CN118568241A

CN118568241A - User dialogue and portrait intention prediction method based on pre-training model

Info

Publication number: CN118568241A
Application number: CN202411038205.9A
Authority: CN
Inventors: 张亶; 项锋; 徐宙杰
Original assignee: Zhejiang University ZJU
Current assignee: Zhejiang University ZJU
Priority date: 2024-07-31
Filing date: 2024-07-31
Publication date: 2024-08-30

Abstract

The invention discloses a user dialogue and portrait intention prediction method based on a pre-training model. The method comprises the following steps: firstly, constructing an intention analysis model by combining training data so as to optimize the prediction effect of user dialogue and portrait intention; then, training a user feedback classification model to divide the user feedback categories into low feedback and high feedback; then, when low feedback occurs in the conversation process of the user and the intelligent sales customer service, predicting the intention of the user to obtain an intention keyword; finally, the corresponding cases and solutions are searched through the intention keywords and fed back to the intelligent sales service to better serve the clients. The invention not only can improve the purchase rate of the user, but also can improve the satisfaction of the user, and has important significance for the intelligent improvement of the traditional intelligent sales customer service.

Description

User dialogue and portrait intention prediction method based on pre-training model

Technical Field

The invention relates to the technical field of computers, in particular to the technical field of artificial intelligence such as natural language processing, deep learning, large models and the like, and particularly relates to an intention prediction method for user dialogue and portraits based on a pre-training model.

Background

The rapid development of internet technology today makes artificial intelligence technology widely used in the fields of user services and customer support. With the increasing popularity of intelligent customer service systems, intelligent customer service based on large models has also been applied in various industries. However, basic intelligent customer service does not work well in the sales area: firstly, because intelligent customer service does not have sales call operation, human sales customer service cannot be well simulated; secondly, the intelligent customer service cannot calculate the real intention of the user and cannot conduct correct guidance. Therefore, the traditional intelligent customer service cannot provide satisfactory service when facing the sales field, and meanwhile, the traditional intelligent customer service cannot finish good conversion for online customers.

Disclosure of Invention

Aiming at the defects of the prior art, the invention aims to provide a user dialogue and portrait intention prediction method based on a pre-training model, which aims to improve the user satisfaction degree of intelligent customer service in the sales field and improve the conversion rate of users.

The aim of the invention is realized by the following technical scheme: a method for predicting intent of a user dialog and portrayal based on a pre-trained model, comprising the steps of:

1. Constructing an intention analysis model through reinforcement learning based on a pre-trained LLama-7 b model in combination with training data, wherein the training data comprises user dialogue and user portraits;

2. Based on a pretrained Bert model, training a user feedback classification model by combining training data, and classifying user feedback into low feedback and high feedback;

3. According to the feedback condition of the user in the current dialogue of the user and the intelligent sales customer service, whether an intention analysis mechanism is triggered or not is evaluated through a user feedback classification model, wherein the intention analysis mechanism is triggered when the feedback is low and is not triggered when the feedback is high;

4. after triggering the mechanism described in the step 3, reading the dialogue histories of the user and the intelligent sales customer service and the portrait features of the user through an intention analysis model, and analyzing the intention of the user through the intention analysis model to obtain intention keywords;

5. And (3) searching cases and solutions corresponding to the intent keywords according to the intent keywords obtained in the step (4), returning to the intelligent sales service, and combining the received cases and solutions with the current conversation by the intelligent sales service to generate new replies and returning the new replies to the user.

Further, said step 1 is realized by the following sub-steps:

1.1. And (3) data processing: matching and cleaning dialogue data of a user and sales customer service with user portrait data at the same moment;

1.2. Data marking: taking the data in the step 1.1 as training data, and classifying the intention of the data according to the dialogue content of the user and the intelligent sales customer service and the purchasing condition in the last seven days to form marked data;

1.3. and (3) building a reinforcement learning environment: defining the target (through continuous training and optimization, the intention analysis model predicts the intention of the user more accurately on the basis of the user dialogue and the portrait data) state, action and rewarding function of reinforcement learning by using the marked data in the step 1.2;

1.4. model training: using the preprocessed data, in conjunction with reinforcement learning algorithms, an intent analysis model is trained on LLama3-7b models to optimize the predictive effects of user dialog and portrayal intent.

