CN117473457B - Big data mining method and system based on digital service - Google Patents
Big data mining method and system based on digital service Download PDFInfo
- Publication number
- CN117473457B CN117473457B CN202311791425.4A CN202311791425A CN117473457B CN 117473457 B CN117473457 B CN 117473457B CN 202311791425 A CN202311791425 A CN 202311791425A CN 117473457 B CN117473457 B CN 117473457B
- Authority
- CN
- China
- Prior art keywords
- text
- preference
- business
- semantic
- activity
- 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 92
- 238000007418 data mining Methods 0.000 title claims abstract description 25
- 230000000694 effects Effects 0.000 claims abstract description 478
- 238000012549 training Methods 0.000 claims abstract description 267
- 238000005065 mining Methods 0.000 claims abstract description 209
- 230000008569 process Effects 0.000 claims abstract description 26
- 238000005457 optimization Methods 0.000 claims description 90
- 230000004927 fusion Effects 0.000 claims description 37
- 230000015654 memory Effects 0.000 claims description 18
- 230000002787 reinforcement Effects 0.000 claims description 14
- 238000005070 sampling Methods 0.000 claims description 12
- 230000006870 function Effects 0.000 claims description 10
- 238000004891 communication Methods 0.000 claims description 4
- 238000010835 comparative analysis Methods 0.000 claims description 4
- 230000004044 response Effects 0.000 claims description 4
- 239000013598 vector Substances 0.000 abstract description 44
- 238000004458 analytical method Methods 0.000 abstract description 13
- 238000005516 engineering process Methods 0.000 abstract description 6
- 238000012545 processing Methods 0.000 description 55
- 239000011159 matrix material Substances 0.000 description 41
- 230000006399 behavior Effects 0.000 description 19
- 238000013528 artificial neural network Methods 0.000 description 16
- 238000011156 evaluation Methods 0.000 description 8
- 238000004422 calculation algorithm Methods 0.000 description 7
- 238000004590 computer program Methods 0.000 description 7
- 239000000284 extract Substances 0.000 description 7
- 238000000605 extraction Methods 0.000 description 5
- 238000007670 refining Methods 0.000 description 5
- 238000013473 artificial intelligence Methods 0.000 description 4
- 238000001914 filtration Methods 0.000 description 4
- 238000003058 natural language processing Methods 0.000 description 4
- 238000004364 calculation method Methods 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000007717 exclusion Effects 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 230000009471 action Effects 0.000 description 2
- 238000013459 approach Methods 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000009223 counseling Methods 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 230000001965 increasing effect Effects 0.000 description 2
- 230000000977 initiatory effect Effects 0.000 description 2
- 230000003993 interaction Effects 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 238000011176 pooling Methods 0.000 description 2
- 230000004913 activation Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000018109 developmental process Effects 0.000 description 1
- 230000002708 enhancing effect Effects 0.000 description 1
- 230000007787 long-term memory Effects 0.000 description 1
- 238000012423 maintenance Methods 0.000 description 1
- 238000007726 management method Methods 0.000 description 1
- 230000003340 mental effect Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000007781 pre-processing Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000012552 review Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 230000033772 system development Effects 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/26—Discovering frequent patterns
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F40/00—Handling natural language data
- G06F40/30—Semantic analysis
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Computation (AREA)
- Evolutionary Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Bioinformatics & Computational Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a big data mining method and system based on digital service. Compared with the traditional preference viewpoint analysis technology, the embodiment of the invention can improve the accuracy and timeliness of preference viewpoint mining. In detail, the semantic moments of the first session activity training texts of the plurality of business session activity training texts are fused, so that the business session activity training texts without business preference requirements and the business session activity training texts with business preference requirements or the derived business preference requirement vectors between the business session activity training texts with business preference requirements and the business session activity training texts with business preference requirements can be obtained, the derived expansion of the business preference requirement vectors can be realized, the diversity and the comprehensiveness of debugging samples in the network debugging process are further ensured, and the adaptability of the target preference view mining network to different digital businesses is improved.
Description
Technical Field
The invention relates to the technical field of digitalization and big data, in particular to a big data mining method and system based on digital service.
Background
Digital business refers to changing traditional business models and processes by using digital technologies (such as cloud computing, big data, artificial intelligence, etc.), improving efficiency, increasing revenue, and creating new value generation modes. Such transformation typically involves an automated process that makes it faster, flexible and accurate.
In the digital service, a new and personalized user experience can be provided by means of a digital tool and a platform, and better customer relationship management is realized. At the same time, they can also gain in-depth business insight through analysis and utilization of large amounts of data to better understand market trends and customer needs, thereby making more intelligent business decisions.
Digital business is not limited to technological improvements, but also requires cultural and organizational changes in enterprises, including developing new business strategies, developing digital ideas, and constructing an organizational structure that accommodates rapidly changing digital environments.
In order to further promote the development of the digital service, a big data mining technology for the digital service plays an important role. Among them, user preference mining is an important point of attention for most digitized services. However, the conventional technology has a disadvantage of poor adaptability in performing the user preference mining process using artificial intelligence.
Disclosure of Invention
The invention provides a big data mining method and system based on digital service.
The technical scheme of the invention is realized by at least partial embodiments as follows.
A big data mining method based on digitized services, applied to a big data mining system, the method comprising:
responding to a user preference viewpoint mining request aiming at a target business project, and obtaining a business session activity text to be mined; wherein, the service session activity text to be mined contains a target user activity event;
loading the service session activity text to be mined into a target preference viewpoint mining network to obtain target preference viewpoint mining information of the service session activity text to be mined, wherein the target preference viewpoint mining information is used for representing the possibility that service preference demands exist on the target user activity event in the service session activity text to be mined; the target preference viewpoint mining network is determined based on a first preference viewpoint mining network, the first preference viewpoint mining network is obtained by debugging the base preference viewpoint mining network through a target optimization variable, the target optimization variable is determined based on a first optimization variable, and the first optimization variable is determined based on global learning annotation and a preference viewpoint prediction result corresponding to a first global session activity text semantic moment; the global learning annotation is obtained by fusing priori learning annotations of a plurality of business session activity training texts, and the priori learning annotations are used for representing whether business preference demands exist on user activity events in the business session activity training texts; the first global conversation activity text semantic moment is obtained by fusing first conversation activity training text semantic moments of the business conversation activity training texts, and the first conversation activity training text semantic moment is obtained by carrying out text semantic recognition on the business conversation activity training texts through the basic preference viewpoint mining network; the preference viewpoint predicting result is obtained by performing preference viewpoint mining on the first global conversation activity text semantic moment through the basic preference viewpoint mining network, and the preference viewpoint predicting result is used for representing the judging possibility that the business preference requirement exists for the user activity event corresponding to the first global conversation activity text semantic moment.
In some aspects, the target optimization variable is determined based on the first optimization variable and a second optimization variable, the second optimization variable being determined based on a first target conversation activity text semantic moment and a second target conversation activity text semantic moment; the first target conversation activity text semantic moment is determined based on the first global conversation activity text semantic moment; the second target session activity text semantic moment is determined based on a second global session activity text semantic moment, the second global session activity text semantic moment is obtained by fusing second session activity training text semantic moments of the business session activity training texts, and the second session activity training text semantic moment is obtained by carrying out text semantic recognition on the business session activity training texts through a second preference viewpoint mining network; the second preference viewpoint mining network is obtained by debugging a debugging sample set, wherein the debugging sample set comprises a plurality of business session activity text samples and priori learning notes of each business session activity text sample; the second preferred view mining network and the first preferred view mining network are matched in network architecture.
In some aspects, the method further comprises:
acquiring prior learning notes of the plurality of business session activity training texts and the business session activity training texts;
text semantic recognition is carried out on the plurality of business session activity training texts according to the basic preference viewpoint mining network to obtain first session activity training text semantic moments of the business session activity training texts, the first session activity training text semantic moments of the business session activity training texts are fused according to fusion factors corresponding to the business session activity training texts to obtain first global session activity text semantic moments, preference viewpoint mining is carried out on the first global session activity text semantic moments according to the basic preference viewpoint mining network to obtain preference viewpoint prediction results corresponding to the first global session activity text semantic moments;
fusing prior learning notes of the business session activity training texts according to fusion factors corresponding to the business session activity training texts to obtain global learning notes, wherein the global learning notes are used for representing the reference possibility that business preference demands exist for user activity events corresponding to the semantic moment of the first global session activity text;
Text semantic recognition is carried out on the plurality of business session activity training texts according to the second preference viewpoint mining network to obtain second session activity training text semantic moments of the business session activity training texts, and the second session activity training text semantic moments of the business session activity training texts are fused according to fusion factors corresponding to the business session activity training texts to obtain second global session activity text semantic moments;
determining the first optimization variable according to the preference viewpoint prediction result and the global learning annotation, obtaining the first target session active text semantic moment determined according to the first global session active text semantic moment and the second target session active text semantic moment determined according to the second global session active text semantic moment, performing comparative analysis on the first target session active text semantic moment and the second target session active text semantic moment, and determining the second optimization variable according to the comparative analysis result;
and determining the target optimization variable according to the first optimization variable, the second optimization variable and the confidence corresponding to the second optimization variable, and updating the network configuration parameters of the basic preference viewpoint mining network according to the target optimization variable to obtain the first preference viewpoint mining network.
In some aspects, the base preference point of view mining network includes a text semantic recognition branch including X text semantic recognition nodes cascaded with X being a positive integer; the text semantic recognition is carried out on the plurality of business session activity training texts by the mining network according to the basic preference viewpoint to obtain first session activity training text semantic moments of the business session activity training texts, and the text semantic recognition comprises the following steps:
loading a target business session activity training text to the text semantic recognition branch so that a target text semantic recognition node carries out text semantic recognition on the target business session activity training text to obtain a session activity text semantic moment generated by the target text semantic recognition node;
the target text semantic recognition node is any text semantic recognition node in the first Y text semantic recognition nodes in the X text semantic recognition nodes, Y is a positive integer not greater than X, and the target business session activity training text is any one of the plurality of business session activity training texts;
and determining the session activity text semantic moment generated by a Y-th text semantic recognition node in the X text semantic recognition nodes as a first session activity training text semantic moment of the target business session activity training text.
