An information processing method, device and readable storage medium
By analyzing question-and-answer pairs in conversation records, a high-quality dialogue summary is generated, solving the problem of summary generation in multi-intent dialogues and achieving clear display of questions and solutions even in a chaotic conversation sequence.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- LENOVO (BEIJING) LTD
- Filing Date
- 2022-06-13
- Publication Date
- 2026-07-31
AI Technical Summary
In multi-intent dialogue scenarios, existing technologies struggle to generate high-quality dialogue summaries, especially when users ask multiple questions simultaneously. The interference in the dialogue order caused by the overlapping solutions provided by the client is difficult to handle.
By analyzing conversation records, identifying and combining questions and steps in question-answer pairs, generating target summary content based on features and sequence, removing noisy data, and ensuring the quality of the summary.
In multi-intent sessions, it generates high-quality summaries that are unaffected by disordered session order, effectively removes noisy data, and ensures the accuracy and clarity of the summary content.
Smart Images
Figure CN115033671B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of information technology, and more specifically, to an information processing method, apparatus and readable storage medium. Background Technology
[0002] One-on-one customer service is prevalent across a wide range of industries, such as product after-sales service consultations, online chatbot customer service in the medical field, and human customer service consultations in telecommunications companies. Compared to reading a long conversation log, conversation summaries can effectively help us quickly understand the user's concerns and corresponding solutions.
[0003] However, users may ask multiple questions simultaneously in a single conversation, leading to overlap in their language. Similarly, on the customer service side, different solutions are offered alongside the questions. This dialogue scenario is called "multi-intent dialogue," and due to the involvement of multiple topics and the interference of the dialogue order, it is difficult to generate high-quality dialogue summaries using existing technologies. Summary of the Invention
[0004] In view of this, this application provides an information processing method, as follows:
[0005] An information processing method, comprising:
[0006] Obtain session records, wherein the session records include at least a first session generated by a first session object and a second session generated by a second session object;
[0007] Analyzing the conversation records yields at least two question-answer pairs. Each question-answer pair includes a first conversation as a question and a second conversation as a step. The step is a step of the solution corresponding to the question in the question-answer pair. A question corresponds to at least one solution, and a solution includes at least one step.
[0008] Based on the characteristics of the steps, determine the solution corresponding to each question in the session record;
[0009] Based on the process of how the questions were generated in the session record, the questions and their corresponding solutions are displayed in sequence to obtain the target summary content of the session record.
[0010] Optionally, in the above method, the analysis of the conversation records to obtain at least two sets of question-answer pairs includes:
[0011] Based on the pairing model, a matching analysis is performed on the first and second sessions in the session records. The first and second sessions that meet the question-and-answer matching conditions are grouped into a question-and-answer pair, wherein the first session that meets the question-and-answer matching conditions is used as the question in the question-and-answer pair, and the second session that meets the question-and-answer matching conditions is used as the question in the question-and-answer pair.
[0012] Optionally, in the above method, determining the solution corresponding to each question in the session record based on the features of the steps includes:
[0013] Obtain at least two question-answer pairs corresponding to the same question to obtain the target question-answer pair set;
[0014] Based on the occurrence sequence of steps in the conversation record, the steps of each question-answer pair in the target question-answer pair set are arranged in order to obtain the solution corresponding to the question.
[0015] Optionally, in the above method, determining the solution corresponding to each question in the session record based on the features of the steps includes:
[0016] Based on the continuous model, at least two second sessions in the session records are analyzed, and at least two second sessions that meet the continuous condition are grouped into a group of continuous steps. Any group of continuous steps corresponds to the same problem, and the group of continuous steps is a solution.
[0017] Arrange the consecutive steps corresponding to the same problem in sequence to obtain the solution to the problem.
[0018] Optionally, in the above method, if there are at least two sets of consecutive steps corresponding to the same problem, the consecutive steps corresponding to the same problem are arranged in sequence to obtain the solution corresponding to the problem, including:
[0019] Based on the question-answer pair to which a step in any consecutive step belongs, determine the question corresponding to the consecutive step;
[0020] Arrange at least two consecutive sets of steps corresponding to the problem in sequence to obtain the solution to the problem.
[0021] Optionally, in the above method, if there are at least two sets of consecutive steps corresponding to the same problem, the consecutive steps corresponding to the same problem are arranged in sequence to obtain the solution corresponding to the problem, including:
[0022] Arrange at least two consecutive sets of steps corresponding to the same problem in sequence to obtain the target solution;
[0023] Based on the question-answer pair to which any step belongs in at least two consecutive sets of steps corresponding to the same question, determine the question corresponding to the step;
[0024] Based on the problem corresponding to the steps and the target solution, the solution corresponding to the problem is determined as the target solution.
[0025] Optionally, the above method, before obtaining the session record, further includes:
[0026] Obtain at least two sets of training question-answer pairs from a domain-specific knowledge base;
[0027] Add question tags to the training questions and step tags to the training steps in each training question-answer pair.
[0028] The original pairing model is trained based on at least the labeled training question-answer pairs to obtain the pairing model.
[0029] Optionally, the above method, before obtaining the session record, further includes:
[0030] Obtain at least two sets of training schemes from a specific domain knowledge base, each set of training schemes containing at least one step;
[0031] Label the training steps in each training scheme sequentially to identify at least one set of consecutive training steps;
[0032] The original continuous model is trained based on at least one set of training sequential steps to obtain the continuous model.
[0033] An information processing device, comprising:
[0034] The acquisition module is used to acquire session records, wherein the session records include at least a first session generated by a first session object and a second session generated by a second session object;
[0035] An analysis module is used to analyze the conversation records to obtain at least two sets of question-answer pairs. Each set of question-answer pairs includes a first conversation as a question and a second conversation as a step. The step is a step of the solution corresponding to the question in the question-answer pair. Among them, a question corresponds to at least one solution, and a solution includes at least one step.
