Recommendation method and device, model training method and device, equipment, medium

Abstract

The present application relates to the technical field of recommendation, and discloses a recommendation method and apparatus, a training method and apparatus of a model, equipment, and a medium. The recommendation method includes: obtaining the conversation content, the user portrait corresponding to the user who generates the conversation, and the historical conversation data corresponding to the user; the conversation content includes the user's first work requirement; the user's position information exists in the user portrait; according to the user portrait, the historical conversation data, and the first work requirement, the first work requirement is identified using a preset intention recognition model to obtain the work intention; the work content is determined according to the work intention; the work content is information related to the user's position and office; and the work content is recommended. In this way, the user portrait can reflect what work content the user does in what job position, and the user portrait, the first work requirement, and the historical conversation data can be combined to determine the user's real work requirement, and then the work content recommended according to the real work requirement can be more accurate to meet the user's expectations.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and in particular to a recommendation method and device, a model training method and device, equipment, and medium. Background Art

[0002] As enterprises continue to increase their requirements for work efficiency and accuracy, it is very important to integrate automation and intelligent office work. In related technologies, users directly input their work requirements into the big language model, and then the big language model recommends work content according to the work requirements. However, the work content recommended in this way often only has some general workflow steps, which is very broad and not targeted, and may not meet the user's expectations, resulting in a poor user experience.

[0003] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present application, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0004] In order to provide a basic understanding of some aspects of the disclosed embodiments, a brief summary is given below. The summary is not an extensive review, nor is it intended to identify key / critical elements or delineate the scope of protection of these embodiments, but rather serves as a prelude to the detailed description that follows.

[0005] The embodiments of the present application provide a recommendation method and apparatus, a model training method and apparatus, equipment, and medium to recommend work content that meets user expectations.

[0006] An embodiment of the present application provides a method for recommending work content, including: obtaining conversation content, a user portrait corresponding to the user who generates the conversation, and historical conversation data corresponding to the user; the conversation content includes the user's first work requirement; the user portrait contains the user's position information; according to the user portrait, the historical conversation data and the first work requirement, the first work requirement is identified using a preset intention recognition model to obtain a work intention; according to the work intention, the work content is determined; the work content is information related to the user's position and office; and the work content is recommended.

[0007] It is understandable that different users may have different real intentions when saying the same sentence. For example, when an employee in the marketing department says "I need to schedule a meeting", his real intention may be to plan an event. When an employee in the finance department says "I need to schedule a meeting", his real intention may be to review the budget. User portraits can reflect what work a user does in what position. Historical conversation data can reflect the user's actual work needs. Therefore, the user's real work needs can be determined by combining the user portrait, the first work need and the historical conversation data, and the work content recommended based on the real work needs can be more accurate and targeted to meet the user's expectations.

[0008] Furthermore, the intent recognition model is obtained in the following manner: obtaining sample data corresponding to multiple users respectively; the sample data includes: sample user portraits, sample historical conversation data and sample work requirements; using the sample data with work intention labels to train a first preset model to obtain an intent recognition model.

[0009] In the above embodiment, the model can capture the underlying rules, patterns and structures behind the data through the learning process. The model is trained by combining work intention, sample user portrait, sample historical conversation data and sample work requirements, so that the model can accurately identify the actual work intention behind the work requirements proposed by the user.

[0010] Furthermore, the first work requirement is identified using a preset intention recognition model according to the user portrait, the historical conversation data and the first work requirement to obtain the work intention, including: extracting first key information from the user portrait; the first key information includes: the user's position information; extracting second key information from the historical conversation data; the second key information includes: the conversation topic and the second work requirement; obtaining the work intention using the intention recognition model according to the position information, the conversation topic, the second work requirement and the first work requirement.

[0011] In the above embodiment, different job information may have different work requirements. At the same time, in the user's historical conversation data, there may also be users who have mentioned their actual work requirements in a certain scenario. Therefore, the user's job information, conversation topic, and second work requirement are extracted for intention recognition, so as to more accurately determine the user's work intention.

[0012] Furthermore, determining the work content according to the work intention includes: determining the work ability corresponding to the work intention; assembling the user portrait, the work ability, the work intention and a preset first prompt template to obtain a first text to be input; the first prompt template includes a sentence that recommends the work content using a collaborative filtering algorithm; inputting the first text to be input into a preset first language model to obtain the work content output by the first language model.

[0013] In the above embodiment, different work intentions may require different work capabilities. By first determining the actual required work capabilities, and then directly using the large language model to use collaborative filtering algorithms to recommend work content based on user portraits, work capabilities, and work intentions, it can be more targeted and facilitate the final recommended work content to better meet user expectations.

[0014] Furthermore, recommending the work content includes: determining whether the work content is compliant; if the work content is compliant, recommending the work content; if the work content is non-compliant, replacing or deleting the non-compliant work content, and recommending the work content after replacement or deletion.

[0015] In the above embodiment, different regions may have different customs or regulations. Therefore, by screening whether the work content is compliant and replacing or deleting it, the generated work content can be made legal and compliant, thereby reducing disputes.

[0016] Furthermore, after recommending the work content, the method also includes: extracting third key information from the conversation content and the work content; the third key information includes: work requirements, professional terms, emotions, three elements of events, and potential needs; assembling the third key information with a preset second prompt template to obtain a second text to be input; inputting the second text to be input into a preset second largest language model to obtain a recommended question output by the second largest language model.

[0017] In the above embodiment, by extracting the third key information from the conversation content and the work content, it is possible to infer questions that the user may need to further ask, thereby reducing the burden of the user in formulating questions and improving the user experience.