In some aspects, the base preference viewpoint mining network further comprises a preference viewpoint decision branch, an input of the preference viewpoint decision branch being an output of the text semantic recognition branch;
the step of mining the preference viewpoint of the first global conversation activity text semantic moment according to the basic preference viewpoint mining network to obtain the preference viewpoint prediction result corresponding to the first global conversation activity text semantic moment, including:
responding Y < X, carrying out text semantic recognition on the first global conversation activity text semantic moment according to the rear X-Y text semantic recognition nodes in the X text semantic recognition nodes, loading the conversation activity text semantic moment generated by the X text semantic recognition nodes into the preference viewpoint decision branch for preference viewpoint mining, and obtaining the preference viewpoint prediction result;
and responding to Y=X, loading the first global conversation activity text semantic moment into the preference viewpoint decision branch to perform preference viewpoint mining, and obtaining the preference viewpoint prediction result.
In some aspects, the obtaining the first target conversation activity text semantic moment determined from the first global conversation activity text semantic moment includes:
Responding to Y < X, and determining the session activity text semantic moment generated by the X-th text semantic recognition node as the first target session activity text semantic moment;
in response to y=x, determining the first global conversation activity text semantic moment as the first target conversation activity text semantic moment.
In some aspects, the number of business session activity training texts includes a first business session activity training text and a second business session activity training text, the method further comprising:
obtaining a preset factor and a sampling factor, and determining the sampling factor as a fusion factor corresponding to the first business session activity training text;
and performing difference between the preset factor and the sampling factor, and determining a difference result as a fusion factor corresponding to the second business session activity training text.
In some aspects, the fusing the semantic moment of the first session activity training text of each business session activity training text according to the fusion factor corresponding to each business session activity training text to obtain the semantic moment of the first global session activity text includes:
carrying out semantic feature reinforcement on the semantic moment of the first conversation activity training text of the first business conversation activity training text according to the fusion factor corresponding to the first business conversation activity training text to obtain the semantic moment of the first conversation activity training text after the semantic feature reinforcement corresponding to the first business conversation activity training text;
Carrying out semantic feature reinforcement on a second conversation activity training text semantic moment of the second business conversation activity training text according to a fusion factor corresponding to the second business conversation activity training text to obtain a first conversation activity training text semantic moment after the semantic feature reinforcement corresponding to the second business conversation activity training text;
and summing the semantic moment of the first conversation activity training text after the semantic feature corresponding to the first business conversation activity training text is reinforced and the semantic moment of the first conversation activity training text after the semantic feature corresponding to the second business conversation activity training text is reinforced, so as to obtain the semantic moment of the first global conversation activity text.
A big data mining system, comprising: a processor, a memory, and a network interface; the processor is connected with the memory and the network interface; the network interface is for providing data communication functions, the memory is for storing program code, and the processor is for invoking the program code to perform the above-described method.
A computer readable storage medium having stored thereon a computer program which, when run, performs a digital business based big data mining method.
A computer program product comprising a computer program or computer-executable instructions which, when executed by a processor, implement a digital business based big data mining method.
According to the embodiment of the invention, the to-be-mined business session activity text containing the target user activity event is loaded to the target preference viewpoint mining network to obtain the target preference viewpoint mining information for representing the possibility of the business preference demand of the target user activity event. The target preference viewpoint mining network in the embodiment of the invention is determined based on a first preference viewpoint mining network obtained by debugging according to a target optimization variable, and the target optimization variable is determined based on a first optimization variable determined according to a global learning annotation and a first global session activity text semantic moment corresponding preference viewpoint prediction result; the semantic moments of the first session activity training texts of the plurality of business session activity training texts are fused, business session activity training texts without business preference requirements and business session activity training texts with business preference requirements or derived business preference requirement vectors between the business session activity training texts with business preference requirements and the business session activity training texts with business preference requirements can be obtained, so that the derived expansion of the business preference requirement vectors can be realized, the diversity and the comprehensiveness of debugging samples in the network debugging process are further ensured, and the adaptability of a target preference view mining network to different digital businesses is improved.
It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention as claimed.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are necessary for the embodiments to be used are briefly described below, the drawings being incorporated in and forming a part of the description, these drawings showing embodiments according to the present invention and together with the description serve to illustrate the technical solutions of the present invention. It is to be understood that the following drawings illustrate only certain embodiments of the invention and are therefore not to be considered limiting of its scope, for the person of ordinary skill in the art may admit to other equally relevant drawings without inventive effort.
Fig. 1 shows a schematic diagram of a big data mining system according to an embodiment of the present invention.
Fig. 2 shows a flowchart of a big data mining method based on a digitalized service according to an embodiment of the present 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 only some embodiments of the present invention, not all embodiments. The components of the embodiments of the invention generally described and illustrated herein may be arranged and designed in a wide variety of different configurations.
Thus, the following detailed description of the embodiments of the invention is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. All other embodiments, which can be made by a person skilled in the art without making any inventive effort, are intended to be within the scope of the present invention.
Fig. 1 shows a schematic diagram of a big data mining system according to an embodiment of the present invention, where the big data mining system 100 includes a processor 110, a memory 120, and a network interface 130. The processor 110 is connected to the memory 120 and the network interface 130. Further, the network interface 130 is configured to provide data communication functions, the memory 120 is configured to store program codes, and the processor 110 is configured to invoke the program codes to perform a digital service-based big data mining method.
Fig. 2 is a flow diagram illustrating a digital service-based big data mining method that may be implemented by the big data mining system 100 shown in fig. 1, and the digital service-based big data mining method illustratively includes S110-S120, in which embodiments of the present invention may be implemented.
S110, responding to a user preference viewpoint mining request aiming at a target business project, and obtaining a business session activity text to be mined; and the service session activity text to be mined contains a target user activity event.
In S110, the target business item refers to a specific business item that we wish to perform data mining. It may be a product, service or activity such as online shopping, movie recommendation, etc. The user preference perspective mining request is a query request for a particular user behavior pattern or trend in order to learn and analyze the user's preferences, interests, or needs. The active text of the service session to be mined refers to the original data which needs to be processed and analyzed, and is usually a text record generated by a user in the process of using the service, such as comments, feedback, discussion or chat records. A target user activity event refers to a specific operation or behavior performed by a user while using a certain service, such as clicking, browsing, purchasing, etc.
Taking the e-commerce platform as an example, assuming that our target business item is "book sales", the user preference viewpoint mining request may come from the platform operator, hopefully by mining user comments to know their preferences for various books. In this example, the business session activity text to be mined is the comments that the user leaves after purchasing the book, including their views of various aspects of book content, author, publishing quality, etc. The target user activity event may then include the user's actions of searching for a book, viewing details, adding a shopping cart, completing a purchase, leaving comments, etc. In this process we pay special attention to the behavior they are completing the purchase and leaving comments, as the text information in these activities can provide us with important clues about the user's preferences. In this way, we can better understand the needs and preferences of the user, thereby optimizing the product or service and improving the user satisfaction and business effect.
Taking intelligent office as an example, the target business item may be an "online meeting system". User preference view mining requests may come from system development and maintenance teams that wish to optimize the online conferencing system by mining the user's feedback information. The business session activity text to be mined may include chat records, feedback information, and usage of specific functions that the user generates when using the online conferencing system. For example, the user may discuss problems with the meeting in a chat window or make ratings and suggestions to the system in a feedback area. The target user activity event may include a user initiating a meeting, switching meeting modes (e.g., audio, video, etc.), sharing a screen, sending a message, ending a meeting, etc. For example, receiving a user preference view mining request requires knowledge of the user's satisfaction with the online conferencing system. Firstly, collecting the active text of the business session to be mined, including chat records and feedback information of users when online conferences are carried out. Descriptions of various operations performed by the user (e.g., initiating a meeting, switching meeting modes, sharing screens, etc.) in these texts are then analyzed to obtain the user's experience and satisfaction with the online meeting system. In this way, the problems and needs encountered by users in using online conferencing systems can be appreciated for optimization and improvement.
Taking digital government enterprise services as an example, the target business project may be an "online tax system". The user preference view mining requests may come from a system operation team who wish to optimize the online tax service by mining user feedback information. The business session activity text to be mined may include query records, filled-in content, question feedback, etc. left by the user when using the online tax system. For example, a user may generate relevant records when querying tax information, or submit problem feedback to the system when operational challenges are encountered. The target user activity event may include user actions of querying tax information, filling tax forms, submitting feedback, online consultation, etc.
For example, receiving a user preference view mining request requires knowledge of the user's experience of using the online tax service. Firstly, collecting the active text of the business session to be mined, including text records generated when users inquire tax information, fill tax forms and submit feedback. Descriptions of various operations performed by the user (such as querying tax information, filling tax forms, etc.) in these texts are then analyzed, thereby obtaining the user's use experience and satisfaction with the online tax service. In this way, the problems and needs encountered by the user in using the online tax service can be appreciated, thereby optimizing and improving the service.
S120, loading the service session activity text to be mined into a target preference viewpoint mining network to obtain target preference viewpoint mining information of the service session activity text to be mined, wherein the target preference viewpoint mining information is used for representing the possibility that service preference demands exist on the target user activity event in the service session activity text to be mined.
The target preference viewpoint mining network is determined based on a first preference viewpoint mining network, the first preference viewpoint mining network is obtained by debugging the base preference viewpoint mining network through a target optimization variable, the target optimization variable is determined based on a first optimization variable, and the first optimization variable is determined based on global learning annotation and a preference viewpoint prediction result corresponding to a first global session activity text semantic moment; the global learning annotation is obtained by fusing priori learning annotations of a plurality of business session activity training texts, and the priori learning annotations are used for representing whether business preference demands exist on user activity events in the business session activity training texts; the first global conversation activity text semantic moment is obtained by fusing first conversation activity training text semantic moments of the business conversation activity training texts, and the first conversation activity training text semantic moment is obtained by carrying out text semantic recognition on the business conversation activity training texts through the basic preference viewpoint mining network; the preference viewpoint predicting result is obtained by performing preference viewpoint mining on the first global conversation activity text semantic moment through the basic preference viewpoint mining network, and the preference viewpoint predicting result is used for representing the judging possibility that the business preference requirement exists for the user activity event corresponding to the first global conversation activity text semantic moment.
