RAG complaint work order question and answer method and system fusing thinking tree

Abstract

The invention provides an RAG complaint work order question answering method and system fused with a thinking tree in the technical field of artificial intelligence, and the method comprises the steps: S1, obtaining a complaint work order, constructing a vocabulary sequence, and constructing a work order thinking tree based on the vocabulary sequence; s2, calculating intention probability distribution of each node based on semantic information of each node in the work order thinking tree, and obtaining a complaint intention of the complaint work order based on the intention probability distribution; s3, creating a task tree based on the complaint work order, performing recursive decomposition on the task tree based on the complaint intention to obtain sub-tasks, and calculating the priority of each sub-task through a priority evaluation function; s4, performing recursive retrieval on the task tree based on the priority to obtain subtasks related to the complaint work order and a retrieval result; and S5, constructing a cue word thinking tree based on each related subtask, and automatically generating a processing suggestion of the complaint work order based on the cue word thinking tree. The method has the advantages that the accuracy and efficiency of complaint work order processing are greatly improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and particularly to a RAG complaint work order question-answering method and system integrating a thought tree. Background Art

[0002] With the continuous progress and popularization of mobile communication technology, the business volume and user scale of mobile communication operators have been continuously expanding. However, with the growth of the number of users, the number of complaints has also risen, which poses higher requirements for the customer service level of mobile communication operators. Traditionally, mobile communication operators have the following methods for handling complaint work orders:

[0003] Method 1: Rule-based complaint work order handling method, which relies on manually preset rules and templates to identify user intentions and generate corresponding answers. Although this method is relatively effective in handling some common problems, because the rules in the rule base are usually formulated based on historical data and common problems and can cover some common complaint situations; however, with the continuous progress of mobile communication technology and the diversification of services, complaint problems have become more and more complex and changeable, and the following disadvantages of this method have gradually emerged:

[0004] 1. Limitations of the rule base: The rules in the rule base are difficult to cover all possible complaint situations. Especially when the complaint problem involves multiple fields or multiple factors, the rule base often cannot provide accurate solutions; 2. Lack of flexibility: It is difficult to adapt to changing complaint situations. Once encountering new complaint problems or when the rules in the rule base are outdated, it often cannot give satisfactory answers; 3. High maintenance cost: With the continuous expansion and update of the rule base, the maintenance cost also increases; manually formulating and updating rules requires a lot of time and effort and is prone to errors and omissions.

[0005] Method 2: RAG (Retrieval-Augmented Generation)-based complaint work order handling method, which combines two technologies of retrieval and generation, aiming to utilize the content generation ability of large models and assist in generating answers by retrieving relevant information; this method improves the accuracy and diversity of answers to a certain extent because it can retrieve historical records and solutions related to the current complaint from the knowledge base; however, this method still has the following disadvantages:

[0006] 1. Shallow understanding of depth: Lack of the ability to understand deeply, making it difficult to accurately capture the deep intentions and specific needs of user complaints. This results in responses that are often lacking in pertinence and coherence, and are difficult to meet the actual needs of users. 2. Limited task decomposition ability: When dealing with complex complaints, it is often difficult to effectively decompose the problem into multiple specific subtasks, which causes the system to be unable to retrieve accurate information related to each subtask from the knowledge base, thus affecting the accuracy and completeness of the response. 3. Single solution: The response method based on preset rules or templates limits the diversity and personalization of solutions.

[0007] In summary, the traditional complaint work order processing method has problems of poor accuracy and efficiency. Therefore, how to provide a RAG complaint work order question and answer method and system that integrates a thinking tree to improve the accuracy and efficiency of complaint work order processing has become an urgent technical problem to be solved. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to provide a RAG complaint work order question and answer method and system that integrates a thinking tree to improve the accuracy and efficiency of complaint work order processing.

[0009] In the first aspect, the present invention provides a RAG complaint work order question and answer method that integrates a thinking tree, including the following steps:

[0010] Step S1: Obtain a complaint work order, preprocess the complaint work order to construct a vocabulary sequence, and construct a work order thinking tree based on the vocabulary sequence;

[0011] Step S2: Calculate the intention probability distribution of each node in the work order thinking tree based on the semantic information of each node in the work order thinking tree, and obtain the complaint intention of the complaint work order based on the intention probability distribution;

[0012] Step S3: Create a task tree based on the complaint work order, recursively decompose the task tree based on the complaint intention to obtain several subtasks, and calculate the priority of each subtask through a priority evaluation function;

[0013] Step S4: Recursively retrieve the task tree based on the priority to obtain the subtasks related to the complaint work order and the retrieval results corresponding to the subtasks;

[0014] Step S5: Construct a prompt word thinking tree based on the retrieval results of each related subtask, and automatically generate a processing suggestion for the complaint work order based on the prompt word thinking tree.

[0015] Further, step S1 is specifically:

[0016] Obtain a complaint work order, perform preprocessing on the complaint work order including at least word segmentation, part-of-speech tagging, and stop word removal to obtain a number of words, construct a word sequence based on each of the words, and construct a work order thinking tree including a number of nodes based on the word sequence;

[0017] The nodes are generated based on each word in the word sequence and carry semantic information; the semantic information is constructed based on the grammatical structure and semantic relationship.

