Method and system for obtaining a continuation word based on a joint decision model

Abstract

The application discloses a method and system for obtaining a taking word based on a joint decision model, and belongs to the technical field of artificial intelligence. The method comprises the following steps: obtaining a user behavior category by dividing user behaviors according to business experience, obtaining a taking word candidate list according to historical dialogue records, and obtaining an attribution table by classifying taking words in the taking word candidate list into the user behavior category; obtaining a behavior data set by obtaining a next user behavior corresponding to the historical dialogue records through a supervision model, and then obtaining a scoring data set by obtaining a taking word score under the next user behavior through the attribution table; and training a joint decision model based on the behavior data set and the scoring data set, and obtaining a final taking word through the successfully trained joint decision model. The technical problem that the prior art is difficult to improve the taking word acquisition efficiency and scene adaptability is overcome.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for acquiring a continuation word based on a joint decision-making model. Background Art

[0002] With the development of artificial intelligence, in order to improve the user interaction experience, the existing technology mainly generates reply information through a large language model to meet the real-time requirements. For example, a voice interaction method, device, electronic device and storage medium with patent publication number CN119649816A involves the field of large language models. The specific implementation scheme is: in response to receiving input voice, initiating multiple first generation requests according to the input voice and historical interaction information, and obtaining multiple candidate continuation words corresponding to the multiple first generation requests, wherein the historical interaction information includes historical input text, historical continuation words and historical reply text for replying to the historical input text; determining from the multiple candidate continuation words a candidate continuation word whose repetition with the historical continuation word meets the preset requirements as the target continuation word; and outputting the voice of the target continuation word. However, the performance of the large language model in the above method is highly dependent on the quality and scale of training data. To improve the quality and scale of training data, existing technologies mainly obtain training data through manual labeling, but this has limitations. On the one hand, the efficiency of manual labeling is limited by labor costs and time investment. On the other hand, the inevitable subjective bias and standard inconsistency in the labeling process will directly affect the accuracy of data labeling. Therefore, the accuracy of the large language model is difficult to guarantee, which in turn makes it difficult to improve the efficiency of obtaining transition words and scene adaptability. Summary of the Invention

[0003] To address the technical problem that existing technologies have difficulty improving the efficiency of acquiring continuation words and their adaptability to specific scenarios, the present invention provides a method and system for acquiring continuation words based on a joint decision-making model. This method uses the powerful semantic understanding capabilities of a supervised model to conduct in-depth analysis of historical conversation records to obtain high-quality behavioral and scoring datasets, overcoming the limitations of manually annotating datasets. The joint decision-making model is then trained using these high-quality behavioral and scoring datasets. The final continuation words are acquired from the successfully trained joint decision-making model, improving the accuracy of user behavior predicted by the joint decision-making model and the adaptability of the final continuation words to specific scenarios. This overcomes the technical problem that existing technologies have difficulty improving the efficiency of acquiring continuation words and their adaptability to specific scenarios.

[0004] To solve the above technical problems, the present invention provides a method for obtaining a transition word based on a joint decision model, comprising the following steps:

[0005] Divide user behaviors into categories based on business experience, obtain a list of candidate continuation words based on historical conversation records, and classify the continuation words in the candidate continuation word list into user behavior categories to obtain an attribution table;

[0006] The supervised model is used to obtain the next user behavior corresponding to the historical conversation record to obtain a behavior dataset, and then the attribution table is used to obtain the continuation word score of the next user behavior to obtain a score dataset;

[0007] The joint decision model is trained based on the behavior dataset and the rating dataset, and the final continuation word is obtained through the successfully trained joint decision model.

[0008] After adopting the above technical solution, the present invention has the following advantages:

[0009] Considering the limitation of existing technologies that require manual annotation of conversation records and the next user behavior corresponding to the conversation records, and the strong semantic understanding capabilities of supervised models, we use supervised models to conduct in-depth analysis of historical conversation records to obtain high-quality behavior datasets and scoring datasets, breaking through the limitations of manually annotating datasets. The joint decision-making model is trained with high-quality behavior datasets, which improves the accuracy of user behavior predicted by the joint decision-making model. On this basis, the joint decision-making model is trained with high-quality scoring datasets, which also improves the scene adaptability of the final transition words, overcoming the technical problem of existing technologies that make it difficult to improve the efficiency of transition word acquisition and scene adaptability.

[0010] By obtaining user behavior categories through business experience and obtaining a list of candidate continuation words through historical conversation records, the practicality of the attribution table is improved. At the same time, by obtaining the attribution table, the continuation words under the user behavior can be quickly located through the attribution table based on the user behavior, further improving the efficiency of obtaining the continuation words.