In S120, the target preference viewpoint mining network is a specific machine learning model trained for analyzing and understanding user preferences. For example, it may be a deep neural network that predicts the user's preferences by learning a large amount of user behavior data. The target preference viewpoint mining information is extracted from the text of the business session activity to be mined and is used for representing the possibility that the business preference requirement exists in the target user activity event. Business preference requirements refer to specific requirements or preferences of a user when using a business, such as what type of merchandise is preferred when shopping on an e-commerce platform. The first preference viewpoint mining network is a preference viewpoint mining network subjected to preliminary debugging and optimization. The target optimization variables are parameters for adjusting and optimizing the preference view mining network. The base preference view mining network (untrained preference view mining network) is an initial state preference view mining network, which has not been subjected to any training and optimization. The first optimization variable is an optimization variable determined based on the global learning annotation and the first global session activity text semantic moment. Global learning annotation (prior fusion label) is a label obtained by fusing prior learning annotations of a plurality of business session activity training texts and is used for guiding the training of a network. The first global conversation activity text semantic moment is a matrix obtained by fusing the first conversation activity training text semantic moment of each business conversation activity training text. The preference viewpoint predicting result is a result obtained by mining the semantic moment of the first global session activity text through a preference viewpoint mining network and is used for representing the judging possibility that the business preference demands exist in the user activity event. Business session activity training text is raw data used to train a preference point of view mining network. A priori learning annotation is a tag used to identify whether a business preference requirement exists for a user activity event in business session activity training text. The first session activity training text semantic moment is a matrix obtained by text semantic recognition of the business session activity training text through a base preference viewpoint mining network. Text semantic recognition is a natural language processing technique used to extract and understand its meaning from text. The discriminant likelihood (prediction likelihood) refers to a probabilistic prediction of occurrence of a certain event (e.g., a user activity event has a business preference requirement) by a model.
Further, a text semantic matrix (also referred to as a text semantic feature matrix or a text semantic feature relationship net) is a data structure used to represent semantic information in text. It is typically a feature variable matrix consisting of word vectors, phrase vectors, or sentence vectors, where each vector represents a semantic feature of a word, phrase, or sentence.
Based on the above, the first global session activity text semantic matrix is a matrix obtained by fusing the first session activity training text semantic matrices of the respective business session activity training texts. For example, if we have two business session activity training texts a and B, each text generates a semantic matrix. In the semantic matrix of A, the vector that might correspond to "good" is [0.7,0.2,0.1], while in the semantic matrix of B, the vector that might correspond to "good" is [0.6,0.3,0.1]. Then, in the first global conversational activity text semantic matrix, the "good" vector may be the average of the two vectors, i.e. [0.65,0.25,0.1].
In addition, the first session activity training text semantic matrix is a matrix obtained by performing text semantic recognition on single business session activity training texts through a basic preference viewpoint mining network. For a simple example, if we have a training text "the phone is very good", then each word (e.g., "the phone", "very good", "good") may be converted into a vector, e.g., "good" may be converted into a vector [0.7,0.2,0.1], which represents the semantic information of the word "good" in the semantic matrix.
In addition, a preference view mining network, which is a deep neural network, typically contains multiple hidden layers and a large number of neurons. It is capable of learning and extracting complex patterns and features in text, which is very helpful for understanding the user's preference perspective. Deep neural networks are usually trained by algorithms such as back propagation and gradient descent, by continuously adjusting the network parameters (weights and biases) so that the difference (loss function) between the predicted and actual results of the network is as small as possible. In processing text data, preference-point mining networks typically employ some natural language processing techniques such as Word Embedding (Word Embedding), convolutional Neural Networks (CNN), recurrent Neural Networks (RNN), and more advanced variants such as long and short term memory networks (LSTM), gated loop units (GRU), etc. These techniques are capable of efficiently processing text sequence data, capturing context and semantic information in text.
For example, a basic preference-perspective mining network might first use word embedding techniques to convert each word into a dense vector, and then capture local and global features in the text through network structures such as CNNs or RNNs. Finally, the network may output the predicted likelihood that there is a business preference requirement for each user activity event through the full connectivity layer and activation function. During the training process, the network continuously adjusts the parameters according to the feedback, so that the predicted result is as close to the actual user preference as possible. This approach makes deep neural networks a powerful tool for processing text-mining tasks.
Taking e-commerce as an example, assume that we have a basic preference view mining network, which is a deep neural network, that can extract information expressing preference for goods from user comments.
First, we load the business session activity text to be mined (e.g., user's commodity reviews) into the target preference viewpoint mining network. This network has been optimized by training data, which contains a large number of users' shopping behaviors and their comment information. The network converts each word in the text into vectors, and then processes the vectors through the network hierarchy to capture and understand patterns and context relationships in the text, ultimately resulting in a set of target preference perspective mining information. Such information may include the likelihood that each item is liked by the user, the user's shopping habits, etc. In this process, the target optimization variables that we need to adjust may include the weights and biases of the network, as well as some super parameters for regularization. The initial values of these variables are determined based on a first optimization variable, which may be determined by global learning annotations and preferred perspective predictions corresponding to the first global conversational activity text semantic moments. In this process, the global learning annotation may be fused from a priori learning annotations of the plurality of business session activity training texts. For example, if we know that some users have strong preference for a certain commodity, this information can be annotated as a priori learning. Finally, the preference viewpoint predicting result is obtained by mining the first global conversation activity text semantic moment through the basic preference viewpoint mining network. This result is used to characterize the likelihood that the user has business preference needs for various goods, thereby helping the e-commerce platform to better understand the user needs, optimize the recommendation system and enhance sales.
In the process of S120, the step of loading the service session activity text to be mined into the target preference viewpoint mining network is that user comments (the service session activity text to be mined) are actually input into the deep neural network (the target preference viewpoint mining network) which we have trained.
When we enter text into the network, the network will perform a complex series of calculations, process vectors for each word, capture and understand patterns and context in the text. Eventually, the network will output a set of target preference viewpoint mining information. Such information may include the likelihood that each item is liked by the user, the user's shopping habits, etc.
The term "obtaining the target preference viewpoint mining information of the service session activity text to be mined" refers to a result obtained by processing user comments through a neural network, and the result is a quantitative representation of the commodity preference of the user.
Finally, "the target preference viewpoint mining information is used for representing the possibility that the target user activity event in the to-be-mined business session activity text has business preference demands," which means that we can understand and predict the preference degree of the user for commodities and the shopping behaviors possibly generated in the future by analyzing the preference viewpoint mining information output by the network.
Taking intelligent office as an example, the text of the business session activity to be mined may include various chat records, problem feedback, or operation records generated by staff when using the online conference system. The target preference viewpoint mining network is a deep neural network, which has been optimized and adjusted by a large amount of training data, and can effectively extract information about employee usage habits and preferences from text.
First, we enter these business session activity text to be mined into the target preference viewpoint mining network. The network processes the input text, and finally obtains a group of target preference viewpoint mining information through calculation of a plurality of layers such as word vectors, convolution layers, pooling layers, full connection layers and the like.
This information may help us understand the specific behavior and preferences of employees in using the online conferencing system. For example, the network may tell us that a certain employee prefers to use video conferencing over audio conferencing, or that they prefer to use which functions (e.g., screen sharing, chat functions, etc.) when in a meeting.
These target preference view mining information are used to characterize the likelihood that a target user activity event in the business session activity text to be mined has business preference requirements. For example, if the network predicts that a particular employee will be likely to use the screen sharing feature in future meetings, we can optimize for this requirement, such as checking in advance and ensuring the stability and ease of use of the feature.
By the method, the requirements and preferences of staff can be better understood, so that continuous optimization is performed, and the working efficiency and satisfaction are improved.
Taking digital government enterprise service as an example, the text of the business session to be mined may include various query records, filled-in content, problem feedback or operation records, etc. generated when online users use the online tax handling system. The target preference viewpoint mining network is a deep neural network that has been optimized and adjusted by a large amount of training data, and can effectively extract information on online user usage habits and preferences from text.
First, we enter these business session activity text to be mined into the target preference viewpoint mining network. The network processes the input text, and finally obtains a group of target preference viewpoint mining information through calculation of a plurality of layers such as word vectors, convolution layers, pooling layers, full connection layers and the like.
This information may help us understand the specific behavior and preferences of online users when using the online tax handling system. For example, the network may tell us that an online user prefers to conduct online tax at which time period, or that they need more assistance and support in transacting which type of tax service.
These target preference view mining information are used to characterize the likelihood that a target user activity event in the business session activity text to be mined has business preference requirements. For example, if the network predicts that an online user will likely need more guidance and assistance in future tax business transactions, we can optimize for this requirement, such as enhancing the system's boot function, or providing more detailed operational instructions.
In this way, the needs and preferences of the online user can be better understood, thereby performing continuous optimization and improving quality of service and satisfaction.
In the embodiment of the invention, the service session activity text to be mined containing the target user activity event is loaded to the target preference viewpoint mining network to obtain the target preference viewpoint mining information for representing the possibility of the service preference demand of the target user activity event. The target preference viewpoint mining network in the embodiment of the invention is determined based on a first preference viewpoint mining network obtained by debugging according to a target optimization variable, and the target optimization variable is determined based on a first optimization variable determined according to a global learning annotation and a first global session activity text semantic moment corresponding preference viewpoint prediction result; the semantic moments of the first session activity training texts of the plurality of business session activity training texts are fused, business session activity training texts without business preference requirements and business session activity training texts with business preference requirements or derived business preference requirement vectors between the business session activity training texts with business preference requirements and the business session activity training texts with business preference requirements can be obtained, so that the derived expansion of the business preference requirement vectors can be realized, the diversity and the comprehensiveness of debugging samples in the network debugging process are further ensured, and the adaptability of a target preference view mining network to different digital businesses is improved.
