CN117407595B - Home decoration designer recommendation method integrating large language model and dynamic dialogue intention - Google Patents
Home decoration designer recommendation method integrating large language model and dynamic dialogue intention Download PDFInfo
- Publication number
- CN117407595B CN117407595B CN202311716405.0A CN202311716405A CN117407595B CN 117407595 B CN117407595 B CN 117407595B CN 202311716405 A CN202311716405 A CN 202311716405A CN 117407595 B CN117407595 B CN 117407595B
- Authority
- CN
- China
- Prior art keywords
- designer
- dialogue
- user
- history
- recommendation
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 33
- 238000005034 decoration Methods 0.000 title claims abstract description 22
- 238000012549 training Methods 0.000 claims abstract description 42
- 238000013461 design Methods 0.000 claims description 13
- 230000008569 process Effects 0.000 claims description 9
- 230000004044 response Effects 0.000 claims description 9
- 238000006243 chemical reaction Methods 0.000 claims description 8
- 238000011156 evaluation Methods 0.000 claims description 7
- 230000006872 improvement Effects 0.000 claims description 4
- 238000010606 normalization Methods 0.000 claims description 3
- 238000012795 verification Methods 0.000 claims description 3
- 238000010586 diagram Methods 0.000 description 7
- 230000008901 benefit Effects 0.000 description 3
- 230000007613 environmental effect Effects 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 238000012545 processing Methods 0.000 description 2
- 238000007670 refining Methods 0.000 description 2
- 238000013459 approach Methods 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 238000009966 trimming Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/27—Regression, e.g. linear or logistic regression
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
- G06F40/35—Discourse or dialogue representation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/02—Knowledge representation; Symbolic representation
- G06N5/022—Knowledge engineering; Knowledge acquisition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/16—Real estate
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F2216/00—Indexing scheme relating to additional aspects of information retrieval not explicitly covered by G06F16/00 and subgroups
- G06F2216/03—Data mining
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Business, Economics & Management (AREA)
- Databases & Information Systems (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- General Health & Medical Sciences (AREA)
- Tourism & Hospitality (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computing Systems (AREA)
- Mathematical Physics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Life Sciences & Earth Sciences (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a home decoration designer recommendation method integrating a large language model and dynamic dialogue intentions, which comprises the steps of firstly constructing user intention training corpus and designer history training corpus, further generating a user intention set and a designer personal history set, then generating a user history dialogue graph based on the user intention set and the designer personal history set, converting the user history dialogue graph into a natural language context for the large language model, then randomly combining customer intentions in the natural language context, normalizing the designer histories, finally training a MOSS dialogue recommendation system, and recommending by the designer through the trained MOSS dialogue recommendation system.
Description
Technical Field
The invention relates to the technical field of data processing, in particular to a home decoration designer recommendation method integrating a large language model and dynamic dialogue intention.
Background
The recommendation system is widely focused and studied as one of important tools for acquiring information in the big data age. In the home decoration industry, in order to improve customer conversion rate and business income, home decoration enterprises recommend proper designers to potential decoration customers according to the intention of users and histories of the designers. As the demand for efficient, personalized indoor design styles increases by customers, it becomes increasingly important to develop efficient customer-designer recommendation systems.
In the prior art, home decoration enterprises recommend designers to customers mainly through human intervention, which is time-consuming and inefficient, although some designer recommendation systems based on big data exist. However, dynamic, accurate designer recommendations face significant challenges due to the need and intent of customer finishing that may vary during the stages of different conversations, as well as the inherent bias and impression of the designer of the home improvement business.
Disclosure of Invention
Therefore, the embodiment of the invention provides a home decoration designer recommendation method integrating a large language model and dynamic dialogue intention to realize dynamic and accurate designer recommendation.
