Multi-round Dialogue Management System and Method Based on Machine Learning Technology

Abstract

The present invention belongs to the technical field of dialogue management. A multi-round dialogue management system and method based on machine learning technology are provided. The method includes: during the process of the user inputting the current question information in the first dialogue column, using a first machine learning model to predict the current question information of the user to obtain the current dialogue intention of the user; performing topic matching on the dialogue history information in the first dialogue column according to the current dialogue intention to obtain a number of matching historical dialogue information, and performing streamlining processing on each historical dialogue information according to a set rule to obtain each target historical dialogue information; determining a sorting number according to the matching degree between each target historical dialogue information and the current dialogue intention, and splicing each target historical dialogue information and the current question information according to the sorting number to obtain a second dialogue column. The present invention can improve the overall dialogue quality and efficiency.

Description

Technical Field

[0001] The present invention relates to the technical field of dialogue management, and in particular, to a multi-round dialogue management system and method based on machine learning technology. Background Art

[0002] Intelligent dialogue is the mainstream trend of current technological development. Through intelligent dialogue, logical and accurate answers to various consulting questions can be provided for users, significantly improving the consulting efficiency and experience of users. AI dialogue is mainly used in, for example, e-commerce customer service, medical system customer service, or dialogue systems in other application scenarios, such as chat rooms, study rooms, etc.

[0003] When a user communicates with a dialogue agent, discontinuous situations, i.e., multi-round dialogues, often occur. For example, the user communicates with an intelligent robot in the customer service system of a certain shopping APP on the first day to consult information such as the functions and sizes of product A. On the second and third days, the user consults the price discount information of other products respectively. On the Nth day, the user consults the relevant information of product A again. However, due to multiple consultations by the user, the dialogue column at this time contains a lot of content unrelated to product A. If the user wants to view the consultation content on the first day, they must manually scroll up to find it, resulting in the user being unable to conveniently refer to the consultation record on the first day to understand product A.

[0004] Therefore, how to splice the associated dialogue content belonging to a certain theme in multi-round dialogues so that users can more fully and conveniently understand the content of this theme is a technical problem that needs to be solved currently. Summary of the Invention

[0005] In response to this, the present invention provides a multi-round dialogue management method, system, electronic device, computer storage medium, and computer program product based on machine learning technology to solve the above technical problems.

[0006] The present invention discloses a multi-round dialogue management method based on machine learning technology, which is applied to an agent. The method includes the following steps: during the process of a user inputting the current problem information in a first dialogue column, using a first machine learning model to predict the current problem information of the user to obtain the current dialogue intention of the user.

[0007] Perform topic matching in the dialogue history information corresponding to the first dialogue column according to the current dialogue intention to obtain several matching historical dialogue information, and perform refinement processing on each piece of the historical dialogue information according to a set rule to obtain each target historical dialogue information.

[0008] Determine the sorting serial numbers according to the matching degrees of the respective target historical dialogue information and the current dialogue intention, and splice the respective target historical dialogue information and the current question information according to the sorting serial numbers to obtain a second dialogue column; wherein, the current question information in the second dialogue column is located at the first page position, and the smaller the sorting serial number, the smaller the position distance between the target historical dialogue information and the current question information.

[0009] In some embodiments, the process of obtaining each target historical dialogue information by performing a refinement process on each historical dialogue information according to a set rule includes: obtaining the content input characteristics during the process of the user inputting the current question information in the first dialogue column, where the content input characteristics at least include character input characteristics and input content refinement characteristics, and the character input characteristics include character input rate and number of character modifications; analyzing the obvious level of the user's dialogue intention based on the content input characteristics, and determining the refinement level according to the obvious level of the dialogue intention; wherein, the refinement level is positively correlated with the obvious level of the dialogue intention; performing a refinement process on each historical dialogue information according to the refinement level to obtain each target historical dialogue information.