Further, the states in step 1.3 include: the environment information at a certain moment comprises the current dialogue content, the user portrait at a certain moment and the historical intent prediction result.

Further, the act of defining reinforcement learning in step 1.3 is specifically: classifying the intention of the current dialogue and the user portrait data, and judging the current intention of the user, wherein the intention comprises the following steps: strong purchase intention, weak purchase intention, strong counseling intention, weak counseling intention.

Further, the reward function defining reinforcement learning in step 1.3 is specifically: and according to the accuracy of prediction of the intent analysis model, giving or deducting points according to corresponding weight coefficients, evaluating the analysis effect of the intent analysis model according to the points, and finally selecting the model coefficient with the highest total points as the final coefficient of the intent analysis model.

Further, said step2 is realized by the following sub-steps:

2.1. and (3) data processing: manually splitting dialogue data of a user and sales customer service into question-answer pairs, and marking each question-answer pair with a data tag, wherein the tag is low feedback or high feedback;

2.2. Fine tuning of the model: based on a pre-trained Bert model, training a user feedback classification model by using the data tag in the step 2.1 so as to judge whether the feedback of the user is low feedback or high feedback according to the input question-answer pairs and the output probability value of high-low feedback, wherein the probability value is between 0 and 1; when the probability value of the feedback is not lower than 0.6, the feedback is judged to be high, and when the probability value of the feedback is lower than 0.6, the feedback is judged to be low (in order to reduce the risk of misjudgment, the feedback is judged to be low more preferably so as to better analyze the intention of the user and improve the satisfaction degree of the user).

Further, said step4 is realized by the following sub-steps:

4.1. And (3) data processing: after the low feedback is determined, pulling the dialogue between the user and the intelligent customer service, and extracting and processing the historical portrait of the user, wherein the method comprises the following steps: the method comprises the steps of simplifying and refining dialogue history, analyzing user purchase history, analyzing user browsing times and residence time;

4.2. intent analysis: and (3) taking the data processed in the step (4.1) as input of an intention analysis model, and analyzing the current intention of the user to obtain intention keywords.

Further, said step 5 is realized by the following sub-steps:

5.1. Through the classification of the intention keywords and the keywords, the RAG retrieval enhancement optimization is based, and cases in a knowledge base are retrieved in multiple ways;

5.2. according to the contents of the multi-path retrieval recall, training a sort model based on Bert to select knowledge so as to screen cases and solutions highly related to the dialogue context;

5.3. and returning the cases and the solutions to the intelligent sales service, wherein the intelligent sales service uses the returned contents in combination with the current dialogue as new input, generates new output contents and returns the new output contents to the user.

The invention also provides an electronic device comprising a memory and a processor, the memory being coupled to the processor; wherein the memory is configured to store program data and the processor is configured to execute the program data to implement the pre-trained model-based user dialog and portrayal intent prediction method.

The present invention also provides a computer readable storage medium having stored thereon a computer program which when executed by a processor implements the pre-trained model based user dialog and portrayal intent prediction method.

The beneficial effects of the invention are as follows: the invention predicts the real intention of the user by combining the user dialogue history and the user portrait; the user experience can be remarkably improved by simulating human sales behaviors instead of a single one-to-one answer mode, and the sales effect is enhanced; through continuous optimization driven by data, the method becomes the core competitiveness of a marketing large model.

Drawings

Fig. 1 is a general flow chart of the present invention.

Detailed Description

Reference will now be made in detail to exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, the same numbers in different drawings refer to the same or similar elements, unless otherwise indicated. The implementations described in the following exemplary examples do not represent all implementations consistent with the invention. Rather, they are merely examples of apparatus and methods consistent with aspects of the invention as detailed in the accompanying claims.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used in this specification and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used herein refers to and encompasses any or all possible combinations of one or more of the associated listed items.

It should be understood that although the terms first, second, third, etc. may be used herein to describe various information, these information should not be limited by these terms. These terms are only used to distinguish one type of information from another. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of the invention. The word "if" as used herein may be interpreted as "at … …" or "at … …" or "in response to a determination" depending on the context.

The present invention will be described in detail with reference to the accompanying drawings. The features of the examples and embodiments described below may be combined with each other without conflict.