[0018] Furthermore, after recommending the work content, the method also includes: obtaining a first score of the user's feedback on the work intention; optimizing the intention recognition model based on the first score; and / or obtaining a second score of the user's feedback on the work content; and optimizing the first language model based on the second score.

[0019] In the above embodiment, by obtaining the user's feedback score to optimize the model, the model can be more accurate in subsequent recognition.

[0020] An embodiment of the present application provides a method for training an intent recognition model, including: obtaining sample data corresponding to multiple users respectively; the sample data includes: sample user portraits, sample historical conversation data and sample work requirements; using the sample data with work intention labels to train a first preset model to obtain an intent recognition model.

[0021] In the above embodiment, the model can capture the underlying rules, patterns and structures behind the data through the learning process. The model is trained by combining work intention, sample user portrait, sample historical conversation data and sample work requirements, so that the model can accurately identify the actual work intention behind the work requirements proposed by the user.

[0022] Furthermore, after obtaining the intention recognition model, it also includes: obtaining the conversation content, the user portrait corresponding to the user who generated the conversation, and the historical conversation data corresponding to the user; the conversation content includes the user's first work requirement; the user's position information exists in the user portrait; according to the user portrait, the historical conversation data and the first work requirement, using a preset intention recognition model to identify the first work requirement and obtain the work intention; obtaining a first score of the user's feedback on the work intention; and optimizing the intention recognition model based on the first score.

[0023] In the above embodiment, by obtaining the user's feedback score to optimize the model, the model can be more accurate in subsequent recognition.

[0024] An embodiment of the present application provides a work content recommendation device, including: a first acquisition module, used to obtain conversation content, a user portrait corresponding to the user who generates the conversation, and historical conversation data corresponding to the user; the conversation content includes the user's first work requirement; the user portrait contains the user's job information; an intention recognition module, used to identify the first work requirement according to the user portrait, the historical conversation data and the first work requirement using a preset intention recognition model to obtain a work intention; a work content determination module, used to determine the work content according to the work intention; the work content is information related to the user's job and office; a recommendation model, used to recommend the work content.

[0025] An embodiment of the present application provides a training device for an intent recognition model, including: a second acquisition module, used to obtain sample data corresponding to multiple users respectively; the sample data includes: sample user portraits, sample historical conversation data and sample work requirements; a training module, used to train a first preset model using the sample data with work intention labels to obtain an intent recognition model.

[0026] An embodiment of the present application provides an electronic device, including a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the above-mentioned work content recommendation method or intent recognition model training method.

[0027] In some embodiments, the storage medium stores computer executable instructions, and when the computer executable instructions are called and executed by the processor, the computer executable instructions enable the processor to implement the above-mentioned work content recommendation method or intent recognition model training method.

[0028] The above general description and the following description are exemplary and explanatory only and are not intended to limit the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] One or more embodiments are exemplarily described by corresponding drawings, and these exemplary descriptions and drawings do not constitute limitations on the embodiments. Elements with the same reference numerals in the drawings are considered similar elements, and the drawings do not constitute a scale limitation, and wherein:

[0030] Figure 1 is a schematic diagram of a work content recommendation method provided in an embodiment of the present application;

[0031] Figure 2 is a schematic diagram of a training method for an intent recognition model provided in an embodiment of the present application;

[0032] Figure 3 is a schematic diagram of a work content recommendation device provided in an embodiment of the present application;

[0033] Figure 4 is a schematic diagram of a training device for an intent recognition model provided in an embodiment of the present application;

[0034] Figure 5 It is a schematic diagram of an electronic device provided in an embodiment of the present application.

[0035] Reference numerals:

[0036] 1: First acquisition module; 2: Intent recognition module; 3: Work content determination module; 4: Recommendation model; 5: Second acquisition module; 6: Training module; 7: Memory; 8: Communication interface; 9: Bus; 10: Processor. DETAILED DESCRIPTION

[0037] In order to be able to understand the features and technical contents of the embodiments of the present application in more detail, the implementation of the embodiments of the present application is described in detail below in conjunction with the accompanying drawings. The attached drawings are for reference only and are not used to limit the embodiments of the present application. In the following technical description, for the convenience of explanation, a full understanding of the disclosed embodiments is provided through multiple details. However, one or more embodiments can still be implemented without these details. In other cases, to simplify the drawings, well-known structures and devices can be simplified for display.

[0038] The terms "first", "second", etc. in the specification and claims of the embodiments of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the embodiments of the present application described here. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions.

[0039] Unless otherwise stated, the term "plurality" means two or more.

[0040] The term "correspondence" may refer to an association relationship or a binding relationship. The correspondence between A and B means that there is an association relationship or a binding relationship between A and B.

[0041] Embodiment 1

[0042] This application embodiment provides a method for recommending work content. Figure 1 As shown, Figure 1 The flowchart of the work content recommendation method provided in the embodiment of the present application includes:

[0043] Step S101, obtaining the conversation content, the user portrait corresponding to the user who generated the conversation, and the historical conversation data corresponding to the user.

[0044] Among them, the content of the conversation includes the user's first work demand.

[0045] For example, the first work requirement is what the user needs to do at work, for example, the conversation content includes "I need to analyze last month's sales data" input by the user, where "need to analyze last month's sales data" is the user's first work requirement.

[0046] Among them, the user portrait includes the user's basic information. The user's basic information includes: the user's job information. In addition, the user's basic information may also include: gender, age, geographic location, educational background, etc. It can be understood that the user portrait may also include: user behavior data, user preferences, etc. Among them, behavior data, such as: browsing history, purchase history, click behavior, interaction history, etc. Preferences, such as: content preference, time preference, style preference, brand preference, communication method preference, etc.

[0047] It can be understood that job information, such as project manager, salesperson, marketing specialist, etc. Different job information may affect the user's demand for work content. Gender can be used to understand the preference differences that may be caused by gender differences. Age helps to analyze the user behavior patterns of different age groups. Users are located in different geographical locations, and factors such as regional culture and economic conditions brought by geographical locations may also affect user behavior and preferences. The user's education level helps to analyze the user's professional ability and knowledge level. Browsing history, that is, the user's browsing history on the system or network, such as: visited websites, types of articles read, etc. Click behavior, that is, the distribution of user clicks in the interface. For example, the buttons that users often click, the page dwell time, etc., can be used to reflect user interests and concerns. Interaction record, that is, the interaction history with the system or others. For example, comments, likes, sharing, etc., can reflect the user's social preferences and activity. Content preference, that is, the type of content that users like to read or watch, such as technology news, financial analysis, healthy living guides, etc. Time preference, that is, the time period when users are active, for example, users prefer to read news in the morning and participate in online courses in the evening. Style preference, that is, the content presentation style preferred by users, such as: simple, professional, humorous, etc. Brand preference, that is, the user's favorability towards different brands, which may affect their consumption choices. Communication method preference, that is, the communication channel preferred by users, such as: email, instant messaging or telephone communication, etc.

[0048] In some embodiments, a user profile can be formed by collecting and analyzing user data such as basic information, behavioral data, preferences, etc. of the user.

[0049] Exemplarily, the user data is preprocessed, and the preprocessed user data is calculated using a clustering algorithm and / or a decision tree algorithm to obtain a user portrait.

[0050] As another example, the user data is preprocessed, the preprocessed user data is screened again according to specific categories, and the preprocessed and screened user data is calculated using a clustering algorithm and / or a decision tree algorithm to obtain a user portrait.

[0051] Among them, the user data is filtered according to specific categories, for example, filtering user data belonging to age, gender, geographic location, and preference. It is understood that specific categories can be set by engineers according to needs.

[0052] Among them, preprocessing includes: cleaning, normalization, etc.

[0053] The clustering algorithm may be a K-means clustering algorithm (K-means clustering algorithm), a hierarchical clustering algorithm, a DBSCAN (density-based clustering algorithm), or the like.

[0054] It can be understood that the K-means clustering algorithm is an unsupervised learning method that forms clusters by iteratively assigning data points to the nearest center point. Exemplarily, the working process of the K-means clustering algorithm can be: (1) Input the number of clusters K. (2) Initialize the cluster center K. (3) Assign each data object to the class with the closest distance. (4) Recalculate the center of each class. (5) Determine whether it converges. If not, execute (4). If converged, execute (6). (6) Output the clustering result.

[0055] A decision tree is a supervised learning method that simulates a decision path by creating a tree, where each node represents a feature and each branch represents a decision rule. For example, the working process of a decision tree can be: input the training data set into the decision tree model to obtain the prediction result of the decision tree model.

[0056] Exemplarily, the user data after preprocessing and feature screening is clustered by the K-means clustering algorithm. After the clustering algorithm, users will be divided into different groups. Users within each group have similar behavior patterns or characteristics, which will help to provide more personalized and accurate services to users in different groups in the future. After clustering, the decision tree divides the features layer by layer by constructing a tree structure model. Each internal node represents a test on a feature, each branch represents a test result, and the final leaf node represents a category or output. The decision tree helps to extract more accurate user features. These features can more carefully characterize the user's behavioral habits and preferences, and provide more refined input information for subsequent personalized recommendations or content generation. It can be understood that the features in this paragraph are the feature vectors corresponding to the user data.

[0057] In some embodiments, the user profile can be constructed by collecting user data of multiple users, but the user data of the user who generated the conversation must be included. The user profile can also be constructed using only the user data of the user who generated the conversation.

[0058] In some embodiments, the user portrait may also be updated based on the conversation content.

[0059] In some embodiments, when the user is using the platform or system with the job content recommendation method of the present application for the first time, the historical conversation data may be empty, that is, there is no historical conversation data. And the conversation content at this time may only be the sentence text input by the user containing the first job requirement. If the user has had multiple conversations, the conversation content may include the user's questions and the system's responses.

[0060] It can be understood that the user's conversation content can be saved to a preset storage location. When the user uses it next time, the conversation content saved in the above storage location is all historical conversation data.

[0061] Exemplarily, the historical conversation data may include the previous interaction records between the user and the system. The interaction records may include: the user's questions, the system's responses, and the user's feedback on the responses.

[0062] Step S102: identifying the first work requirement by using a preset intention recognition model according to the user portrait, historical conversation data and the first work requirement to obtain the work intention.

[0063] In some embodiments, step S102 may include: extracting first key information from the user portrait; extracting second key information from historical conversation data; and obtaining work intention using an intention recognition model based on job information, conversation topic, second work requirement and first work requirement.

[0064] The first key information includes: the user's position information. The second key information includes: the conversation topic and the second job requirement.

[0065] It can be understood that the first key information includes not only job information. The first key information can also include other categories that constitute the user portrait, such as age, gender, geographic location, preferences, etc. The second key information can be used to represent data points that can reveal the user's true intentions, clarify needs, or indicate the direction of the conversation. Exemplarily, the second key information includes not only the conversation topic and the second work requirement. The second key information can also include emotional tendencies, specific task requests, etc. Among them, regarding the conversation topic: For example, in a conversation about project management, the topic may be "project progress tracking". This conversation topic shows that the user is currently concerned about how to effectively monitor and report the progress of each stage of the project. Regarding the second user demand: Assuming that the user expressed in the conversation "I need an automated tool to help me organize the weekly team work report", then the second user demand is specifically that the automated tool assists in the preparation of the weekly report. The second user demand reflects that the user wants to improve work efficiency and reduce the time consumption of manually organizing reports, and also expects the report content to fully and accurately reflect the progress of the team's work.