[0010] In some embodiments, the process of analyzing the obvious level of the user's dialogue intention based on the content input characteristics includes: using a second machine learning model to analyze the content input characteristics to obtain the first obvious level of the user's dialogue intention; retrieving several second obvious levels of the dialogue intention associated with the user, and determining a fine-tuning value according to each second obvious level of the dialogue intention; wherein, the second obvious level of the dialogue intention is the obvious level of the dialogue intention stored in the database that belongs to the same or related topic type as the current question information input by the user; using the fine-tuning value to fine-tune the first obvious level of the dialogue intention to obtain the second obvious level of the dialogue intention.

[0011] In some embodiments, before determining the sorting serial numbers according to the matching degrees of the respective target historical dialogue information and the current dialogue intention, it further includes: outputting a prompt message in the first dialogue column, where the prompt message is used to ask whether the user enables multi-round dialogue integration processing; if a confirmation signal from the user for the prompt message is received, generating an integration processing trigger signal, and the integration processing trigger signal is used to trigger the determination of the sorting serial numbers according to the matching degrees of the respective target historical dialogue information and the current dialogue intention.

[0012] In some embodiments, after obtaining the second dialogue column, it further includes: having a dialogue with the user in the second dialogue column, and after the dialogue ends, supplementing the new dialogue content in the second dialogue column to the first dialogue column according to the dialogue time.

[0013] The present invention also discloses a multi-turn dialogue management system based on machine learning technology, which is applied to an intelligent agent. The system includes a processor and a memory, and a computer program stored in the memory can run on the processor. When the computer program is run by the processor, the following steps are executed: during the process of the user inputting the current problem information in the first dialogue column, use a first machine learning model to predict the current problem information of the user to obtain the current dialogue intention of the user; perform topic matching in the dialogue history information corresponding to the first dialogue column according to the current dialogue intention, obtain several matching historical dialogue information, and perform a refinement process on each of the historical dialogue information according to a set rule to obtain each target historical dialogue information; determine a sorting serial number according to the matching degree between each target historical dialogue information and the current dialogue intention, and splice and process each target historical dialogue information and the current problem information according to the sorting serial number to obtain a second dialogue column; wherein, the current problem information in the second dialogue column is located at the home page position, and the smaller the sorting serial number, the smaller the position distance between the target historical dialogue information and the current problem information.

[0014] In some embodiments, the step of performing a refinement process on each of the historical dialogue information according to a set rule to obtain each target historical dialogue information includes: obtaining the content input features during the process of the user inputting the current problem information in the first dialogue column, where the content input features at least include character input features and input content refinement features, and the character input features include character input rate and number of character modifications; analyzing the content input features to obtain the obvious level of the user's dialogue intention, and determining the refinement level according to the obvious level of the dialogue intention; wherein, the refinement level is positively correlated with the obvious level of the dialogue intention; perform a refinement process on each of the historical dialogue information according to the refinement level to obtain each target historical dialogue information.

[0015] In some embodiments, the step of analyzing the content input features to obtain the obvious level of the user's dialogue intention includes: using a second machine learning model to analyze the content input features to obtain the first obvious level of the user's dialogue intention; retrieving several second obvious levels of the dialogue intention associated with the user, and determining a fine-tuning value according to each of the second obvious levels of the dialogue intention; wherein, the second obvious level of the dialogue intention is the obvious level of the dialogue intention stored in the database that belongs to the same or related topic type as the current problem information input by the user; use the fine-tuning value to fine-tune the first obvious level of the dialogue intention to obtain the second obvious level of the dialogue intention.

[0016] In some embodiments, before determining the sorting sequence number according to the matching degree between each piece of the target historical dialogue information and the current dialogue intention, the following steps are further included: outputting a prompt message in the first dialogue column, where the prompt message is used to ask the user whether to enable multi-turn dialogue integration processing; if a confirmation signal from the user for the prompt message is received, an integration processing trigger signal is generated, and the integration processing trigger signal is used to trigger the determination of the sorting sequence number according to the matching degree between each piece of the target historical dialogue information and the current dialogue intention.

[0017] In some embodiments, after obtaining the second dialogue column, the following steps are further included: having a dialogue with the user in the second dialogue column, and after the dialogue ends, supplementing the newly added dialogue content in the second dialogue column to the first dialogue column according to the dialogue time.