[0036] The module determines the solution corresponding to each question in the session record based on the characteristics of the steps.
[0037] The summary generation module is used to display the questions and corresponding solutions in sequence based on the process of question generation in the session record, so as to obtain the target summary content of the session record.
[0038] A readable storage medium having a computer program stored thereon, the computer program being invoked and executed by a processor to implement the steps of the information processing method as described in any of the preceding claims.
[0039] As can be seen from the above technical solution, this application provides an information processing method. After obtaining a session record, the session record is analyzed to obtain multiple sets of question-and-answer pairs. Each set of question-and-answer pairs includes a first session as a question and a second session as a step. The step is a step of the solution corresponding to the question in the question-and-answer pair. Each question corresponds to at least one solution, and each solution includes at least one step. Based on the characteristics of the steps, the solution corresponding to each question in the session record is determined. Based on the generation process of the questions in the session record, the questions and their corresponding solutions are displayed in sequence to obtain the target summary content of the session record. In this solution, in the face of multi-intent session scenarios, each set of question-and-answer pairs is first determined, and then the question-and-answer pairs are sorted as a whole. The resulting target summary content is not affected by the disordered order of the sessions in the session record, and the quality of the target summary is high. Moreover, noisy data in the session record does not belong to any question-and-answer pair. Therefore, dialogue information that does not belong to any question-and-answer pair is not added to the target summary content, achieving the effect of removing noisy data from the session record. Attached Figure Description
[0040] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0041] Figure 1 This is a flowchart of an embodiment 1 of an information processing method provided in this application;
[0042] Figure 2 This is a schematic diagram of a session record in Embodiment 1 of an information processing method provided in this application;
[0043] Figure 3 This is a schematic diagram of the target abstract content in Embodiment 1 of an information processing method provided in this application;
[0044] Figure 4 This is a flowchart of an embodiment 2 of an information processing method provided in this application;
[0045] Figure 5 This is a flowchart of an embodiment 3 of an information processing method provided in this application;
[0046] Figure 6 This is a flowchart of embodiment 4 of the information processing method provided in this application;
[0047] Figure 7 This is a flowchart of embodiment 5 of the information processing method provided in this application;
[0048] Figure 8 This is a flowchart of an embodiment 6 of an information processing method provided in this application;
[0049] Figure 9 This is a flowchart of embodiment 7 of an information processing method provided in this application;
[0050] Figure 10 This is a flowchart of embodiment 8 of the information processing method provided in this application;
[0051] Figure 11 This is a schematic diagram of an embodiment of an information processing device provided in this application. Detailed Implementation
[0052] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0053] like Figure 1 The flowchart shown is a sample of an embodiment 1 of an information processing method provided in this application. The method is applied to an electronic device, which is a device with information processing capabilities and has an output structure. The method includes the following steps:
[0054] Step S101: Obtain session records;
[0055] The session records include at least a first session generated by a first session object and a second session generated by a second session object.
[0056] The session record can be a record of a conversation between a user and customer service. This conversation can take place between the user and the customer service server via a personal device (such as a mobile terminal or personal computer) or via a customer terminal (such as an ATM).
[0057] This involves obtaining the session record, which contains the session generated by both parties.
[0058] Specifically, the session record includes a first session generated by a first session object and a second session generated by a second session object.
[0059] As an example, the first conversation object is the user, and the corresponding second conversation object is the customer service representative. The user consults on one or more issues, and the customer service representative provides feedback on the issues. The content generated in this process is the conversation record involved in this embodiment.
[0060] Step S102: Analyze the conversation records to obtain at least two sets of question-answer pairs;
[0061] Each question-and-answer pair includes a first session as a question and a second session as a step, wherein the step is a step of a solution corresponding to the question in the question-and-answer pair; wherein a question corresponds to at least one solution, and a solution includes at least one step.
[0062] This session record is a record of a multi-intent scenario session. Specifically, this session record contains multiple questions and their corresponding responses.
[0063] The conversation record contains multiple questions and answers to those questions, forming question-and-answer pairs.
[0064] Each question-and-answer pair contains one question and one step.
[0065] In this approach, each question corresponds to a solution. When a solution has only one step, a single question-and-answer pair can represent both the question and the corresponding solution. When a solution has multiple steps, the multiple question-and-answer pairs obtained in a given step relate to the same question.
[0066] The questions and feedback steps in this conversation record overlap in terms of language expression.
[0067] like Figure 2 The diagram shows a conversation log. Based on the timeline, the user first asks question Q1, and the customer service replies with solution 1, which has only one step, A1. The user then asks question Q2, and the customer service replies with solution 2, which has two steps, A21 and A22. After the customer service replies with step A21 and before step A22, the user asks question Q3, and the customer service replies with solution 3, which has only one step, A3, after replying with step A22.
[0068] It should be noted that if the session records are analyzed, some sessions (the first session and / or the second session) cannot form question-and-answer pairs, and these sessions that cannot form question-and-answer pairs can be considered as noise.
[0069] As an example, the conversations at the beginning of the conversation record, such as "Hello!" for the first conversation object and "Hello! How can I help you?" for the second conversation object, and "Okay, thank you" for the first conversation object and "You're welcome, it's a pleasure to help you" for the second conversation object, are irrelevant to the question and solution and cannot form a question-and-answer pair. They can be ignored or deleted as noise.
[0070] Step S103: Based on the characteristics of the steps, determine the solution corresponding to each question in the session record;
[0071] In this process, once multiple question-answer pairs are identified, if the steps in these multiple question-answer pairs belong to the same question, then the multiple steps can be combined to obtain a solution to the question; if a question-answer pair corresponds to only one step, then the steps in that question-answer pair can be determined as a solution to the question.
[0072] Specifically, the order of steps in the solution is fixed. Therefore, in the session generated by the second session object, the steps of the same solution are arranged in chronological order. Thus, the solution to the corresponding problem can be obtained by combining the steps based on their chronological order of appearance in the session record.