[0018] Further, the specific content of step S2 is as follows:

[0019] Based on the semantic information of each node in the work order thinking tree, calculate the intention probability distribution of each node from bottom to top in turn, and select the intention with the highest probability as the complaint intention of the complaint work order based on the intention probability distribution.

[0020] Further, the specific content of step S3 is as follows:

[0021] Create a task tree with the complaint work order as the root node, and parse the complaint work order through a key information parsing model pre-trained with the complaint intention to obtain key information including at least the complaint type, problem description, and impact degree;

[0022] Extract at least a complaint task event including an event type, a trigger word, and an argument from the key information through a pre-trained event extraction model;

[0023] Recursively decompose the task tree based on the complaint task event to obtain a number of subtasks, and calculate the priorities of each subtask through a priority evaluation function.

[0024] Further, the specific content of step S5 is as follows:

[0025] Construct a prompt word thinking tree based on the retrieval results of relevant subtasks, guide a pre-trained complaint handling model to automatically generate handling suggestions for the complaint work order through the prompt word thinking tree, integrate and sort each handling suggestion to construct a handling suggestion list and display it.

[0026] In a second aspect, the present invention provides a RAG complaint work order question-answering system integrating a thinking tree, including the following modules:

[0027] A work order thinking tree construction module, configured to obtain a complaint work order, perform preprocessing on the complaint work order to construct a word sequence, and construct a work order thinking tree based on the word sequence;

[0028] A complaint intention acquisition module, configured to calculate the intention probability distribution of each node based on the semantic information of each node in the work order thinking tree, and obtain the complaint intention of the complaint work order based on the intention probability distribution;

[0029] The task tree recursive decomposition module is used to create a task tree based on the complaint work order, recursively decompose the task tree based on the complaint intention to obtain several subtasks, and calculate the priorities of the subtasks through a priority evaluation function;

[0030] The recursive retrieval module is used to recursively retrieve the task tree based on the priority, and obtain the subtasks related to the complaint work order and the retrieval results corresponding to the subtasks;

[0031] The processing suggestion generation module is used to construct a prompt word thinking tree based on the retrieval results of the relevant subtasks, and automatically generate a processing suggestion for the complaint work order based on the prompt word thinking tree.

[0032] Furthermore, the work order thinking tree construction module is specifically used for:

[0033] Obtain the complaint work order, perform preprocessing on the complaint work order including at least word segmentation, part-of-speech tagging, and stop word removal to obtain several words, construct a word sequence based on the words, and construct a work order thinking tree containing several nodes based on the word sequence;

[0034] The nodes are generated based on the words in the word sequence and carry semantic information; the semantic information is constructed based on the grammatical structure and semantic relationship.

[0035] Furthermore, the complaint intention acquisition module is specifically used for:

[0036] Based on the semantic information of the nodes in the work order thinking tree, calculate the intention probability distribution of each node from bottom to top in turn, and select the intention with the highest probability as the complaint intention of the complaint work order.

[0037] Furthermore, the task tree recursive decomposition module is specifically used for:

[0038] Create a task tree with the complaint work order as the root node, and parse the complaint work order through a key information parsing model pre-trained by the complaint intention to obtain key information including at least the complaint type, problem description, and impact degree;

[0039] Extract at least complaint task events including event type, trigger word, and argument from the key information through a pre-trained event extraction model;

[0040] Recursively decompose the task tree based on the complaint task events to obtain several subtasks, and calculate the priorities of the subtasks through a priority evaluation function.

[0041] Furthermore, the processing suggestion generation module is specifically used for:

[0042] Construct a prompt word thinking tree based on the retrieval results of each of the related subtasks, and use the prompt word thinking tree to guide a pre-trained complaint handling model to automatically generate handling suggestions for complaint work orders. Integrate and sort the handling suggestions to construct a handling suggestion list and display it.

[0043] The advantages of the present invention are as follows:

[0044] By obtaining a complaint work order for preprocessing to construct a vocabulary sequence, constructing a work order thinking tree based on the vocabulary sequence, and then based on the semantic information of each node in the work order thinking tree, calculating the intention probability distribution of each node to obtain the complaint intention of the complaint work order; then creating a task tree based on the complaint work order, recursively decomposing the task tree based on the complaint intention to obtain several subtasks, calculating the priority of each subtask through a priority evaluation function, and recursively retrieving the task tree based on the priority to obtain subtasks related to the complaint work order and the retrieval results corresponding to the subtasks; finally, constructing a prompt word thinking tree based on the retrieval results of each related subtask, and automatically generating handling suggestions for the complaint work order based on the prompt word thinking tree; that is, by constructing a hierarchical work order thinking tree, recursively analyzing and reasoning the semantic information of the complaint work order, so as to accurately understand the user's complaint intention, the complaint work order can be gradually disassembled into finer-grained subtasks (semantic units). Through bottom-up traversal and calculation, the information of each node can be integrated, and finally the understanding of the overall complaint intention can be output, which helps to reduce misunderstandings and omissions and improve the accuracy of complaint work order handling; it can dynamically adjust the division of subtasks according to different situations, decompose complex complaint work orders (complex problems) into multiple subtasks and solve them step by step. This method of constructing a task tree and recursively decomposing can automatically and flexibly decompose the complaint work order into multiple simple and executable subtasks according to the complexity and hierarchical structure of the complaint work order, thus greatly improving the efficiency and accuracy of complaint work order handling; on the basis of the constructed task tree, by integrating the generation capabilities of recursive retrieval, prompt word thinking tree, and complaint handling model, efficient and accurate generation of handling suggestions is achieved, improving the accuracy of retrieval and the efficiency of generation; by introducing a thinking tree structure (work order thinking tree, task tree, prompt word thinking tree), the answer (handling suggestion) to the complaint work order is made more logical and well-organized, and finally the accuracy and efficiency of complaint work order handling are greatly improved. Brief Description of the Drawings