According to the home decoration designer recommendation method integrating a large language model and dynamic dialogue intention, the method comprises the following steps:
step 1, constructing user intention training corpus according to user ID, historical dialogue text between the user and a sales consultant and historical recommendation relation between the user and a designer; constructing a designer history training corpus according to the ID of the designer, demographic information, business expertise information and historical successful recommended case information of the designer;
step 2, generating a user intention set based on the user intention training corpus, and generating a designer personal history set based on the designer history training corpus;
step 3, generating a user history dialogue graph based on a user intention set and a designer personal history set, wherein the user history dialogue graph is a heterogeneous graph, the user history dialogue graph comprises two types of node object sets and two types of edge sets, the two types of node object sets are respectively the user intention set and the designer personal history set, the two types of edge sets are respectively an edge set in a conversion process and an edge set which is successfully recommended, the edge set marked in the conversion process indicates that final recommendation is not achieved, the edge set marked as successfully recommended indicates that final recommendation is successful, and then the user history dialogue graph is converted into a natural language context for a large language model;
step 4, carrying out random combination on the client intention in the natural language context to obtain a natural language context after random combination;
step 5, based on the designer personal record set, carrying out designer record normalization to obtain a normalized designer personal record set;
and step 6, training the MOSS dialogue recommendation system by adopting the random combined natural language context and the normalized designer personal history set, and then recommending the designer through the trained MOSS dialogue recommendation system.
According to the home decoration designer recommendation method integrating the large language model and the dynamic dialogue intention, firstly, user intention training corpus is built, designer history training corpus is built, user intention set and designer personal history set are generated, then, user history dialogue diagrams are generated based on the user intention set and the designer personal history set, the user history dialogue diagrams are converted into natural language contexts for the large language model, then, the client intention in the natural language contexts is randomly combined, the designer histories are normalized, finally, a MOSS dialogue recommendation system is trained, dynamic changes of the user decoration intention are mined from the historical dialogue corpus of a user and a sales consultant based on the capability of extracting fine text features and massive Chinese knowledge base, and the dynamic changes are matched with the personal histories of the designer.
In addition, the home decoration designer recommendation method integrating the large language model and the dynamic dialogue intention according to the embodiment of the invention can also have the following additional technical characteristics:
further, in step 1, the user intent training corpus is text in txt format, and the history recommendation relationship between the user and the designer satisfies the following conditional expression:
；
wherein,representing a set of users who were successfully recommended to the designer, < >>Represents the kth user, +.>Designer indicating that there is a history recommendation relationship with the kth user, < >>Representing a set of historical users>Representing a set of designers;
the historic training corpus of the designer is text in txt format, and the historic successful recommended case information of the designer meets the following conditional expression:
；
wherein,representing a set of designers that were successfully recommended to the user, < +.>Representing the j-th designer->Representing users who have a history of recommended relationships with the j-th designer.
Further, in step 2, the designer's personal history is collected fromA plurality of pairs of special tuples, wherein a pair of special tuples is expressed asWherein->Representing i-th design expertise, +.>Representing the weight corresponding to the i-th design expertise, each designer having +.>For specialized tuples.
Further, step 4, performing random combination on the client intention in the natural language context to obtain a natural language context after random combination, which specifically includes:
acquiring a history dialogue text between a user and a sales consultant in a natural language context, and marking user intention, problems and preferences contained in the history dialogue text in a preset evaluation mode to serve as a verification sample;
randomly disturbing sentences or dialogue segments in the sample, and changing the sequence of the context in the original dialogue;
and recombining the disturbed sentences or dialogue segments into a new sentence or dialogue segment to realize random combination of client intentions, thereby obtaining the natural language context after random combination.
Further, the following conditional expression is satisfied in step 5:
；
wherein,weights representing normalized i-th design expertise, < >>Representing design wThe weight corresponding to the expertise.