[0011] Preferably, obtaining a candidate list of connecting words based on historical conversation records includes:

[0012] Obtain the first continuation word candidate list based on the frequency of reply prefixes in historical conversation records;

[0013] Calculating the similarity of the connecting words in the initial connecting word candidate list through the supervision model, and preliminarily grouping the connecting words in the initial connecting word candidate list according to the similarity;

[0014] A regrouping threshold is obtained according to the similarity of the continuation words in each group and the user behavior category, the continuation words in each group are regrouped based on the regrouping threshold, and a continuation word candidate list is obtained according to the regrouping result.

[0015] Preferably, the behavior dataset corresponding to the next user behavior obtained by the supervised model through the historical dialogue record comprises:

[0016] The user behavior category is taken as an output constraint condition of the supervised model.

[0017] The historical dialogue record is structured into a structured dialogue record according to the roles of the dialogue parties, the dialogue sequence and the dialogue content in the historical dialogue record.

[0018] The prompt word is constructed according to the output constraint condition, the prompt word and the structured dialogue record are spliced as input text of the supervised model, and the next user behavior corresponding to the historical dialogue record is obtained.

[0019] The next user behavior corresponding to the historical dialogue record is associated with the historical dialogue record, and then the behavior dataset is obtained.

[0020] Preferably, the score dataset of the taking word under the next user behavior obtained by the attribution table comprises:

[0021] The taking word under the next user behavior is obtained through the attribution table.

[0022] The next user behavior, the taking word under the next user behavior and the context information of the historical dialogue record are vectorized through the supervised model to obtain a historical behavior vector, a historical taking word vector and a historical context vector.

[0023] The first semantic similarity and the second semantic similarity of the historical taking word vector and the historical behavior vector, the historical context vector are obtained respectively.

[0024] The score of the taking word under the next user behavior is obtained based on the first semantic similarity and the second semantic similarity to obtain the score dataset.

[0025] Preferably, before the joint decision model is trained based on the behavior dataset and the score dataset, the method further comprises:

[0026] The behavior dataset is divided into a first behavior dataset and a second behavior dataset according to the acquisition time of the historical dialogue record, and the supervised model is trained once through the second behavior dataset.

[0027] The historical dialogue record corresponding to the first behavior dataset is input into the supervised model trained once to obtain a predicted user behavior, the predicted user behavior is compared with the next user behavior corresponding to the first behavior dataset to obtain a first comparison result, the first behavior dataset is modified through the first comparison result, and the behavior dataset is updated through the modified first behavior dataset.

[0028] The scoring data set is divided into a first scoring data set and a second scoring data set according to the acquisition time of the historical dialogue record, and the supervised model is trained again through the second scoring data set;

[0029] The historical dialogue record corresponding to the first scoring data set is input into the supervised model after the second training to obtain a predicted score, the predicted score is compared with the taking word score corresponding to the first scoring data set to obtain a second comparison result, the first scoring data set is modified through the second comparison result, and the scoring data set is updated through the modified first scoring data set.

[0030] Preferably, the training of the joint decision model based on the behavior data set and the scoring data set comprises:

[0031] The classification model in the joint decision model is trained through the behavior data set, and the fine arrangement model in the joint decision model is trained through the scoring data set.

[0032] Preferably, the training of the classification model in the joint decision model through the behavior data set and the training of the fine arrangement model in the joint decision model through the scoring data set comprise:

[0033] The behavior data set is divided into several parts, each part of the behavior data set is input into the classification model in sequence to obtain a classification model output, and the classification model is modified according to the difference between the classification model output and the first actual result;

[0034] The scoring data set is divided into several parts, each part of the scoring data set is input into the fine arrangement model in sequence to obtain a fine arrangement model output, and the fine arrangement model is modified according to the difference between the fine arrangement model output and the second actual result.

[0035] Preferably, the method further comprises:

[0036] When the difference between the classification model output and the first actual result and the difference between the fine arrangement model output and the second actual result both satisfy a preset condition, it indicates that the joint decision model is successfully trained.

[0037] Preferably, the method for obtaining the final taking word through the successfully trained joint decision model comprises:

[0038] The real-time dialogue record is input into the classification model to obtain real-time user behavior, the taking word under the real-time user behavior is obtained through the attribution table, the taking word under the real-time user behavior is input into the fine arrangement model to obtain a taking word score under the real-time user behavior, and the taking word with the highest score under the real-time user behavior is taken as the final taking word.

[0039] The method has the following beneficial effects:

[0040] Considering that the prior art needs to manually annotate the dialogue record and the next user behavior corresponding to the dialogue record, and the supervised model has strong semantic understanding ability, the historical dialogue record is analyzed in depth by the supervised model to obtain a high-quality behavior data set and a scoring data set, thereby breaking through the limitation of manually annotating the data set, training the joint decision model through the high-quality behavior data set, improving the accuracy of the user behavior predicted by the joint decision model, and on this basis, training the joint decision model through the high-quality scoring data set, and also improving the scene adaptability of the final takeover word, thereby overcoming the technical problems that the prior art is difficult to improve the efficiency and scene adaptability of the takeover word;

[0041] The behavior data set and the scoring data set are divided, and the supervised model is trained respectively, the predicted user behavior and the predicted score are obtained through the supervised model trained respectively, and the actual result is compared, and the behavior data set and the scoring data set are updated, the quality of the data set is significantly improved through the data batch and cycle inspection mode, and the scene adaptability of the takeover word is further improved;

[0042] The user behavior category is obtained through business experience, the takeover word candidate list is obtained through the historical dialogue record, the practicability of the attribution table is improved, and the attribution table is obtained, so that the takeover word under the user behavior can be quickly located through the attribution table according to the user behavior in the future, and the efficiency of obtaining the takeover word is further improved.

[0043] The application also provides a takeover word obtaining system based on a joint decision model, which is suitable for the takeover word obtaining method based on the joint decision model and comprises:

[0044] An attribution table obtaining module is configured to divide the user behavior to obtain a user behavior category according to business experience, obtain a takeover word candidate list according to historical dialogue records, and obtain an attribution table by classifying the takeover words in the takeover word candidate list into the user behavior category.

[0045] A training set obtaining module is configured to obtain a behavior data set by obtaining the next user behavior corresponding to the historical dialogue record through a supervised model, and obtain a scoring data set by obtaining the scoring of the takeover word under the next user behavior through the attribution table.

[0046] A takeover word obtaining module is configured to train a joint decision model based on the behavior data set and the scoring data set, and obtain a final takeover word through the trained joint decision model.

[0047] The beneficial effects of the present application are as follows:

[0048] Considering the limitation of existing technologies that require manual annotation of conversation records and the next user behavior corresponding to the conversation records, and the strong semantic understanding capabilities of supervised models, we use supervised models to conduct in-depth analysis of historical conversation records to obtain high-quality behavior datasets and scoring datasets, breaking through the limitations of manually annotating datasets. The joint decision-making model is trained with high-quality behavior datasets, which improves the accuracy of user behavior predicted by the joint decision-making model. On this basis, the joint decision-making model is trained with high-quality scoring datasets, which also improves the scene adaptability of the final transition words, overcoming the technical problem of existing technologies that make it difficult to improve the efficiency of transition word acquisition and scene adaptability.

[0049] By obtaining user behavior categories through business experience and obtaining a list of candidate continuation words through historical conversation records, the practicality of the attribution table is improved. At the same time, by obtaining the attribution table, the continuation words under the user behavior can be quickly located through the attribution table based on the user behavior, further improving the efficiency of obtaining the continuation words. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Other features, objects, and advantages of the present invention will become more apparent upon reading the detailed description of non-limiting embodiments made with reference to the following drawings. The drawings are provided for the purpose of illustrating preferred embodiments only and are not to be construed as limiting the present invention. Like reference characters are used throughout the drawings to designate like parts.

[0051] Figure 1 Schematic diagram of the flow of the method for obtaining a connecting word based on the joint decision model of the present invention;

[0052] Figure 2 This is a flow chart for obtaining a behavior data set in the method for obtaining a linking word based on a joint decision-making model of the present invention;

[0053] Figure 3 This is a flowchart for obtaining a scoring data set in the method for obtaining a linking word based on a joint decision-making model of the present invention. DETAILED DESCRIPTION

[0054] In order to make the objectives, technical solutions and advantages of the present invention more clear, the present invention is further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific implementation method described herein is only an optimal embodiment of the present invention, which is only used to explain the present invention and does not limit the scope of protection of the present invention. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0055] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flow charts. Although the flow charts describe the operations (or steps) as sequential processes, many of the operations (or steps) therein can be performed in parallel, concurrently, or simultaneously. In addition, the order of the operations can be rearranged. The process can be terminated when its operations are completed, but can also have additional steps not included in the figures; the process can correspond to a method, function, procedure, subroutine, subprogram, etc.

[0056] Example 1: Figure 1 As shown, the method for obtaining the continuation word based on the joint decision model includes the following steps:

[0057] S1: Divide user behaviors into categories based on business experience, obtain a list of candidate continuation words based on historical conversation records, and classify the continuation words in the candidate continuation word list into user behavior categories to obtain an attribution table.

[0058] Specifically, user behavior categories include: regaining customers through interest points, collecting (asking for) information, making suggestions (solutions), introducing activities (interest points), soothing emotions, expressing understanding / resonance, clarifying issues (removing doubts), ending the conversation, guiding the next step, changing the topic, returning to the main thread, expressing gratitude, requesting repetition or unclear hearing, text noise, irrelevant text, etc.

[0059] As a preferred embodiment, the step of obtaining a candidate list of continuation words based on historical conversation records includes:

[0060] Obtain the first continuation word candidate list based on the frequency of reply prefixes in historical conversation records;

[0061] Calculating the similarity of the connecting words in the initial connecting word candidate list through the supervision model, and preliminarily grouping the connecting words in the initial connecting word candidate list according to the similarity;

[0062] A regrouping threshold is obtained according to the similarity of the continuation words in each group and the user behavior category, the continuation words in each group are regrouped based on the regrouping threshold, and a continuation word candidate list is obtained according to the regrouping result.

[0063] In this embodiment, obtaining a list of candidate initial continuation words based on the frequency of reply prefixes in historical conversation records is performed by, for example, using SQL statements to collect 300,000 historical conversation records from a business database. The top 150 most frequent customer service reply prefixes in these historical conversation records are then counted to generate a list of candidate initial continuation words. By filtering frequently occurring continuation words from actual user usage scenarios, redundant information is reduced while also improving the practicality of the continuation words. The supervised model can be GPT-O1, also known as the Strawberry model, which has demonstrated outstanding performance in natural language processing, enabling efficient and accurate text generation and understanding. Through large-scale data training, this model possesses strong generalization capabilities and excels in a variety of language tasks. Determining a regrouping threshold based on the similarity of continuation words within each group and user behavior categories is performed by, for example, using the number of user behavior categories as the final number of groups, adjusting the initial grouping threshold corresponding to the initial grouping by combining the final number of groups with the similarity of the continuation words within each group, and thereby obtaining a regrouping threshold. The final number of groups is the number of groups in the regrouping result. By utilizing the powerful semantic understanding ability of the supervised model for grouping, the accuracy of grouping is improved. In this embodiment, by obtaining an attribution table, the subsequent attribution table can quickly locate the continuation word under the user behavior based on the user behavior, further improving the efficiency of obtaining the continuation word.

[0064] S2: Obtain the next user behavior corresponding to the historical conversation record through the supervision model to obtain a behavior dataset, and then obtain the continuation word score under the next user behavior through the attribution table to obtain a score dataset.

[0065] Specifically, such as Figure 2 As shown, the behavior dataset is obtained by obtaining the next user behavior corresponding to the historical conversation record through the supervision model, including:

[0066] S21a: Use user behavior categories as output constraints for the supervision model;

[0067] S21b: Structuring the historical conversation records according to the roles, conversation order, and conversation content of the two parties in the historical conversation records to obtain structured conversation records;

[0068] S21c: Construct a prompt word based on the output constraints, concatenate the prompt word with the structured conversation record as the input text of the supervision model, and obtain the next user behavior corresponding to the historical conversation record;

[0069] S21d: Associating the next user behavior corresponding to the historical conversation record with the historical conversation record to obtain a behavior dataset.

[0070] Taking the e-commerce customer service conversation scenario as an example, the structured conversation record can be specifically expressed as follows:

[0071] [User 1]: The size of the clothes I bought is too small. Can you handle it?

[0072] [Customer Service 1]: Please take a photo of the clothes' tag. I will arrange a replacement after checking.

[0073] [User 2]: I have sent the photo. Please take a look.

[0074] [Customer Service 2]: I have received the photo. I have applied for a replacement. The new item will be shipped within 3 days.

[0075] The prompt word can be specifically represented as:

[0076] According to the structured dialogue record, the next behavior of the user is predicted, which is constrained by the output constraint condition.

[0077] The input text can be specifically represented as:

[0078] [User 1]: The size of the clothes I bought is too small. Can you handle it?

[0079] [Customer Service 1]: Please take a photo of the clothes' tag. I will arrange a replacement after checking.

[0080] [User 2]: I have sent the photo. Please take a look.

[0081] [Customer Service 2]: I have received the photo. I have applied for a replacement. The new item will be shipped within 3 days.

[0082] According to the above structured dialogue record, the next behavior of the user is predicted, which is constrained by the output constraint condition.

[0083] At this time, the supervision model can infer the next user behavior corresponding to the historical dialogue record as "confirming the order logistics" according to the input text. By binding the confirmation of the order logistics to the above historical dialogue record through the unique identifier, a set of first binding data can be obtained. It can be understood that each set of historical dialogue record corresponds to a set of first binding data, so each set of first binding data constitutes a behavior data set. By deeply analyzing the historical dialogue record through the supervision model, a high-quality behavior data set and a scoring data set are obtained, which breaks through the limitations of manually labeling data sets.

[0084] As shown in Figure 3 , the scoring data set obtained by the next user behavior through the attribution table includes:

[0085] S22a: Obtain the next user behavior through the attribution table;

[0086] S22b: Using the supervised model, the next user action, the subsequent word under the next user action, and the contextual information of the historical conversation records are vectorized to obtain a historical action vector, a historical subsequent word vector, and a historical context vector.

[0087] S22c: Obtaining first semantic similarity and second semantic similarity between the historical continuation word vector and the historical behavior vector and the historical context vector respectively;

[0088] S22d: Obtain a score of the continuation word under the next user behavior based on the first semantic similarity and the second semantic similarity to obtain a score dataset.

[0089] S3: Train the joint decision model based on the behavior dataset and the rating dataset, and obtain the final continuation word through the successfully trained joint decision model.

[0090] As a preferred embodiment, before training the joint decision model based on the behavior dataset and the scoring dataset, the method further includes:

[0091] The behavior dataset is divided into a first behavior dataset and a second behavior dataset based on the acquisition time of the historical conversation records. The supervision model is trained once using the second behavior dataset.

[0092] Inputting the historical conversation records corresponding to the first behavior dataset into a trained supervised model to obtain predicted user behavior, comparing the predicted user behavior with the next user behavior corresponding to the first behavior dataset to obtain a first comparison result, correcting the first behavior dataset based on the first comparison result, and updating the behavior dataset based on the corrected first behavior dataset;

[0093] The scoring dataset is divided into a first scoring dataset and a second scoring dataset based on the acquisition time of the historical conversation records, and the supervision model is trained twice using the second scoring dataset;

[0094] The historical conversation records corresponding to the first scoring dataset are input into the supervised model that has completed secondary training to obtain predicted scores, and the predicted scores are compared with the continuation word scores corresponding to the first scoring dataset to obtain a second comparison result. The first scoring dataset is corrected according to the second comparison result, and the scoring dataset is updated according to the corrected first scoring dataset.

[0095] In this embodiment, before dividing the behavioral dataset into the first and second behavioral datasets based on the acquisition time of the historical conversation records, the process also includes cleaning the historical conversation records and removing duplicate historical conversation records using the edit distance. This removal of redundant data further improves the training efficiency of the supervised model.

[0096] In this embodiment, the user can flexibly set the number of groups of first-bound data in the first behavior dataset as needed. If the number of groups of first-bound data in the first behavior dataset is 1000 and the number of behavior dataset groups is 6000, the behavior dataset is divided into a first behavior dataset and a second behavior dataset based on the time when the historical conversation records were acquired. Specifically, each group of first-bound data is labeled according to the time when the historical conversation records were acquired, and the behavior dataset is divided into two major groups according to the labels, namely the first behavior dataset and the second behavior dataset. The first behavior dataset consists of the first-bound data numbered 1-1000, and the second behavior data consists of the first-bound data numbered 1001-6000. Training the supervised model once using the second behavior dataset specifically means inputting the first-bound data numbered 1001-2000 into the supervised model to train the supervised model. When the degree of match between the output of the supervised model and the next user behavior in the first-bound data is greater than a first preset value, it indicates that the training is successful. The first preset value can be flexibly set by the user as needed. Updating the behavior dataset using the corrected first behavior dataset includes: using the corrected first behavior dataset as the first behavior dataset in the behavior dataset, and using the first binding data numbered 1001-6000 as the first behavior dataset, and using the first binding data numbered 2001-3000 as the second behavior dataset, training the supervision model again using the second behavior dataset, inputting the historical conversation records corresponding to the first behavior dataset into the once trained supervision model to obtain predicted user behavior, comparing the predicted user behavior with the next user behavior corresponding to the first behavior dataset to obtain a first comparison result, correcting the first behavior dataset using the first comparison result, and updating the behavior dataset using the corrected first behavior dataset, and continuously looping the above steps until the first binding data constituting the first behavior data is numbered 5001-6000, and then stopping the loop.