[0045] The present invention will be further described below with reference to the accompanying drawings in conjunction with embodiments.

[0046] Figure 1 is a flowchart of a method for answering complaint work orders of RAG integrated with a thinking tree according to the present invention.

[0047] Figure 2 is a schematic structural diagram of a system for answering complaint work orders of RAG integrated with a thinking tree according to the present invention. Detailed implementation manners

[0048] The technical solutions in the embodiments of the present application generally have the following ideas: By constructing a hierarchical work order thinking tree, recursively parsing and reasoning the semantic information of complaint work orders, so as to accurately understand the user's complaint intention, the complaint work orders can be gradually decomposed into finer-grained subtasks. Through bottom-up traversal and calculation, the information of each node can be integrated, and finally the understanding of the overall complaint intention can be output, which helps to reduce misunderstandings and omissions and improve the accuracy of complaint work order processing; It can dynamically adjust the division of subtasks according to different situations, decompose complex complaint work orders into multiple subtasks and solve them step by step. This method of constructing a task tree and recursively decomposing can automatically and flexibly decompose complaint work orders into multiple simple and executable subtasks according to the complexity and hierarchical structure of complaint work orders, thereby improving the efficiency and accuracy of complaint work order processing; On the basis of the constructed task tree, by integrating the generation capabilities of recursive retrieval, prompt word thinking tree, and complaint handling model, efficient and accurate generation of processing suggestions can be realized, and the accuracy of retrieval and the efficiency of generation can be improved; By introducing the thinking tree structure, the answers to complaint work orders are made more logical and well-organized, thereby improving the accuracy and efficiency of complaint work order processing.

[0049] Please refer to Figures 1 to 2 As shown in the figure, a preferred embodiment of the RAG complaint work order question-answering method integrating a thinking tree according to the present invention includes the following steps:

[0050] Step S1: Obtain a complaint work order, preprocess the complaint work order to construct a vocabulary sequence, and construct a work order thinking tree based on the vocabulary sequence;

[0051] Step S2: Calculate the intention probability distribution of each node in the work order thinking tree, and obtain the complaint intention of the complaint work order based on the intention probability distribution;

[0052] Step S3: Create a task tree based on the complaint work order, recursively decompose the task tree based on the complaint intention to obtain several subtasks, and calculate the priorities of the subtasks through a priority evaluation function;

[0053] Step S4: Recursively retrieve the task tree based on the priority to obtain subtasks related to the complaint work order and the retrieval results corresponding to the subtasks; The specific process of recursive retrieval is as follows:

[0054] Using the hierarchical structure of the task tree, recursively retrieve subtasks (task nodes) related to the complaint work order starting from the root node; During the retrieval process, the retrieval of each node will combine the retrieval results of its parent node to ensure the accuracy and depth of the retrieval. Define the constructed task tree as T:

[0055] T = {n1 , n 2 ,..., n n};

[0056] Among them, each node ni represents a subtask, having attributes type i (task type), description i (task description), and children i (list of child nodes).

[0057] Define a recursive function:

[0058] result i = retrieve_task(n i , query, parent_result);

[0059] During the recursion, the retrieval of each child node will combine the retrieval result of its parent node, that is:

[0060] combined_result = merge(parent_result, local_result i );

[0061] Among them, local_result i is the local retrieval result of the current child node; parent_result is the retrieval result of the parent node; merge is a merging function used to integrate the retrieval results of the parent node and the child node.

[0062] Step S5, construct a prompt word thinking tree based on the retrieval results of each of the relevant subtasks, and automatically generate a handling suggestion for the complaint work order.

[0063] The present invention ensures that even in complex or ambiguous situations, the true complaint intention of the user can be correctly interpreted by integrating advanced technologies such as thinking trees, RAG, and large models; through more flexible task decomposition (decomposing the complaint work order into subtasks), the division of subtasks can be dynamically adjusted according to different situations; and the effect of task retrieval is improved, making the retrieval results both comprehensive and accurate; the thinking tree structure enhances the quality of generating handling suggestions, making the answer (handling suggestion) more logical and well-organized, that is, achieving accurate understanding, efficient handling, and personalized response to the complaint intention of the complaint work order.