Further, in step 6, when training the MOSS dialogue recommendation system by using the randomly combined natural language context and the normalized designer personal history set, performing prompt word fine adjustment and instruction fine adjustment, guiding the MOSS dialogue recommendation system to understand the randomly combined natural language context through the prompt word fine adjustment, and refining the execution of the recommendation task through the instruction fine adjustment, wherein the prompt word fine adjustment and the instruction fine adjustment share the same final target lossFinal target loss->The expression of (2) is:
；
wherein,representing the initial parameters of the package->And->Respectively representing prompt input and prompt response in the customized prompt message,/->Is training set, is->Indicates the corresponding times->Representation->Is>Secondary response->Representation->Before response +_>Is indicated at prompt input +.>And->Under the condition->Is a conditional distribution probability of (c).
Drawings
The foregoing and/or additional aspects and advantages of embodiments of the invention will be apparent and readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:
fig. 1 is a flow chart of a method for recommending home decoration designer by fusing large language model and dynamic dialogue intention according to an embodiment of the invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
Referring to fig. 1, an embodiment of the present invention provides a home decoration designer recommendation method integrating a large language model and a dynamic dialogue intention, the method includes steps 1 to 6:
step 1, constructing user intention training corpus according to user ID, historical dialogue text between the user and a sales consultant and historical recommendation relation between the user and a designer; and constructing a designer history training corpus according to the ID of the designer, the demographic information, the business expertise information and the history successful recommended case information of the designer.
In step 1, the history recommendation relationship between the user and the designer satisfies the following conditional expression:
；
wherein,representing a set of users who were successfully recommended to the designer, < >>Represents the kth user, +.>Designer indicating that there is a history recommendation relationship with the kth user, < >>Representing a set of historical users>Representing a collection of designers.
Historical user collectionThe ID of each user, the historical dialog text, and the information of the successful recommendation designer are all recorded in the database of the call center or user relationship management system of the corresponding enterprise.
The historical dialog text is presented in the form of natural language, historical dialog text for the user with the sales advisor, such as "customer: i want to decorate a house. "," sales advisor: please ask you if the house is a blank or old house "," customer: is a new house. "," sales advisor: how much area there is, several rooms, several halls, several guards "," customer: 127 square meters, 3 rooms, 2 toilets, 1 living room, 1 restaurant, and 2 balconies.
The user intent training corpus is specifically text in txt format, examples are as follows:
successful recommendation: i want to decorate a set of houses, please you's house is blank or old house, how much area is new house, several rooms, several halls, several guards, 127 square meters, 3 rooms, 2 toilets, 1 living room, 1 restaurant, 2 balconies, how much you want to spend money to dress the repairing woolen, 2, about 3 hundred thousand bars, mainly the quality is good, you see the simplified Chinese style like, the material used for decoration is environment-friendly and cheap.
Unsuccessful recommendation: i want to decorate a set of houses, please ask you what the blank is or what the old houses, second houses, the area is big, what the house type is, 87 square meters, 2 rooms, 1 guest's dining room, 1 bathroom, 1 kitchen, old houses still live, what the furniture home appliances need not to be replaced with new ones, we still live all home, find places to live, how long they need to be loaded, how long the materials used by us are all environment-friendly, have detection, feel confident, need about 5 months from dismantling to loading.
In step 1, the designer's history success recommended case information satisfies the following conditional expression:
；
wherein,representing a set of designers that were successfully recommended to the user, < +.>Representing the j-th designer->Representing users who have a history of recommended relationships with the j-th designer.
Designer setDemographic information, business expertise information, and information on their success cases for each designer are recorded in the database of the marketing management system or user relationship management system of the corresponding business.
The designer's demographic information, business expertise information, may be refined by business experts into the form of natural language, for example: "Zhang San, ID, man, 37 years old, practitioner's experience 12 years, success case Wanke Garden 19 ten thousand, roman garden 21 ten thousand, shengda garden 12 ten thousand, and Binjiang garden 31 ten thousand. Good at the style of Chinese classical, 80 minutes; a new house, 90 minutes; a second room, 40 minutes; apartment, 70 minutes; leveling and 90 minutes; high affinity, high customer evaluation, 100 minutes … … set charging, 1500 yuan per square meter. ".