[0097] It is understandable that the scores of the next user and the next word of the next user in the scoring dataset constitute the second binding data. The scoring dataset is composed of several groups of second binding data. The user can flexibly set the number of groups of second binding data in the first scoring dataset according to needs. If the number of groups of second binding data in the first scoring dataset is 1000, and the number of groups of the scoring dataset is 6000, then the scoring dataset is divided into the first scoring dataset and the second scoring dataset according to the acquisition time of the historical conversation records. Specifically, it means: each group of second binding data is numbered according to the acquisition time of the historical conversation records, and the scoring dataset is divided into two major groups according to the number. The first scoring dataset consists of second binding data numbered 1-1000, and the second scoring data consists of second binding data numbered 1001-6000. Secondary training of the supervised model through the second scoring dataset specifically refers to training the supervised model through second binding data numbered 1001-2000. When the degree of match between the output of the supervised model and the score of the continuation word in the second binding data is greater than a second preset value, it indicates that the secondary training is successful, wherein the second preset value can be flexibly set by the user according to needs. Updating the scoring dataset using the revised first scoring dataset includes: using the revised first scoring dataset as the first scoring dataset in the scoring dataset, using the second binding data numbered 1001-6000 as the first scoring dataset, and using the second binding data numbered 2001-3000 as the second scoring dataset, retraining the supervised model using the second scoring dataset, inputting the historical conversation records corresponding to the first scoring dataset into the supervised model after the secondary training to obtain a predicted score, comparing the predicted score with the score of the continuation word corresponding to the first scoring dataset to obtain a second comparison result, correcting the first scoring dataset based on the second comparison result, and updating the scoring dataset using the revised first scoring dataset, and continuously looping the above steps until the second binding data constituting the first scoring dataset is numbered 5001-6000, at which time the loop stops. By implementing a batch-by-batch cyclic quality inspection of data sets, the quality of the behavioral and scoring data sets has been greatly improved, thereby improving the scenario adaptability of the transition words.

[0098] The training of the joint decision-making model based on the behavior dataset and the scoring dataset includes:

[0099] The classification model in the joint decision-making model is trained using the behavior dataset, and the ranking model in the joint decision-making model is trained using the scoring dataset.

[0100] As a preferred embodiment, the training of the classification model in the joint decision-making model using the behavior dataset and the training of the refinement model in the joint decision-making model using the scoring dataset include:

[0101] Divide the behavior dataset into several parts, input each behavior dataset into the classification model in turn to obtain the classification model output, and modify the classification model according to the difference between the classification model output and the first actual result;

[0102] The scoring data set is divided into several parts, and each scoring data set is input into the refined ranking model in turn to obtain the refined ranking model output, and the refined ranking model is corrected according to the gap between the refined ranking model output and the second actual result.

[0103] In this embodiment, the classification model is specifically the LLM classification model, which is also called the large language model. It is a model built based on the pre-trained large language model and used to solve classification tasks. The refinement model is specifically the BERT refinement model.

[0104] The BERT ranking model mainly uses the powerful semantic understanding ability of the BERT model to capture the semantic relevance between query terms and documents, thereby scoring and ranking documents. Each behavioral data set is input into the classification model in turn to obtain the classification model output, and the classification model is corrected according to the gap between the classification model output and the first actual result. Specifically, the following steps are used: input the historical conversation records in each behavioral data set into the classification model in turn to obtain the classification model output, and the conditional probability distribution function in the classification model is corrected according to the gap between the classification model output and the next user behavior corresponding to the historical conversation record in each behavioral data set, where the next user behavior corresponding to the historical conversation record in each behavioral data set is the first actual result. The expression of the conditional probability distribution function is: , Indicates the category of the next user behavior predicted by the classification model, Indicates that the historical conversation record is When , the category of the next user behavior is The conditional probability of Indicates that the historical conversation record is When the prompt word, Indicates that gpt-o1 is based on the prompt word , is the probability of the answer, and A, B, and K represent user behavior categories.

[0105] In this embodiment, each scoring data set is sequentially input into the refined ranking model to obtain the refined ranking model output, and the refined ranking model is corrected according to the difference between the refined ranking model output and the second actual result. Specifically, the following steps are performed: sequentially input the continuation words in each scoring data set into the refined ranking model to obtain the refined ranking model output, and the parameters in the refined ranking model are corrected according to the difference between the refined ranking model output and the continuation word score in each scoring data set, wherein the continuation word score in each scoring data set is the second actual result. The formula for correcting the parameters in the refined ranking model is: , represents the parameters in the refined model, It represents the mean square error between the output of the refined ranking model and the score of the continuation word in each scoring dataset. Indicates the total number of scoring datasets, represents the connecting word in the i-th scoring dataset, It represents the output of the refined sorting model corresponding to the continuation word in the i-th scoring dataset, , represents the score of the continuation word in the i-th scoring dataset, By training the joint decision-making model with a high-quality behavioral dataset, the accuracy of the user behavior predicted by the joint decision-making model is improved. On this basis, by training the joint decision-making model with a high-quality rating dataset, the scene adaptability of the final follow-up word is also improved, overcoming the technical problem of the existing technology that it is difficult to improve the efficiency of follow-up word acquisition and scene adaptability.