[0064] The specific content of the said step S1 is:

[0065] Obtain a complaint work order, perform preprocessing on the complaint work order, including at least word segmentation, part-of-speech tagging, and stop word removal, to obtain a number of words, construct a word sequence based on each of the words, and construct a work order thinking tree containing a number of nodes based on the word sequence;

[0066] The nodes are generated based on each word in the word sequence and carry semantic information; the semantic information is constructed based on the syntactic structure and semantic relationships.

[0067] The formula for the word sequence is:

[0068] f preprocess (text) = [word 1 , word 2 ,..., word n ;

[0069] Where text represents the complaint work order (input text); word n represents the nth word; [word 1 , word 2 ,..., word n represents the word sequence; f preprocess () represents the preprocessing function of the complaint work order.

[0070] The formula for the work order thinking tree is:

[0071] Tree = f build_base_tree ([word 1 , wor 2 ,..., wod n );

[0072] Where Tree represents the work order thinking tree; f build_base_tree () represents a function for recursively constructing a tree structure based on the semantic and syntactic relationships between words. The core calculation principle is as follows:

[0073] For each word word in the word sequence i , create a new leaf node (nodes are divided into parent nodes / root nodes, child nodes / leaf nodes based on dependency relationships), and assign word i to the leaf node, and find the parent node of word i according to the syntactic relationship (if it exists):

[0074] If word i is a dependent word of a certain word word j (for example, word i is the object of word j ), then word iThe node is regarded as wodn j The child node of the node. If wodn i is independent (without direct syntactic dependencies), it is temporarily attached to the root node or a temporary set of "isolated nodes".

[0075] According to the semantic relationship, further optimize the tree structure of the work order thinking tree, and calculate the semantic similarity between the word i node and its potential parent node (found based on syntactic relationships). If the semantic similarity is higher than the preset threshold and the syntactic relationship also supports this merger, the current parent-child relationship is maintained; otherwise, re-consider the parent node of the word i and merge it with other nodes to form a new semantic node.

[0076] The specific steps of step S2 are as follows:

[0077] Based on the semantic information of each node in the work order thinking tree, calculate the intention probability distribution of each node from bottom to top in turn, and select the intention with the highest probability as the complaint intention of the complaint work order based on the intention probability distribution.

[0078] By constructing a hierarchical work order thinking tree, recursively analyze and reason the semantic information of the complaint work order, so as to accurately understand the complaint intention.

[0079] For any node node i in the work order thinking tree, its intention probability distribution P(intent|node i ) can be calculated by integrating the intention probabilities and semantic relationships of its child nodes.

[0080] Let {child 1 , child 2 ,..., child m} be the set of child nodes of node i , then P(intent|node i ) can be expressed as:

[0081] where P(intent|child k ) represents the intention probability of the child node child k ; P(child k |node i ) represents the conditional probability of the child node child k under the parent node node i ;

[0082] Finally, on the root node root of the work order thought tree, select the intent with the highest probability as the final complaint intent:

[0083] final_intent = argmax(P(intent|root)).

[0084] Introducing a complaint work order thought tree to achieve the understanding of complaint intents, compared with conventional intent understanding and classification methods, shows the following significant advantages in the scenario characteristics of complaint work orders: 1. Comprehensiveness and accuracy: The integrated work order thought tree can integrate information from multiple channels and dimensions, including user descriptions, historical records, system logs, etc., so as to more comprehensively understand the user's complaint intent, reduce misunderstandings and omissions, and improve the accuracy of complaint work order processing; 2. Deep mining and correlation analysis: The integrated work order thought tree can deeply mine the underlying reasons and correlation factors behind user complaints, reveal the internal connections and laws between problems, help discover the root causes of problems, formulate targeted solutions, and thus more effectively solve complaint work orders.

[0085] The specific steps of step S3 are as follows:

[0086] Create a task tree with the complaint work order as the root node, and parse the complaint work order through the key information parsing model pre-trained with the complaint intent to obtain key information including at least the complaint type, problem description, and impact degree;

[0087] Extract complaint task events including at least event type, trigger word, and argument from the key information through the pre-trained event extraction model;

[0088] Based on the complaint task events, recursively decompose the task tree to obtain several subtasks, and calculate the priorities of each subtask through the priority evaluation function.

[0089] In the scenario of complaint work orders, each complaint work order contains multiple aspects and details, which need to be decomposed into actionable subtasks in order to quickly and effectively solve the problem. Based on the task tree (thought tree) complaint work order task decomposition algorithm, the complaint work order is regarded as the root node, and by analyzing the content of the complaint work order, it is decomposed into multiple subtasks (sub-nodes), and each subtask can be further decomposed into finer subtasks until each subtask is clear and executable enough.

[0090] Let the event set constructed by complaint task events be E, and the complaint task event be e i , e i ∈E, which can be expressed as:

[0091] e i = {type, trigger, args};

[0092] Among them, type represents the event type (such as "complaint", "fault", etc.); trigger represents the keyword or phrase that triggers the complaint task event; args represents the argument set of the complaint task event, including the key information of the subject, object, time, and location of the complaint task event.