[0066] In one implementation of the above embodiment, the intent recognition model is obtained in the following manner: obtaining sample data corresponding to multiple users respectively; the sample data includes: sample user portraits, sample historical conversation data and sample work requirements; using the sample data with work intention labels to train the first preset model to obtain the intent recognition model.

[0067] In the above implementation, the first preset model may be a natural language model. Accordingly, the first preset model is trained using sample data with work intention labels to obtain the intention recognition model by inputting sample data with work intention labels into the natural language model for training to obtain the intention recognition model.

[0068] Among them, the natural language model can be a convolutional neural network model, a recurrent neural network, a Transformer model (a deep learning model architecture for natural language processing tasks), a BERT (Bidirectional Encoder Representations from Transformers, a pre-trained language model based on a bidirectional Transformer encoder) model, etc., without limitation here.

[0069] Exemplarily, the sample data with work intention labels are input into a natural language model for training, and the intent recognition model is obtained by: for each piece of sample data with work intention labels: extract the first key information in the sample user portrait, extract the second key information in the sample historical conversation data, integrate the first key information and the second key information, and obtain integrated information. The integrated information with work intention labels is input into a preset natural language model for training to obtain an intent recognition model.

[0070] The first key information and the second key information are integrated by converting the first key information into a first feature vector, converting the second key information into a second feature vector, and adding the first feature vector to the second feature vector.

[0071] Exemplarily, the natural language model used in this application is a BERT model with an additional classification layer. The classification layer usually contains a softmax (normalized exponential function) function, which is used to convert the output of the BERT model without a classification layer into the probability distribution of each intent category. The processing process of this natural language model is as follows: the BERT model without a classification layer converts the input text into an output vector, receives the output vector through the classification layer, and maps these output vectors to the output space required for the task through linear transformation and nonlinear activation function, and then calculates the probability of each intent category through the softmax function, thereby predicting the intent category of the input text. Among them, linear transformation, for example: weight matrix multiplication and bias addition. Non-linear activation function, for example: ReLU (Rectified Linear Unit, linear rectification function).

[0072] In the above implementation, the first preset model may be a preset third language model. Accordingly, the first preset model is trained using sample data with work intention labels to obtain an intention recognition model by: using a preset training template, assembling sample user portraits, sample historical conversation data, and sample work requirements into the training template, and inputting the assembled training template into the third language model for training to obtain an intention recognition model.

[0073] It can be understood that the training template is a preset text, and the specific positions in the text need to be filled with work intention, sample user portrait, sample historical conversation data and sample work requirements. For example, in the case of a training template where the portrait is "sample user portrait", the historical conversation is "sample historical conversation data", and the work requirement is "sample work requirement", the user's intention is "work intention".

[0074] For example, for each piece of sample data with a work intention label: extract the first key information in the sample user portrait, and extract the second key information in the sample historical conversation data. Assemble the first key information, the second key information, the sample work requirements, and the work intention into a training template, and input the assembled training template into a large language model for training to obtain an intent recognition model.

[0075] In an optional manner of the above embodiment, the intention recognition model can be obtained by training a natural speech model. Accordingly, the first work requirement is identified using a preset intention recognition model according to the user portrait, historical conversation data, and the first work requirement, and the work intention can be obtained by: extracting the first key information from the user portrait. Extracting the second key information from the historical conversation data. Inputting the first key information, the second key information, and the first work requirement into the intention recognition model to obtain the work intention.

[0076] Exemplarily, the intention recognition model is obtained by training a natural speech model. The first key information is the job information, and the second key information is the conversation topic and the second work requirement. Accordingly, the intention recognition model is used according to the job information, the conversation topic, the second work requirement and the first work requirement to obtain the work intention, which can be: the job information, the conversation topic, the second work requirement and the first work requirement are input into the intention recognition model to obtain the work intention.

[0077] In another optional manner of the above embodiment, the intent recognition model can be obtained by training a large language model. Accordingly, according to the user portrait, historical conversation data and the first work requirement, the first work requirement is identified using a preset intent recognition model to obtain the work intention, which can be: extracting the first key information from the user portrait. Extracting the second key information from the historical conversation data. Assembling the first key information, the second key information, the first work requirement and the preset third prompt template to obtain the third text to be input; inputting the third text to be input into the intent recognition model to obtain the work intention output by the intent recognition model. Among them, the third prompt template is used to describe the work intention of the user judged in combination with the first key information, the second key information and the first work requirement.

[0078] Exemplarily, the intent recognition model is obtained by training a large speech model. The first key information is the job information, and the second key information is the conversation topic and the second work requirement. Accordingly, the work intent can be obtained by using the intent recognition model according to the job information, the conversation topic, the second work requirement and the first work requirement: assembling the job information, the conversation topic, the second work requirement, the first work requirement and the preset third prompt template to obtain the third text to be input; inputting the third text to be input into the intent recognition model to obtain the work intent output by the intent recognition model.

[0079] For example, a Soft Prompt template can be designed to prompt the user's input. Soft Prompt template, for example: Hello, I am your office assistant. Based on your position [position name], I understand that you may need the following types of help: For [position-related tasks], you may need [specific tasks], for example: "I need to arrange [specific tasks]." For cross-departmental collaboration, you may need to perform [specific tasks] with [collaborating department], for example: "I need to collaborate with [collaborating department] on [specific tasks]." Please provide more detailed information based on your current needs so that I can provide you with accurate support.