[0018] The present invention also discloses an electronic device, which is applied to an intelligent agent. The electronic device includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor. When the computer program is executed, the method described in any of the previous items is implemented.

[0019] The present invention also discloses a computer storage medium, which is applied to an intelligent agent. The computer-readable storage medium stores a computer program. When the computer program is executed, the method described in any of the previous items is implemented.

[0020] The present invention also discloses a computer program product, which is applied to an intelligent agent. The computer program product is pre-packaged with computer program code. When the computer program code is executed, the method described in any of the previous items is implemented.

[0021] The beneficial effects of the present invention are as follows: Based on the predicted dialogue intention, the present invention can quickly retrieve matching information from the multi-turn dialogue history, saving the user's time in searching for relevant past topics, enhancing the dialogue coherence, and enabling the rapid reuse of key details in the past dialogue, avoiding the user from repeating the description, optimizing the interaction experience, and thus improving the overall dialogue quality and efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required to be used in the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the scope. For those of ordinary skill in the art, other relevant drawings can be obtained based on these drawings without creative efforts.

[0023] Figure 1 It is a structural schematic diagram of a multi-turn dialogue management method based on machine learning technology disclosed in an embodiment of the present invention.

[0024] Figure 2 It is a schematic structural diagram of a multi-turn dialogue management system based on machine learning technology disclosed in an embodiment of the present invention. Specific embodiments

[0025] The following specific embodiments illustrate the implementation manners of the present application. Those skilled in this technology can easily understand other advantages and effects of the present application from the content disclosed in this specification. Obviously, the described embodiments are part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present application.

[0026] In addition, the technical features involved in different implementation manners of the present application described below can be combined with each other as long as they do not conflict with each other.

[0027] As Figure 1 shown, an embodiment of the present invention discloses a multi-turn dialogue management method based on machine learning technology, which is applied to a dialogue agent. The method includes the following steps: S1, during the process of the user inputting the current problem information in the first dialogue column, use a first machine learning model to predict the current problem information of the user to obtain the current dialogue intention of the user.

[0028] In this step, the dialogue agent monitors the input operation of the user in the first dialogue column and captures the input stream character by character or word by word. Once the user starts to input the current problem information, the preprocessing process is immediately started.

[0029] First, remove special characters, such as "#", "&", "*", etc., which do not carry key semantic information but will interfere with subsequent analysis. For the cleaned problem information, split it into individual words or sub-word units.

[0030] Then, adopt a pre-trained machine learning model, such as an architecture combining a multi-layer bidirectional long short-term memory network (Bi-LSTM) and an attention mechanism (Attention). Such models are good at capturing long-distance dependencies in text sequences and perform well in the dialogue intention classification task. Pre-load the model parameters that have been fine-tuned on a large-scale dialogue dataset, and the dataset covers various dialogue scenarios such as daily, business, and customer service to ensure the generalization ability of the model.

[0031] Tokenize the preprocessed text. Use word embedding techniques (such as GloVe) to map each word into a vector of a fixed dimension and input it into the Bi-LSTM model. The model traverses and learns the text sequence forward and backward. The attention mechanism focuses on the key semantic parts. The final output layer uses the softmax function to give different dialogue intention categories, and the category with the highest probability is the predicted dialogue intention for this conversation. Examples of this dialogue intention are "the type of motor used in washing machine A", "how technology B is used for its chat intention analysis", etc.

[0032] S2. Perform topic matching in the dialogue history information corresponding to the first dialogue column according to the said dialogue intention for this conversation, obtain several matching historical dialogue information, and perform refinement processing on each said historical dialogue information according to the set rules to obtain each target historical dialogue information.

[0033] In this step, use the predicted dialogue intention for this conversation as an index to query the dialogue history information stored in the database that belongs to the first dialogue column. The database uses a relational database or a document database to store structured dialogue records, and each record has fields such as the belonging dialogue column, timestamp, dialogue role, dialogue content, etc. Use the fuzzy query function of SQL statements or the full-text search function of the document database to search for paragraphs related to this intention in past conversations and extract a batch of matching historical dialogue information. For example, if the current intention is "how technology B is used for its chat intention analysis", search for historical conversations containing keywords such as "technology B", "chat", "intention". Among them, each obtained matching historical dialogue information is also associated with a corresponding matching degree, which is used to represent the level of semantic matching between this historical dialogue information and the dialogue intention for this conversation.

[0034] At the same time, the obtained matching historical dialogue information may also contain a lot of irrelevant information, such as "Are you there?", "Hello", overly long background description information, etc. It is necessary to first streamline this information that is irrelevant to the core topic to ensure that the finally spliced dialogue page is more concise and efficient.

[0035] The refinement rules include, for example: 1) Remove duplicate semantic fragments. If several consecutive sentences are describing the same store address, only keep the most complete and clear one; 2) Screen out meaningless modal particles and pet phrases, such as "um", "ah", "that"; 3) Trim overly long background descriptions and only keep the key information directly related to the core topic. Use natural language processing tools, such as part-of-speech tagging and dependency syntactic analysis, to assist in judging semantic relevance and complete the refinement of each historical dialogue information to obtain the target historical dialogue information.

[0036] After the above-mentioned streamlining process, each piece of matched historical dialogue information is trimmed into target historical dialogue information. Among them, each piece of matched historical dialogue information refers to the historical dialogue information with a matching degree higher than a preset matching threshold.

[0037] It should be noted that for the matching degree between each piece of target historical dialogue information and the current dialogue intention, a semantic similarity algorithm can be used, such as the cosine similarity based on the semantic vector space. First, vectorize the current dialogue intention and the target historical dialogue respectively, and then calculate the cosine value of the included angle between the two vectors. The closer the value is to 1, the higher the matching degree. At the same time, combined with the vocabulary overlap rate, count the proportion of the number of common key nouns and verbs between the two, and comprehensively obtain the matching degree score. Furthermore, the historical dialogue information with a matching degree higher than the preset matching threshold is screened out.

[0038] S3. Determine the sorting serial numbers according to the matching degrees between each piece of the target historical dialogue information and the current dialogue intention, and splice each piece of the target historical dialogue information and the current question information according to the sorting serial numbers to obtain a second dialogue column; wherein, the current question information in the second dialogue column is located at the home page position, and the smaller the sorting serial number of a piece of the target historical dialogue information, the smaller the position distance from the current question information.

[0039] In this step, after obtaining each piece of target historical dialogue information with a high relevance to the current dialogue intention, each piece of target historical dialogue information and the current question information can be spliced to obtain a second dialogue column. In the second dialogue column, the historical turns of dialogue irrelevant to the current dialogue intention are hidden, and only the current question information and each piece of target historical dialogue information are retained, wherein the current question information is located at the home page position of the page, that is, the "latest" position of the current dialogue page.

[0040] For each piece of target historical dialogue information, its position in the second dialogue column can be determined according to the level of its matching degree with the current dialogue intention. Specifically, for the target historical dialogue information with a higher matching degree, it is closer to the current question information, that is, closer to the above-mentioned "latest" position. In other words, it is ranked in the front; while for the target historical dialogue information with a lower matching degree, it is farther from the current question information, that is, closer to the above-mentioned "latest" position. In other words, it is ranked in the back.

[0041] Based on the predicted dialogue intention, the present invention can quickly retrieve the matching information from the multi-turn dialogue history, saving the user's time in searching for relevant past topics, enhancing the dialogue coherence, and enabling the quick reuse of key details in the past dialogue, avoiding the user from repeating the elaboration, optimizing the interaction experience, and thus improving the overall dialogue quality and efficiency.

[0042] In some embodiments, the process of streamlining each piece of historical conversation information according to a set rule to obtain each piece of target historical conversation information includes: obtaining the content input characteristics during the process of the user inputting the current question information in the first conversation column, where the content input characteristics at least include character input characteristics and input content streamlining characteristics, and the character input characteristics include character input rate and number of character modifications; analyzing the obvious level of the user's conversation intention based on the content input characteristics, and determining the streamlining level according to the obvious level of the conversation intention; wherein, the streamlining level is positively correlated with the obvious level of the conversation intention; streamlining each piece of historical conversation information according to the streamlining level to obtain each piece of target historical conversation information.