In this embodiment, the designer history training corpus is also text in txt format, and examples are as follows:
zhang three, ID, man, 37 years old, success in case of 12 years old from experience of 19 ten thousand of Van gardens, 21 ten thousand of Roman gardens, 12 ten thousand of grand gardens, 31 ten thousand of coastal gardens, 31 ten thousand\n, 80 minutes\new houses, 90 minutes\second houses, 40 minutes\apartments, 70 minutes\flat, 90 minutes\affinity high, customer evaluation high, 100 minutes\charge, 1500 yuan per square meter.
And 2, generating a user intention set based on the user intention training corpus, and generating a designer personal history set based on the designer history training corpus.
In step 2, the set of user intents may be represented asWherein q represents the total number of user intentions, +.>The 1 st, 2 nd, q th, user intention are respectively indicated.
For example, the number of the cells to be processed,price sensitive>Attention qualityQuantity (S)>Attention is paid to environmental protection. For any given user, the user's specific intent may be described as a subset +.>Wherein->. The intent in the user intent set is refined from the user intent training corpus by the business expert.
The designer's personal history set consists of pairs of specialized tuples, where one pair of specialized tuples is represented asWherein->Representing i-th design expertise, +.>Representing the weight corresponding to the i-th design expertise, each designer having +.>For specialized tuples.
For example, the number of the cells to be processed,，/>for a particular designer, the set of personal histories may be described as a subset +.>Wherein->. The designer personal histories in the designer personal histories collection are extracted from the designer histories training corpus by business specialists.
And 3, generating a user history dialog diagram based on the user intention set and the designer personal history set, wherein the user history dialog diagram is a heterogeneous diagram, the user history dialog diagram comprises two types of node object sets and two types of edge sets, the two types of node object sets are respectively the user intention set and the designer personal history set, the two types of edge sets are respectively an edge set in the conversion process and an edge set which is successfully recommended, the edge set marked in the conversion process indicates that final recommendation is not achieved, the edge set marked as successfully recommended indicates that final recommendation is successful, and then the user history dialog diagram is converted into a natural language context for a large language model.
Wherein the user history dialogue graph is a heterogeneous graphDescribed as->Comprising two types of node object setsV(i.e., user intent set and designer personal biography set) and two types of edge setsE(i.e., the set of edges in the conversion process and the set of edges that have been successfully recommended). All user history dialogs may constitute a plurality of heterograms, each heterogram being a star-shaped structure with a designer personal set of histories centered around and user intent sets around.
Because each type of edge in the heterogram has unique business semantics, the invention converts the customer history dialog graph into a natural language context suitable for a large language model. First, each heterogeneous graph is broken down into a plurality of tuples 〈 customer intents, edges, designer personal histories 〉, one customer owns one tuple, one designer corresponding to a plurality of tuples.
And 4, carrying out random combination on the client intention in the natural language context to obtain the natural language context after random combination.
In a realistic offline designer recommendation scenario, many factors may limit or even reduce the efficiency and effectiveness of the recommendation. For example, within the same time period, a single sales advisor interacting with multiple different customers can result in a bias in recommendations. In addition, the process of interaction may be intermittent due to external environmental factors, preventing the advisor from capturing the customer's intent accurately. Still further, they may reveal one-sided, erroneous intentions and preferences, subject to user-subjective factors. This is the case, and the present invention proposes a customer intent random combination method.
In this approach, the historical dialog corpus of the customer and sales advisor is randomly cluttered and then reassembled into a new multi-turn dialog. There are three benefits to doing so. First, random interference to the original conversation breaks the constraint of the fixed conversation pattern. This change provides a broader potential interaction path for clients and consultants. Second, random combining introduces new dialog patterns. These combinations help to capture the diverse needs of customers, and can facilitate the capture of different intents of these customers. Finally, for interrupted and abrupt dialogs, the stochastic combining method can make efficient use of each dialog segment, ensuring that no valuable customer feedback is ignored. In summary, the customer intent random combination method is a technique for processing and analyzing customer requirements, and is particularly suitable for the dialog corpus of complex customer-sales consultants.