[0080] Assume that the user input is, for example: Employee A in the marketing department: I need to schedule a meeting. After the intention recognition model recognizes, the output is: Employee A's work intention: plan an event.

[0081] Assume that the user input is, for example: Employee B of the Finance Department: I need to schedule a meeting. After the intent recognition model recognizes, the output is: Employee B's work intention: budget review.

[0082] It can be understood that [Position Name]: the specific position of the employee, for example: Marketing Manager, Financial Analyst, etc. [Position-related Tasks]: Common tasks related to the position, for example: Event Planning of the Marketing Department, Budget Review of the Finance Department, etc. [Collaborative Department]: The names of other departments within the company, which may be different from the department of the requesting employee. [Specific Task]: The specific task description proposed by the employee based on the job requirements. In this way, by integrating the employee's user profile into the prompt, employees can be guided to use similar request language, and the large language model can recognize different intentions according to their different positions. The SoftPrompt template allows prompts to be dynamically adjusted according to different job data, so that the model can more accurately understand the specific needs of employees in different positions.

[0083] Step S103, determining the work content according to the work intention.

[0084] Among them, the work content is information related to the user's position and office.

[0085] Exemplarily, the work content can be the following aspects: text reports, task lists, schedules, data analysis reports, email drafts, decision support materials, training materials, etc. Among them, text reports, for example: according to the monthly summary requirements of the sales department, the system can automatically generate detailed reports containing sales performance analysis, customer feedback summary, market trend forecasts, etc. Task list, for example: based on the project manager's project plan, the system can generate a task list containing task assignments, deadlines, and priority rankings to ensure that team members are clear about their respective responsibilities and timelines. Schedule, for example: based on employees' work habits and upcoming meetings, business trips and other activities, the system can automatically generate personal schedules, including meeting reminders, time arrangements, preparations, etc. Data analysis reports, for example: for the finance department, the system can analyze historical financial data and automatically generate charts and tables containing key indicators such as income and expenditure comparison, cost-benefit analysis, and budget execution. Email draft, for example: according to user instructions, if you need to send an email about project progress report, the system can generate an email draft containing the latest progress, problem overview, and next action plan. Decision-making support materials, for example: for senior managers, the system can integrate market research, competition analysis, SWOT (SWOT analytical method) analysis and other information to generate decision-making support reporting materials, helping to quickly understand the company's operating conditions and external environment. Training materials, for example: for new employees or specific skills training needs, the system can automatically generate training plans and materials including course outlines, learning resource links, and assessment standards.

[0086] In some embodiments, step S103 may include: searching a preset content recommendation database for work content corresponding to the work intention. The preset content recommendation database stores the correspondence between the work intention and the work content.

[0087] In some embodiments, step S103 may include: determining the work ability corresponding to the work intention; assembling the user portrait, work ability, work intention and a preset first prompt template to obtain a first text to be input; the first prompt template includes a sentence that recommends work content using a collaborative filtering algorithm and / or a content-based recommendation algorithm; inputting the first text to be input into a preset first language model to obtain the work content output by the first language model.

[0088] It can be understood that the first prompt template can be used to describe the recommendation of work content using a collaborative filtering algorithm and / or a content-based recommendation algorithm in combination with user portrait, work ability, and work intention.

[0089] In an optional manner of the above embodiment, determining the work capability corresponding to the work intention includes: searching for the work capability corresponding to the work intention in a preset knowledge graph. The knowledge graph stores the correspondence between the work intention and the work capability.

[0090] In another optional manner of the above embodiment, determining the work capability corresponding to the work intention includes: filling the work intention into a preset fourth prompt template to obtain a fourth text to be inputted. Inputting the fourth text to be inputted into a preset fourth language model to obtain the work capability outputted by the fourth language model.

[0091] Among them, the fourth prompt template is used to describe the recommended work ability in combination with the work intention.

[0092] It can be understood that work capability is a set of skills or permissions that a user needs to complete a specific work task. For example, work capability can be one or more of software operation capability, professional skills, collaboration capability, and any other specific capabilities related to job responsibilities. Among them, software operation capability, for example, using Excel (spreadsheet) to perform complex data pivoting. Professional skills, for example, financial analysis. Collaboration capability, for example, cross-departmental communication and coordination.

[0093] For example, if the intent identifies that the user needs to prepare a project report, then the work capabilities may involve document writing, PPT (slide) production, data visualization, etc. These work capabilities will be used to automatically generate or recommend work content that meets the reporting requirements, such as project summary reports, presentations, etc.

[0094] It is worth noting that engineers can change the content of the above prompt templates according to their own needs.

[0095] Step S104: recommending work content.

[0096] In some embodiments, step S104 includes: determining whether the work content is compliant; if the work content is compliant, recommending the work content; if the work content is not compliant, replacing or deleting the non-compliant work content, and recommending the replaced or deleted work content.

[0097] In an optional manner of the above embodiment, determining whether the work content is compliant may include: segmenting the work content to obtain each word, comparing each word with a preset sensitive word database, and determining that the work content is not compliant when there is a word recorded in the sensitive word database.

[0098] Accordingly, replacing or deleting non-compliant work content may be: determining the sensitivity level of the illegal words, and modifying the work content according to the operation method corresponding to the sensitivity level. The operation method includes: replacing or deleting. The sensitive word database records sensitive words and the sensitivity levels corresponding to the sensitive words.

[0099] Exemplarily, modifying the work content according to the operation mode corresponding to the sensitivity level may be: when the sensitivity level of the offending word is a first preset level, deleting the word. When the sensitivity level of the offending word is a second preset level, replacing the word with a preset word. The first preset level is higher than the second preset level.

[0100] In the above example, corresponding preset words can be established for each sensitive word in advance, and the corresponding relationship between the sensitive word and the preset word can be stored in a preset replacement database. The preset word can be an approximate word, a synonym or a hypernym of the word to be replaced.

[0101] In another optional manner of the above embodiment, determining whether the work content is compliant may be: inputting the work content into a preset violation judgment model to obtain a violation result output by the violation judgment model. The violation result includes compliance or non-compliance.

[0102] In some embodiments, after recommending the work content, it also includes: extracting third key information from the conversation content and the work content; the third key information includes: work requirements, professional terms, emotions, three elements of events, and potential needs; assembling the key information with a preset second prompt template to obtain a second text to be input; inputting the second text to be input into a preset second largest language model to obtain a recommended question output by the second largest language model.

[0103] The second prompt template is used to describe the problem of recommending a preset number in combination with the third key information.

[0104] In some embodiments, after recommending work content, it also includes: obtaining a first score of the user's feedback on the work intention; optimizing the intention recognition model based on the first score; and / or obtaining a second score of the user's feedback on the work content; optimizing the first language model based on the second score.

[0105] Exemplarily, the range or type of the first score can be preset, and the range or type of the second score can also be preset. For example, the first score can be very poor, bad, average, good, and very good, and the intent recognition model is optimized by the first score selected by the user. The second score can be bad and good, and the first language model is optimized by the second score selected by the user.

[0106] In the above embodiment, optimizing the intent recognition model according to the first score includes: using a preset reinforcement learning or online learning algorithm to optimize the intent recognition model using the first score.

[0107] In the above embodiment, optimizing the first large language model according to the second score includes: optimizing the first large language model using the second score using a preset reinforcement learning or online learning algorithm.

[0108] It can be understood that the specific scheme of adjusting the recommendation strategy through reinforcement learning mainly includes the following core links: state definition, action definition, reward design, strategy update, exploration and utilization. In the context of recommendation systems, state can be defined as the user's historical behavior, user characteristics, characteristics of the current recommended content, and other relevant information of the environment. For example, the user's past click-through rate, purchase history, current browsing page, etc. Action refers to the choice made by the recommendation system, that is, deciding what content to recommend to the user. This can be a specific item ID (Identity document), content category, or change in content ranking. Rewards are formulated based on user feedback on the recommended content and directly reflect the quality of the recommendation strategy. Common reward signals include positive feedback such as user clicks, purchases, likes, and shares, as well as negative feedback such as ignores and negative comments. Designing a reasonable reward function is the key to the success of reinforcement learning. For example, high rewards can be given for clicks or purchases, while low or negative scores can be given for ignored recommendations. Using reinforcement learning algorithms such as policy gradient methods or Q-learning, the recommendation strategy is updated based on the rewards received after each recommendation. This means that the model will learn what actions to take in a specific state to maximize the long-term cumulative rewards. During the learning process, the algorithm needs to balance exploring unknown recommendation strategies and using known best strategies. Common methods include the ε-greedy strategy, which randomly selects actions for exploration with a certain probability, and selects the action currently considered to be the best for the rest of the time.

[0109] The specific scheme for adjusting the recommendation strategy through online learning mainly includes the following core links: continuous model update, incremental learning, feedback loop, and model fusion. Continuous model update means that whenever there is new user feedback, such as clicks and ratings, the model updates its parameters immediately or regularly to reflect the latest user preferences. Through incremental learning, only the newly added data is trained instead of retraining the entire model, which greatly improves the efficiency of model update and reduces the consumption of computing resources. Feedback loop means establishing a fast feedback mechanism so that the model can respond quickly to changes in user behavior. For example, by monitoring user activities in real time, the effect of the recommendation strategy can be quickly verified and the model parameters can be fine-tuned accordingly. Model fusion means that in some cases, a combination of online learning and offline learning can be adopted, that is, a stable offline model is maintained as the basis, while the online learning model is used to capture instant changes, and the results of the two are fused to achieve a stable and flexible effect.

[0110] For example, user A is a project manager of a company and often needs to write daily, weekly and monthly project reports. In the interaction with the intelligent assistant, the system collects data about user A, including his position, work content, writing habits and historical conversation records. Through analysis, the system finds that user A often asks how to quickly write daily reports on Friday afternoons and has certain needs for data analysis and chart display. In one conversation, user A said: "Today is Friday again, and I have to write a daily report." The intelligent assistant analyzes the user's position data, the user's historical conversation data and the current conversation content, and identifies that user A's real work intention is to get help to quickly complete the writing of the daily report. The intelligent assistant searches for the capabilities corresponding to the work intention and finds that "daily report writing ability", "data analysis ability" and "chart making ability" are required. The intelligent assistant further analyzes the user portrait data of user A and finds that user A has "daily report writing ability" but lacks "data analysis ability" and "chart making ability", so the system recommends that user A improve these two abilities. Based on the recommended abilities, the intelligent assistant provides the following services to user A: Improvement of data analysis ability: The intelligent assistant provides a series of data analysis tutorials and tools to help user A learn how to extract key information from project data. Improved chart making ability: The smart assistant recommends some chart making software and provides a quick start guide so that user A can easily make professional charts. Daily report writing assistance: Combining user A's daily report writing ability and newly learned data analysis and chart making skills, the smart assistant provides a daily report template, which includes a data analysis section and a chart display area. User A only needs to fill in the corresponding data and text description, and the smart assistant can automatically generate a daily report.

[0111] Embodiment 2

[0112] The present application embodiment provides a method for training an intent recognition model. Figure 2 As shown, Figure 2 A flowchart of a training method for an intent recognition model provided in an embodiment of the present application includes:

[0113] Step S201, obtaining sample data corresponding to a plurality of users respectively.