[0073] Specifically, the order of steps in the solution is fixed, and the steps are continuous. Based on the continuity of the steps in the solution, multiple steps are combined to obtain a solution to the corresponding problem.
[0074] If steps belonging to the same solution have corresponding characteristics, then based on these characteristics, the problem pair corresponding to the session record is processed to obtain the solution corresponding to the problem.
[0075] If a step in a question-and-answer pair does not satisfy the corresponding feature with the steps in any other question-and-answer pair, then the step in that question-and-answer pair is determined to be the solution to the question, and the question has only one step.
[0076] If a step in a question-and-answer pair satisfies a certain characteristic with the steps in one or more other question-and-answer pairs, then the steps in those other question-and-answer pairs together constitute a solution to the problem.
[0077] Step S104: Based on the process of how the questions were generated in the session record, display the questions and their corresponding solutions in sequence to obtain the target summary content of the session record.
[0078] The target summary content, after removing noise from the conversation records, contains records with substantial question-and-answer content, including the questions and corresponding solutions.
[0079] Specifically, the order in which questions arise in the session record is used to determine the display order of the questions in the session record. The questions and their corresponding solutions are then displayed according to this display order to obtain the target summary content of the session record.
[0080] like Figure 3 The diagram shows the target summary content, which includes, in sequence: step A1 of problem 1 (Q1) and solution 1; steps A21 and A22 of problem 2 (Q2) and solution 2; and step A3 of problem 3 (Q3) and solution 3.
[0081] In summary, this embodiment provides an information processing method comprising: after obtaining a session record, analyzing the session record to obtain multiple sets of question-and-answer pairs, each set of question-and-answer pairs including a first session as a question and a second session as a step, wherein the step is a step of the solution corresponding to the question in the question-and-answer pair; wherein, a question corresponds to at least one solution, and a solution includes at least one step; determining the solution corresponding to each question in the session record based on the characteristics of the steps; and displaying the questions and corresponding solutions in sequence based on the generation process of the questions in the session record to obtain the target summary content of the session record. In this solution, in the face of multi-intent session scenarios, each set of question-and-answer pairs is determined first, and then the question-and-answer pairs are sorted as a whole. The obtained target summary content is not affected by the disordered order of the sessions in the session record, and the quality of the target summary is high. Moreover, the noisy data in the session record does not belong to any question-and-answer pair; therefore, the dialogue information that does not belong to any question-and-answer pair is not added to the target summary content, thus achieving the effect of removing noisy data in the session record.
[0082] like Figure 4 The diagram shown is a flowchart of an embodiment 2 of an information processing method provided in this application. The method includes the following steps:
[0083] Step S401: Obtain session records;
[0084] Step S401 is the same as step S101 in Embodiment 1, and will not be described again in this embodiment.
[0085] Step S402: Based on the pairing model, perform matching analysis on the first and second sessions in the session records, and form a question-answer pair with the first and second sessions that meet the question-answer matching conditions;
[0086] The steps are as follows: the first session that meets the question-and-answer matching conditions is taken as the question of the question-and-answer pair, and the second session that meets the question-and-answer matching conditions is taken as the question-and-answer pair.
[0087] The electronic device has a pre-set pairing model that performs matching analysis on the first and second sessions in the session record to obtain question-answer pairs.
[0088] As an example, the pairing model could employ an ispairs model, which is capable of identifying whether any first session and any second session can form a question-and-answer pair.
[0089] Specifically, you can choose one in the first session and one in the second session. The pairing model determines whether the first and second sessions meet the question-and-answer matching conditions. If they do, the two form a question-and-answer pair; if they do not, they cannot form a question-and-answer pair.
[0090] In practice, since the second session object is a response to the first session of the first session object, in order to reduce the data processing volume of the pairing model, it can be selected from the second session that appears after the first session. The pairing model determines whether the first session and the second session meet the question-and-answer matching conditions. If they meet the conditions, they form a question-and-answer pair; if they do not meet the conditions, they cannot form a question-and-answer pair.
[0091] If a first session is selected and no second session matches the first session, the first session can be ignored and treated as noise. Another first session is then selected, and it is determined whether the second session matches the selected first session.
[0092] Step S403: Based on the characteristics of the steps, determine the solution corresponding to each question in the session record;
[0093] Step S404: Based on the process of how the questions were generated in the session record, display the questions and their corresponding solutions in sequence to obtain the target summary content of the session record.
[0094] Steps S403-404 are the same as steps S103-104 in Example 1, and will not be described again in this example.
[0095] In summary, the information processing method provided in this embodiment includes: performing matching analysis on the first and second sessions in the session records based on a pairing model; forming a question-and-answer pair by combining the first and second sessions that meet the question-and-answer matching conditions; wherein the first session that meets the question-and-answer matching conditions is used as the question in the question-and-answer pair, and the second session that meets the question-and-answer matching conditions is used as the question in the question-and-answer pair. In this solution, the first and second sessions that meet the question-and-answer matching conditions are combined into a question-and-answer pair, while the first or second sessions that do not meet the question-and-answer matching conditions cannot be combined into a question-and-answer pair. This ensures that sessions with valid question-and-answer content can enter the target summary content, while sessions with invalid content will not enter the target content summary, thus achieving the effect of removing noise data from the session records.
[0096] like Figure 5 The diagram shown is a flowchart of an embodiment 3 of an information processing method provided in this application. The method includes the following steps:
[0097] Step S501: Obtain session records;
[0098] Step S502: Analyze the conversation records to obtain at least two sets of question-answer pairs;
[0099] Steps S501-502 are the same as steps S101-102 in Example 1, and will not be described again in this example.
[0100] Step S503: Obtain at least two question-answer pairs corresponding to the same question to obtain the target question-answer pair set;
[0101] In this process, the multiple sets of question-and-answer pairs are divided into sets according to the corresponding questions.