In detail, by applying the embodiment, the accuracy and timeliness of preference viewpoint mining can be improved: conventional preference perspective analysis techniques often rely on manually set rules or simple machine learning models, which may not adequately capture complex patterns and context relationships in text, resulting in poor accuracy of the mining results. By using the deep neural network to mine the preference views, the method can automatically learn and extract complex features in the text, and greatly improves the accuracy of mining results. In addition, because the neural network can process a large amount of data in parallel, the method has high processing speed, can analyze newly generated data in real time, and improves timeliness of mining preference views.
Further, by applying the above embodiment, derivative expansion of the service preference demand vector can be realized: by fusing semantic moments of first conversation activity training texts of a plurality of business conversation activity training texts, different types of business preference demand vectors, such as derived business preference demand vectors between business conversation activity training texts without business preference demands and business conversation activity training texts with business preference demands, can be obtained, so that our training data can be expanded, and diversity and comprehensiveness of samples are increased.
In addition, by applying the embodiment, the adaptability of the target preference viewpoint mining network can be improved: the target preference viewpoint mining network is determined by the first optimization variable determined based on the global learning annotation and the corresponding preference viewpoint prediction result of the first global conversation activity text semantic moment and the first preference viewpoint mining network obtained by debugging according to the target optimization variable, so that the network can be self-adjusted and optimized, and the adaptability of the network to different digital services is improved.
It can be seen that the embodiment of the invention can help us to understand the preference demands of users more accurately and more quickly in various scenes such as electronic commerce, intelligent office, digital government and enterprise services, and the like, thereby providing better quality services.
It should be noted that, first, the method of S110-S120 includes using deep neural networks for preference viewpoint mining, which obviously involves specific technical implementation, such as natural language processing, deep learning, etc. These are all technical means in the fields of artificial intelligence and computer science, and need to be implemented by computer hardware and software. Second, the method of S110-S120 requires processing and analysis of a large number of business session activity training texts, which involves technical means of data processing and analysis, such as text preprocessing, word vectorization, semantic analysis, etc. Again, the method of S110-S120 debugs and optimizes the preference perspective mining network by means of target optimization variables, which involve model optimization techniques in machine learning, such as gradient descent, back propagation, etc. Therefore, the method of S110-S120 is not a simple rule and method of mental activities, but is specifically applied to technical solutions in the technical field.
It should be noted that, first, the method of S110-S120 is obviously a new technical solution. The method comprises the steps of mining the preferred views through a deep neural network, determining a target preferred view mining network by using a brand new first optimization variable based on global learning annotation and a first global conversation activity text semantic moment corresponding to a preferred view prediction result, and performing debugging according to the target optimization variable to obtain the first preferred view mining network. This approach is clearly different from conventional preference point of view analysis techniques. Second, the methods of S110-S120 also involve improvements over existing methods. Compared with the traditional preference viewpoint analysis technology, the method of S110-S120 can more accurately and quickly understand the preference demands of the users, and the accuracy and timeliness of preference viewpoint mining are greatly improved. Finally, the method of S110-S120 can be considered either as a specific product (i.e., preference view mining system) or as a specific method (i.e., preference view mining through deep neural networks). Thus, the method of S110-S120 should obviously be considered as a patentable invention.
It should be noted that, as can be seen from the foregoing discussion, S110 and S120 have been described in detail, clearly and completely. They describe how to obtain a first global conversation activity text semantic moment by merging the first conversation activity training text semantic moment of the plurality of business conversation activity training texts (S110), and how to load the business conversation activity text to be mined into a target preference viewpoint mining network, and obtain target preference viewpoint mining information of the text (S120), respectively. These descriptions are all operative and provide sufficient guidance for implementation. In the art, i.e. in the fields of artificial intelligence and natural language processing, etc., the skilled person should be able to implement the described method on the basis of these descriptions. Technical concepts such as deep neural networks, semantic matrices, optimization variables, etc., are well known and widely used in the art. Furthermore, the solution also explicitly proposes specific steps and methods for preference view mining using these techniques. Accordingly, the contents of S110 and S120 have satisfied the requirement that the specification should make clear and complete description of the utility model or the utility model, so as to be able to be realized by those skilled in the art.
In some examples, the target optimization variable is determined based on the first optimization variable and a second optimization variable, the second optimization variable being determined based on a first target conversation activity text semantic moment and a second target conversation activity text semantic moment. The first target conversation activity text semantic moment is determined based on the first global conversation activity text semantic moment. The second target session activity text semantic moment is determined based on a second global session activity text semantic moment, the second global session activity text semantic moment is obtained by fusing second session activity training text semantic moments of the business session activity training texts, and the second session activity training text semantic moment is obtained by carrying out text semantic recognition on the business session activity training texts through a second preference viewpoint mining network. Further, the second preference viewpoint mining network is obtained through debugging a debugging sample set, wherein the debugging sample set comprises a plurality of business session activity text samples and a priori learning notes of each business session activity text sample. The second preferred view mining network and the first preferred view mining network are matched in network architecture.
Taking digital government enterprise services as an example, we obtain a first global session activity text semantic moment by fusing the first session activity training text semantic moment of each business session activity training text. Such text may come from query records, feedback information, or operational records of online users using the online government enterprise service platform.
Based on this first global conversation activity text semantic moment, we can get a first target conversation activity text semantic moment. For example, this matrix may be of particular concern to online users' behavior and needs in using certain specific services (e.g., transacting business licenses, paying tax fees, etc.).
Then, the second session activity training text semantic moment of each business session activity training text is fused, and the second global session activity text semantic moment is obtained. This matrix may contain more extensive user behavior and demand information, such as in which time periods, and under which circumstances, the online user is more likely to need to use a certain service.
Then, based on this second global conversation activity text semantic moment, we can get a second target conversation activity text semantic moment. For example, this matrix may further focus on the service needs of online users in particular holidays or emergency situations.
With the two target conversation activity text semantic moments (the first target conversation activity text semantic moment and the second target conversation activity text semantic moment), we can determine a second optimization variable based on the first optimization variable and then determine a target optimization variable based on the first optimization variable and the second optimization variable. This target optimization variable will be used to debug and optimize the preference-point-of-view mining network so that it predicts the business preference needs of online users more accurately.
To obtain the second target conversation activity text semantic moment, we need to use a second preference point mining network to text semantic recognition of business conversation activity training text. This second preferred view mining network is obtained by tuning a sample set comprising a number of business session activity text samples and a priori learning notes of the individual business session activity text samples.
Note that the network architecture of the second preference point of view mining network and the first preference point of view mining network are matched, meaning that they have the same number of network layers and number of neurons, but may have different parameter variables. The purpose of this design is to ensure that the second preferred view mining network is able to inherit the advantages of the first preferred view mining network, while also allowing further optimization based on new training data.
In further embodiments, the method further comprises S210-S260.
S210, acquiring prior learning notes of the service session activity training texts.
S220, text semantic recognition is conducted on the plurality of business session activity training texts according to the basic preference viewpoint mining network, first session activity training text semantic moments of the business session activity training texts are obtained, the first session activity training text semantic moments of the business session activity training texts are fused according to fusion factors corresponding to the business session activity training texts, first global session activity text semantic moments are obtained, preference viewpoint mining is conducted on the first global session activity text semantic moments according to the basic preference viewpoint mining network, and preference viewpoint prediction results corresponding to the first global session activity text semantic moments are obtained.
S230, according to fusion factors corresponding to the business session activity training texts, prior learning annotations of the business session activity training texts are fused to obtain global learning annotations, and the global learning annotations are used for representing the reference possibility that business preference demands exist for user activity events corresponding to the semantic moment of the first global session activity text.
S240, performing text semantic recognition on the plurality of business session activity training texts according to the second preference viewpoint mining network to obtain second session activity training text semantic moments of the business session activity training texts, and fusing the second session activity training text semantic moments of the business session activity training texts according to fusion factors corresponding to the business session activity training texts to obtain second global session activity text semantic moments.
S250, determining the first optimization variable according to the preference viewpoint prediction result and the global learning annotation, obtaining the first target session active text semantic moment determined according to the first global session active text semantic moment and the second target session active text semantic moment determined according to the second global session active text semantic moment, comparing and analyzing the first target session active text semantic moment and the second target session active text semantic moment, and determining the second optimization variable according to a comparison and analysis result.
And S260, determining the target optimization variable according to the first optimization variable, the second optimization variable and the confidence corresponding to the second optimization variable, and updating the network configuration parameters of the basic preference viewpoint mining network according to the target optimization variable to obtain the first preference viewpoint mining network.
First, S210 is to obtain a number of business session activity training texts and a priori learning notes of the respective business session activity training texts. In an e-commerce scenario, these training texts may come from the user's shopping records, rating information, etc. on the e-commerce platform; in a smart office scenario, it may come from an employee's operational records, feedback information, etc. while using the online office system; in a digital government enterprise service scenario, the query records, feedback information, etc. may come from online users when using a business online service platform.
Next, in S220, we use the basic preference viewpoint mining network to perform text semantic recognition on the training texts, and fuse the obtained first session activity training text semantic moments according to the fusion factors corresponding to the respective business session activity training texts, so as to obtain first global session activity text semantic moments. Then, the base preference viewpoint mining network is used for mining the preference viewpoints of the global semantic matrix, and a preference viewpoint prediction result is obtained.
In S230, we fuse the prior learning comments of the training texts according to the fusion factors corresponding to the training texts of each business session activity, so as to obtain global learning comments. This global learning annotation is used to characterize the reference likelihood that the business preference requirement exists for the user activity event corresponding to the first global session activity text semantic moment.
In S240, we use the second preference viewpoint mining network to identify text semantics of the training texts, and fuse the obtained second session activity training text semantics of the second global session activity according to the fusion factors corresponding to the session activity training texts.
Then, in S250, we determine a first optimization variable from the preference viewpoint prediction result and the global learning annotation, and obtain a first target conversation activity text semantic moment determined from the first global conversation activity text semantic moment and a second target conversation activity text semantic moment determined from the second global conversation activity text semantic moment. Next, we compare the two target session activity text semantic moments and determine a second optimization variable based on the comparison analysis results.
Finally, in S260, we determine a target optimization variable according to the first optimization variable, the second optimization variable, and the confidence level corresponding to the second optimization variable, and then update the network configuration parameters of the basic preference viewpoint mining network according to the target optimization variable, so as to obtain the first preference viewpoint mining network.
It can be seen that the method steps of S210-S260 provide a new network training technology path for preference viewpoint mining, and have the following beneficial effects.
(1) Improving accuracy: through fusing semantic matrixes and priori learning notes of different business session activity training texts and utilizing a basic preference viewpoint mining network and a second preference viewpoint mining network to conduct deep semantic recognition and preference viewpoint mining on the texts, behaviors and demands of users can be more accurately understood.
(2) Flexibility is improved: according to the method, not only is a single training text considered, but also the semantic matrix of each training text and the priori learning annotation are fused, so that the system can process diversified business scenes and user demands.
(3) And (3) training an optimization model: by introducing the first optimization variable and the second optimization variable and determining the target optimization variable according to the two optimization variables, network configuration parameters of the preference viewpoint mining network can be effectively adjusted, so that a model training process is optimized, and the prediction performance of a model is improved.
(4) And the user experience is improved: through deep understanding of the preference demands of the users, more accurate and personalized services can be provided for the users, and user experience is greatly improved. For example, in an e-commerce scenario, items that are more likely to be of interest to the user may be recommended; in the intelligent office scene, the intelligent office system can provide services more conforming to the working habit of staff; in a digital government enterprise service scene, more convenient and more careful service can be provided for online users.
(5) The decision efficiency is improved: for enterprises or business departments, users can be helped to make more accurate business decisions by understanding the preference demands of the users, and the decision efficiency is improved, so that the overall operation effect is improved.
In still other possible embodiments, the base preference point of view mining network includes a text-semantic-recognition branch including X text-semantic-recognition nodes cascaded with X being a positive integer. Based on this, the mining network in S220 performs text semantic recognition on the plurality of business session activity training texts according to the basic preference viewpoint, so as to obtain a first session activity training text semantic moment of each business session activity training text, including S221-S222.
S221, loading a target business session activity training text to the text semantic recognition branch, so that a target text semantic recognition node carries out text semantic recognition on the target business session activity training text to obtain a session activity text semantic moment generated by the target text semantic recognition node; the target text semantic recognition node is any text semantic recognition node in the first Y of the X text semantic recognition nodes, Y is a positive integer not greater than X, and the target business session activity training text is any one of the plurality of business session activity training texts.
S222, determining the session activity text semantic moment generated by a Y-th text semantic recognition node in the X text semantic recognition nodes as a first session activity training text semantic moment of the target business session activity training text.
In the above embodiment, the text semantic recognition branch is an integral part of the basic preference viewpoint mining network, and is dedicated to processing and understanding semantic information of text data. The text semantic recognition node is a unit in the text semantic recognition branch and is responsible for processing a part of specific text data and generating a corresponding semantic matrix. The conversational active text semantic moment is an output generated by a text semantic recognition node after processing text data, and is typically a two-dimensional array or matrix containing text semantic information.
Taking electronic commerce as an application scene, it is assumed that a recommendation system of an electronic commerce platform is designed, and the recommendation system needs to understand shopping records and evaluation information (namely business session activity training texts) of users and predict shopping preferences of the users.
In S221 we first load a target business session activity training text (e.g. a user' S shopping record) into the text semantic recognition branch. The text-semantic-recognition branch comprises X text-semantic-recognition nodes which are cascaded together in a certain order. Then, we select any one of the first Y text semantic recognition nodes as the target text semantic recognition node, let it perform text semantic recognition on this shopping record, and generate a conversational active text semantic moment. In S222 we determine the session activity text semantic moment generated by the Y-th text semantic recognition node as the first session activity training text semantic moment for this shopping record. The semantic matrix contains shopping behavior and preference information of the user and can be used for subsequent preference viewpoint mining and commodity recommendation. Through the steps, a large amount of service session activity training texts can be effectively processed and understood, valuable user preference information is extracted, and therefore recommendation effect and user experience of an e-commerce platform are improved.
Taking the smart office application scenario as an example, a large amount of business session activity training text may include mail traffic between employees, meeting records, work reports, etc.
In S221, a target business session activity training text, such as a conference recording, is first selected. Then, the conference record is loaded to a text semantic recognition branch of the basic preference viewpoint mining network, and the branch is formed by cascading X text semantic recognition nodes. Then, any one of the first Y text semantic recognition nodes is selected as a target text semantic recognition node, which performs text semantic recognition on the conference record, understands and extracts key information therein, and represents the information as a session activity text semantic matrix. In S222, the session activity text semantic matrix generated by the Y-th text semantic recognition node is determined as the first session activity training text semantic matrix of the meeting record. This semantic matrix contains important information about the content of the meeting, such as topics, participants, discussion content, etc. Through the operation, a large amount of business session activity training texts can be efficiently processed, valuable information is extracted, and corresponding semantic matrixes are generated. The method has important significance for tasks such as decision analysis, workflow optimization and the like in the intelligent office environment.
In digital government enterprise services, the business session activity training text may come from records of interactions between the online user and the digital government platform system, such as consultations or complaints submitted by the online user on the digital government platform system, business replies, and the like.
In S221, a target business session activity training text, such as an online user consultation for city planning, is first selected. This consultation is loaded into the text semantic recognition branch of the underlying preference point of view mining network. This branch is formed by cascading X text semantic recognition nodes. Then, any node from the previous Y text semantic recognition nodes is selected as the target text semantic recognition node. The node carries out text semantic recognition on consultation of the online user, understands and extracts key information in the consultation, and represents the information as a session activity text semantic matrix. Next, in S222, the session activity text semantic matrix generated by the yh text semantic recognition node is determined as the first session activity training text semantic matrix consulted by the online user. This semantic matrix contains important information about the online user's counseling, such as counseling topics, specific content, etc.
Through the processing, the interaction between the online user and the digital government affair platform system can be effectively understood in depth, so that the quality of digital government affair service is improved. For example, based on these conversational activity text semantic matrices, the digital government platform system can more accurately learn about the points of interest and needs of the online user, thereby making decisions that better meet the user's expectations.
In some examples, the base preference viewpoint mining network further includes a preference viewpoint decision branch, an input of the preference viewpoint decision branch being an output of the text semantic recognition branch. Based on this, the step S220 of mining the preference viewpoint of the first global session active text semantic moment according to the basic preference viewpoint mining network obtains the preference viewpoint prediction result corresponding to the first global session active text semantic moment, including step S223 or S224.
S223, responding to Y < X, carrying out text semantic recognition on the first global conversation activity text semantic moment according to the X-Y text semantic recognition nodes after X text semantic recognition nodes, loading the conversation activity text semantic moment generated by the X text semantic recognition nodes into the preference viewpoint decision branch for preference viewpoint mining, and obtaining the preference viewpoint prediction result.
And S224, loading the first global conversation activity text semantic moment into the preference viewpoint decision branch to perform preference viewpoint mining in response to Y=X, and obtaining the preference viewpoint prediction result.
In the above embodiment, the preference viewpoint decision branch is an integral part of the basic preference viewpoint mining network, and is dedicated to mining the preference viewpoint of the user from the text semantic matrix input.
In the following, this embodiment will be described by way of example in detail using e-commerce, smart office and digital government services as application scenarios.
Taking e-commerce as an example, assume that we are processing a user's shopping record that has generated a first global conversational activity text semantic moment through a text semantic recognition branch. In S223, if Y is smaller than X, we will use the post-X-Y text semantic recognition nodes to further perform text semantic recognition on the semantic matrix, and load the session active text semantic moment generated by the X-th text semantic recognition node to the preference viewpoint decision branch to perform preference viewpoint mining, so as to obtain a preference viewpoint prediction result. This result may help us understand the shopping preferences of the user, thereby providing more accurate merchandise recommendations.
Taking wisdom office as an example, assume we are processing a section of employee's meeting record that has generated a first global conversational activity text semantic moment through a text semantic recognition branch. In S224, if Y is equal to X, we will directly load this semantic matrix into the preference viewpoint decision branch to perform preference viewpoint mining, and obtain the preference viewpoint prediction result. This result can help us understand the employee's perspective and attitude to the meeting topic, thereby optimizing the internal communication and decision making process.
Taking the digital government enterprise service as an example, assume that we are processing a consultation record of an online user, which has generated a first global conversational activity text semantic moment through a text semantic recognition branch. Similar to the intelligent office scenario, in S224, if Y is equal to X, we will directly load this semantic matrix into the preference viewpoint decision branch to perform preference viewpoint mining, and obtain the preference viewpoint prediction result. This result may help the digital government enterprise business platform system understand the needs and desires of online users, thereby providing more careful, efficient services.
In some possible embodiments, the obtaining of the first target conversation activity text semantic moment determined from the first global conversation activity text semantic moment in S250 includes S251 or S252.
S251, responding to Y < X, and determining the session activity text semantic moment generated by the X-th text semantic recognition node as the first target session activity text semantic moment.
And S252, determining the first global conversation activity text semantic moment as the first target conversation activity text semantic moment in response to Y=X.
In S250, assume we have generated a first global session activity text semantic moment for a user shopping record. In S251, if Y is less than X, we determine the session active text semantic moment generated by the xth text semantic recognition node as the first target session active text semantic moment. The target semantic matrix contains shopping behavior and preference information of the user and can be used for subsequent commodity recommendation and personalized service.
Also in S250, assume that we have generated a first global conversation activity text semantic moment for a piece of employee meeting record. In S252, if Y is equal to X, we will directly determine the first global conversation activity text semantic moment as the first target conversation activity text semantic moment. The target semantic matrix reflects information such as the theme, the content and the employee's view of the conference, and can be used for improving the conference efficiency and improving the decision process.
Also in S250, assume that we have generated a first global session activity text semantic moment for an online user consultation record. In S252, if Y is equal to X, we will directly determine the first global conversation activity text semantic moment as the first target conversation activity text semantic moment. The target semantic matrix contains the consultation content and the demand information of the online user, and can help the digital government enterprise business platform system to provide more accurate service meeting public demands.
Through the steps, no matter the e-commerce, the intelligent office or the digital government enterprise service is realized, accurate business decision and service optimization can be performed according to the target conversation activity text semantic matrix generated in the respective scenes, so that the user experience and the business efficiency are greatly improved.
In some other possible embodiments, the number of business session activity training texts includes a first business session activity training text and a second business session activity training text. Based on this, the method further comprises S310-S320.
S310, obtaining a preset factor and a sampling factor, and determining the sampling factor as a fusion factor corresponding to the first business session activity training text.
S320, the preset factor and the sampling factor are subjected to difference, and the difference result is determined to be a fusion factor corresponding to the second business session activity training text.
Assuming that the preset factor is 0.7 and the sampling factor is 0.4, in S310 we determine the sampling factor as the fusion factor corresponding to the first service session activity training text, i.e. the fusion factor of the first service session activity training text is 0.4. Then, in S320, we perform the difference between the preset factor and the sampling factor, that is, 0.7-0.4=0.3, and determine the difference result as the fusion factor corresponding to the second service session activity training text, so that the fusion factor of the second service session activity training text is 0.3. Thus, the fusion factors corresponding to the first business session activity training text and the second business session activity training text are obtained. These fusion factors can be used to adjust the weights of different business session activity training texts in subsequent processing, thereby optimizing the training and prediction effects of the model.
In some other embodiments, the fusing factor corresponding to each business session activity training text in S220 fuses the first session activity training text semantic moment of each business session activity training text to obtain the first global session activity text semantic moment, including S220a-S220c.
S220a, carrying out semantic feature reinforcement on the semantic moment of the first conversation activity training text of the first business conversation activity training text according to the fusion factor corresponding to the first business conversation activity training text, and obtaining the semantic moment of the first conversation activity training text after the semantic feature reinforcement corresponding to the first business conversation activity training text.
S220b, carrying out semantic feature reinforcement on the semantic moment of the second session activity training text of the second service session activity training text according to the fusion factor corresponding to the second service session activity training text, and obtaining the semantic moment of the first session activity training text after the semantic feature reinforcement corresponding to the second service session activity training text.
S220c, summing the semantic moment of the first conversation activity training text after the semantic feature corresponding to the first business conversation activity training text is strengthened and the semantic moment of the first conversation activity training text after the semantic feature corresponding to the second business conversation activity training text is strengthened, and obtaining the semantic moment of the first global conversation activity text.
Assume that we have the following two session activity training texts:
First session activity training text of first business session activity training text semantic moment (Matrix 1): [[0.1,0.2],[0.3,0.4]].
First session activity training text semantic moment of second business session activity training text (Matrix 2): [[0.5,0.6],[0.7,0.8]].
According to the fusion factor obtained in the previous step, the fusion factor of the first business session activity training text is assumed to be 0.4, and the fusion factor of the second business session activity training text is assumed to be 0.3.
S220a: carrying out semantic feature reinforcement on Matrix1, namely multiplying each element in Matrix1 by 0.4 to obtain Matrix1 with reinforced semantic features: [[0.04,0.08],[0.12,0.16]].
S220b: carrying out semantic feature reinforcement on Matrix2, namely multiplying each element in Matrix2 by 0.3 to obtain Matrix2 with reinforced semantic features: [[0.15,0.18],[0.21,0.24]].
S220c: summing the Matrix1 and the Matrix2 with the reinforced semantic features to obtain a first global session activity text semantic moment (GlobalMatrix): [ [0.04+0.15,0.08+0.18], [0.12+0.21,0.16+0.24] ] = [ [0.19,0.26], [0.33,0.40] ].
In this way, the first global session activity text semantic moment, globalMatrix, can be obtained. The matrix contains the fusion information of the first business session activity training text and the second business session activity training text, and can be used for subsequent preference viewpoint mining and decision analysis.
In some independent embodiments, after loading the to-be-mined business session activity text into the target preference viewpoint mining network as described in S120 to obtain the target preference viewpoint mining information of the to-be-mined business session activity text, the method further includes S130.
S130, combining the target preference viewpoint mining information and the service session active text to be mined to determine an associated storage session text; the associated storage session text is used for explaining a preference viewpoint mining process of the target preference viewpoint mining network.
In an embodiment of the present invention, the associative memory session text is a text for recording and interpreting the preference viewpoint mining process. It typically contains target preference viewpoint mining information and related content of the business session activity text to be mined.
In the e-commerce application scenario, we assume that we are processing a user's shopping record, which is the text of the business session activity to be mined. The target preference viewpoint mining information may be shopping history or rating information of the user. In S130, we combine this information to determine the associated storage session text. For example, the associated storage session text may include information about the time, number, price, etc. that the user purchased a good, and the user's rating of the good. This associative storage of session text may help us understand and interpret the shopping preferences of the user.
In the smart office application scenario, we assume that we are processing a section of employee's meeting records, which is the text of the business session activity to be mined. The target preference view mining information may be the role of the employee, the task allocation, or the work report content. In S130, we combine this information to determine the associated storage session text. For example, this associated storage session text may include information about the employee's speech content in the meeting, completion of the task, and so on. This associative storage of session text may help us understand and interpret the employee's job attitudes and efficiencies.
In the digital government enterprise service application scenario, it is assumed that we are processing a consultation record of an online user, and the record is the text of the business session to be mined. The target preference viewpoint mining information may be identity information of the online user, previous consultation history, or feedback conditions. In S130, we combine this information to determine the associated storage session text. For example, this associative storage session text may include information about the time the online user presented a consultation, the content of the question, the reply status of the digital government enterprise business platform system, etc. This associative storage of session text may help us understand and interpret the needs and satisfaction of online users.
In some independent embodiments, the determining of the associated stored session text by combining the target preference viewpoint mining information and the to-be-mined business session activity text as described in S130 includes S131-S133.
S131, inputting structural semantic refining branches of a structural semantic association recognition algorithm by text pairs consisting of the target preference viewpoint mining information and the to-be-mined business conversation activity text, and obtaining structural semantic knowledge features output by the structural semantic refining branches.
In this step, the target preference viewpoint mining information and the business session activity text to be mined are combined and input into a structural semantic refining branch of a structural semantic association recognition algorithm to obtain structural semantic knowledge features.
S132, inputting the structural semantic knowledge features into a plurality of target semantic association processing cores in the structural semantic association recognition algorithm to obtain semantic association suggestion labels output by each target semantic association processing core, wherein each target semantic association processing core is obtained by debugging an initial semantic association processing core corresponding to the target semantic association processing core through text pairs corresponding to a target association storage mode of the target semantic association processing core, each target association storage mode corresponding to the target semantic association processing core is determined from a plurality of setting association storage modes according to mutually exclusive feature sets corresponding to the plurality of setting association storage modes, and each feature variable in the mutually exclusive feature set represents mutual exclusivity between two setting association storage modes corresponding to the feature variable.
The structural semantic knowledge features obtained from the previous step are input into a plurality of target semantic association processing cores, and each processing core outputs a semantic association suggestion label.
S133, acquiring and counting semantic association suggestion labels output by each target semantic association processing core, and taking the semantic association suggestion labels as current semantic association suggestion labels corresponding to the text pairs; and carrying out association processing based on structural semantics on the target preference viewpoint mining information and the service session active text to be mined through the current semantic association suggestion tag to obtain an association storage session text.
And counting semantic association suggestion labels output by each target semantic association processing core, taking the semantic association suggestion labels as current semantic association suggestion labels of corresponding text pairs, carrying out structural semantic-based association processing on target preference viewpoint mining information and the to-be-mined business session active text based on the labels, and finally obtaining associated storage session text.
The following is a specific description taking an e-commerce recommendation system as an example: suppose we are processing a user's shopping records (text of business session activity to be mined) while having the user's shopping history and rating information (target preference view mining information).
In S131, we combine the shopping record of the user with the shopping history and the evaluation information, and input the combined shopping record and the shopping history and the evaluation information into the structural semantic refining branch of the structural semantic association recognition algorithm to obtain the structural semantic knowledge feature. For example, the features may include the type of merchandise purchased by the user, price intervals, frequency of purchase, content of evaluation, and the like.
In S132, we input the structured semantic knowledge features acquired in the previous step into a plurality of target semantic association processing cores. Each processing core may correspond to a particular recommendation policy or model, such as collaborative filtering models based on user historical behavior, content filtering models based on merchandise attributes, and so forth. Each processing core generates a semantic association suggestion tag according to its internal policies or models. For example, collaborative filtering models based on historical behavior of users may generate a label that "resembles historical purchase behavior" while content filtering models based on merchandise attributes may generate a label that "matches merchandise attributes.
In S133, we count the semantic association suggestion tags output by each target semantic association processing core as the current semantic association suggestion tags of the corresponding users. And then, carrying out association processing based on structural semantics on shopping records, shopping histories and evaluation information of the user based on the labels, and finally obtaining an association storage session text. For example, this associative storage session text may include a user's shopping records, shopping history, and rating information, and semantic association suggestion tags associated therewith.
In this way, we can perform in-depth structured semantic analysis and processing of business session activity text while understanding user preferences. This not only can improve the accuracy and efficiency of the recommendation system, but also helps to provide a richer and personalized user experience. In addition, as each target semantic association processing core is obtained through a specific debugging and optimizing process, the target semantic association processing cores can be well adapted to different application scenes and requirements, and the flexibility and the expandability of the system are improved.
In some independent embodiments, the method for obtaining the target association storage mode corresponding to the target semantic association processing core includes S1321-S1324.
S1321, acquiring a plurality of text pairs in each set associated storage mode as a plurality of target text pairs corresponding to each set associated storage mode.
In this step, a plurality of text pairs in each set associative memory mode are obtained as a plurality of target text pairs corresponding to each set associative memory mode.
S1322, processing the cases through the structured semantic extraction branches to a plurality of target texts corresponding to each set associative storage mode to obtain associative storage mode vectors corresponding to each set associative storage mode.
And processing the cases through the structured semantic refining branches for a plurality of target texts corresponding to each set associated storage mode to obtain associated storage mode vectors corresponding to each set associated storage mode.
S1323, determining mutually exclusive feature sets corresponding to the plurality of setting association storage modes according to the association storage mode vectors of the plurality of setting association storage modes.
And determining mutual exclusion feature sets corresponding to the set association storage modes according to the association storage mode vectors of the set association storage modes.
S1324, determining a target association storage mode corresponding to each target semantic association processing core according to the mutual exclusion feature set and the number of the target semantic association processing cores.
And determining a target association storage mode corresponding to each target semantic association processing core according to the mutual exclusion feature set and the number of the target semantic association processing cores.
For example, if we are processing an e-commerce application scenario, in S1321 we may obtain a plurality of text pairs in each set associative storage mode (such as "recommendation based on user history behavior", "recommendation based on commodity attributes", etc.), as a target text pair corresponding to each set associative storage mode.
In S1322, we will process the cases with these target text through the structured semantic refinement branch, resulting in an associative storage pattern vector corresponding to each set associative storage pattern. These vectors may contain features such as "purchase frequency", "price range", "rating content", etc.
In S1323, we determine, according to the association storage pattern vector obtained in the previous step, mutually exclusive feature sets corresponding to each set association storage pattern. For example, "recommendations based on historical behavior of the user" and "recommendations based on properties of the merchandise" may be mutually exclusive in the feature of "purchase frequency".
Finally, in S1324, we will determine, according to the mutually exclusive feature set obtained in the previous step and the number of target semantic association processing cores (e.g. 5 processing cores), a target association storage mode corresponding to each target semantic association processing core. This step may be implemented by various optimization algorithms to ensure that each processing core gets a target associative memory pattern that best suits its characteristics and needs.
By this method, a large amount of business session activity texts can be more effectively organized and managed, and the preference and the requirement of the user can be better understood and mined. This not only helps to improve the accuracy and efficiency of the recommendation system, but also helps to provide a richer and personalized user experience.
In some independent embodiments, the processing of the cases by the structured semantic extraction branch in S1322 on the target text corresponding to each set associative storage pattern results in an associative storage pattern vector corresponding to each set associative storage pattern, including S13221-S13222.
S13221, identifying each target text pair case corresponding to each set association storage mode through the structural semantic extraction branch to obtain a target structural semantic vector case of each target text pair case in each set association storage mode.
In this step, each target text corresponding to each set associative memory mode is identified by the structural semantic extraction branch to obtain a target structural semantic vector case of each target text to case in each set associative memory mode.
S13222, performing splicing processing on the target structured semantic vector cases of the cases on each target text in each set associated storage mode to obtain associated storage mode vectors corresponding to each set associated storage mode.
And performing splicing processing on the target structured semantic vector cases of the cases by each target text in each set associated storage mode to obtain associated storage mode vectors corresponding to each set associated storage mode.
For example, in an e-commerce application scenario, setting the associative storage mode may be "recommendation based on user history behavior" or "recommendation based on merchandise properties". Assume that a target text pair case is the user's evaluation of a good.
In S13221, we will identify the evaluation content through the structural semantic extraction branch, extract the structural semantic information therein, such as "good quality of commodity", "reasonable price", etc., and then convert the structural semantic information into a vector form, so as to obtain the target structural semantic vector case of the evaluation content.
In S13222, we will splice all the target text in the same set associative memory mode to the target structured semantic vector cases of the cases, and finally obtain the associative memory mode vector corresponding to each set associative memory mode. This vector may reflect the overall preferences and needs of the user in the set associative storage mode.
In this way, we can not only organize and manage large volumes of business session activity text more efficiently, but also better understand and mine the preferences and needs of the user. This not only helps to improve the accuracy and efficiency of the recommendation system, but also helps to provide a richer and personalized user experience.
Based on the same or similar technical ideas described above, the embodiments of the present invention also provide a computer-readable storage medium having stored thereon a computer program that performs a digital service-based big data mining method at runtime.
Based on the same or similar technical concept, the embodiment of the invention also provides a computer program product, which comprises a computer program or a computer executable instruction, wherein the computer program or the computer executable instruction realizes the big data mining method based on the digital service when being executed by a processor.
Those of ordinary skill in the art will appreciate that the elements and algorithm steps described in connection with the embodiments disclosed herein may be embodied in electronic hardware, in computer software, or in a combination of the two, and that the elements and steps of the examples have been generally described in terms of function in the foregoing description to clearly illustrate the interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.
The foregoing disclosure is merely illustrative of the presently preferred embodiments of the present invention, and it is to be understood that the scope of the invention is not limited thereto, but is intended to cover modifications as fall within the scope of the present invention.
Claims (10)
1. A method for mining big data based on digitized services, characterized in that the method is applied to a big data mining system, the method comprising:
responding to a user preference viewpoint mining request aiming at a target business project, and obtaining a business session activity text to be mined; wherein, the service session activity text to be mined contains a target user activity event;
loading the service session activity text to be mined into a target preference viewpoint mining network to obtain target preference viewpoint mining information of the service session activity text to be mined, wherein the target preference viewpoint mining information is used for representing the possibility that service preference demands exist on the target user activity event in the service session activity text to be mined; the target preference viewpoint mining network is determined based on a first preference viewpoint mining network, the first preference viewpoint mining network is obtained by debugging the base preference viewpoint mining network through a target optimization variable, the target optimization variable is determined based on a first optimization variable, and the first optimization variable is determined based on global learning annotation and a preference viewpoint prediction result corresponding to a first global session activity text semantic moment; the global learning annotation is obtained by fusing priori learning annotations of a plurality of business session activity training texts, and the priori learning annotations are used for representing whether business preference demands exist on user activity events in the business session activity training texts; the first global conversation activity text semantic moment is obtained by fusing first conversation activity training text semantic moments of the business conversation activity training texts, and the first conversation activity training text semantic moment is obtained by carrying out text semantic recognition on the business conversation activity training texts through the basic preference viewpoint mining network; the preference viewpoint predicting result is obtained by performing preference viewpoint mining on the first global conversation activity text semantic moment through the basic preference viewpoint mining network, and the preference viewpoint predicting result is used for representing the judging possibility that the business preference requirement exists for the user activity event corresponding to the first global conversation activity text semantic moment.
2. The method of claim 1, wherein the target optimization variable is determined based on the first optimization variable and a second optimization variable, the second optimization variable being determined based on a first target conversation activity text semantic moment and a second target conversation activity text semantic moment; the first target conversation activity text semantic moment is determined based on the first global conversation activity text semantic moment; the second target session activity text semantic moment is determined based on a second global session activity text semantic moment, the second global session activity text semantic moment is obtained by fusing second session activity training text semantic moments of the business session activity training texts, and the second session activity training text semantic moment is obtained by carrying out text semantic recognition on the business session activity training texts through a second preference viewpoint mining network; the second preference viewpoint mining network is obtained by debugging a debugging sample set, wherein the debugging sample set comprises a plurality of business session activity text samples and priori learning notes of each business session activity text sample; the second preferred view mining network and the first preferred view mining network are matched in network architecture.
3. The method according to claim 2, wherein the method further comprises:
acquiring prior learning notes of the plurality of business session activity training texts and the business session activity training texts;
text semantic recognition is carried out on the plurality of business session activity training texts according to the basic preference viewpoint mining network to obtain first session activity training text semantic moments of the business session activity training texts, the first session activity training text semantic moments of the business session activity training texts are fused according to fusion factors corresponding to the business session activity training texts to obtain first global session activity text semantic moments, preference viewpoint mining is carried out on the first global session activity text semantic moments according to the basic preference viewpoint mining network to obtain preference viewpoint prediction results corresponding to the first global session activity text semantic moments;
fusing prior learning notes of the business session activity training texts according to fusion factors corresponding to the business session activity training texts to obtain global learning notes, wherein the global learning notes are used for representing the reference possibility that business preference demands exist for user activity events corresponding to the semantic moment of the first global session activity text;
Text semantic recognition is carried out on the plurality of business session activity training texts according to the second preference viewpoint mining network to obtain second session activity training text semantic moments of the business session activity training texts, and the second session activity training text semantic moments of the business session activity training texts are fused according to fusion factors corresponding to the business session activity training texts to obtain second global session activity text semantic moments;
determining the first optimization variable according to the preference viewpoint prediction result and the global learning annotation, obtaining the first target session active text semantic moment determined according to the first global session active text semantic moment and the second target session active text semantic moment determined according to the second global session active text semantic moment, performing comparative analysis on the first target session active text semantic moment and the second target session active text semantic moment, and determining the second optimization variable according to the comparative analysis result;
and determining the target optimization variable according to the first optimization variable, the second optimization variable and the confidence corresponding to the second optimization variable, and updating the network configuration parameters of the basic preference viewpoint mining network according to the target optimization variable to obtain the first preference viewpoint mining network.
4. A method according to claim 3, wherein the base preference point of view mining network comprises a text semantic recognition branch comprising X text semantic recognition nodes, the X text semantic recognition nodes being cascaded, X being a positive integer; the text semantic recognition is carried out on the plurality of business session activity training texts by the mining network according to the basic preference viewpoint to obtain first session activity training text semantic moments of the business session activity training texts, and the text semantic recognition comprises the following steps:
loading a target business session activity training text to the text semantic recognition branch so that a target text semantic recognition node carries out text semantic recognition on the target business session activity training text to obtain a session activity text semantic moment generated by the target text semantic recognition node;
the target text semantic recognition node is any text semantic recognition node in the first Y text semantic recognition nodes in the X text semantic recognition nodes, Y is a positive integer not greater than X, and the target business session activity training text is any one of the plurality of business session activity training texts;
and determining the session activity text semantic moment generated by a Y-th text semantic recognition node in the X text semantic recognition nodes as a first session activity training text semantic moment of the target business session activity training text.