Specifically, step 4 includes:
acquiring a history dialogue text between a user and a sales consultant in a natural language context, and marking user intention, problems and preferences contained in the history dialogue text in a preset evaluation mode to serve as a verification sample;
randomly disturbing sentences or dialogue segments in the sample, and changing the sequence of the context in the original dialogue;
and recombining the disturbed sentences or dialogue segments into a new sentence or dialogue segment to realize random combination of client intentions, thereby obtaining the natural language context after random combination.
A specific example is as follows:
the user: i like modern conclusive styles, with budgets around 10 ten thousand.
Sales advisor: it is known what particular requirements you have for space utilization.
The user: i pay more attention to functionality and want living room space to be spacious.
Sales advisor: that me recommends someone designer to you, who is experienced in budget cost control and space utilization.
After application of the random combination of customer intents, the dialog may be reorganized as:
sales advisor: it is known what particular requirements you have for space utilization.
The user: i like modern conclusive styles, with budgets around 10 ten thousand.
Sales advisor: that me recommends someone designer to you, who is experienced in budget cost control and space utilization.
The user: i pay more attention to functionality and want living room space to be spacious.
In this reorganized dialogue, the user's question precedes the customer's statement regarding style and budget, which may reveal deeper considerations of the customer's intent in finishing, such as specific needs for space and preferences for functionality. In this way, the user's intent and needs can be more fully captured.
And 5, normalizing the designer's personal histories based on the designer's personal histories, and obtaining the normalized designer's personal histories.
In the designer's personal history sets, each set has a differentAnd corresponding weight->. In the real home improvement industry, a designer is totally lacking in a certain +>It is not possible. Therefore, eliminating such expertise offset is critical to ensure stability of the recommended results. In view of this, the present inventionThe history normalization method of the designer is divided into two parts.
On the one hand, knowledge elements of all home improvement industries are brought into the resume collection of each designer and are uniformly distributed in the resume collection. If a designer lacks specific expertise, then these elements will be assigned a minimum weight.
On the other hand, the professional knowledge weights in the personal history set of each designer are normalized, specifically, the following conditional expression is satisfied in step 5:
；
wherein,weights representing normalized i-th design expertise, < >>Representing the weight corresponding to the w-th design expertise. The sum of the weights of all expertise in each designer's personal biography set is equal to 1.
And step 6, training the MOSS dialogue recommendation system by adopting the random combined natural language context and the normalized designer personal history set, and then recommending the designer through the trained MOSS dialogue recommendation system.
In step 6, in order to make the large language model correspond to the application in the vertical field of home decoration, the invention adopts prompt word fine adjustment and instruction fine adjustment for the MOSS dialogue recommendation system. Specifically, when training the MOSS dialogue recommendation system by adopting a random combined natural language context and a normalized designer personal history set, performing prompt word fine adjustment and instruction fine adjustment, guiding the MOSS dialogue recommendation system to understand the random combined natural language context through the prompt word fine adjustment, and refining the execution of recommendation tasks through the instruction fine adjustment, wherein the prompt word fine adjustment and the instruction fine adjustment share the same final target lossFinal target loss->The expression of (2) is:
；
wherein,representing the initial parameters of the package->And->Respectively representing prompt input and prompt response in the customized prompt message,/->Is training set, is->Indicates the corresponding times->Representation->Is>Secondary response->Representation->In response to the previous response,is indicated at prompt input +.>And->Under the condition->Is a conditional distribution probability of (c).
As a specific example, the step of hinting word fine tuning is as follows:
1) Defining a prompt structure: the format and structure of the hints to be used in the fine tuning process is determined. For example, determining prompt inputAnd prompt response->The hint information is output in a specified nature and structure.
2) Loading a model and initializing parameters: loading initial parametersThe fine adjustment is prepared.