[0102] Specifically, select a question, determine all its corresponding question-answer pairs, and use all the question-answer pairs corresponding to that question as the target question-answer pair set for that question.
[0103] When a question has only one question-answer pair, that single question-answer pair can be directly regarded as the question-answer pair set for that question, and this question-answer pair set is a set of single elements.
[0104] As an example, multiple question-answer pairs are filtered to obtain the question-answer pairs for question Q2, including Q2-A21 and Q2-A22, and the question-answer pairs for question Q4, including Q4-A41, Q4-A42, Q4-A43, and Q4-A44. The question-answer pairs Q4-A41, Q4-A42, Q4-A43, and Q4-A44 are combined to obtain a question-answer pair set, and the question-answer pairs Q2-A21 and Q2-A22 are combined to obtain a question-answer pair set.
[0105] Step S504: Based on the occurrence process of steps in the conversation record, arrange the steps of each question-answer pair in the target question-answer pair set in order to obtain the solution corresponding to the question;
[0106] In a conversation log, if a solution to a problem involves multiple steps, the customer service representative will respond according to the order in which the steps are performed. Therefore, the order of the steps in the solution for the same problem in the conversation log is determined based on the order in which they were generated, thus obtaining the solution for that problem.
[0107] As an example, consider the question-answer pair set Q4-A41, Q4-A42, Q4-A43, and Q4-A44, where the steps appear in the conversation record in the order of A41, A42, A43, and A44. Arranging these steps in sequence yields the solution corresponding to question Q4, which includes steps A41, A42, A43, and A44.
[0108] Step S505: Based on the process of how the questions were generated in the session record, display the questions and their corresponding solutions in sequence to obtain the target summary content of the session record.
[0109] Step S505 is the same as step S104 in Example 1, and will not be described again in this example.
[0110] In summary, the information processing method provided in this embodiment includes: obtaining at least two question-answer pairs corresponding to the same question to obtain a target question-answer pair set; and arranging the steps of each question-answer pair in the target question-answer pair set in sequence based on the order of appearance of steps in the conversation record to obtain a solution corresponding to the question. In this solution, multiple question-answer pairs corresponding to the same question are first combined into a target question-answer pair set. Based on the order of appearance of each step in the conversation record, the steps in each question-answer pair in the target question-answer pair set are arranged in sequence to obtain a solution corresponding to the question, thus realizing the process of determining the solution corresponding to the question.
[0111] like Figure 6 The diagram shown is a flowchart of an embodiment 4 of an information processing method provided in this application. The method includes the following steps:
[0112] Step S601: Obtain session records;
[0113] Step S602: Analyze the conversation records to obtain at least two sets of question-answer pairs;
[0114] Steps S601-602 are the same as steps S101-102 in Example 1, and will not be described again in this example.
[0115] Step S603: Based on the continuous model, analyze at least two second sessions in the session records, and group at least two second sessions that meet the continuous condition into a set of continuous steps;
[0116] In this context, any set of consecutive steps corresponds to the same problem, and the set of consecutive steps constitutes a solution.
[0117] The electronic device has a preset continuous model. For any two second sessions in the session record, the continuous condition is determined, and the second sessions that meet the continuous condition are grouped into a continuous step.
[0118] As an example, the continuous model can take the isnext model, which is able to identify whether any two second sessions are consecutive steps.
[0119] In practice, since the second session object is a response to the first session of the first session object, it will sequentially feed back the steps in the response plan. Therefore, the order of the steps generated by the second session object for the same plan is determined sequentially. So, in order to reduce the amount of data processing in the continuous model, after selecting a certain second session, a process can be performed to determine whether the two satisfy the continuity condition from the second sessions that appear after that second session.
[0120] For example, the continuity condition includes two consecutive steps, three consecutive steps, or even more consecutive steps.
[0121] In this embodiment, the two steps are described as consecutive.
[0122] The second sessions contained in the session records are steps A1, A21, A22, A3, A41, A42, A43, and A44. Any two second sessions are obtained in sequence, and the second sessions that satisfy the consecutive condition are determined to be steps A21 and A22, A41 and A42, A42 and A43, and A43 and A44.
[0123] Step S604: Arrange the consecutive steps corresponding to the same problem in sequence to obtain the solution corresponding to the problem;
[0124] Among them, consecutive steps corresponding to the same problem are transitive. By sorting consecutive steps corresponding to the same problem, the solution corresponding to the problem can be obtained.
[0125] As an example, consecutive steps A41 and A42, consecutive steps A42 and A43, and consecutive steps A43 and A44 all correspond to problem 4. Based on the transitivity of consecutive steps, the above three sets of consecutive steps are ordered in sequence to obtain steps A41, A42, A43, and A44. The above four consecutive steps are the solutions corresponding to problem 4.
[0126] Step S605: Based on the process of how the questions were generated in the session record, display the questions and their corresponding solutions in sequence to obtain the target summary content of the session record.
[0127] Step S605 is the same as step S104 in Example 1, and will not be described again in this example.
[0128] In summary, the information processing method provided in this embodiment includes: based on a continuity model, analyzing at least two second sessions in the session records; grouping at least two second sessions that satisfy the continuity condition into a set of continuous steps, where any set of continuous steps corresponds to the same problem, and the set of continuous steps constitutes a solution; and sequentially arranging the continuous steps corresponding to the same problem to obtain the solution corresponding to the problem. In this solution, based on the transitivity of continuous steps, sorting multiple continuous steps yields the solution corresponding to the problem.
[0129] like Figure 7 The flowchart shown is a 5th embodiment of an information processing method provided in this application. The method includes the following steps:
[0130] Step S701: Obtain session records;
[0131] Step S702: Analyze the conversation records to obtain at least two sets of question-answer pairs;
[0132] Step S703: Based on the continuous model, analyze at least two second sessions in the session records, and group at least two second sessions that meet the continuous condition into a set of continuous steps;
[0133] Steps S701-703 are the same as steps S601-603 in Example 4, and will not be described again in this example.