[0093] The formula of the priority evaluation function is as follows:

[0094]

[0095] Among them, P(n) represents the priority of node n; T(n) represents the processing time of node m; U(n) represents the urgency of node n; H(n) represents the historical processing difficulty of node n; both α and β represent weight coefficients, which are used to adjust the influence of urgency and historical resolution difficulty on the priority.

[0096] By recursively decomposing to construct a task tree, not only can the complaint work order be automatically and flexibly decomposed into multiple simple and executable subtasks according to the complexity and hierarchical structure of the complaint work order, thus greatly improving the processing efficiency and accuracy; but also it can ensure that all tasks are thoroughly decomposed, avoiding the situation of omission and duplicate processing; in addition, the recursive decomposition also has good scalability and maintainability, and can easily adapt to the requirements of different scenarios and task types.

[0097] The specific content of step S5 is as follows:

[0098] Based on the retrieval results of each of the related subtasks, construct a prompt word thinking tree, and through the prompt word thinking tree, guide the pre-trained complaint processing model to automatically generate processing suggestions for the complaint work order, integrate and sort the processing suggestions to construct a processing suggestion list and display it.

[0099] The prompt word thinking tree also integrates additional information related to the complaint work order (such as historical processing cases, common problems, etc.) to form prompt words for guiding the complaint processing model to generate processing suggestions.

[0100] Define the prompt word thinking tree as C: C = {c 1 , c 2 ,..., c m};

[0101] Among them, each node c i represents a prompt word or sub-prompt word, and has attributes prompt i (prompt word), related_cases i (related cases) and common_solutions i (common solutions).

[0102] Each constructed node c i is added to the prompt thought tree C for complaint handling.

[0103] Each node c in the prompt thought tree C for complaint handling i is input into the complaint handling model LLM to generate corresponding handling suggestions suggestion i : suggestion i = LLM(c i ).

[0104] Through the recursive retrieval of the task tree and the construction of the prompt thought tree, the structured retrieval and generation of handling suggestions for complaint work orders are realized, improving the accuracy of retrieval and the efficiency of generation; combined with the specific situation of the complaint work order and the characteristics of task nodes (sub-tasks), a personalized prompt thought tree is constructed, and accurate handling suggestions are generated using a large model (complaint handling model), enhancing the practicality and satisfaction of the handling suggestions.

[0105] A preferred embodiment of the RAG complaint work order Q&A system integrating a thought tree of the present invention includes the following modules:

[0106] A work order thought tree construction module for obtaining a complaint work order, preprocessing the complaint work order to construct a vocabulary sequence, and constructing a work order thought tree based on the vocabulary sequence;

[0107] A complaint intention acquisition module for calculating the intention probability distribution of each node in the work order thought tree based on the semantic information of each node, and obtaining the complaint intention of the complaint work order based on the intention probability distribution;

[0108] A task tree recursive decomposition module for creating a task tree based on the complaint work order, recursively decomposing the task tree based on the complaint intention to obtain a number of sub-tasks, and calculating the priorities of the sub-tasks through a priority evaluation function;

[0109] A recursive retrieval module for recursively retrieving the task tree based on the priorities to obtain sub-tasks related to the complaint work order and the retrieval results corresponding to the sub-tasks; the specific process of recursive retrieval is as follows:

[0110] Using the hierarchical structure of the task tree, recursively retrieve sub-tasks (task nodes) related to the complaint work order starting from the root node; during the retrieval process, the retrieval of each node will combine the retrieval results of its parent node to ensure the accuracy and depth of the retrieval. Define the constructed task tree as T:

[0111] T = {n 1 , n 2 ,..., n n};

[0112] Among them, each node ni represents a subtask and has attributes type i (task type), description i (task description), and children i (list of child nodes).

[0113] Define a recursive function:

[0114] result i = retrieve_task(n i , query, parent_result);

[0115] During the recursion, the retrieval of each child node will combine the retrieval result of its parent node, that is:

[0116] combined_result = merge(parent_result, local_result i );

[0117] Among them, loacal_result i is the local retrieval result of the current child node; parent_result is the retrieval result of the parent node; merge is a merging function used to integrate the retrieval results of the parent node and the child node.

[0118] A processing suggestion generation module, which is used to construct a prompt word thinking tree based on the retrieval results of each of the relevant subtasks, and automatically generate a processing suggestion for the complaint work order based on the prompt word thinking tree.

[0119] Through the integration of advanced technologies such as thinking trees, RAG, and large models, the present invention ensures that even in complex or ambiguous situations, the true complaint intention of the user can be correctly interpreted; through more flexible task decomposition (decomposing the complaint work order into subtasks), the division of subtasks can be dynamically adjusted according to different situations; and the effect of task retrieval is improved, making the retrieval results both comprehensive and accurate; the thinking tree structure enhances the quality of processing suggestion generation, making the answer (processing suggestion) more logical and well-organized, that is, realizing the accurate understanding, efficient processing, and personalized answer of the complaint intention of the complaint work order.

[0120] The work order thinking tree construction module is specifically used for:

[0121] Obtain a complaint work order, perform preprocessing on the complaint work order including at least word segmentation, part-of-speech tagging, and removal of stop words to obtain a number of words, construct a word sequence based on each of the words, and construct a work order thinking tree containing a number of nodes based on the word sequence;

[0122] The nodes are generated based on the words in the word sequence and carry semantic information; the semantic information is constructed based on the syntactic structure and semantic relationships.

[0123] The formula for the word sequence is:

[0124] f preprocess (text) = [word 1 , word 2 ,..., word n ;

[0125] where text represents the complaint work order (input text); word n represents the nth word; [word 1 , word 2 ,..., word n represents the word sequence; f prerocess () represents the preprocessing function of the complaint work order.

[0126] The formula for the work order thinking tree is:

[0127] Tree = f build_base_tree ([word 1 , word 2 ,..., word n );

[0128] where Tree represents the work order thinking tree; f build_base_tree () represents the function of recursively constructing the tree structure according to the semantic and syntactic relationships between words. The core calculation principle is as follows:

[0129] For each word word i in the word sequence, create a new leaf node (nodes are divided into parent nodes / root nodes, child nodes / leaf nodes based on dependency relationships), and assign word i to this leaf node. Find the parent node of word i according to the syntactic relationship (if it exists):

[0130] If word i is the dependent word of a certain word word j (for example, word i is the object of word j ), then take the word i node as the child node of the word j node. If word i is independent (without direct syntactic dependency relationship), temporarily mount it under the root node or in a temporary "isolated node" set.

[0131] According to the semantic relationship, further optimize the tree structure of the work order thinking tree, and calculate the word based on the pre-trained text vector model i The semantic similarity between the node and its potential parent node (found based on the syntactic relationship). If the semantic similarity is higher than the preset threshold and the syntactic relationship also supports this merger, then maintain the current parent-child relationship; otherwise, reconsider the parent node of the word i and merge it with other nodes to form a new semantic node.

[0132] The complaint intention acquisition module is specifically used for:

[0133] Based on the semantic information of each node in the work order thinking tree, calculate the intention probability distribution of each node from bottom to top in turn, and select the intention with the highest probability as the complaint intention of the complaint work order based on the intention probability distribution.

[0134] By constructing a hierarchical work order thinking tree, recursively analyze and infer the semantic information of the complaint work order, so as to accurately understand the complaint intention.

[0135] For any node node in the work order thinking tree i , its intention probability distribution P(intent|node i ) can be calculated by integrating the intention probabilities and semantic relationships of its child nodes.

[0136] Let {child 1 , child 2 ,..., child m} be the set of child nodes of node i , then P(intent|node i ) can be expressed as:

[0137] where P(intent|child k ) represents the intention probability of the child node child k ; P(child k |node i ) represents the conditional probability of the child node child k under the parent node node i ;

[0138] Finally, at the root node root of the work order thinking tree, select the intention with the highest probability as the final complaint intention:

[0139] final_intent = argmax(P(intent|root)).

[0140] Introduce a complaint work order thinking tree to achieve the understanding of complaint intentions. Compared with the conventional intention understanding and classification methods, in the scenario characteristics of complaint work orders, the following significant advantages are demonstrated: 1. Comprehensiveness and accuracy: The integrated work order thinking tree can integrate information from multiple channels and dimensions, including user descriptions, historical records, system logs, etc., so as to more comprehensively understand the user's complaint intentions, reduce misunderstandings and omissions, and improve the accuracy of complaint work order processing; 2. In-depth mining and correlation analysis: The integrated work order thinking tree can deeply mine the underlying causes and correlation factors behind user complaints, reveal the internal connections and laws between problems, help discover the root causes of problems, formulate targeted solutions, and thus more effectively solve complaint work orders.

[0141] The task tree recursive decomposition module is specifically used for:

[0142] Create a task tree with the complaint work order as the root node, and parse the complaint work order through the key information parsing model pre-trained for complaint intentions to obtain key information including at least the complaint type, problem description, and impact degree;

[0143] Extract complaint task events including at least event type, trigger word, and arguments from the key information through a pre-trained event extraction model;

[0144] Based on the complaint task events, recursively decompose the task tree to obtain several subtasks, and calculate the priorities of each subtask through a priority evaluation function.

[0145] In the scenario of complaint work orders, each complaint work order contains multiple aspects and details, which need to be decomposed into actionable subtasks in order to quickly and effectively solve the problem. Based on the complaint work order task decomposition algorithm of the task tree (thinking tree), the complaint work order is regarded as the root node, and by analyzing the content of the complaint work order, it is decomposed into multiple subtasks (sub-nodes), and each subtask can be further decomposed into finer subtasks until each subtask is clear and executable enough.

[0146] Let the event set constructed by complaint task events be E, and the complaint task event be e i , e i ∈E, which can be expressed as:

[0147] e i ={type, trigger, args};

[0148] Among them, type represents the event type (such as "complaint", "fault", etc.); trigger represents the keyword or phrase that triggers the complaint task event; args represents the argument set of the complaint task event, including the key information of the subject, object, time, and place of the complaint task event.

[0149] The formula of the priority evaluation function is as follows:

[0150]

[0151] Wherein, P(n) represents the priority of node n; T(n) represents the processing time of node m; U(n) represents the urgency of node n; H(n) represents the historical processing difficulty of node n; both α and β represent weight coefficients, which are used to adjust the influence of urgency and historical solution difficulty on the priority.

[0152] By recursively decomposing to construct a task tree, not only can the complaint work order be automatically and flexibly decomposed into multiple simple and executable subtasks according to the complexity and hierarchical structure of the complaint work order, thereby greatly improving the processing efficiency and accuracy; but also it can ensure that all tasks are thoroughly decomposed, avoiding the situation of omission and repeated processing; in addition, the recursive decomposition also has good scalability and maintainability, and can easily adapt to the requirements of different scenarios and task types.

[0153] The processing suggestion generation module is specifically used for:

[0154] Based on the retrieval results of the relevant subtasks, construct a prompt word thinking tree, and through the prompt word thinking tree, guide the pre-trained complaint handling model to automatically generate the handling suggestions for the complaint work order, integrate and sort the handling suggestions to construct a handling suggestion list and display it.

[0155] The prompt word thinking tree also integrates additional information related to the complaint work order (such as historical handling cases, common problems, etc.) to form prompt words for guiding the complaint handling model to generate handling suggestions.

[0156] Define the prompt word thinking tree as C: C = {c 1 , c 2 ,..., c m};

[0157] Wherein, each node c i represents a prompt word or sub-prompt word, and has attributes prompt i (prompt word), related_cases i (related cases) and common_solutions i (common solutions).

[0158] Add each constructed node c i to the prompt word thinking tree C.

[0159] For each node c iInput into the complaint handling model LLM to generate corresponding handling suggestions i : suggestion i = LLM(c i ).

[0160] Through the recursive retrieval of the task tree and the construction of the prompt word thinking tree, the structured retrieval and generation of handling suggestions for complaint work orders are realized, improving the accuracy of retrieval and the efficiency of generation; combined with the specific situation of the complaint work order and the characteristics of task nodes (sub-tasks), a personalized prompt word thinking tree is constructed, and a precise handling suggestion is generated using a large model (complaint handling model), enhancing the practicality and satisfaction of the handling suggestion.

[0161] In summary, the advantages of the present invention are as follows:

[0162] By obtaining the complaint work order for preprocessing to construct a vocabulary sequence, constructing a work order thinking tree based on the vocabulary sequence, and then calculating the intention probability distribution of each node based on the semantic information of each node in the work order thinking tree to obtain the complaint intention of the complaint work order; then creating a task tree based on the complaint work order, recursively decomposing the task tree based on the complaint intention to obtain several sub-tasks, calculating the priority of each sub-task through a priority evaluation function, and recursively retrieving the task tree based on the priority to obtain the sub-tasks related to the complaint work order and the retrieval results corresponding to the sub-tasks; finally, constructing a prompt word thinking tree based on the retrieval results of each related sub-task, and automatically generating a handling suggestion for the complaint work order; that is, by constructing a hierarchical work order thinking tree, recursively analyzing and reasoning the semantic information of the complaint work order, so as to accurately understand the user's complaint intention, the complaint work order can be gradually decomposed into finer-grained sub-tasks (semantic units), and through bottom-up traversal and calculation, the information of each node can be integrated, and finally the understanding of the overall complaint intention can be output, which helps to reduce misunderstandings and omissions and improve the accuracy of complaint work order handling; it can dynamically adjust the division of sub-tasks according to different situations, decompose complex complaint work orders (complex problems) into multiple sub-tasks and solve them step by step. This method of constructing a task tree and recursively decomposing it can automatically and flexibly decompose the complaint work order into multiple simple and executable sub-tasks according to the complexity and hierarchical structure of the complaint work order, thus greatly improving the efficiency and accuracy of complaint work order handling; on the basis of the constructed task tree, by integrating the generation capabilities of recursive retrieval, prompt word thinking tree, and complaint handling model, the efficient and accurate generation of handling suggestions is realized, improving the accuracy of retrieval and the efficiency of generation; by introducing the thinking tree structure (work order thinking tree, task tree, prompt word thinking tree), the answer (handling suggestion) to the complaint work order is made more logical and well-organized, and finally the accuracy and efficiency of complaint work order handling are greatly improved.

[0163] Although the specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments we described are illustrative rather than used to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should all be covered by the scope protected by the claims of the present invention.

Claims

1. A RAG complaint ticket question and answer method integrating a thinking tree, characterized by: The steps include: Step S1, obtaining a complaint work order, preprocessing the complaint work order to construct a vocabulary sequence, and constructing a work order thinking tree based on the vocabulary sequence; Step S2: based on the semantic information of each node in the work order thinking tree, calculate the intention probability distribution of each node, and obtain the complaint intention of the complaint work order based on the intention probability distribution; Step S3: creating a task tree based on the complaint work order, recursively decomposing the task tree based on the complaint intent to obtain a number of subtasks, and calculating the priority of each subtask through a priority evaluation function; Step S4: recursively search the task tree based on the priority to obtain subtasks related to the complaint work order and search results corresponding to the subtasks; Step S5: construct a prompt word thinking tree based on the search results of the relevant subtasks, and automatically generate processing suggestions for the complaint work order based on the prompt word thinking tree.

