node. The next target downstream node is determined based on the second target directed edge, and the process returns to the step of determining the second directed edge content of at least one directed edge connected to the target downstream node. This process continues until a preset condition is met, at which point the response node is determined.

6. The method according to claim 1, characterized in that, Before the query intent based on the query statement is queried in a pre-constructed question-answering decision tree to obtain an intent matching node that matches the query intent, the method further includes: M question statements are determined from a pre-defined query database, and corresponding question statement answers are determined for the M question statements; the query database includes question statements and corresponding question statement answers; M is an integer greater than 1. A node table is determined based on the M question statements, and a directed edge table is determined based on the question statement responses to the M question statements. Construct a question-answering decision tree based on the node table and the directed edge table.

7. The method according to claim 1, characterized in that, The query intent based on the query statement is queried in a pre-constructed question-answering decision tree to obtain intent matching nodes that match the query intent, including: The query intent is vectorized to obtain an intent vector representation of the query intent, and the node content of the multiple nodes is vectorized to obtain a content vector representation of the multiple nodes. Similarity matching is performed between the intent vector representation and the content vector representation of the multiple nodes to obtain the similarity matching result of the multiple nodes; The nodes corresponding to the similarity matching results that meet the preset matching conditions among the multiple nodes are determined as the intent matching nodes.

8. A response determination device based on large model decision-making, characterized in that, include: The acquisition module is used to acquire the query statements sent by the user. The query module is used to perform a query in a pre-constructed question-and-answer decision tree based on the query intent of the query statement, and obtain intent matching nodes that match the query intent; wherein, the question-and-answer decision tree includes: multiple nodes; each node includes: node content; The matching module is used to match the query statement with the first node content of the intent matching node and the second node content of the downstream node of the intent matching node to obtain the matching result; the node content includes: the first node content and the second node content; The determination module is used to determine the response node that matches the matching result from the question-and-answer decision tree, starting from the intent matching node, and to determine the response content of the query statement based on the node content of the response node.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored program, wherein the program, when executed, performs the method of any one of claims 1 to 7.

10. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to execute the method of any one of claims 1 to 7 through the computer program.

Images