3) Training a model: through training setThe model is trained according to the defined prompt format, and an autoregressive training method is used to optimize the model to generate a response that meets the expectations of business experts.
4) Evaluation and adjustment: and evaluating the performance of the fine-tuned model, and iteratively adjusting.
As a specific example, the steps of instruction trimming are as follows:
1) Defining an instruction structure: the instruction format to be used in the instruction fine-tuning is specified.
2) Loading a model and initializing parameters: loading initial parametersThe fine adjustment is prepared.
3) Training a model: through training setThe model is trained according to the defined hint format. The model is optimized using an autoregressive training method to generate a response that meets business expert expectations.
4) Evaluation and adjustment: and evaluating the performance of the fine-tuned model, and iteratively adjusting.
In summary, according to the home decoration designer recommendation method integrating a large language model and dynamic dialogue intentions provided by the embodiment of the invention, firstly, user intention training corpus is built, designer history training corpus is built, then user intention collection and designer personal history collection are generated, then, a user history dialogue graph is generated based on the user intention collection and the designer personal history collection, then, the user history dialogue graph is converted into a natural language context for the large language model, then, the client intentions in the natural language context are randomly combined, the designer histories are normalized, finally, a MOSS dialogue recommendation system is trained, dynamic changes of the user decoration intentions are mined from the historical dialogue corpus of a user and a sales consultant based on the capability of extracting fine text features and massive Chinese knowledge base, and the dynamic changes are matched with the personal histories of the designer.
In the description of the present specification, a description referring to terms "one embodiment," "some embodiments," "examples," "specific examples," or "some examples," etc., means that a particular feature, structure, material, or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the present invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiments or examples. Furthermore, the particular features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.
While embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that: many changes, modifications, substitutions and variations may be made to the embodiments without departing from the spirit and principles of the invention, the scope of which is defined by the claims and their equivalents.
Claims (5)
1. A home decoration designer recommendation method integrating a large language model and dynamic dialogue intentions, which is characterized by comprising the following steps:
step 1, constructing user intention training corpus according to user ID, historical dialogue text between the user and a sales consultant and historical recommendation relation between the user and a designer; constructing a designer history training corpus according to the ID of the designer, demographic information, business expertise information and historical successful recommended case information of the designer;
step 2, generating a user intention set based on the user intention training corpus, and generating a designer personal history set based on the designer history training corpus;
step 3, generating a user history dialogue graph based on a user intention set and a designer personal history set, wherein the user history dialogue graph is a heterogeneous graph, the user history dialogue graph comprises two types of node object sets and two types of edge sets, the two types of node object sets are respectively the user intention set and the designer personal history set, the two types of edge sets are respectively an edge set in a conversion process and an edge set which is successfully recommended, the edge set marked in the conversion process indicates that final recommendation is not achieved, the edge set marked as successfully recommended indicates that final recommendation is successful, and then the user history dialogue graph is converted into a natural language context for a large language model;
step 4, carrying out random combination on the client intention in the natural language context to obtain a natural language context after random combination;
step 5, based on the designer personal record set, carrying out designer record normalization to obtain a normalized designer personal record set;
step 6, training the MOSS dialogue recommendation system by adopting the random combined natural language context and the normalized designer personal history set, and then recommending the designer by the trained MOSS dialogue recommendation system;
in step 4, the client intention in the natural language context is randomly combined to obtain a natural language context after random combination, which specifically includes:
acquiring a history dialogue text between a user and a sales consultant in a natural language context, and marking user intention, problems and preferences contained in the history dialogue text in a preset evaluation mode to serve as a verification sample;
randomly disturbing sentences or dialogue segments in the sample, and changing the sequence of the context in the original dialogue;
and recombining the disturbed sentences or dialogue segments into a new sentence or dialogue segment to realize random combination of client intentions, thereby obtaining the natural language context after random combination.