[0114] Among them, the sample data includes: sample user portraits, sample historical conversation data and sample work requirements.

[0115] Step S202: training the first preset model using sample data with work intention labels to obtain an intention recognition model.

[0116] In some embodiments, after obtaining the intent recognition model, it also includes: obtaining the conversation content, the user portrait corresponding to the user who generated the conversation, and the historical conversation data corresponding to the user; the conversation content includes the user's first work requirement; the user's position information exists in the user portrait; according to the user portrait, the historical conversation data and the first work requirement, the first work requirement is identified using a preset intent recognition model to obtain the work intention; obtaining the user's first score for feedback on the work intention; and optimizing the intent recognition model based on the first score.

[0117] It can be understood that the various embodiments described in the first embodiment are also applicable to the second embodiment if there is no conflict. Based on the consideration of brevity of description, they will not be repeated here.

[0118] Embodiment 3

[0119] Based on the same inventive concept, the present application embodiment provides a work content recommendation device, such as Figure 3 As shown, the work content recommendation device includes: a first acquisition module 1, an intention recognition module 2, a work content determination module 3 and a recommendation model 4. Among them, the first acquisition module is used to obtain the conversation content, the user portrait corresponding to the user who generates the conversation and the historical conversation data corresponding to the user; the conversation content includes the user's first work requirement; the user portrait contains the user's position information; the intention recognition module is used to identify the first work requirement according to the user portrait, the historical conversation data and the first work requirement using a preset intention recognition model to obtain the work intention; the work content determination module is used to determine the work content according to the work intention; the work content is the information related to the user's position and office; the recommendation model is used to recommend the work content.

[0120] In some embodiments, the work content recommendation device also includes: a model training module, which is used to obtain an intent recognition model in the following manner: obtaining sample data corresponding to multiple users respectively; the sample data includes: sample user portraits, sample historical conversation data and sample work requirements; using sample data with work intention labels to train a first preset model to obtain an intent recognition model.

[0121] In some embodiments, the intention recognition module 2 is used to identify the first work requirement and obtain the work intention by using a preset intention recognition model based on the user portrait, historical conversation data and the first work requirement in the following manner: extracting the first key information from the user portrait; the first key information includes: the user's job information; extracting the second key information from the historical conversation data; the second key information includes: the conversation topic and the second work requirement; obtaining the work intention by using the intention recognition model based on the job information, conversation topic, the second work requirement and the first work requirement.

[0122] In some embodiments, the work content determination module 3 is used to determine the work content according to the work intention in the following manner: determine the work ability corresponding to the work intention; assemble the user portrait, work ability, work intention and a preset first prompt template to obtain a first text to be input; the first prompt template includes a sentence that recommends the work content using a collaborative filtering algorithm and / or a content-based recommendation algorithm; input the first text to be input into a preset first language model to obtain the work content output by the first language model.

[0123] In some embodiments, the recommendation model 4 is used to recommend work content in the following manner: determining whether the work content is compliant; if the work content is compliant, recommending the work content; if the work content is not compliant, replacing or deleting the non-compliant work content, and recommending the replaced or deleted work content.

[0124] In some embodiments, the work content recommendation device also includes: a question recommendation module, which is used to extract third key information from the conversation content and the work content after recommending the work content; the third key information includes: work requirements, professional terms, emotions, three elements of events, and potential needs; assembling the key information with a preset second prompt template to obtain a second text to be input; inputting the second text to be input into a preset second largest language model to obtain a recommended question output by the second largest language model.

[0125] In some embodiments, the work content recommendation device also includes: a first model optimization module, which is used to obtain a first score of the user's feedback on the work intention after recommending the work content; optimize the intention recognition model based on the first score; and / or obtain a second score of the user's feedback on the work content; and optimize the first language model based on the second score.

[0126] It can be understood that the various embodiments described in the first embodiment are also applicable to the third embodiment if there is no conflict. Based on the consideration of brevity of description, they will not be repeated here.

[0127] Embodiment 4

[0128] Based on the same inventive concept, the present application embodiment provides a training device for an intent recognition model, such as Figure 4 As shown, the training device of the intention recognition model includes: a second acquisition module 5 and a training module 6. The second acquisition module 5 is used to obtain sample data corresponding to multiple users respectively; the sample data includes: sample user portraits, sample historical conversation data and sample work requirements; the training module 6 is used to train the first preset model using the sample data with work intention labels to obtain the intention recognition model.

[0129] In some embodiments, the training device of the intent recognition model also includes: a second model optimization module, which is used to obtain the conversation content, the user portrait corresponding to the user who generated the conversation, and the historical conversation data corresponding to the user after obtaining the intent recognition model; the conversation content includes the user's first work requirement; the user's job information exists in the user portrait; according to the user portrait, historical conversation data and the first work requirement, the first work requirement is identified using a preset intent recognition model to obtain the work intention; obtain the user's first score for feedback on the work intention; and optimize the intention recognition model based on the first score.

[0130] It can be understood that the various embodiments described in the first embodiment are also applicable to the fourth embodiment if there is no conflict. Based on the consideration of brevity of description, they will not be repeated here.

[0131] Combination Figure 5 As shown, an embodiment of the present application provides an electronic device, including a processor 10 and a memory 7. Optionally, the device may also include a communication interface 8 and a bus 9. Among them, the processor 10, the communication interface 8, and the memory 7 can communicate with each other through the bus 9. The communication interface 8 can be used for information transmission. The processor 10 can call the logic instructions in the memory 7 to execute the above-mentioned work content recommendation method or the training method of the intent recognition model.

[0132] In addition, the logic instructions in the above-mentioned memory 7 can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product.

[0133] The memory 7 is a computer-readable storage medium that can be used to store software programs and computer executable programs, such as program instructions / modules corresponding to the method in the embodiment of the present application. The processor 10 executes the function application and data processing by running the program instructions / modules stored in the memory 7, that is, the above-mentioned work content recommendation method or the training method of the intention recognition model is realized.

[0134] The memory 7 may include a program storage area and a data storage area, wherein the program storage area may store an operating system and an application required for at least one function; the data storage area may store data created according to the use of the terminal device, etc. In addition, the memory 7 may include a high-speed random access memory and may also include a non-volatile memory.

[0135] The electronic device may be a computer or a server, etc.

[0136] An embodiment of the present application provides a storage medium storing computer executable instructions, wherein the computer executable instructions are configured to execute the above-mentioned work content recommendation method or the intention recognition model training method.

[0137] An embodiment of the present application provides a computer program product, which includes a computer program stored on a storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the above-mentioned work content recommendation method or the intention recognition model training method.

[0138] The computer-readable storage medium mentioned above may be a transient computer-readable storage medium or a non-transitory computer-readable storage medium.

[0139] The technical solution of the embodiment of the present application can be embodied in the form of a software product, which is stored in a storage medium and includes one or more instructions for a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiment of the present application. The aforementioned storage medium may be a non-transient storage medium, including: a USB flash drive, a mobile hard disk, a read-only memory, a random access memory, a magnetic disk or an optical disk, and other media that can store program codes, or a transient storage medium.

[0140] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed.

[0141] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application. At the same time, the above embodiments can be combined with each other to form new embodiments without conflict.

Claims

1. A method for recommending work content, characterized in that: include: Obtaining the conversation content, the user portrait corresponding to the user who generated the conversation, and the historical conversation data corresponding to the user; The conversation content includes the first work requirement of the user; The user portrait contains the user's job information; the first job requirement is what the user needs to do at work; Identify the first work requirement using a preset intention recognition model according to the user portrait, the historical conversation data and the first work requirement to obtain a work intention; Determine the work content according to the work intention; the work content is information related to the user's position and office; Recommend the content of the work described; Determining the work content according to the work intention includes: determining the work ability corresponding to the work intention; assembling the user portrait, the work ability, the work intention and a preset first prompt template to obtain a first text to be input; the first prompt template includes a sentence that recommends the work content using a collaborative filtering algorithm and / or a content-based recommendation algorithm; inputting the first text to be input into a preset first language model to obtain the work content output by the first language model; the work ability is a set of skills necessary for the user to complete a specific work task.

2. The method according to claim 1, characterized in that The intent recognition model is obtained in the following way: Obtain sample data corresponding to multiple users respectively; the sample data includes: sample user portraits, sample historical conversation data and sample work requirements; The first preset model is trained using the sample data with work intention labels to obtain an intention recognition model.

3. The method according to claim 1, characterized in that: The first work requirement is identified by using a preset intention recognition model according to the user portrait, the historical conversation data and the first work requirement to obtain a work intention, including: Extracting first key information from the user portrait; the first key information includes: user position information; Extracting second key information from the historical conversation data; the second key information includes: the conversation topic and the second work requirement; The work intention is obtained by using the intention recognition model according to the position information, the conversation topic, the second work requirement and the first work requirement.

4. The method according to claim 1, characterized in that The recommended work content includes: Determine whether the work content is in compliance with regulations; If the work content is in compliance with regulations, recommend the work content; In the case where the work content is not compliant, the non-compliant work content is replaced or deleted, and the work content after replacement or deletion is recommended.

5. The method according to any one of claims 1 to 4, characterized in that: After recommending the work content, the method further includes: Extracting third key information from the conversation content and the work content; the third key information includes: work requirements, professional terms, emotions, three elements of events, and potential requirements; Assembling the third key information with a preset second prompt template to obtain a second text to be input; The second text to be input is input into a preset second largest language model to obtain a recommendation question output by the second largest language model.

6. The method according to claim 1, characterized in that After recommending the work content, the method further includes: Obtaining a first score of the user's feedback on the work intention; optimizing the intention recognition model according to the first score; and / or, Obtain a second score of the user's feedback on the work content; and optimize the first language model according to the second score.

7. A work content recommendation device, characterized in that: include: A first acquisition module is used to acquire the conversation content, the user portrait corresponding to the user who generated the conversation, and the historical conversation data corresponding to the user; The conversation content includes the first work requirement of the user; the user portrait contains the user's position information; the first work requirement is what the user needs to do at work; An intention recognition module, configured to recognize the first work requirement by using a preset intention recognition model according to the user portrait, the historical conversation data and the first work requirement, and obtain a work intention; A work content determination module, used to determine the work content according to the work intention; the work content is information related to the user's position and office; A recommendation model for recommending the work content; The work content determination module is specifically used to determine the work ability corresponding to the work intention; assemble the user portrait, the work ability, the work intention and a preset first prompt template to obtain a first text to be input; The first prompt template includes a sentence that recommends work content using a collaborative filtering algorithm and / or a content-based recommendation algorithm; the first input text is input into a preset first language model to obtain the work content output by the first language model; the work capability is a set of skills that a user must have to complete a specific work task.

8. An electronic device, characterized in that: The invention comprises a processor and a memory, wherein the memory stores computer executable instructions that can be executed by the processor, and the processor executes the computer executable instructions to implement the work content recommendation method according to any one of claims 1 to 6.

9. A storage medium, characterized in that: The storage medium stores computer-executable instructions, and when the computer-executable instructions are called and executed by the processor, the computer-executable instructions enable the processor to implement the work content recommendation method described in any one of claims 1 to 6.

Images