5. The method of claim 4, wherein the base preference point of view mining network further comprises a preference point of view decision branch, an input of the preference point of view decision branch being an output of the text semantic recognition branch;
the step of mining the preference viewpoint of the first global conversation activity text semantic moment according to the basic preference viewpoint mining network to obtain the preference viewpoint prediction result corresponding to the first global conversation activity text semantic moment, including:
responding Y < X, carrying out text semantic recognition on the first global conversation activity text semantic moment according to the rear X-Y text semantic recognition nodes in the X text semantic recognition nodes, loading the conversation activity text semantic moment generated by the X text semantic recognition nodes into the preference viewpoint decision branch for preference viewpoint mining, and obtaining the preference viewpoint prediction result;
and responding to Y=X, loading the first global conversation activity text semantic moment into the preference viewpoint decision branch to perform preference viewpoint mining, and obtaining the preference viewpoint prediction result.
6. The method of claim 5, wherein the obtaining the first target conversation activity text semantic moment determined from the first global conversation activity text semantic moment comprises:
Responding to Y < X, and determining the session activity text semantic moment generated by the X-th text semantic recognition node as the first target session activity text semantic moment;
in response to y=x, determining the first global conversation activity text semantic moment as the first target conversation activity text semantic moment.
7. The method of claim 3, wherein the number of business session activity training texts comprises a first business session activity training text and a second business session activity training text, the method further comprising:
obtaining a preset factor and a sampling factor, and determining the sampling factor as a fusion factor corresponding to the first business session activity training text;
and performing difference between the preset factor and the sampling factor, and determining a difference result as a fusion factor corresponding to the second business session activity training text.
8. The method of claim 7, wherein the fusing the first session activity training text semantic moments of each of the business session activity training texts according to the fusion factor corresponding to each of the business session activity training texts to obtain the first global session activity text semantic moment comprises:
Carrying out semantic feature reinforcement on the semantic moment of the first conversation activity training text of the first business conversation activity training text according to the fusion factor corresponding to the first business conversation activity training text to obtain the semantic moment of the first conversation activity training text after the semantic feature reinforcement corresponding to the first business conversation activity training text;
carrying out semantic feature reinforcement on a second conversation activity training text semantic moment of the second business conversation activity training text according to a fusion factor corresponding to the second business conversation activity training text to obtain a first conversation activity training text semantic moment after the semantic feature reinforcement corresponding to the second business conversation activity training text;
and summing the semantic moment of the first conversation activity training text after the semantic feature corresponding to the first business conversation activity training text is reinforced and the semantic moment of the first conversation activity training text after the semantic feature corresponding to the second business conversation activity training text is reinforced, so as to obtain the semantic moment of the first global conversation activity text.
9. The method according to claim 1, wherein the method further comprises:
combining the target preference viewpoint mining information and the service session active text to be mined to determine an associated storage session text; the associated storage session text is used for explaining a preference viewpoint mining process of the target preference viewpoint mining network.
10. A big data mining system, comprising: a processor, a memory, and a network interface; the processor is connected with the memory and the network interface; the network interface is for providing data communication functions, the memory is for storing program code, and the processor is for invoking the program code to perform the digital service based big data mining method of any of claims 1-8.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311791425.4A CN117473457B (en) | 2023-12-25 | 2023-12-25 | Big data mining method and system based on digital service |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311791425.4A CN117473457B (en) | 2023-12-25 | 2023-12-25 | Big data mining method and system based on digital service |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN117473457A CN117473457A (en) | 2024-01-30 |
| CN117473457B true CN117473457B (en) | 2024-02-27 |
Family
ID=89627717
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311791425.4A Active CN117473457B (en) | 2023-12-25 | 2023-12-25 | Big data mining method and system based on digital service |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN117473457B (en) |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115357804A (en) * | 2022-07-19 | 2022-11-18 | 大连民族大学 | Two-stage collaborative filtering multi-behavior recommendation method based on graph convolution network |
| CN116070027A (en) * | 2023-02-21 | 2023-05-05 | 李雪峰 | Online pushing topic analysis method based on artificial intelligence and big data system |
| CN116308551A (en) * | 2023-03-16 | 2023-06-23 | 江苏擎虎智能科技有限公司 | Content recommendation method and system based on digital financial AI platform |
| CN116361643A (en) * | 2023-02-22 | 2023-06-30 | 阿里巴巴（中国）有限公司 | Model training method for realizing object recommendation, object recommendation method and related device |
| CN116467415A (en) * | 2023-04-21 | 2023-07-21 | 西安工业大学 | Bidirectional cross-domain session recommendation method based on GCNsformer hybrid network and multi-channel semantics |
| CN117112909A (en) * | 2023-09-12 | 2023-11-24 | 厦门登谷网络科技有限公司 | Big data recommendation method and big data mining system applied to cloud digital service |
| CN117149996A (en) * | 2023-09-19 | 2023-12-01 | 武汉腾翔越科技有限公司 | Man-machine interface digital conversation mining method and AI system for artificial intelligence application |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20230245651A1 (en) * | 2020-03-14 | 2023-08-03 | Polypie Inc. | Enabling user-centered and contextually relevant interaction |
-
2023
- 2023-12-25 CN CN202311791425.4A patent/CN117473457B/en active Active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115357804A (en) * | 2022-07-19 | 2022-11-18 | 大连民族大学 | Two-stage collaborative filtering multi-behavior recommendation method based on graph convolution network |
| CN116070027A (en) * | 2023-02-21 | 2023-05-05 | 李雪峰 | Online pushing topic analysis method based on artificial intelligence and big data system |
| CN116361643A (en) * | 2023-02-22 | 2023-06-30 | 阿里巴巴（中国）有限公司 | Model training method for realizing object recommendation, object recommendation method and related device |
| CN116308551A (en) * | 2023-03-16 | 2023-06-23 | 江苏擎虎智能科技有限公司 | Content recommendation method and system based on digital financial AI platform |
| CN116467415A (en) * | 2023-04-21 | 2023-07-21 | 西安工业大学 | Bidirectional cross-domain session recommendation method based on GCNsformer hybrid network and multi-channel semantics |
| CN117112909A (en) * | 2023-09-12 | 2023-11-24 | 厦门登谷网络科技有限公司 | Big data recommendation method and big data mining system applied to cloud digital service |
| CN117149996A (en) * | 2023-09-19 | 2023-12-01 | 武汉腾翔越科技有限公司 | Man-machine interface digital conversation mining method and AI system for artificial intelligence application |
Non-Patent Citations (5)
| Title |
|---|
| Huizhao Wang 等.DMFP：A Dynamic Multi-faceted Fine-Grained Preference Model for Recommendation.《2019 IEEE International Conference on Data Mining（ICDM）》.2020,608-617. * |
| Yan Xiao 等.User preference mining based on fine-grained sentiment analysis.《Journal of Retailing and Consumer Services》.2022,1-10. * |
| Yu Liu 等.Long- and Short-Term Preference Model Based on Graph Embedding for Sequential Recommendation.《Database Systems for Advanced Applications DASFAA 2020 International Workshops》.2020,241-257. * |
| 何昊晨 ; 张丹红 ; .基于多维社交关系嵌入的深层图神经网络推荐方法.《计算机应用》.2020,第40卷(第10期),2795-2803. * |
| 沈冬东 ; 汪海涛 ; 姜瑛 ; 陈星 ; .基于知识图谱嵌入与多神经网络的序列推荐算法.《计算机工程与科学》.2020,第42卷(第09期),1661-1669. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN117473457A (en) | 2024-01-30 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Sohaib et al. | Cloud computing model selection for e-commerce enterprises using a new 2-tuple fuzzy linguistic decision-making method | |
| US10725827B2 (en) | Artificial intelligence based virtual automated assistance | |
| Kazak et al. | Artificial intelligence in the tourism sphere | |
| US7574379B2 (en) | Method and system of using artifacts to identify elements of a component business model | |
| CN111737546B (en) | Method and device for determining entity service attribute | |
| Rizk et al. | A conversational digital assistant for intelligent process automation | |
| CN112669096B (en) | Object recommendation model training method and device | |
| US20220300850A1 (en) | End-to-end machine learning pipelines for data integration and analytics | |
| CN117608650B (en) | Business flow chart generation method, processing device and storage medium | |
| US20230034820A1 (en) | Systems and methods for managing, distributing and deploying a recursive decisioning system based on continuously updating machine learning models | |
| Kim et al. | Document-based workflow modeling: a case-based reasoning approach | |
| CN113792134B (en) | User service method and system based on digital twin technology | |
| Li et al. | Research on the application of multimedia entropy method in data mining of retail business | |
| Chen | Iot-enabled supply chain finance risk management performance big data analysis using Fuzzy Qfd | |
| Chae et al. | Discovering latent topics of digital technologies from venture activities using structural topic modeling | |
| US20200043019A1 (en) | Intelligent identification of white space target entity | |
| Mustroph et al. | Social Network Mining from Natural Language Text and Event Logs for Compliance Deviation Detection | |
| CN117473457B (en) | Big data mining method and system based on digital service | |
| Proletarsky et al. | Approach to forecasting the development of crisis situations in complex information networks | |
| US11776006B2 (en) | Survey generation framework | |
| US20140149186A1 (en) | Method and system of using artifacts to identify elements of a component business model | |
| Turktarhan et al. | Re-architecting the firm for increased value: How business models are adapting to the new AI environment | |
| Gao | Implementation of a dynamic planning algorithm in accounting information technology administration | |
| Wang et al. | Semi-automatic service value network modeling approach based on external public data | |
| CN112328899A (en) | Information processing method, information processing apparatus, storage medium, and electronic device |
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 |