As shown in FIG. 1, the method for predicting the intention of a user dialogue and a portrait based on a pre-training model comprises the following specific steps:

1. An intent analysis model is built by reinforcement learning based on a pre-trained LLama-7 b model in combination with training data, including user dialog and user portrayal.

1.1, Data processing: and matching the dialogue data of the user and the sales customer service with the user portrait data at the same time, and cleaning. The dialogue data of the user and the sales customer service come from historical data stored in a database; the user portrait data mainly includes: the user consumption times, user potential layering, user purchase intention, user last consumption time, personal characteristics of the user and the like cover the personal tags of the user and the behavior tags of the user. Matching dialogue data and portrait data means: the dialogue data is combined with the history of the user at the time of the expiration of the dialogue (judging the characteristics of the user) to form (dialogue data, keywords, user tags) data.

1.2, Data marking: and (3) taking the data in the step (1.1) as training data, and classifying the intention of the data according to the dialogue content of the user and the intelligent sales customer service and the purchasing condition in the last seven days to form marked data. The step is mainly to mark the data manually. The data is labeled manually, rather than according to a certain rule, so that the data can be more in line with human expectations, and the model can be more accurately classified. The way in which data is marked mainly is roughly as follows: dialog data, keywords and user tags are labeled as user intent (tags), i.e. the data of 1.1 is tagged with intent tags.

1.3, Building a reinforcement learning environment: using the labeled training data of step 1.2 and the user intent, user image data, a reinforcement learning objective, state, action, and reward function is defined, wherein:

Defining a target: the goal of reinforcement learning is to maximize the accuracy of intent predictions. Specifically, by continuously training and optimizing the model, the user's intent can be predicted more accurately based on the user's dialog and portrayal data.

Defining a state: the state represents environmental information that is located at a certain time. In the context of user dialog and portrait intent prediction, the states may include: current dialog content, user portraits up to a certain moment, historical intent prediction results.

Defining actions: the action is mainly the intention classification operation of the current dialogue and the user portrait data, and the main intention classification is as follows: strong purchase intention, weak purchase intention, strong counseling intention, weak counseling intention, commodity understanding intention, commodity comparison intention, and the like.

Defining a reward function: the reward function is a standard for evaluating the quality of a certain action by the reinforcement learning model, and is mainly to give points or deduct points according to the correctness of prediction and the weighting coefficient. The weighting coefficient is mainly calibrated manually, and the manual judgment is carried out: the current dialogue content and the user portrait of the user are mainly between 0.6 and 1.5 for analyzing the difficulty coefficient of the user intention. The reward function is:

Where x represents training data, x=1 represents first piece of training data, x=n represents nth piece of training data, α represents a weighting coefficient, and E _x represents the intended prediction case of the xth input (completely correct 1, large classification correct 0.5, error 0).

1.4, Model training: using the preprocessed data, in conjunction with reinforcement learning algorithms, an intent analysis model is trained on LLama3-7b models to optimize the predictive effects of user dialog and portrayal intent. Model training generally requires 30 rounds, with 30 rounds being able to reach the highest accuracy for data prediction. One round: all training data is traversed.

2. Based on the pretrained Bert model, the user feedback classification model is trained in combination with training data, and the user feedback categories are classified into low feedback and high feedback.

2.1, Data processing: the dialogue data of the user and the sales customer service are manually split into question-answer pairs, and each question-answer pair is marked with a data tag, wherein the tag is low feedback or high feedback. The data of the part only comprises the dialogue between the user and the sales customer service, and the multi-round dialogue is split into a plurality of question-answer pairs; the labeling criteria were: the subjective judgment of the person is made whether the user's reply is positive (high feedback) or negative (low feedback).

2.2, Fine tuning of a model: based on a pretrained Bert model, training a user feedback classification model by using the labeled data in the step 2.1 so as to judge whether the feedback of the user is low feedback or high feedback according to the probability value of the output high-low feedback and the input question-answer pair by the user feedback classification model, wherein the probability value is between 0 and 1; when the probability value of feedback is not lower than 0.6, high feedback is determined, and when the probability value of feedback is lower than 0.6, low feedback is determined. The fine tuning mode used is the Lora fine tuning, wherein the Lora fine tuning: the local parameters of the model, but not all the parameters, can be adjusted to obtain new capacity without losing original capacity, and the cost of computing resources required by training is low (the requirement on hardware equipment is lower).