[0134] Step S704: Based on the question-answer pair to which any step in a series of steps belongs, determine the question corresponding to the series of steps;
[0135] In this process, once multiple sets of consecutive steps are obtained from the second session based on the session records, it is necessary to further determine the relationship between these multiple sets of consecutive steps and whether they can form a solution to the same problem.
[0136] The solution in this embodiment is provided to illustrate the situation where there are multiple sets of consecutive steps corresponding to the same problem.
[0137] Specifically, since each step in a series of consecutive steps corresponds to the same question, we first determine the question corresponding to the question-answer pair to which any step belongs in the series of consecutive steps. This question is the question corresponding to the series of consecutive steps.
[0138] As an example, the second sessions in the session record that satisfy the continuity condition are steps A21 and A22, steps A41 and A42, steps A42 and A43, and steps A43 and A44. Among them, steps A21 and A22 correspond to question 2, steps A41 and A42 correspond to question 4, steps A42 and A43 correspond to question 4, and steps A43 and A44 correspond to question 4.
[0139] Step S705: Arrange at least two consecutive sets of steps corresponding to the problem in sequence to obtain the solution to the problem;
[0140] If there are two or more sets of consecutive steps corresponding to the same problem, then according to the transitivity of the multiple sets of consecutive steps, the consecutive steps corresponding to the same problem are sorted to obtain the solution corresponding to the problem.
[0141] As an example, consecutive steps A41 and A42, consecutive steps A42 and A43, and consecutive steps A43 and A44 all correspond to problem 4. Based on the transitivity of consecutive steps, the above three sets of consecutive steps are ordered in sequence to obtain steps A41, A42, A43, and A44. The above four consecutive steps are the solutions corresponding to problem 4.
[0142] Step S706: Based on the process of how the questions were generated in the session record, display the questions and their corresponding solutions in sequence to obtain the target summary content of the session record.
[0143] Step S706 is the same as step S605 in Example 4, and will not be described again in this example.
[0144] In summary, the information processing method provided in this embodiment includes: determining the question corresponding to any consecutive step based on the question-answer pair to which the step belongs; and arranging at least two sets of consecutive steps corresponding to the question in sequence to obtain the solution corresponding to the question. In this solution, firstly, based on the question-answer pair to which any step belongs in the consecutive steps, the question corresponding to that step is determined, thus obtaining the question corresponding to that consecutive step. Then, based on the transitivity of the consecutive steps, multiple sets of consecutive steps corresponding to the question are arranged in sequence to obtain the solution corresponding to the question.
[0145] like Figure 8 The flowchart shown is a sample of embodiment 6 of an information processing method provided in this application. The method includes the following steps:
[0146] Step S801: Obtain session records;
[0147] Step S802: Analyze the conversation records to obtain at least two sets of question-answer pairs;
[0148] Step S803: Based on the continuous model, analyze at least two second sessions in the session records, and group at least two second sessions that meet the continuous condition into a set of continuous steps;
[0149] Steps S801-803 are the same as steps S601-603 in Example 4, and will not be described again in this example.
[0150] Step S804: Arrange at least two consecutive sets of steps corresponding to the same problem in sequence to obtain the target solution;
[0151] The solution in this embodiment is provided to illustrate the situation where there are multiple sets of consecutive steps corresponding to the same problem.
[0152] In this process, after obtaining multiple sets of consecutive steps in the second session based on the session records, the multiple sets of consecutive steps corresponding to the same problem are sorted based on the transitivity of the consecutive steps.
[0153] It should be noted that in this step, the multiple sets of consecutive steps can be sorted based on the transitivity of consecutive steps. This only determines that the consecutive steps correspond to the same problem, but it does not determine which specific problem the consecutive steps correspond to.
[0154] As an example, the second sessions in the session record that satisfy the continuity condition are steps A21 and A22, steps A41 and A42, steps A42 and A43, and steps A43 and A44. Based on the transitivity of continuous steps, it is determined that multiple sets of continuous steps A41 and A42, steps A42 and A43, and steps A43 and A44 corresponding to the same problem can be combined to obtain a solution, which includes steps A41, A42, A43, and A44. Another solution can be obtained directly from steps A21 and A22.
[0155] Step S805: Based on the question-answer pair to which any step belongs in at least two sets of consecutive steps corresponding to the same question, determine the question corresponding to the step;
[0156] Among them, the multiple consecutive steps belonging to the same target solution correspond to the same question. Select any one step, determine the question-answer pair to which the step belongs, and obtain the question corresponding to the step based on the question-answer pair to which the step belongs.
[0157] Step S806: Based on the problem corresponding to the step and the target solution, determine the solution corresponding to the problem as the target solution;
[0158] Since multiple consecutive steps belonging to the same objective solution correspond to the same problem, determining the corresponding problem based on a certain step is the problem corresponding to that objective solution.
[0159] Accordingly, based on the problem corresponding to this step and the target solution, it can be determined that the solution corresponding to this problem is the target solution.
[0160] In practice, when there are multiple target solutions in the session record, the correspondence between the multiple issues involved in the session record and the target solutions can be determined.
[0161] Step S807: Based on the process of how the questions were generated in the session record, display the questions and their corresponding solutions in sequence to obtain the target summary content of the session record.
[0162] Step S807 is the same as step S605 in Example 4, and will not be described again in this example.
[0163] In summary, the information processing method provided in this embodiment includes: sequentially arranging at least two sets of consecutive steps corresponding to the same question to obtain a target solution; determining the question corresponding to the step based on the question-answer pair to which any step in the at least two sets of consecutive steps corresponding to the same question belongs; and determining the solution corresponding to the question as the target solution based on the question corresponding to the step and the target solution. In this solution, multiple sets of consecutive steps belonging to the same solution are first determined, then the question corresponding to any step in the solution is determined to identify the question corresponding to the solution, and the solution corresponding to the question is obtained as the target solution based on the question and the target solution.
[0164] like Figure 9 The flowchart shown is a 7th embodiment of an information processing method provided in this application. The method includes the following steps:
[0165] Step S901: Obtain at least two sets of training question-answer pairs from a specific domain knowledge base;
[0166] Specifically, based on the domain involved in this method, a corresponding domain knowledge base is determined, which stores the problems in this domain and the corresponding solution steps.
[0167] In this process, a set of training question-answer pairs is obtained from a question and a step in the solution corresponding to the question from the knowledge base of the domain. Similarly, multiple sets of training question-answer pairs are obtained from the knowledge base of the domain.
[0168] It should be noted that the questions and corresponding solution steps stored in the knowledge base of this domain can be pre-written into the data, and can provide feedback on user inquiries.
[0169] Step S902: Add question tags to the training questions and step tags to the training steps in each training question-answer pair in sequence;
[0170] In this set of training question-and-answer teams, there is a training question and a training step, and the two have a correct question-and-answer relationship.
[0171] Specifically, add question tags to the training questions in each training question-answer pair and add step tags to the training steps.
[0172] Step S903: Train the original pairing model based at least on the labeled training question-answer pairs to obtain the pairing model;
[0173] The original pairing model is trained by adding question and step labels to the training question-answer pairs, so that the original pairing model can be trained to obtain a pairing model, which can identify whether the first and second sessions in the session record are a question-answer pair.
[0174] In practice, to improve the accuracy of training, the training question-answer pairs are used as positive samples, and negative samples are also obtained. The original pairing model is trained based on the positive and negative samples to obtain the pairing model.
[0175] As an example, for problem Q, there are multiple solutions, S1 and S2. Solution S1 includes steps S11 and S12, and solution S2 includes steps S21, S22 and S23. Positive samples are obtained from [(Q, S11), (Q, S12), (Q, S21), (Q, S22), (Q, S23)]. One or more steps of other problems are randomly obtained from the domain data. These steps are unrelated to the steps in the solutions to problem Q, such as S'. Negative samples (Q, S') are obtained.
[0176] It should be noted that, in order to improve the accuracy of training, when generating negative samples, the selected step S' needs to be different from the step in the positive samples.
[0177] It should be noted that for some domain databases with a small amount of data, all the data can be used as training samples to train the original pairing model. For some domain databases with a large amount of data, a portion of the data can be selected as training samples. This application does not impose any restrictions on the amount of data sampled from the domain data.
[0178] Step S904: Obtain session records;
[0179] Step S905: Based on the pairing model, perform matching analysis on the first and second sessions in the session records, and form a question-answer pair with the first and second sessions that meet the question-answer matching conditions;
[0180] Step S906: Based on the characteristics of the steps, determine the solution corresponding to each question in the session record;
[0181] Step S907: Based on the process of how the questions were generated in the session record, display the questions and their corresponding solutions in sequence to obtain the target summary content of the session record.
[0182] Steps S904-907 are the same as steps S401-404 in Example 2, and will not be described again in this example.
[0183] In summary, the information processing method provided in this embodiment further includes: obtaining at least two sets of training question-answer pairs in a specific domain knowledge base; sequentially adding question tags to the training questions and step tags to the training steps in each set of training question-answer pairs; and training an original pairing model based at least on the labeled training question-answer pairs to obtain a pairing model. In this solution, the original pairing model is trained based on data in a specific domain knowledge base to obtain a pairing model, and training based on real data results in high accuracy.
[0184] like Figure 10 The flowchart shown is an embodiment 8 of an information processing method provided in this application. The method includes the following steps:
[0185] Step S1001: Obtain at least two sets of training schemes from a specific domain knowledge base, each set of training schemes containing at least one step;
[0186] Specifically, based on the domain involved in this method, a corresponding domain knowledge base is determined, which stores the problems in this domain and the corresponding solution steps.
[0187] Among them, multiple sets of schemes are obtained from the knowledge base of this field, and these multiple sets of schemes are used as training schemes.
[0188] It should be noted that the solution steps stored in this domain knowledge base can be pre-written into the data, and can provide feedback on user inquiries.
[0189] Step S1002: Add labels to the training steps in each training scheme in sequence to identify at least one set of consecutive training steps;
[0190] In this process, a label is added to each training step in the training scheme. This label is used to indicate that multiple training steps are in the same training scheme and to indicate the order in which these multiple training steps are performed.
[0191] In this process, a label is added to each training step in the multiple training schemes obtained in sequence. Based on the label, multiple steps belonging to the same training scheme with a sequential order can be identified.
[0192] As an example, the training scheme S has ordered steps [t1,t2,t3,t4], and based on this training scheme, three sets of consecutive training steps [t1,t2], [t2,t3], and [t3,t4] are obtained.
[0193] It should be noted that in this embodiment, the number of steps in a set of training successive steps is 2, which is from the perspective of computational cost. The computational cost of 2 steps is relatively small, and the overall scheme can be determined based on the transitivity of the steps.
[0194] It should be noted that the number of steps in a set of consecutive training steps can be set according to the actual situation, such as 2, 3, 4, etc. This application does not impose any restrictions on the value of this number.
[0195] Step S1003: Train the original continuous model based on at least one set of training successive steps to obtain the continuous model;
[0196] The process involves training the original continuous model with labeled training steps, thereby training the original continuous model to obtain a continuous model that can identify whether any two or more second sessions in the session record are consecutive steps.
[0197] In practice, to improve the accuracy of training, the continuous training steps are used as positive samples, and negative samples are also obtained. The original continuous model is trained based on the positive and negative samples to obtain the continuous model.
[0198] As an example, training scheme S has ordered steps [t1,t2,t3,t4]. Based on this training scheme, three sets of consecutive training steps [t1,t2], [t2,t3], and [t3,t4] are obtained, resulting in positive samples containing [t1,t2], [t2,t3], and [t3,t4]. In the domain data, one or more steps from other schemes are randomly obtained, which are unrelated to the steps in scheme S, such as t1' and t2', resulting in negative samples (t1', t2').
[0199] It should be noted that, in order to improve the accuracy of training, the steps (t1', t2') selected when generating negative samples need to be different from those in positive samples.
[0200] It should be noted that for some domain databases with a small amount of data, all the data can be used as training samples to train the original continuous model. For some domain databases with a large amount of data, a portion of the data can be selected as training samples. This application does not impose any restrictions on the amount of data sampled from the domain data.
[0201] Step S1004: Obtain session records;
[0202] Step S1005: Analyze the conversation records to obtain at least two sets of question-answer pairs;
[0203] Step S1006: Based on the continuous model, analyze at least two second sessions in the session records, and group at least two second sessions that meet the continuous condition into a set of continuous steps;
[0204] Step S1007: Arrange the consecutive steps corresponding to the same problem in sequence to obtain the solution corresponding to the problem;
[0205] Step S1008: Based on the process of how the questions were generated in the session record, display the questions and their corresponding solutions in sequence to obtain the target summary content of the session record.
[0206] Steps S1004-1008 are the same as steps S601-605 in Example 4, and will not be described again in this example.
[0207] In summary, the information processing method provided in this embodiment further includes: obtaining at least two sets of training schemes in a specific domain knowledge base, each set of training schemes containing at least one step; sequentially adding labels to the training steps in each set of training schemes to determine at least one set of training continuous steps; and training an original continuous model based on at least the at least one set of training continuous steps to obtain a continuous model. In this solution, the original continuous model is trained based on data in a specific domain knowledge base to obtain a continuous model, and training based on real data results in high accuracy.
[0208] Corresponding to the above-described embodiment of an information processing method provided in this application, this application also provides an embodiment of an apparatus for applying the information processing method.
[0209] like Figure 11 The diagram shown is a structural schematic of an embodiment of an information processing device provided in this application. The device includes the following structure: an acquisition module 1101, an analysis module 1102, a determination module 1103, and a summary generation module 1104.
[0210] The obtaining module 1101 is used to obtain session records, wherein the session records include at least a first session generated by a first session object and a second session generated by a second session object;
[0211] The analysis module 1102 is used to analyze the conversation records to obtain at least two sets of question-answer pairs. Each set of question-answer pairs includes a first conversation as a question and a second conversation as a step. The step is a step of the solution corresponding to the question in the question-answer pair. A question corresponds to at least one solution, and a solution includes at least one step.
[0212] The determining module 1103 determines the solution corresponding to each question in the session record based on the characteristics of the steps.
[0213] The summary generation module 1104 is used to display the questions and corresponding solutions in sequence based on the process of generating the questions in the session record, so as to obtain the target summary content of the session record.
[0214] Optionally, this analysis module is specifically used for:
[0215] Based on the pairing model, a matching analysis is performed on the first and second sessions in the session records. The first and second sessions that meet the question-and-answer matching conditions are grouped into a question-and-answer pair, wherein the first session that meets the question-and-answer matching conditions is used as the question in the question-and-answer pair, and the second session that meets the question-and-answer matching conditions is used as the question in the question-and-answer pair.
[0216] Optionally, determine the specific use of the module for:
[0217] Obtain at least two question-answer pairs corresponding to the same question to obtain the target question-answer pair set;
[0218] Based on the occurrence sequence of steps in the conversation record, the steps of each question-answer pair in the target question-answer pair set are arranged in order to obtain the solution corresponding to the question.
[0219] Optionally, determine the specific use of the module for:
[0220] Based on the continuous model, at least two second sessions in the session records are analyzed, and at least two second sessions that meet the continuous condition are grouped into a group of continuous steps. Any group of continuous steps corresponds to the same problem, and the group of continuous steps is a solution.
[0221] Arrange the consecutive steps corresponding to the same problem in sequence to obtain the solution to the problem.
[0222] Optionally, if there are at least two consecutive sets of steps corresponding to the same problem, the summary generation module is specifically used for:
[0223] Based on the question-answer pair to which a step in any consecutive step belongs, determine the question corresponding to the consecutive step;
[0224] Arrange at least two consecutive sets of steps corresponding to the problem in sequence to obtain the solution to the problem.
[0225] Optionally, if there are at least two consecutive sets of steps corresponding to the same problem, the summary generation module is specifically used for:
[0226] Arrange at least two consecutive sets of steps corresponding to the same problem in sequence to obtain the target solution;
[0227] Based on the question-answer pair to which any step belongs in at least two consecutive sets of steps corresponding to the same question, determine the question corresponding to the step;
[0228] Based on the problem corresponding to the steps and the target solution, the solution corresponding to the problem is determined as the target solution.
[0229] Optional, also includes:
[0230] The first training module is used to obtain at least two sets of training question-answer pairs from a specific domain knowledge base; to add question labels to the training questions and step labels to the training steps in each set of training question-answer pairs; and to train the original pairing model based on at least the labeled training question-answer pairs to obtain the pairing model.
[0231] Optional, also includes:
[0232] The second training module is used to obtain at least two sets of training schemes from a specific domain knowledge base, each set of training schemes containing at least one step; to add labels to the training steps in each set of training schemes in turn to determine at least one set of training continuous steps; and to train the original continuous model based on at least the at least one set of training continuous steps to obtain a continuous model.
[0233] It should be noted that the structural and functional descriptions of the information processing device involved in this embodiment are explained in the method embodiment and will not be repeated in this embodiment.
[0234] In summary, this application provides an information processing apparatus that, in multi-intent conversation scenarios, first identifies each question-and-answer pair, then sorts the question-and-answer pairs as a whole. The resulting target summary content is unaffected by the disordered order of individual conversations in the conversation record, resulting in high-quality target summaries. Furthermore, noisy data in the conversation record does not belong to any question-and-answer pair; therefore, dialogue information that does not belong to any question-and-answer pair is not added to the target summary content, achieving the effect of removing noisy data from the conversation record.
[0235] Corresponding to the above-described embodiment of an information processing method provided in this application, this application also provides an electronic device and a readable storage medium corresponding to the information processing method.
[0236] The electronic device includes: a memory and a processor;
[0237] The memory stores the processing program;
[0238] The processor is used to load and execute the processing program stored in the memory to implement the steps of the information processing method as described in any of the preceding claims.
[0239] For details on the specific information processing method implemented in this electronic device, please refer to the aforementioned information processing method embodiments.
[0240] The readable storage medium stores a computer program that is invoked and executed by a processor to implement the steps of the information processing method as described in any of the preceding claims.
[0241] For details on the specific implementation of the information processing method by executing the computer program stored in the readable storage medium, please refer to the aforementioned information processing method embodiments.
[0242] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. The apparatus provided in the embodiments is described simply because it corresponds to the method provided in the embodiments; relevant parts can be found in the method section.
[0243] The above description of the provided embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features provided herein.
Claims
1. An information processing method, comprising: Obtain session records, wherein the session records include at least a first session generated by a first session object and a second session generated by a second session object; Analyzing the conversation records yields at least two question-and-answer pairs. Each pair includes a first conversation as a question and a second conversation as a step. The step is a step in the solution corresponding to the question in the question-and-answer pair. Each question corresponds to at least one solution, and each solution includes at least one step. The questions and feedback steps in the conversation records overlap in terms of language expression. Based on the continuity of steps, the solution corresponding to each question in the conversation record is determined; if the solution corresponding to a question includes multiple steps, the steps belonging to the same question in multiple question-answer pairs are combined to obtain the solution corresponding to that question. Based on the process of how the questions were generated in the session record, the questions and their corresponding solutions are displayed in sequence to obtain the target summary content of the session record.
2. The method according to claim 1, wherein analyzing the conversation record to obtain at least two sets of question-answer pairs includes: Based on the pairing model, a matching analysis is performed on the first and second sessions in the session records. The first and second sessions that meet the question-and-answer matching conditions are grouped into a question-and-answer pair, wherein the first session that meets the question-and-answer matching conditions is used as the question in the question-and-answer pair, and the second session that meets the question-and-answer matching conditions is used as the question in the question-and-answer pair.
3. The method according to claim 1, wherein determining the scheme corresponding to each question in the session record based on the features of the step includes: Obtain at least two question-answer pairs corresponding to the same question to obtain the target question-answer pair set; Based on the occurrence sequence of steps in the conversation record, the steps of each question-answer pair in the target question-answer pair set are arranged in order to obtain the solution corresponding to the question.
4. The method according to claim 1, wherein determining the scheme corresponding to each question in the session record based on the features of the steps includes: Based on the continuous model, at least two second sessions in the session records are analyzed, and at least two second sessions that meet the continuous condition are grouped into a group of continuous steps. Any group of continuous steps corresponds to the same problem, and the group of continuous steps is a solution. Arrange the consecutive steps corresponding to the same problem in sequence to obtain the solution to the problem.
5. The method according to claim 4, wherein if at least two sets of consecutive steps correspond to the same problem, the consecutive steps corresponding to the same problem are arranged in sequence to obtain a solution corresponding to the problem, comprising: Based on the question-answer pair to which a step in any consecutive step belongs, determine the question corresponding to the consecutive step; Arrange at least two consecutive sets of steps corresponding to the problem in sequence to obtain the solution to the problem.
6. The method according to claim 4, wherein if at least two sets of consecutive steps correspond to the same problem, the consecutive steps corresponding to the same problem are arranged in sequence to obtain a solution corresponding to the problem, comprising: Arrange at least two consecutive sets of steps corresponding to the same problem in sequence to obtain the target solution; Based on the question-answer pair to which any step belongs in at least two consecutive sets of steps corresponding to the same question, determine the question corresponding to the step; Based on the problem corresponding to the steps and the target solution, the solution corresponding to the problem is determined as the target solution.
7. The method according to claim 2, further comprising, before obtaining the session record: Obtain at least two sets of training question-answer pairs from a domain-specific knowledge base; Add question tags to the training questions and step tags to the training steps in each training question-answer pair. The original pairing model is trained based on at least the labeled training question-answer pairs to obtain the pairing model.
8. The method according to claim 4, further comprising, before obtaining the session record: Obtain at least two sets of training schemes from a specific domain knowledge base, each set of training schemes containing at least one step; Label the training steps in each training scheme sequentially to identify at least one set of consecutive training steps; The original continuous model is trained based on at least one set of training sequential steps to obtain the continuous model.
9. An information processing apparatus, comprising: The acquisition module is used to acquire session records, wherein the session records include at least a first session generated by a first session object and a second session generated by a second session object; An analysis module is used to analyze the conversation records to obtain at least two sets of question-answer pairs. Each set of question-answer pairs includes a first conversation as a question and a second conversation as a step. The step is a step of the solution corresponding to the question in the question-answer pair. Among them, a question corresponds to at least one solution, and a solution includes at least one step. The questions and feedback steps in the conversation records have topic overlap in terms of language expression. The determination module determines the solution corresponding to each question in the session record based on the continuity characteristics of the steps; if the solution corresponding to a question includes multiple steps, the steps belonging to the same question in multiple question-answer pairs are combined to obtain the solution corresponding to that question. The summary generation module is used to display the questions and corresponding solutions in sequence based on the process of question generation in the session record, so as to obtain the target summary content of the session record.
10. A readable storage medium having a computer program stored thereon, the computer program being invoked and executed by a processor to implement the steps of the information processing method as described in any one of claims 1-8.