[0043] In the embodiments of the present invention, for the character input rate: the conversation agent starts the timing module from the moment the user starts to input the current question information in the first conversation column. The character input rate is obtained by counting the number of characters input by the user per unit time (such as per second). If the user quickly types 5 characters in 1 second, it indicates a faster input rhythm, perhaps already having a clear idea about the question; conversely, if it takes 3 seconds to input 1 character, it is likely that the user is still thinking about the core content involved in this question.

[0044] For the number of character modifications: the system monitors the user's modification operations on the already input characters in real time, and every action of deleting or replacing characters is recorded. A higher number of modifications may imply that the user is still organizing a precise expression and has not fully clarified their own demands.

[0045] The above-mentioned character input rate and number of character modifications constitute the character input characteristics.

[0046] After the user's final input is completed, the essence of the user's input content is analyzed and refined. By comparing the ratio of the extracted essence content to the characters of the entire input content, this character ratio is the input content streamlining characteristic. Obviously, the higher the character ratio, the more streamlined the user's input content, and vice versa (i.e., it contains too many irrelevant vocabulary contents).

[0047] The above-mentioned character input characteristics and input content streamlining characteristics extracted constitute the content input characteristics of the user, based on which the obvious level of the user's conversation intention can be analyzed. The obvious level of the conversation intention here refers to the clarity of the user about the core content or related content of the question to be consulted this time when inputting content. In other words, the higher the obvious level of the conversation intention, the clearer the user knows the core content or related content of the question to be consulted this time, and vice versa.

[0048] After determining the obviousness level of the dialogue intention, the simplification level can be determined according to the positive correlation. The simplification level here refers to the intensity of the simplification process for each piece of historical dialogue information. The higher the simplification level, the higher the intensity of the simplification process, and the lower the simplification level, the lower the intensity of the simplification process. The simplification process includes, but is not limited to, deletion and semantic integration. For example, the historical dialogue information is as follows:

[0049] Agent: Hello, how can I help you?

[0050] User: I am currently researching Technology B and specifically analyzing the possibility of its application in chat intention analysis. Can you give me some relevant guidance?

[0051] Agent: Okay, Technology B is... Chat intention analysis is...

[0052] User: I know these background contents. Please directly tell me how to use Technology B for chat intention analysis.

[0053] Agent: Okay,... (Specific suggestions on the application of Technology B in chat intention analysis).

[0054] For low-level simplification, the above historical dialogue information can be simplified to:

[0055] (Delete: Agent: Hello, how can I help you?)

[0056] User: I am currently researching Technology B and specifically analyzing the possibility of its application in chat intention analysis. (Delete: Can you give me some relevant guidance?)

[0057] Agent: (Delete: Okay,) Technology B is... Chat intention analysis is...

[0058] User: (Delete: I know these background contents.) Please directly tell me how to use Technology B for chat intention analysis.

[0059] Agent: (Delete: Okay,)... (Specific suggestions on the application of Technology B in chat intention analysis).

[0060] For high-level simplification, the above historical dialogue information can be simplified to:

[0061] (Delete: Agent: Hello, how can I help you?)

[0062] (Delete: User: I am currently researching Technology B and specifically analyzing the possibility of its application in chat intention analysis. Can you give me some relevant guidance?)

[0063] (Deleted: Agent: Okay, B technology is... Chat intention analysis is...)

[0064] User: (Deleted: I know this background information.) Please directly tell me how to use B technology for chat intention analysis.

[0065] Agent: (Deleted: Okay,... (Specific suggestions for applying B technology to chat intention analysis).)

[0066] The present invention analyzes the obvious level of the user's conversation intention to correspondingly determine the degree of streamlining processing of each historical conversation information, thereby realizing personalized streamlining of historical conversation information, so as to ensure that more valid information is included in the second conversation column constructed subsequently without missing key information. Specifically, when the user knows more clearly the core content or related content of the question to be consulted this time, a higher-intensity streamlining processing is performed on the historical conversation information, so that the second conversation column can be more streamlined; when the user knows less clearly the core content or related content of the question to be consulted this time, a lower-intensity streamlining processing is performed on the historical conversation information, so that more detailed content of the historical conversation information can be included in the second conversation column, which is beneficial to giving the user more conversation references.