2. A RAG complaint ticket question and answer method integrating a thinking tree as claimed in claim 1, characterized in that: The step S1 is specifically as follows: Obtain a complaint work order, perform preprocessing on the complaint work order including at least word segmentation, part-of-speech tagging, and stop word removal to obtain a plurality of words, construct a word sequence based on each of the words, and construct a work order thinking tree containing a plurality of nodes based on the word sequence; The nodes are generated based on each word in the word sequence and carry semantic information; the semantic information is constructed based on the grammatical structure and the semantic relationship.

3. A RAG complaint ticket question and answer method integrating a thinking tree as claimed in claim 1, characterized in that: The step S2 is specifically as follows: Based on the semantic information of each node in the work order thinking tree, the intention probability distribution of each node is calculated from bottom to top, and based on the intention probability distribution, the intention with the highest probability is selected as the complaint intention of the complaint work order.

4. A RAG complaint ticket question and answer method integrating a thinking tree as claimed in claim 1, characterized in that: The step S3 is specifically as follows: A task tree is created with the complaint work order as the root node, and the complaint work order is parsed through the key information parsing model pre-trained with the complaint intention to obtain key information including at least the complaint type, problem description and impact degree; Extracting complaint task events including at least event types, trigger words, and arguments from the key information through a pre-trained event extraction model; Based on the complaint task event, the task tree is recursively decomposed to obtain a number of subtasks, and the priority of each subtask is calculated through a priority evaluation function.

5. A RAG complaint ticket question and answer method integrating a thinking tree as claimed in claim 1, characterized in that: The step S5 is specifically as follows: A prompt word thinking tree is constructed based on the retrieval results of the relevant subtasks. The prompt word thinking tree is used to guide the pre-trained complaint handling model to automatically generate handling suggestions for the complaint work order. The handling suggestions are integrated and sorted to construct a handling suggestion list and displayed.

6. A RAG complaint ticket question and answer system integrating a mind tree, characterized by: Includes the following modules: A work order thinking tree construction module is used to obtain a complaint work order, pre-process the complaint work order to construct a vocabulary sequence, and construct a work order thinking tree based on the vocabulary sequence; A complaint intention acquisition module, used to calculate the intention probability distribution of each node in the work order thinking tree based on the semantic information of each node, and obtain the complaint intention of the complaint work order based on the intention probability distribution; A task tree recursive decomposition module is used to create a task tree based on the complaint work order, recursively decompose the task tree based on the complaint intention to obtain a number of subtasks, and calculate the priority of each subtask through a priority evaluation function; A recursive search module, used to recursively search the task tree based on the priority level to obtain subtasks related to the complaint work order and search results corresponding to the subtasks; The processing suggestion generation module is used to construct a prompt word thinking tree based on the search results of the relevant subtasks, and automatically generate processing suggestions for the complaint work order based on the prompt word thinking tree.

7. A RAG complaint ticket question and answer system integrating a thinking tree as claimed in claim 6, characterized in that: The work order thinking tree construction module is specifically used for: Obtain a complaint work order, perform preprocessing on the complaint work order including at least word segmentation, part-of-speech tagging, and stop word removal to obtain a plurality of words, construct a word sequence based on each of the words, and construct a work order thinking tree containing a plurality of nodes based on the word sequence; The nodes are generated based on each word in the word sequence and carry semantic information; the semantic information is constructed based on the grammatical structure and the semantic relationship.

8. A RAG complaint ticket question and answer system integrating a thinking tree as claimed in claim 6, characterized in that: The complaint intention acquisition module is specifically used for: Based on the semantic information of each node in the work order thinking tree, the intention probability distribution of each node is calculated from bottom to top, and based on the intention probability distribution, the intention with the highest probability is selected as the complaint intention of the complaint work order.

9. A RAG complaint ticket question and answer system integrating a thinking tree as claimed in claim 6, characterized in that: The task tree recursive decomposition module is specifically used for: A task tree is created with the complaint work order as the root node, and the complaint work order is parsed through the key information parsing model pre-trained with the complaint intention to obtain key information including at least the complaint type, problem description and impact degree; Extracting complaint task events including at least event types, trigger words, and arguments from the key information through a pre-trained event extraction model; Based on the complaint task event, the task tree is recursively decomposed to obtain a number of subtasks, and the priority of each subtask is calculated through a priority evaluation function.

10. The RAG complaint ticket question and answer system integrating the thinking tree as claimed in claim 6, characterized in that: The processing suggestion generating module is specifically used for: A prompt word thinking tree is constructed based on the retrieval results of the relevant subtasks. The prompt word thinking tree is used to guide the pre-trained complaint handling model to automatically generate handling suggestions for the complaint work order. The handling suggestions are integrated and sorted to construct a handling suggestion list and displayed.

Images