3. According to the feedback condition of the user in the conversation of the user current and intelligent sales customer service, whether the intention analysis mechanism is triggered is evaluated through a user feedback classification model, the intention analysis mechanism is triggered when the user is in low feedback, and the intention analysis mechanism is not triggered when the user is in high feedback (in order to reduce the risk of misjudgment, the user intention is better analyzed by rather judging the user to be in low feedback, and the satisfaction degree of the user is improved).

4. After the mechanism in the step 3 is triggered, the dialogue histories of the user and the intelligent sales customer service and the portrait features of the user are read through an intention analysis model, and the intention of the user is analyzed through the intention analysis model to obtain intention keywords.

4.1, Data processing: after the low feedback is determined, the dialogue between the user and the intelligent customer service is extracted and processed, and the method comprises the following steps: the method comprises the steps of simplifying and refining dialogue history, extracting dialogue keywords, analyzing user purchase history, analyzing user browsing times and residence time. Data in such a form is formed (dialogue data, keywords, user tags).

4.2, Intention analysis: and (3) taking the data processed in the step (4.1) as input of an intention analysis model, and analyzing intention such as purchase, state and the like of a user to obtain intention keywords.

5.1, Through the classification of the intended keywords and the keywords, the RAG retrieval enhancement optimization is based on, and cases in a knowledge base corresponding to the intended keywords are retrieved in multiple ways. Wherein, RAG retrieval enhancement: searching in a knowledge base; multipath searching: a plurality of similar cases are retrieved.

And 5.2, according to the recalled content of the multi-path retrieval, training a sort model based on Bert to select knowledge so as to screen cases and solutions highly related to the dialogue context (select the most relevant knowledge of top 3).

And 5.3, returning the cases and the solutions to an intelligent sales customer service (sales large model), and enabling the returned contents to be combined with the current dialogue to serve as new input, generating new output contents and returning the new output contents to the user. The returned content will be used as a prompt word, i.e. giving historical case knowledge or related background knowledge, enabling a large model to better understand how well the conversation with the user is handled.

The embodiment of the invention also provides electronic equipment, which comprises a memory and a processor, wherein the memory is coupled with the processor; wherein the memory is configured to store program data and the processor is configured to execute the program data to implement the pre-training model-based user dialog and portrayal intent prediction method of the above-described embodiments.

The embodiment of the invention also provides a computer readable storage medium, on which a computer program is stored, which when being executed by a processor, implements the user dialogue and portrait intention prediction method based on the pre-training model of the above embodiment.

The computer readable storage medium may be an internal storage unit, such as a hard disk or a memory, of any of the data processing enabled devices described in any of the previous embodiments. The computer readable storage medium may also be any device having data processing capabilities, such as a plug-in hard disk, a smart memory card (SMART MEDIA CARD, SMC), an SD card, a flash memory card (FLASH CARD), or the like, provided on the device. Further, the computer readable storage medium may include both internal storage units and external storage devices of any data processing device. The computer readable storage medium is used for storing the computer program and other programs and data required by the arbitrary data processing apparatus, and may also be used for temporarily storing data that has been output or is to be output.

The above embodiments are merely for illustrating the design concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement the same, the scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications according to the principles and design ideas of the present invention are within the scope of the present invention.

Claims

1. A method for predicting intent of a user dialog and portrayal based on a pre-trained model, comprising the steps of:

(1) Constructing an intention analysis model through reinforcement learning based on a pre-trained LLama-7 b model in combination with training data, wherein the training data comprises user dialogue and user portraits;

(2) Based on a pretrained Bert model, training a user feedback classification model by combining training data, and classifying user feedback into low feedback and high feedback;

(3) According to feedback conditions of a user in a dialogue of the user and intelligent sales customer service, whether an intention analysis mechanism is triggered or not is evaluated through a user feedback classification model, wherein the intention analysis mechanism is triggered when the feedback is low and is not triggered when the feedback is high;

(4) After triggering the mechanism in the step (3), reading the dialogue histories of the user and the intelligent sales customer service and the portrait features of the user through an intention analysis model, and analyzing the intention of the user through the intention analysis model to obtain intention keywords;

(5) And (3) searching cases and solutions corresponding to the intent keywords according to the intent keywords obtained in the step (4), returning to the intelligent sales service, and combining the received cases and solutions with the current conversation by the intelligent sales service to generate a new reply and returning the new reply to the user.