2. The home decoration designer recommendation method integrating a large language model and dynamic dialogue intentions according to claim 1, wherein in step 1, the user intent training corpus is txt format text, and the history recommendation relationship between the user and the designer satisfies the following conditional expression:
；
wherein,representing a set of users who were successfully recommended to the designer, < >>Represents the kth user, +.>Designer indicating that there is a history recommendation relationship with the kth user, < >>Representing a set of historical users>Representing a set of designers;
the historic training corpus of the designer is text in txt format, and the historic successful recommended case information of the designer meets the following conditional expression:
；
wherein,representing a set of designers that were successfully recommended to the user, < +.>Representing the j-th designer->Representing users who have a history of recommended relationships with the j-th designer.
3. The method of claim 1, wherein in step 2, the designer's personal history set consists of a plurality of pairs of expertise tuples, wherein a pair of expertise tuples is expressed asWherein->Representing i-th design expertise, +.>Representing the weight corresponding to the i-th design expertise, each designer having +.>For specialized tuples.
4. The home improvement designer recommendation method fusing large language models and dynamic dialogue intentions as claimed in claim 3, wherein the following conditional expression is satisfied in step 5:
；
wherein,weights representing normalized i-th design expertise, < >>Representing the weight corresponding to the w-th design expertise.
5. The home decoration designer recommendation method integrating large language models and dynamic dialogue intentions according to claim 1, wherein in step 6, when training a MOSS dialogue recommendation system by adopting a random combined natural language context and a normalized designer personal history set, prompt word fine tuning and instruction fine tuning are performed, the MOSS dialogue recommendation system is guided to understand the random combined natural language context through the prompt word fine tuning, and then execution of a recommendation task is refined through the instruction fine tuning, wherein the prompt word fine tuning and the instruction fine tuning share the same final target lossFinal target loss->The expression of (2) is:
；
wherein,representing the initial parameters of the package->And->Respectively represent prompt input and prompt response in the customized prompt message,is training set, is->Indicates the corresponding times->Representation->Is>Secondary response->Representation->Before response +_>Is indicated at prompt input +.>And->Under the condition->Is a conditional distribution probability of (c).
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311716405.0A CN117407595B (en) | 2023-12-14 | 2023-12-14 | Home decoration designer recommendation method integrating large language model and dynamic dialogue intention |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311716405.0A CN117407595B (en) | 2023-12-14 | 2023-12-14 | Home decoration designer recommendation method integrating large language model and dynamic dialogue intention |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117407595A CN117407595A (en) | 2024-01-16 |
| CN117407595B true CN117407595B (en) | 2024-03-08 |
Family
ID=89500331
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311716405.0A Active CN117407595B (en) | 2023-12-14 | 2023-12-14 | Home decoration designer recommendation method integrating large language model and dynamic dialogue intention |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117407595B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117689450B (en) * | 2024-01-29 | 2024-04-19 | 北京一起网科技股份有限公司 | Digital marketing system based on big data |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112883277A (en) * | 2021-03-18 | 2021-06-01 | 杭州太火鸟科技有限公司 | Intelligent designer matching method, device, equipment and storage medium |
| CN116304007A (en) * | 2022-11-11 | 2023-06-23 | 蚂蚁财富(上海)金融信息服务有限公司 | Information recommendation method and device, storage medium and electronic equipment |
| CN116595150A (en) * | 2023-06-16 | 2023-08-15 | 上海微盟企业发展有限公司 | Dialogue recommendation method, device, equipment and storage medium |
| CN116757652A (en) * | 2023-08-17 | 2023-09-15 | 北京华品博睿网络技术有限公司 | Online recruitment recommendation system and method based on large language model |