[0067] In some embodiments, the analyzing the obvious level of the user's conversation intention according to the content input feature includes: using a second machine learning model to analyze the content input feature to obtain the first obvious level of the user's conversation intention; retrieving a plurality of second obvious levels of the conversation intention associated with the user, and determining a fine-tuning value according to each of the second obvious levels of the conversation intention; wherein, the second obvious level of the conversation intention is the obvious level of the conversation intention stored in the database that belongs to the same or related topic type as the current question information input by the user; using the fine-tuning value to fine-tune the first obvious level of the conversation intention to obtain the second obvious level of the conversation intention.

[0068] In the embodiments of the present invention, a machine learning algorithm suitable for processing small data volumes and high-dimensional features is selected to build a second machine learning model, such as a lightweight gradient boosting tree (LightGBM). A large amount of historical user input data is collected as a training set, and these data cover rich content input features, such as various character input rates, character modification times, and sample combinations of input content streamlining features, and the corresponding obvious levels of the conversation intention are marked. After multiple rounds of training, the model learns the potential mapping relationship between the features and the levels.

[0069] After obtaining the content input features during the process of the user entering the current question information in the first dialogue column, organize them into a format that meets the model input requirements. For example, numericalize the character input rate and reshape the number of character modifications into a vector form, and input them into the trained second machine learning model. Based on the learned potential mapping relationship above, the second machine learning model outputs a preliminary obvious level of the user's first dialogue intention. The level division can be qualitative descriptions such as "low", "medium", "high", or specific level values, providing a basis for subsequent fine-tuning.

[0070] Meanwhile, the present invention also obtains the obvious levels of the user's dialogue intentions for the same or similar types of topics stored in the database. These obvious levels of dialogue intentions are analyzed by the agent when determining the second dialogue column of the present invention. That is, each time multi-round dialogue management is performed, the analyzed obvious level of dialogue intention is stored in the database together with the topic type. Statistically extract several obvious levels of the second dialogue intention, for example, calculate the mean of each obvious level of the second dialogue intention, and determine the fine-tuning value based on this mean. If most of these past levels are "high", it indicates that the user has always had a clear intention on such topics, and the fine-tuning value is set to be closer to 1 to reduce the intervention in the obvious level of the current first dialogue intention. If most of the past levels are "low", it indicates that there is a "high" probability that the confidence level of the obvious level of the first dialogue intention obtained in this analysis is "high", and at this time, the fine-tuning value is set to be less than 1. Among them, an association relationship between the fine-tuning value and the above mean can be established in advance, and the corresponding fine-tuning value can be determined by querying this association relationship.

[0071] In some embodiments, before determining the sorting serial number according to the matching degree between each piece of target historical dialogue information and the current dialogue intention, it further includes: outputting a prompt message in the first dialogue column, where the prompt message is used to ask the user whether to enable multi-round dialogue integration processing; if a confirmation signal from the user for the prompt message is received, an integration processing trigger signal is generated, and the integration processing trigger signal is used to trigger the determination of the sorting serial number according to the matching degree between each piece of target historical dialogue information and the current dialogue intention.

[0072] In the embodiments of the present invention, to improve the user experience, it is also possible to consult the user in advance whether there is a need for integrating multi-round dialogue management. Only when it is confirmed that the user has such a need, multi-round dialogue management is performed, and only the above-mentioned second dialogue column is generated. When the second dialogue column is not generated, the user can manually scroll up to find the historical dialogue content.

[0073] In some embodiments, after obtaining the second dialogue column, it further includes: having a dialogue with the user in the second dialogue column, and after the dialogue ends, supplementing the new dialogue content in the second dialogue column to the first dialogue column according to the dialogue time.

[0074] In an embodiment of the present invention, the first conversation column and the second conversation column involved in the present invention may be the same conversation column or different conversation columns. For example, the first conversation column is a separate conversation column opened by the user with a certain merchant, a certain topic person / anchor, which contains multiple rounds of historical conversations; and the second conversation column may also be the first conversation column, but the historical conversations in the first conversation column are subjected to the above-mentioned screening and splicing based on the current conversation intention. The second conversation column may also be a blank conversation column directly opened by the intelligent agent, and the content obtained by the above-mentioned screening and splicing is transmitted to the blank conversation column. After the user ends the current conversation, the "blank conversation column" is deleted, and the content of the current conversation is stored in the "latest" position in the original first conversation column in chronological order.

[0075] As Figure 2 shown, the present invention also discloses a multi-round conversation management system based on machine learning technology, which is applied to an intelligent agent. The system includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is run by the processor, the following steps are executed: determining a sorting serial number according to the matching degree between each piece of target historical conversation information and the current conversation intention, and performing splicing processing on each piece of target historical conversation information and the current question information according to the sorting serial number to obtain a second conversation column; wherein, the current question information in the second conversation column is located at the home page position, and the smaller the sorting serial number, the smaller the position distance between the target historical conversation information and the current question information.

[0076] An embodiment of the present invention also discloses an electronic device, which is applied to an intelligent agent. The electronic device includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor. When the computer program is executed, the method described in any one of the previous items is implemented.

[0077] An embodiment of the present invention also discloses a computer storage medium, which is applied to an intelligent agent. The computer-readable storage medium stores a computer program, and when the computer program is executed, the method described in any one of the previous items is implemented.

[0078] An embodiment of the present invention also discloses a computer program product, which is applied to an intelligent agent. The computer program product is pre-packaged with computer program code, and when the computer program code is executed, the method described in any one of the previous items is implemented.

[0079] It should be noted that in the description of the present invention, the terms "first", "second", etc. are only used for descriptive purposes and cannot be construed as indicating or implying relative importance. In addition, in the description of the present invention, unless otherwise specified, the meaning of "a plurality of" is two or more. Any process or method description in a flowchart or described otherwise herein can be understood to represent a module, segment, or portion of code including one or more executable instructions for implementing a specific logical function or process, and the scope of the preferred embodiments of the present invention includes additional implementations, where the functions can be executed in a substantially simultaneous manner or in the reverse order according to the functions involved, rather than in the order shown or discussed, which should be understood by those skilled in the technical field of the embodiments of the present invention.

[0080] It should be understood that various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application specific integrated circuits having appropriate combinational logic gate circuits, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

[0081] Although the embodiments of the present invention have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those of ordinary skill in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

Claims

1. A multi-turn dialogue management method based on machine learning technology, applied to an intelligent agent, characterized in that Including: During the process that the user inputs the current problem information in the first dialogue column, use the first machine learning model to predict the current problem information, and obtain the current dialogue intention of the user; Perform topic matching in the dialogue history information corresponding to the first dialogue column according to the current dialogue intention, obtain several matching historical dialogue information, and perform refinement processing on each of the historical dialogue information according to the set rules to obtain each target historical dialogue information; Determine the sorting serial number according to the matching degree between each target historical dialogue information and the current dialogue intention, and splice each target historical dialogue information and the current problem information according to the sorting serial number to obtain a second dialogue column; wherein, the current problem information in the second dialogue column is located at the home page position, and the smaller the sorting serial number, the smaller the position distance between the target historical dialogue information and the current problem information; Perform refinement processing on each of the historical dialogue information according to the set rules to obtain each target historical dialogue information, including: obtaining the content input features during the process that the user inputs the current problem information in the first dialogue column, the content input features at least including character input features and input content refinement features, the character input features including character input rate and character modification times; analyzing the obvious level of the user's dialogue intention according to the content input features, and determining the refinement level according to the obvious level of the dialogue intention; the refinement level is positively correlated with the obvious level of the dialogue intention; performing refinement processing on each of the historical dialogue information according to the refinement level to obtain each target historical dialogue information; after the user finishes inputting, extract and calculate the ratio of the essential content of the input content to the characters of the entire input content, which is the input content refinement feature; Analyze the obvious level of the user's dialogue intention according to the content input features, including: using the second machine learning model to analyze the content input features to obtain the first obvious level of the user's dialogue intention; retrieving several second obvious levels of the dialogue intention associated with the user, and determining the fine-tuning value according to each of the second obvious levels of the dialogue intention; wherein, the second obvious level of the dialogue intention is the obvious level of the dialogue intention stored in the database that belongs to the same or related topic type as the current problem information input by the user; using the fine-tuning value to fine-tune the first obvious level of the dialogue intention to obtain the second obvious level of the dialogue intention; After obtaining the second dialogue column, it further includes: having a dialogue with the user in the second dialogue column, and after the dialogue ends, supplementing the new dialogue content in the second dialogue column to the first dialogue column according to the dialogue time.

2. The multi-round dialogue management method based on machine learning technology according to claim 1, characterized in that: Before determining the sorting sequence number according to the matching degree between each piece of the target historical dialogue information and the current dialogue intention, the following steps are also included: outputting a prompt message in the first dialogue column, where the prompt message is used to ask whether the user enables multi-turn dialogue integration processing; if a confirmation signal from the user for the prompt message is received, generating an integration processing trigger signal, where the integration processing trigger signal is used to trigger the determination of the sorting sequence number according to the matching degree between each piece of the target historical dialogue information and the current dialogue intention.

3. A multi-turn dialogue management system based on machine learning technology, applied to an intelligent agent, characterized in that, The system includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the computer program is run by the processor, the following steps are executed: during the process that the user inputs the current problem information in the first dialogue column, using a first machine learning model to predict the current problem information to obtain the user's current dialogue intention; Performing topic matching on the dialogue history information corresponding to the first dialogue column according to the current dialogue intention to obtain several matching historical dialogue information, performing a refinement process on each piece of the historical dialogue information according to a set rule to obtain each piece of target historical dialogue information; determining a sorting sequence number according to the matching degree between each piece of the target historical dialogue information and the current dialogue intention, and splicing each piece of the target historical dialogue information and the current problem information according to the sorting sequence number to obtain a second dialogue column; where the current problem information in the second dialogue column is located at the home page position, and the smaller the sorting sequence number, the closer the position distance between the piece of the target historical dialogue information and the current problem information. Performing a refinement process on each piece of the historical dialogue information according to a set rule to obtain each piece of target historical dialogue information, including: obtaining the content input features during the process that the user inputs the current problem information in the first dialogue column, where the content input features at least include character input features and input content refinement features, and the character input features include character input rate and character modification times; analyzing the obvious level of the user's dialogue intention according to the content input features, and determining the refinement level according to the obvious level of the dialogue intention; where the refinement level is positively correlated with the obvious level of the dialogue intention; performing a refinement process on each piece of the historical dialogue information according to the refinement level to obtain each piece of the target historical dialogue information; Analyzing the obvious level of the user's dialogue intention according to the content input features, including: using a second machine learning model to analyze the content input features to obtain the first obvious level of the user's dialogue intention; retrieving several second obvious levels of the dialogue intention associated with the user, and determining an adjustment value according to each of the second obvious levels of the dialogue intention; where the second obvious level of the dialogue intention is the obvious level of the dialogue intention stored in the database and belonging to the same or related topic type as the current problem information input by the user; using the adjustment value to fine-tune the first obvious level of the dialogue intention to obtain the second obvious level of the dialogue intention; After obtaining the second conversation bar, it further includes: having a conversation with the user in the second conversation bar, and after the conversation ends, supplementing the newly added conversation content in the second conversation bar to the first conversation bar according to the conversation time.

4. An electronic device, applied to an intelligent agent, characterized in that: The electronic device includes: at least one processor, a memory, and a computer program stored in the memory and executable on the at least one processor, and when the computer program is executed, it implements the method according to claim 1 or 2.

5. A computer storage medium, applied to an agent, characterized in that: The computer storage medium stores a computer program, and when the computer program is executed, it implements the method according to claim 1 or 2.

6. A computer program product, applied to an agent, characterized in that: The computer program product is pre-packaged with computer program code, and when the computer program code is executed, it implements the method according to claim 1 or 2.

Images