2. The method for predicting intent of user dialog and portrayal based on pre-training model of claim 1, wherein said step (1) is implemented by the sub-steps of:

(1.1) data processing: matching and cleaning dialogue data of a user and sales customer service with user portrait data at the same moment;

(1.2) data labeling: taking the data in the step (1.1) as training data, and classifying the intention of the data according to the dialogue content of the user and the intelligent sales customer service and the purchasing condition in the last seven days to form marked data;

(1.3) reinforcement learning environment construction: defining the goal, state, action and rewarding function of reinforcement learning by using the marked data of the step (1.2);

(1.4) model training: using the preprocessed data, in conjunction with reinforcement learning algorithms, an intent analysis model is trained on LLama3-7b models to optimize the predictive effects of user dialog and portrayal intent.

3. The pretrained model-based user dialog and portrayal intent prediction method as recited in claim 2, wherein the states in step (1.3) include: the environment information at a certain moment comprises the current dialogue content, the user portrait at a certain moment and the historical intent prediction result.

4. The method for predicting intent of user dialog and portrayal based on pre-training model of claim 2, wherein the act of defining reinforcement learning in step (1.3) is specifically: classifying the intention of the current dialogue and the user portrait data, and judging the current intention of the user, wherein the intention comprises the following steps: strong purchase intention, weak purchase intention, strong counseling intention, weak counseling intention.

5. The method for predicting intent of user dialog and portrayal based on pre-training model of claim 2, wherein the reward function defining reinforcement learning in step (1.3) is specifically: and according to the accuracy of prediction of the intent analysis model, giving or deducting points according to corresponding weight coefficients, evaluating the analysis effect of the intent analysis model according to the points, and finally selecting the model coefficient with the highest total points as the final coefficient of the intent analysis model.

6. The method of pre-training model based user dialog and portrayal intent prediction as claimed in claim 1, wherein step (2) is implemented by the sub-steps of:

(2.1) data processing: manually splitting dialogue data of a user and sales customer service into question-answer pairs, and marking each question-answer pair with a data tag, wherein the tags are classified as low feedback or high feedback;

(2.2) model fine tuning: training a user feedback classification model by using the data tag in the step (2.1) based on the pretrained Bert model so as to judge whether the feedback of the user is low feedback or high feedback according to the probability value of the output high-low feedback and the input question-answer pair by the user feedback classification model, wherein the probability value is between 0 and 1; when the probability value of feedback is not lower than 0.6, high feedback is determined, and when the probability value of feedback is lower than 0.6, low feedback is determined.

7. The method of pre-training model based user dialog and portrayal intent prediction as claimed in claim 1, wherein said step (4) is implemented by the sub-steps of:

(4.1) data processing: after the low feedback is determined, pulling the dialogue between the user and the intelligent customer service, and extracting and processing the historical portrait of the user, wherein the method comprises the following steps: the method comprises the steps of simplifying and refining dialogue history, analyzing user purchase history, analyzing user browsing times and residence time;

(4.2) intention analysis: and (3) taking the data processed in the step (4.1) as input of an intention analysis model, and analyzing the current intention of the user to obtain an intention keyword.

8. The method for predicting the intent of a user's dialog and portrayal based on a pre-trained model according to claim 1, characterized in that said step (5) is implemented by the sub-steps of:

(5.1) through intention keywords and keyword classification, enhancing optimization based on RAG retrieval, and multi-path retrieval of cases in a knowledge base;

(5.2) according to the content of the multi-path retrieval recall, training a sort model based on Bert to select knowledge so as to screen cases and solutions highly relevant to the dialogue context;

(5.3) returning the cases and solutions to the intelligent sales service, wherein the intelligent sales service uses the returned contents in combination with the current dialogue as new input, generates new output contents and returns the new output contents to the user.

9. An electronic device comprising a memory and a processor, wherein the memory is coupled to the processor; wherein the memory is for storing program data and the processor is for executing the program data to implement the pre-trained model based user dialog and portrayal intent prediction method as claimed in any of the preceding claims 1-8.

10. A computer readable storage medium having stored thereon a computer program, wherein the program when executed by a processor implements the pre-trained model based user dialog and portrayal intent prediction method of any of the preceding claims 1-8.