[0106] As a preferred embodiment, it also includes:

[0107] When the difference between the classification model output and the first actual result, and the difference between the refined ranking model output and the second actual result, both meet the preset conditions, the joint decision model training is successful. The preset conditions can be flexibly set by the user according to actual needs.

[0108] Specifically, obtaining the final connecting word through the successfully trained joint decision model includes:

[0109] Input the real-time conversation record into the classification model to obtain the real-time user behavior, obtain the continuation word under the real-time user behavior through the attribution table, input the continuation word under the real-time user behavior into the refined ranking model to obtain the continuation word score under the real-time user behavior, and use the continuation word with the highest score under the real-time user behavior as the final continuation word.

[0110] Embodiment 2: This embodiment further provides a system for acquiring a continuation word based on a joint decision model, which is applicable to the method for acquiring a continuation word based on a joint decision model, and includes:

[0111] The attribution table acquisition module is used to classify user behaviors according to business experience to obtain user behavior categories, obtain a list of candidate continuation words based on historical conversation records, and classify the continuation words in the candidate continuation word list into user behavior categories to obtain an attribution table;

[0112] The training set acquisition module is used to obtain the next user behavior corresponding to the historical conversation record through the supervision model to obtain a behavior dataset, and then obtain the continuation word score of the next user behavior through the attribution table to obtain a score dataset;

[0113] The continuation word acquisition module is used to train the joint decision model based on the behavior dataset and the rating dataset, and obtain the final continuation word through the successfully trained joint decision model.

[0114] In this embodiment, user behavior categories include: customer acquisition through interest points, information collection (inquiry), suggestions (solutions), activity introduction (interest points), emotional soothing, expressing understanding / empathy, clarifying questions (resolving concerns), concluding a conversation, guiding the next step, changing the topic, returning to the main thread, expressing gratitude, requesting repetition or unclear audio, text noise, and irrelevant text. The supervisory model can be GPT-O1. The joint decision-making model consists of a classification model and a refined ranking model.

[0115] The comparison table of the effects of the present invention and the prior art obtained by testing the machine Nvidia L20 is shown in Table 1:

[0116] Table 1. Comparison of the effects of the present invention and the prior art

[0117] ;

[0118] Acc indicates the classification accuracy. f1_score indicates the adaptability of the linker. A larger value indicates higher adaptability. qps@p99 200ms indicates the maximum number of requests that can be achieved while maintaining a p99 latency of less than 200ms.

[0119] The specific implementation described above is a preferred implementation of the method and system for obtaining continuation words based on the joint decision-making model of the present invention, and is not intended to limit the specific implementation scope of the present invention. The scope of the present invention includes but is not limited to this specific implementation. Any equivalent changes made in accordance with the shape and structure of the present invention are within the scope of protection of the present invention.

Claims

1. A method for obtaining a continuation word based on a joint decision model, characterized in that: The following steps are involved: Divide user behaviors into categories based on business experience, obtain a list of candidate continuation words based on historical conversation records, and classify the continuation words in the candidate continuation word list into user behavior categories to obtain an attribution table; The supervised model is used to obtain the next user behavior corresponding to the historical conversation record to obtain a behavior dataset, and then the attribution table is used to obtain the continuation word score of the next user behavior to obtain a score dataset; The joint decision model is trained based on the behavior dataset and the rating dataset, and the final continuation word is obtained through the successfully trained joint decision model; The method of obtaining the next user behavior corresponding to the historical conversation record through the supervision model to obtain the behavior dataset includes: Use user behavior categories as output constraints for the supervision model; Structuring the historical conversation records according to the roles, conversation order, and conversation content of the two parties in the historical conversation records to obtain structured conversation records; Construct prompt words based on the output constraints, concatenate the prompt words with the structured conversation record as the input text of the supervision model, and obtain the next user behavior corresponding to the historical conversation record; Associating the next user behavior corresponding to the historical conversation record with the historical conversation record to obtain a behavior dataset; The scoring dataset is obtained by obtaining the score of the continuation word under the next user behavior through the attribution table, including: Obtain the continuation word for the next user behavior through the attribution table; The supervised model vectorizes the next user action, the subsequent word under the next user action, and the contextual information of the historical conversation records to obtain the historical action vector, the historical subsequent word vector, and the historical context vector. Obtain the first semantic similarity and the second semantic similarity of the historical continuation word vector with the historical behavior vector and the historical context vector respectively; The scores of the continuation words under the next user behavior are obtained based on the first semantic similarity and the second semantic similarity to obtain a score data set.

2. The method for obtaining a transition word based on a joint decision model according to claim 1, characterized in that: The step of obtaining a candidate list of continuation words based on historical conversation records includes: Obtain the first continuation word candidate list based on the frequency of reply prefixes in historical conversation records; Calculating the similarity of the connecting words in the initial connecting word candidate list through the supervision model, and preliminarily grouping the connecting words in the initial connecting word candidate list according to the similarity; A regrouping threshold is obtained according to the similarity of the continuation words in each group and the user behavior category, the continuation words in each group are regrouped based on the regrouping threshold, and a continuation word candidate list is obtained according to the regrouping result.

3. The method for obtaining a transition word based on a joint decision model according to claim 1, characterized in that: Before training the joint decision model based on the behavior dataset and the scoring dataset, the method further includes: The behavior dataset is divided into a first behavior dataset and a second behavior dataset based on the acquisition time of the historical conversation records. The supervision model is trained once using the second behavior dataset. Inputting the historical conversation records corresponding to the first behavior dataset into a trained supervised model to obtain predicted user behavior, comparing the predicted user behavior with the next user behavior corresponding to the first behavior dataset to obtain a first comparison result, correcting the first behavior dataset based on the first comparison result, and updating the behavior dataset based on the corrected first behavior dataset; The scoring dataset is divided into a first scoring dataset and a second scoring dataset based on the acquisition time of the historical conversation records, and the supervision model is trained twice using the second scoring dataset; The historical conversation records corresponding to the first scoring dataset are input into the supervised model that has completed secondary training to obtain predicted scores, and the predicted scores are compared with the continuation word scores corresponding to the first scoring dataset to obtain a second comparison result. The first scoring dataset is corrected according to the second comparison result, and the scoring dataset is updated according to the corrected first scoring dataset.

4. The method for obtaining a transition word based on a joint decision model according to claim 3, characterized in that: The training of the joint decision-making model based on the behavior dataset and the scoring dataset includes: The classification model in the joint decision-making model is trained using the behavior dataset, and the ranking model in the joint decision-making model is trained using the scoring dataset.

5. The method for obtaining a transition word based on a joint decision model according to claim 4, characterized in that: The training of the classification model in the joint decision-making model using the behavior dataset and the training of the refinement model in the joint decision-making model using the scoring dataset include: Divide the behavior dataset into several parts, input each behavior dataset into the classification model in turn to obtain the classification model output, and modify the classification model according to the difference between the classification model output and the first actual result; The scoring data set is divided into several parts, and each scoring data set is input into the refined ranking model in turn to obtain the refined ranking model output, and the refined ranking model is corrected according to the gap between the refined ranking model output and the second actual result.

6. The method for obtaining transition words based on a joint decision model according to claim 5, characterized in that: Also includes: When the gap between the classification model output and the first actual result and the gap between the refined ranking model output and the second actual result both meet the preset conditions, it means that the joint decision model training is successful.

7. The method for obtaining transition words based on a joint decision model according to claim 4, characterized in that: The method of obtaining the final connecting word through the successfully trained joint decision model includes: Input the real-time conversation record into the classification model to obtain the real-time user behavior, obtain the continuation word under the real-time user behavior through the attribution table, input the continuation word under the real-time user behavior into the refined ranking model to obtain the continuation word score under the real-time user behavior, and use the continuation word with the highest score under the real-time user behavior as the final continuation word.

8. A system for acquiring a continuation word based on a joint decision model, applicable to the method for acquiring a continuation word based on a joint decision model according to any one of claims 1 to 7, characterized in that: include: The attribution table acquisition module is used to classify user behaviors according to business experience to obtain user behavior categories, obtain a list of candidate continuation words based on historical conversation records, and classify the continuation words in the candidate continuation word list into user behavior categories to obtain an attribution table; The training set acquisition module is used to obtain the next user behavior corresponding to the historical conversation record through the supervision model to obtain a behavior dataset, and then obtain the continuation word score of the next user behavior through the attribution table to obtain a score dataset; The continuation word acquisition module is used to train the joint decision model based on the behavior dataset and the rating dataset, and obtain the final continuation word through the successfully trained joint decision model.

Images