| CN116821318A (en) * | 2023-08-29 | 2023-09-29 | 中船奥蓝托无锡软件技术有限公司 | Business knowledge recommendation method, device and storage medium based on large language model |
| CN116955613A (en) * | 2023-06-12 | 2023-10-27 | 广州数说故事信息科技有限公司 | Method for generating product concept based on research report data and large language model |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11715111B2 (en) * | 2018-09-25 | 2023-08-01 | Capital One Services, Llc | Machine learning-driven servicing interface |
-
2023
- 2023-12-14 CN CN202311716405.0A patent/CN117407595B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112883277A (en) * | 2021-03-18 | 2021-06-01 | 杭州太火鸟科技有限公司 | Intelligent designer matching method, device, equipment and storage medium |
| CN116304007A (en) * | 2022-11-11 | 2023-06-23 | 蚂蚁财富(上海)金融信息服务有限公司 | Information recommendation method and device, storage medium and electronic equipment |
| CN116955613A (en) * | 2023-06-12 | 2023-10-27 | 广州数说故事信息科技有限公司 | Method for generating product concept based on research report data and large language model |
| CN116595150A (en) * | 2023-06-16 | 2023-08-15 | 上海微盟企业发展有限公司 | Dialogue recommendation method, device, equipment and storage medium |
| CN116757652A (en) * | 2023-08-17 | 2023-09-15 | 北京华品博睿网络技术有限公司 | Online recruitment recommendation system and method based on large language model |
| CN116821318A (en) * | 2023-08-29 | 2023-09-29 | 中船奥蓝托无锡软件技术有限公司 | Business knowledge recommendation method, device and storage medium based on large language model |
Non-Patent Citations (3)
| Title |
|---|
| "Modeling Instant User Intent and Content-Level Transition for Sequential Fashion Recommendation";Yujuan Ding 等;《IEEE TRANSACTIONS ON MULTIMEDIA》;20221231;2867-2700 * |
| "以ChatGPT为代表的大语言模型在临床医学中的应用综述";马武仁等;《医学信息学杂志》;20230725;9-17 * |
| "基于大语言模型的电力***通用人工智能初探:理论与应用";赵俊华等;《电力***》;20231127;1-16 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117407595A (en) | 2024-01-16 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107886949B (en) | Content recommendation method and device | |
| US11210300B2 (en) | System and methods of generating structured data from unstructured data | |
| CN117407595B (en) | Home decoration designer recommendation method integrating large language model and dynamic dialogue intention | |
| CN110619506B (en) | Post image generation method, post image generation device and electronic equipment | |
| US9489625B2 (en) | Rapid development of virtual personal assistant applications | |
| EP3821352A1 (en) | Machine learning tool for navigating a dialogue flow | |
| US10083263B2 (en) | Automatic modeling farmer | |
| Rieser et al. | Learning and evaluation of dialogue strategies for new applications: Empirical methods for optimization from small data sets | |
| KR102100214B1 (en) | Method and appratus for analysing sales conversation based on voice recognition | |
| CN109711871B (en) | Potential customer determination method, device, server and readable storage medium | |
| CN110689261A (en) | Service quality evaluation product customization platform and method | |
| KR20200143991A (en) | Answer recommendation system and method based on text content and emotion analysis | |
| US10115116B2 (en) | Optimizing efficiency and cost of crowd-sourced polling | |
| US10664927B2 (en) | Automation of crowd-sourced polling | |
| Giorgi et al. | Correcting sociodemographic selection biases for population prediction from social media | |
| US20240070726A1 (en) | Automated generation of creative parameters based on approval feedback | |
| US20230367816A1 (en) | Feedback control for automated messaging adjustments | |
| CN115271282A (en) | Customer value determination method and device based on fuzzy logic | |
| KR20210009266A (en) | Method and appratus for analysing sales conversation based on voice recognition | |
| CN115169322A (en) | Natural language data mining statement solving method and device based on knowledge graph | |
| Elsaid et al. | Automatic framework for requirement analysis phase | |
| CN114398495A (en) | Method and device for constructing knowledge graph based on multiple rounds of questions and answers and readable storage medium | |
| Sheng et al. | Feature selection for dynamic range compressor parameter estimation | |
| Riaz et al. | Social Interactions Leading Role in Adopting the Fintech: A Case of Banking Sector | |
| Erman et al. | Impact cultural-quality factors on successes and failures software system |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |