CN116991919B - Service data retrieval method combined with platform database and artificial intelligent system - Google Patents
Service data retrieval method combined with platform database and artificial intelligent system Download PDFInfo
- Publication number
- CN116991919B CN116991919B CN202311243360.XA CN202311243360A CN116991919B CN 116991919 B CN116991919 B CN 116991919B CN 202311243360 A CN202311243360 A CN 202311243360A CN 116991919 B CN116991919 B CN 116991919B
- Authority
- CN
- China
- Prior art keywords
- service data
- data retrieval
- data
- vector
- interference
- 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 51
- 239000013598 vector Substances 0.000 claims abstract description 404
- 230000004927 fusion Effects 0.000 claims abstract description 79
- 238000013473 artificial intelligence Methods 0.000 claims abstract description 20
- 230000004931 aggregating effect Effects 0.000 claims abstract description 12
- 230000001629 suppression Effects 0.000 claims description 218
- 238000005065 mining Methods 0.000 claims description 194
- 238000012545 processing Methods 0.000 claims description 102
- 238000006243 chemical reaction Methods 0.000 claims description 32
- 238000013528 artificial neural network Methods 0.000 claims description 27
- 238000007405 data analysis Methods 0.000 claims description 26
- 230000009467 reduction Effects 0.000 claims description 20
- 238000009412 basement excavation Methods 0.000 claims description 12
- 230000007246 mechanism Effects 0.000 claims description 9
- 238000004590 computer program Methods 0.000 claims description 6
- 238000000605 extraction Methods 0.000 description 29
- 230000008569 process Effects 0.000 description 7
- 238000013527 convolutional neural network Methods 0.000 description 6
- 238000011084 recovery Methods 0.000 description 6
- 239000004973 liquid crystal related substance Substances 0.000 description 3
- 239000011159 matrix material Substances 0.000 description 3
- 238000012549 training Methods 0.000 description 3
- 230000002776 aggregation Effects 0.000 description 2
- 238000004220 aggregation Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 230000006870 function Effects 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 238000003058 natural language processing Methods 0.000 description 2
- 238000012935 Averaging Methods 0.000 description 1
- BQCADISMDOOEFD-UHFFFAOYSA-N Silver Chemical compound [Ag] BQCADISMDOOEFD-UHFFFAOYSA-N 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 238000003491 array Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000005540 biological transmission Effects 0.000 description 1
- 230000019771 cognition Effects 0.000 description 1
- 238000004891 communication Methods 0.000 description 1
- 238000010219 correlation analysis Methods 0.000 description 1
- 238000007499 fusion processing Methods 0.000 description 1
- 230000006872 improvement Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 238000006116 polymerization reaction Methods 0.000 description 1
- 229910052709 silver Inorganic materials 0.000 description 1
- 239000004332 silver Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/245—Query processing
- G06F16/2458—Special types of queries, e.g. statistical queries, fuzzy queries or distributed queries
- G06F16/2465—Query processing support for facilitating data mining operations in structured databases
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/20—Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
- G06F16/24—Querying
- G06F16/245—Query processing
- G06F16/2455—Query execution
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/25—Fusion techniques
- G06F18/253—Fusion techniques of extracted features
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- 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)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Databases & Information Systems (AREA)
- Health & Medical Sciences (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Fuzzy Systems (AREA)
- Probability & Statistics with Applications (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention provides a service data retrieval method and an artificial intelligence system combined with a platform database, and relates to the technical field of artificial intelligence. In the invention, based on the first service data retrieval feature and the second service data retrieval feature, outputting a first service data retrieval vector and a second service data retrieval vector; outputting initial fusion data retrieval characteristics based on the first service data retrieval vector and the second service data retrieval vector, and outputting adjustment service data retrieval vectors according to the initial fusion data retrieval characteristics, the first service data retrieval characteristics and the second service data retrieval vector; aggregating the second service data retrieval vector and adjusting the service data retrieval vector to output an aggregated service data retrieval vector; outputting target fusion data retrieval features based on the first service data retrieval vector and the aggregate service data retrieval vector, and performing service data retrieval based on the target fusion data retrieval features. Based on the above, the reliability of service data retrieval can be improved.
Description
Technical Field
The invention relates to the technical field of artificial intelligence, in particular to a service data retrieval method combining a platform database and an artificial intelligence system.
Background
The storage of the required business data into a database for later use is an important technical means. For the service data stored in the database, correlation analysis is generally performed based on a search request of a user, so as to search out the relevant and matched service data, thereby searching the service data. However, in the prior art, it is general to search based on various features in a search request, and then take an intersection or the like to obtain required service data. In this way, a problem that the reliability of the service data retrieval is relatively low easily occurs.
Disclosure of Invention
In view of the above, the present invention aims to provide a service data retrieval method and an artificial intelligence system combined with a platform database, so as to improve the reliability of service data retrieval.
In order to achieve the above purpose, the embodiment of the present invention adopts the following technical scheme:
a business data retrieval method combined with a platform database comprises the following steps:
Performing feature mining on the first service data retrieval feature, outputting a first service data retrieval vector, performing feature mining on the second service data retrieval feature, and outputting a second service data retrieval vector, wherein the first service data retrieval feature and the second service data retrieval feature belong to retrieval features of two different data dimensions in a target service data retrieval request;
the first service data retrieval vector is subjected to feature restoration according to the second service data retrieval vector, initial fusion data retrieval features are output, and the second service data retrieval vector is adjusted according to the difference between the initial fusion data retrieval features and the first service data retrieval features, so that a corresponding adjustment service data retrieval vector is output;
aggregating the second service data retrieval vector and the adjustment service data retrieval vector to output an aggregated service data retrieval vector;
and carrying out feature restoration according to the aggregate service data retrieval vector on the first service data retrieval vector, outputting target fusion data retrieval features, and carrying out service data retrieval in a target platform database based on the target fusion data retrieval features to obtain target service data corresponding to the target service data retrieval request.
In some preferred embodiments, in the above service data retrieval method combined with the platform database, the data dimensions of the first service data retrieval feature, the initial fusion data retrieval feature and the target fusion data retrieval feature are the same, and all belong to first dimension data, and the second service data retrieval feature belongs to second dimension data, wherein the first dimension data comprises text dimension data, and the second dimension data comprises image dimension data or voice dimension data.
In some preferred embodiments, in the above method for searching service data in combination with a platform database, the step of feature mining the first service data searching feature, outputting a first service data searching vector, and feature mining the second service data searching feature, outputting a second service data searching vector includes:
generating interference data features, wherein the interference data features have the same data dimension as the first service data retrieval features;
feature combination is carried out on the first service data retrieval feature and the interference data feature to form a corresponding combined service data retrieval feature;
feature mining is carried out on the combined service data retrieval features, and a first service data retrieval vector corresponding to the first service data retrieval features is output;
And performing feature mining on the second service data retrieval features, and outputting second service data retrieval vectors corresponding to the second service data retrieval features.
In some preferred embodiments, in the above method for searching service data in conjunction with a platform database, the feature restoration of the first service data search vector according to the second service data search vector is performed by using a text data restoration network, where the text data restoration network includes a first number of interference data suppression units and feature restoration units connected in sequence, and the first number is greater than or equal to 2;
the step of restoring the first service data retrieval vector according to the features of the second service data retrieval vector, outputting an initial fusion data retrieval feature, adjusting the second service data retrieval vector according to the difference between the initial fusion data retrieval feature and the first service data retrieval feature, and outputting a corresponding adjusted service data retrieval vector, includes:
using the ith interference data suppression unit in the first number of successively connected interference data suppression units to perform interference suppression processing on the loading data of the ith interference data suppression unit, loading the ith interference suppression data output by the ith interference data suppression unit to the jth interference data suppression unit to perform interference suppression processing, and outputting the jth interference suppression data corresponding to the jth interference data suppression unit, wherein j=i+1; when i is equal to 1, the loading data of the ith interference data suppression unit is the first service data retrieval vector and the second service data retrieval vector, and when i is greater than or equal to or less than the first quantity, the loading data of the ith interference data suppression unit is the interference suppression data output by the previous interference data suppression unit of the ith interference data suppression unit and the second service data retrieval vector;
The feature reduction unit is utilized to perform feature reduction on the interference suppression data output by the last interference data suppression unit, and corresponding initial fusion data retrieval features are output;
and adjusting the second service data retrieval vector according to the difference between the initial fusion data retrieval feature and the first service data retrieval feature, and outputting a corresponding adjustment service data retrieval vector.
In some preferred embodiments, in the service data retrieval method combined with the platform database, the ith interference data suppression unit includes a second number of successively connected focus mining units, where the second number is greater than or equal to 2;
the step of using the ith interference data suppression unit in the first number of interference data suppression units connected in sequence to perform interference suppression processing on the loading data of the ith interference data suppression unit, and loading the ith interference suppression data output by the ith interference data suppression unit to the jth interference data suppression unit to perform interference suppression processing, and outputting the jth interference suppression data corresponding to the jth interference data suppression unit includes:
Utilizing an a-th focusing excavation unit in the i-th interference data suppression unit to perform focusing characteristic excavation on the loading data of the a-th focusing excavation unit and the second service data retrieval vector, outputting a corresponding focusing service data retrieval vector, and marking the corresponding focusing service data retrieval vector as a-th focusing excavation data of the a-th focusing excavation unit in the i-th interference data suppression unit;
loading the a-th focus mining data of the a-th focus mining unit in the i-th interference data suppression unit to the b-th focus mining unit to perform focus feature mining, outputting the b-th focus mining data of the b-th focus mining unit in the i-th interference data suppression unit, wherein when b is equal to 1, the loading data of the a-th focus mining unit is interference suppression data output by a previous interference data suppression unit of the i-th interference data suppression unit, and when b is greater than or equal to 2 and less than the second quantity, the loading data of the a-th focus mining unit is focus mining data output by a previous focus mining unit;
marking the focus mining data output by the last focus mining unit in the ith interference data suppression unit, so that the focus mining data is marked as the ith interference suppression data;
And loading the ith interference suppression data output by the ith interference data suppression unit to the jth interference data suppression unit to perform interference suppression processing, and outputting the jth interference suppression data corresponding to the jth interference data suppression unit.
In some preferred embodiments, in the above method for searching service data in conjunction with a platform database, the step of using an a-th focus mining unit in the i-th interference data suppression unit to perform focus feature mining on the loading data of the a-th focus mining unit and the second service data search vector, outputting a corresponding focus service data search vector, and marking the corresponding focus service data search vector as an a-th focus mining data of an a-th focus mining unit in the i-th interference data suppression unit includes:
carrying out first linear conversion of an attention mechanism on the loading data of the a-th focusing excavation unit, and outputting a corresponding first linear conversion result;
performing a second linear conversion of the attention mechanism on the second service data retrieval vector, and outputting a corresponding second linear conversion result;
performing third linear conversion of an attention mechanism on the second service data retrieval vector, and outputting a corresponding third linear conversion result;
Based on the first linear conversion result and the second linear conversion result, determining similarity parameters between the loading data of the a-th focusing mining unit and the second service data retrieval vector;
and weighting the third linear conversion result based on the similarity parameter, outputting a corresponding focus service data retrieval vector, and marking the focus service data retrieval vector as the a focus mining data of the a focus mining unit in the i interference data suppression unit.
In some preferred embodiments, in the above service data retrieval method combined with the platform database, the feature restoration of the first service data retrieval vector according to the second service data retrieval vector is performed by using a text data restoration network, and the feature restoration of the first service data retrieval vector according to the aggregate service data retrieval vector is performed by using a target neural network, where the target neural network includes the text data restoration network and a plurality of text data analysis networks;
before the step of restoring the first service data retrieval vector according to the feature of the aggregate service data retrieval vector, outputting a target fusion data retrieval feature, and performing service data retrieval in a target platform database based on the target fusion data retrieval feature to obtain target service data corresponding to the target service data retrieval request, the service data retrieval method of the combined platform database further comprises:
Processing the first service data retrieval vector and a target service data retrieval vector in the target neural network to form corresponding intermediate data retrieval features, wherein the target service data retrieval vector is the second service data retrieval vector or the aggregate service data retrieval vector;
and optimizing network parameters of a plurality of text data analysis networks in the target neural network according to the difference between the intermediate data retrieval feature and the first service data retrieval feature to form an optimized target neural network.
In some preferred embodiments, in the service data retrieval method of the combined platform database, the text data reduction network includes a first number of interference data suppression units and feature reduction units that are sequentially connected, where the first number is greater than or equal to 2, the target neural network is formed by setting a text data analysis network for each of the interference data suppression units on the basis of the text data reduction network, each of the interference data suppression units and the corresponding text data analysis network form a corresponding joint processing unit, and a sequential connection relationship between a plurality of the joint processing units is consistent with a sequential connection relationship between a plurality of the interference data suppression units;
The step of processing the first service data retrieval vector and the target service data retrieval vector in the target neural network to form corresponding intermediate data retrieval features comprises the following steps:
the method comprises the steps of carrying out joint interference suppression processing on loading data of an ith joint processing unit by utilizing an ith joint processing unit in a first number of joint processing units which are connected in sequence, loading the ith joint interference suppression data output by the ith joint processing unit to the jth joint processing unit for joint interference suppression processing, and outputting the jth joint interference suppression data of the jth joint processing unit;
performing feature restoration on the joint interference suppression data output by the last joint processing unit, and outputting corresponding intermediate data retrieval features;
and when i is equal to 1, the loading data of the ith joint processing unit is the first service data retrieval vector and the target service data retrieval vector, and when i is more than or equal to 2 and less than the first quantity, the loading data of the ith joint processing unit is the joint interference suppression data output by the previous joint processing unit and the target service data retrieval vector.
In some preferred embodiments, in the above service data retrieval method combined with the platform database, the step of aggregating the second service data retrieval vector and the adjustment service data retrieval vector to output an aggregate service data retrieval vector includes:
performing vector decimation according to a first size on the second service data retrieval vector to form a corresponding first size decimated vector, wherein the first size decimated vector comprises vector local data of a first size with a front distribution coordinate in the second service data retrieval vector;
performing vector decimation according to a second size on the adjustment service data retrieval vector to form a corresponding second-size decimated vector, wherein the second-size decimated vector comprises vector local data of a second size with a front distribution coordinate in the adjustment service data retrieval vector;
and cascading the second size decimation vector to the tail end position of the first size decimation vector to form a corresponding aggregate service data retrieval vector.
The embodiment of the invention also provides an artificial intelligence system, which comprises a processor and a memory, wherein the memory is used for storing a computer program, and the processor is used for executing the computer program so as to realize the service data retrieval method combined with the platform database.
The service data retrieval method and the artificial intelligence system combined with the platform database provided by the embodiment of the invention can output a first service data retrieval vector and a second service data retrieval vector based on the first service data retrieval feature and the second service data retrieval feature; outputting initial fusion data retrieval characteristics based on the first service data retrieval vector and the second service data retrieval vector, and outputting adjustment service data retrieval vectors according to the initial fusion data retrieval characteristics, the first service data retrieval characteristics and the second service data retrieval vector; aggregating the second service data retrieval vector and adjusting the service data retrieval vector to output an aggregated service data retrieval vector; outputting target fusion data retrieval features based on the first service data retrieval vector and the aggregate service data retrieval vector, and performing service data retrieval based on the target fusion data retrieval features. Based on the foregoing, before service data retrieval, the retrieval features of two different data dimensions in the target service data retrieval request are fully fused, so that the obtained target fusion data retrieval feature can effectively represent the target service data retrieval request, and therefore, the reliability of service data retrieval based on the target fusion data retrieval feature can be ensured, and the problem of relatively low retrieval reliability in the prior art is solved.
In order to make the above objects, features and advantages of the present invention more comprehensible, preferred embodiments accompanied with figures are described in detail below.
Drawings
FIG. 1 is a block diagram of an artificial intelligence system according to an embodiment of the present invention.
Fig. 2 is a flowchart illustrating steps included in a service data retrieval method combined with a platform database according to an embodiment of the present invention.
Fig. 3 is a schematic diagram of each module included in the service data retrieval device combined with the platform database according to the embodiment of the present invention.
Description of the embodiments
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, but not all embodiments of the present invention. The components of the embodiments of the present invention generally described and illustrated in the figures herein may be arranged and designed in a wide variety of different configurations.
Thus, the following detailed description of the embodiments of the invention, as presented in the figures, 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 those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
As shown in FIG. 1, an embodiment of the present invention provides an artificial intelligence system. Wherein the artificial intelligence system may include a memory and a processor.
In detail, the memory and the processor are electrically connected directly or indirectly to realize transmission or interaction of data. For example, electrical connection may be made to each other via one or more communication buses or signal lines. The memory may store at least one software functional module (computer program) that may exist in the form of software or firmware. The processor may be configured to execute the executable computer program stored in the memory, thereby implementing the service data retrieval method combined with the platform database according to the embodiment of the present invention (as described below).
Alternatively, in some embodiments, the Memory may be, but is not limited to, random access Memory (Random Access Memory, RAM), read Only Memory (ROM), programmable Read Only Memory (Programmable Read-Only Memory, PROM), erasable Read Only Memory (Erasable Programmable Read-Only Memory, EPROM), electrically erasable Read Only Memory (Electric Erasable Programmable Read-Only Memory, EEPROM), and the like.
Alternatively, in some embodiments, the processor may be a general purpose processor, including a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP), a System on Chip (SoC), etc.; but also Digital Signal Processors (DSPs), application Specific Integrated Circuits (ASICs), field Programmable Gate Arrays (FPGAs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.
Alternatively, in some embodiments, the artificial intelligence system may be a server with data processing capabilities.
With reference to fig. 2, the embodiment of the invention further provides a service data retrieval method combined with a platform database, which can be applied to the artificial intelligence system. The method steps defined by the flow related to the service data retrieval method combined with the platform database can be realized by the artificial intelligence system.
The specific flow shown in fig. 2 will be described in detail.
Step S110, the first service data retrieval feature is subjected to feature mining, a first service data retrieval vector is output, and the second service data retrieval feature is subjected to feature mining, and a second service data retrieval vector is output.
In the embodiment of the invention, the artificial intelligence system can perform feature mining on the first service data retrieval feature to output a first service data retrieval vector, and perform feature mining on the second service data retrieval feature to output a second service data retrieval vector. The first service data retrieval feature and the second service data retrieval feature belong to retrieval features of two different data dimensions in a target service data retrieval request, such as text dimension, image dimension or voice dimension, so that multi-dimensional data retrieval can be realized, and the retrieval features are richer, for example, for a product providing service, a user can retrieve the product based on images of similar products and text description data of the product at the same time, namely, a target product is retrieved, feature mining can be a quantization-oriented process, namely, a first service data retrieval vector is used for representing the first service data retrieval feature, and a second service data retrieval vector is used for representing the second service data retrieval feature.
Step S120, performing feature restoration according to the second service data search vector on the first service data search vector, outputting an initial fusion data search feature, adjusting the second service data search vector according to the difference between the initial fusion data search feature and the first service data search feature, and outputting a corresponding adjusted service data search vector.
In the embodiment of the invention, the artificial intelligence system can restore the first service data retrieval vector according to the characteristics of the second service data retrieval vector, output initial fusion data retrieval characteristics, adjust the second service data retrieval vector according to the difference between the initial fusion data retrieval characteristics and the first service data retrieval characteristics, and output corresponding adjustment service data retrieval vectors. That is, when the feature recovery (decoding, reconstruction, etc. may be understood as being the inverse of the feature mining process) is performed on the first service data retrieval vector, the second service data retrieval vector may be combined, so that not only the first service data retrieval vector but also the second service data retrieval vector may be reflected in the obtained initial fusion data retrieval feature. In addition, the second service data retrieval vector is adjusted according to the difference between the initial fusion data retrieval feature and the first service data retrieval feature, and in fact, the second service data retrieval vector is finely adjusted, so that cognition of the initial first service data retrieval feature in the processing process is guaranteed, and consistency in the processing process is guaranteed.
Step S130, the second service data retrieval vector and the adjustment service data retrieval vector are aggregated, and an aggregated service data retrieval vector is output.
In the embodiment of the invention, the artificial intelligence system can aggregate the second service data retrieval vector and the adjustment service data retrieval vector to output an aggregate service data retrieval vector. In this way, the second service data retrieval vector and the adjustment service data retrieval vector are aggregated, so that the aggregated service data retrieval vector can be matched with the first service data retrieval feature while representing the second service data retrieval feature.
Step S140, performing feature restoration according to the aggregate service data search vector on the first service data search vector, outputting a target fusion data search feature, and performing service data search in a target platform database based on the target fusion data search feature to obtain target service data corresponding to the target service data search request.
In the embodiment of the invention, the artificial intelligence system can restore the first service data retrieval vector according to the characteristics of the aggregate service data retrieval vector, output target fusion data retrieval characteristics, and perform service data retrieval in a target platform database based on the target fusion data retrieval characteristics to obtain target service data corresponding to the target service data retrieval request. Based on the method, the formed target fusion data retrieval feature can be ensured to effectively represent the semantics of the first service data retrieval feature and the second service data retrieval feature. The searching of the service data in the target platform database based on the target fusion data searching feature may mean that, among the service data included in the target platform database, the target service data matched with the target fusion data searching feature (for example, the service data with the largest matching degree or the service data with the matching degree greater than the preset matching degree) is found, and the specific searching process may refer to the related prior art.
For example:
in one practical application scenario, it is assumed that there is a product search system that contains text descriptions and images as two distinct data dimensions. The user may search by entering a textual description of the product and uploading an image of the product. In this case: first service data retrieval feature mining: for text description data, natural language processing techniques can be used to extract features of keywords, phrases, or sentences and convert them into a vector representation; and (3) mining the second service data retrieval characteristics: for product image data, computer vision techniques may be used to extract features of the image, such as extracting feature vectors of the image using convolutional neural networks. Through the two steps, a first service data retrieval vector and a second service data retrieval vector can be obtained and are respectively used for representing the characteristics of text description and image data; in the above example, a more comprehensive retrieval feature is obtained by fusing the text description and the image data. The method comprises the following specific steps: feature restoration using the second traffic data retrieval vector: using the second service data retrieval vector as a guide, and combining the first service data retrieval vector to obtain an initial fusion data retrieval feature; adjusting according to the difference between the initial fusion data retrieval feature and the first service data retrieval feature: the second business data retrieval vector is finely adjusted by comparing the difference between the initial fusion data retrieval feature and the first business data retrieval feature, and a corresponding adjustment business data retrieval vector is generated; thus, the initial fusion data retrieval feature which fuses the text description and the image data can be obtained, and the second service data retrieval vector is adjusted. Thus, the final retrieval feature can reflect the features of the first service data and fully consider the features of the second service data. By aggregating the second service data retrieval vector and the adjustment service data retrieval vector, an aggregate service data retrieval vector can be obtained which comprehensively considers the characteristics of the text description and the image data. Based on this, further, final service data retrieval can be performed by aggregating the service data retrieval vectors. The method comprises the following specific steps: first service data retrieval vector reduction: according to the aggregate service data retrieval vector and the adjustment service data retrieval vector, performing feature restoration operation; target fusion data retrieval feature: obtaining target fusion data retrieval characteristics from the restored first service data retrieval vector; and (5) searching target service data: and based on the target fusion data retrieval characteristics, service data retrieval is carried out in a target platform database, and target service data matched with the target fusion data retrieval characteristics is found. Through the steps, the text description input by the user and the uploaded image data can be fused, and a target fusion data retrieval feature is obtained. Then, service data retrieval can be performed in the target platform database, and target service data meeting the target service data retrieval request is found. In summary, through the processing of these steps, the data features of different dimensions can be effectively fused, and more accurate and comprehensive service data retrieval can be realized. The method is beneficial to improving the problem of relatively low reliability in the prior art and improving the accuracy and reliability of service data retrieval.
An example of an electronic product is provided as a product search system. In the system, the first business data is a text description of the electronic product and the second business data is an image of the electronic product.
Step S110, feature mining:
first service data retrieval feature mining: for text description of electronic products, natural language processing techniques are used to extract features of keywords, make, model, function, etc. from the text and convert it into a vector representation. For example, for the text description "XXX Pro 13 inch notebook computer, 256GB store, liquid crystal display", keywords [ "XXX", "Pro", "13 inch", "256GB", "liquid crystal display" ] may be extracted and converted into vector representations. And (3) mining the second service data retrieval characteristics: for images of electronic products, feature vectors of the images are extracted using computer vision techniques, such as Convolutional Neural Networks (CNNs). For example, the pre-trained CNN model is used to extract the characteristics of the electronic product image, and a vector representing the characteristics of the image is obtained.
Step S120-feature reduction and adjustment:
feature reduction: based on the text description vector and the image feature vector obtained in step S110, feature restoration operations are performed to combine them into an initial fusion data retrieval feature (e.g., color feature and shape feature are added, or some features such as size feature are corrected, etc.). For example, the initial fused data retrieval feature may be derived by concatenating the text description vector and the image feature vector. Adjusting a service data retrieval vector: and fine-tuning the second service data retrieval vector according to the difference between the initial fusion data retrieval feature and the first service data retrieval feature, and outputting and adjusting the service data retrieval vector. For example, by comparing the differences of the initial fused data retrieval feature with the first business data retrieval feature, the image feature vector is adjusted to better match the features of the first business data.
Step S130-polymerization:
aggregate traffic data retrieval vector: and aggregating the first service data retrieval vector and the adjusted second service data retrieval vector to obtain an aggregated service data retrieval vector. Aggregation may be achieved, for example, by simply stitching or weighted averaging the two vectors.
Step S140, target data retrieval:
and (5) searching target service data: and searching the service data in the target electronic product database based on the aggregate service data search vector, and finding the target electronic product data matched with the aggregate service data search vector. For example, using the aggregate business data retrieval vector as a query condition, the electronic product data closest to the vector is looked up in the database. Through the steps, text description and image data can be fused, and accurate product search (such as 'YYY Pro 13 inch notebook computer, 256GB storage, liquid crystal display screen and silver gray') can be performed in a target electronic product database by utilizing the aggregated service data retrieval vector. In this way, the user can find an electronic product meeting his needs by entering a text description and uploading an image.
The above is only an example and other products or content may be retrieved.
Based on the foregoing, before service data retrieval, the retrieval features of two different data dimensions in the target service data retrieval request are fully fused, so that the obtained target fusion data retrieval feature can effectively represent the target service data retrieval request, and therefore, the reliability of service data retrieval based on the target fusion data retrieval feature can be ensured, and the problem of relatively low retrieval reliability in the prior art is solved.
Optionally, in some embodiments, the data dimensions of the first business data retrieval feature, the initial fusion data retrieval feature, and the target fusion data retrieval feature are the same, e.g., all belong to first dimension data, and the second business data retrieval feature belongs to second dimension data, e.g., the first dimension data comprises text dimension data, and the second dimension data comprises image dimension data or voice dimension data.
Optionally, in some embodiments, step S110 described above may include:
generating interference data features, wherein the interference data features and the first service data retrieval features have the same data dimension, such as all belong to text data, and the interference data features can be randomly generated text data, and the interference data features can be used for improving the overfitting to a certain extent;
Feature combination is performed on the first service data retrieval feature and the interference data feature to form a corresponding combined service data retrieval feature, and illustratively, the first service data retrieval feature and the interference data feature can be respectively embedded to obtain two corresponding embedded features, and then the two embedded features can be subjected to processing such as superposition, so that feature combination can be realized, and the corresponding combined service data retrieval feature can be obtained; alternatively, in some embodiments, the interference data feature may be some characters, and the characters in the first service data retrieval feature may be directly replaced by the characters to implement feature merging;
feature mining is carried out on the combined service data retrieval features, a first service data retrieval vector corresponding to the first service data retrieval features is output, the feature mining can be realized through a convolutional neural network or a feature extraction network, the convolutional neural network or the feature extraction network can comprise a plurality of convolutional kernels and the like, so that deep mining of the features is realized, and the first service data retrieval vector is obtained;
and performing feature mining on the second service data retrieval features, and outputting second service data retrieval vectors corresponding to the second service data retrieval features, as described in the previous related description.
For example:
the following first business data retrieval feature describes the electronic product and the following randomly generated text of the interference data feature description:
first service data retrieval feature: "This smartphone features a high-resolution display and a powerful processor";
interference data characteristics: "Randomly generated text for interference";
for this example, the embedded features may be combined as follows:
building a vocabulary table: likewise, a vocabulary containing all possible words needs to be built;
selecting an embedding model: selecting an appropriate text embedding model (e.g., word2Vec, gloVe, or BERT) to generate an embedding vector for each Word;
embedding the text: the first traffic data retrieval feature and the interference data feature are converted into their respective embedded feature sequences using the selected embedded model. For example:
"This smartphone features a high-resolution display and a powerful processor" can be converted to a sequence having multiple embedded vectors, such as [ [0.12, 0.56, -0.23, 0.89], [0.78, -0.45, 0.67, -0.91], ];
"Randomly generated text for interference" can be converted to another sequence with multiple embedded vectors, such as [ [0.34, -0.67, 0.91, 0.12], [0.56, 0.23, 0.78, -0.45], ];
Feature combination: and combining the embedded features corresponding to the first service data retrieval features and the embedded features corresponding to the interference data features. One approach is to operate on two embedded feature sequences according to element position, e.g., add or join them together, resulting in a combined embedded feature vector. Thus, feature combination of the first service data retrieval feature and the interference data feature is realized. For example, add [ [0.12, 0.56, -0.23, 0.89], [0.78, -0.45, 0.67, -0.91] ] and [ [0.34, -0.67, 0.91, 0.12], [0.56, 0.23, 0.78, -0.45] ] to get a combined embedded feature sequence: [[0.46, -0.11, 0.68, 1.01], [1.34, -0.22, 1.45, -1.36],...].
Optionally, in some embodiments, the feature restoration of the first service data search vector according to the second service data search vector is performed by using a text data restoration network, where the text data restoration network includes a first number of interference data suppression units and feature restoration units connected in series, such as an output of the first interference data suppression unit and an input connection of the second interference data suppression unit, an output of the second interference data suppression unit and an input connection of the third interference data suppression unit, and the first number is greater than or equal to 2, where the step S120 may include:
The method comprises the steps of performing interference suppression processing (overlapping interference data is arranged in front of a first number of interference data suppression units, so that interference suppression needs to be performed firstly, namely interference information is removed) on loading data of the ith interference data suppression unit by utilizing the ith interference data suppression unit in a first number of interference data suppression units which are connected in sequence, loading the ith interference data output by the ith interference data suppression unit to the jth interference data suppression unit to perform interference suppression processing, and outputting jth interference suppression data corresponding to the jth interference data suppression unit, wherein j=i+1; when i is equal to 1, the loading data of the ith interference data suppression unit is the first service data retrieval vector and the second service data retrieval vector, and when i is greater than or equal to or less than the first quantity, the loading data of the ith interference data suppression unit is the interference suppression data output by the previous interference data suppression unit of the ith interference data suppression unit and the second service data retrieval vector, that is, the second service data retrieval vector is combined, interference suppression is carried out on the characteristics of the first dimension, so that the basis of interference suppression is more sufficient, and the reliability of interference suppression is ensured; in addition, the output of the previous interference data suppression unit is suppressed by the next interference data suppression unit, so that cascade of interference suppression processing can be realized, and the accuracy of the interference suppression processing can be further improved;
The feature reduction unit is utilized to perform feature reduction on the interference suppression data output by the last interference data suppression unit, and corresponding initial fusion data retrieval features are output; the feature reduction unit may be a decoding part in an encoded neural network, so that the interference suppression data in the form of vectors may be mapped (generated) into the initial fused data retrieval feature in the form of text data;
according to the difference between the initial fusion data retrieval feature and the first service data retrieval feature, the second service data retrieval vector is adjusted, and a corresponding adjustment service data retrieval vector is output; illustratively, a distance (such as an edit distance, a cosine distance, etc.) between the initial fusion data retrieval feature and the first service data retrieval feature may be calculated, then the distance may be normalized to obtain a corresponding normalized value, then the second service data retrieval vector may be weighted based on the normalized value, and finally the obtained weighted vector and the second service data retrieval vector may be superimposed to obtain an adjusted service data retrieval vector.
Optionally, in some embodiments, the ith interference data suppression unit includes a second number of successively connected focus mining units, where the second number is greater than or equal to 2, based on which, the step of using the ith interference data suppression unit in the first number of successively connected interference data suppression units to perform interference suppression processing on the loading data of the ith interference data suppression unit, and loading the ith interference suppression data output by the ith interference data suppression unit to the jth interference data suppression unit to perform interference suppression processing, and outputting the jth interference suppression data corresponding to the jth interference data suppression unit may include:
the loading data of the a-th focusing mining unit and the second service data retrieval vector are subjected to focusing feature mining by utilizing the a-th focusing mining unit in the i-th interference data suppression unit, a corresponding focusing service data retrieval vector is output and marked as the a-th focusing mining data of the a-th focusing mining unit in the i-th interference data suppression unit, so that the focusing service data retrieval vector can simultaneously carry the loading data of the a-th focusing mining unit and the second service data retrieval vector, and the represented semantic is better;
Loading the a-th focus mining data of the a-th focus mining unit in the i-th interference data suppression units to the b-th focus mining unit to perform focus feature mining, outputting the b-th focus mining data of the b-th focus mining unit in the i-th interference data suppression units, wherein b=a+1, that is, further mining the focus mining data output by the previous focus mining unit through the next focus mining unit, when b is equal to 1, the loading data of the a-th focus mining unit is the interference suppression data output by the previous interference data suppression unit of the i-th interference data suppression unit, and when b is greater than or equal to 2 and less than the second quantity, the loading data of the a-th focus mining unit is the focus mining data output by the previous focus mining unit;
marking the focus mining data output by the last focus mining unit in the ith interference data suppression unit, so that the focus mining data is marked as the ith interference suppression data; that is, the focus mining data output by the last focus mining unit in the i-th interference data suppression unit may be taken as the output of the i-th interference data suppression unit;
The ith interference suppression data output by the ith interference data suppression unit is loaded to the jth interference data suppression unit to perform interference suppression processing, and the jth interference suppression data corresponding to the jth interference data suppression unit is output; illustratively, the 1 st interference suppression data is obtained by performing interference suppression on the first service data search vector and the second service data search vector by using the 1 st interference data suppression unit, the 2 nd interference suppression data is obtained by performing interference suppression on the 1 st interference suppression data and the second service data search vector by using the 2 nd interference data suppression unit, and the 3 rd interference suppression data is obtained by performing interference suppression on the 2 nd interference suppression data and the second service data search vector by using the 3 rd interference data suppression unit.
For example:
the first step: and (3) utilizing an a-th focusing mining unit in the i-th interference data suppression unit to carry out focusing feature mining on the loading data and the second service data retrieval vector. Assuming that the ith interference data suppression unit is the 3 rd one, the a-th focusing and mining unit is the 2 nd one, the loading data is [0.5, 0.4, 0.2], and the second service data retrieval vector is [0.8, 0.7, 0.9]. And generating the 2 nd focus mining data of the 2 nd focus mining unit in the 3 rd interference data suppression unit, wherein the corresponding focus service data retrieval vector is [0.6, 0.55 and 0.3 ].
And a second step of: the 2 nd focus mining data of the 2 nd focus mining unit in the 3 rd interference data suppression unit is loaded to the 3 rd focus mining unit for further focus feature mining. Assume that the focus mining data output by the previous focus mining unit is [0.6, 0.55, 0.3]. By further mining, the 3 rd focus mining unit outputs focus mining data of [0.45, 0.4, 0.25]. Thus, in the first example, the focus traffic data retrieval vector of the 2 nd focus mining unit of the 3 rd interference data suppression unit carries information of both the loading data [0.5, 0.4, 0.2] and the second traffic data retrieval vector [0.8, 0.7, 0.9 ]. In the second example, feature mining is performed by successive focus mining units, layer-by-layer extraction and processing, and finally focus mining data [0.45, 0.4, 0.25] of the 3 rd focus mining unit in the 3 rd interference data suppression unit is obtained.
Optionally, in some embodiments, the step of performing focus feature mining on the loading data of the a-th focus mining unit and the second service data retrieval vector by using the a-th focus mining unit in the i-th interference data suppression unit, outputting a corresponding focus service data retrieval vector, and marking the corresponding focus service data retrieval vector as the a-th focus mining data of the a-th focus mining unit in the i-th interference data suppression unit may include:
Performing first linear conversion of the attention mechanism on the loading data of the a-th focusing and mining unit, outputting a corresponding first linear conversion result, and illustratively multiplying the loading data of the a-th focusing and mining unit by a first parameter matrix included in the focusing and mining unit to obtain a first linear conversion result (Query Vector) for specifying information of interest or a target of Query;
performing a second linear conversion of the attention mechanism on the second service data retrieval Vector, outputting a corresponding second linear conversion result, and illustratively multiplying the second service data retrieval Vector by a second parameter matrix included in the focusing mining unit to obtain a second linear conversion result (Key Vector) for representing an identifier or index of input data;
performing a third linear conversion of the attention mechanism on the second service data search Vector, outputting a corresponding third linear conversion result, and illustratively multiplying the second service data search Vector by a third parameter matrix included in the focus mining unit to obtain a third linear conversion result (Value Vector) containing specific information or features related to the input data;
Based on the first linear conversion result and the second linear conversion result, determining a similarity parameter between the loading data of the a-th focusing mining unit and the second service data retrieval vector, for example, multiplying the transposed data of the first linear conversion result and the second linear conversion result to obtain the similarity parameter;
and weighting the third linear conversion result based on the similarity parameter, namely multiplying, outputting a corresponding focus service data retrieval vector, and marking the vector as the a focus mining data of the a focus mining unit in the i interference data suppression unit.
Optionally, in some embodiments, step S130 described above may include:
performing vector decimation according to a first size on the second service data retrieval vector to form a corresponding first size decimated vector, wherein the first size decimated vector comprises vector local data of a first size with a front distribution coordinate in the second service data retrieval vector; performing vector decimation according to a second size on the adjustment service data retrieval vector to form a corresponding second-size decimated vector, wherein the second-size decimated vector comprises vector local data of a second size with a front distribution coordinate in the adjustment service data retrieval vector; and cascading the second size decimation vector to the tail end position of the first size decimation vector to form a corresponding aggregate service data retrieval vector.
That is, the adjusted service data search vector and the unadjusted second service data search vector are spliced to generate a final aggregate service data search vector, and the splicing rule may be that the unadjusted second service data search vector is located before the adjusted second service data search vector, and if a part of the adjusted second service data search vector and the unadjusted second service data search vector is spliced to form the final aggregate service data search vector, for example, for 128 vectors as a size of subsequent processing (i.e., a size of the aggregate service data search vector), the first 64 vectors of the unadjusted second service data search vector are spliced and combined with the first 64 vectors of the adjusted second service data search vector. Wherein the reason for placing the unadjusted second traffic data retrieval vector in front is that the unadjusted second traffic data retrieval vector retains more of the original information, whereas the adjusted traffic data retrieval vector represents the original first traffic data retrieval feature, so that the aggregate traffic data retrieval vector may be placed in front in order to make it more prone to the original second traffic data retrieval feature.
For example:
the second service data retrieval vector: [0.2, 0.5, 0.4, 0.1]; adjusting a service data retrieval vector: [0.6, 0.3, 0.9, 0.7]; the corresponding first-size decimated vector may be formed according to a first-size vector decimation, where the first-size vector local data with the first-size distribution coordinates preceding the second service data retrieval vector is included. Assuming that the first size is 2, the first size decimation vector is: [0.2, 0.5]; and forming a corresponding second-size decimated vector according to the second-size vector decimation, wherein the second-size vector decimation comprises adjusting the vector local data of the second size with the prior distribution coordinates in the service data retrieval vector. Assuming that the second size is 3, the second size decimation vector is: [0.6, 0.3, 0.9]; finally, cascading the second size decimation vector to the end position of the first size decimation vector to form a corresponding aggregate service data retrieval vector: [0.2, 0.5, 0.6, 0.3, 0.9].
Optionally, in some embodiments, the feature restoration of the first service data retrieval vector according to the second service data retrieval vector is performed by using a text data restoration network, and the feature restoration of the first service data retrieval vector according to the aggregate service data retrieval vector is performed by using a target neural network, where the target neural network includes the text data restoration network and a plurality of text data analysis networks (that is, the latter feature restoration operation further requires processing by a plurality of text data analysis networks compared to the former feature restoration operation, so that the accuracy of the latter feature restoration operation may be higher), and before the step S140, the service data retrieval method of the combined platform database may further include:
Processing the first service data retrieval vector and a target service data retrieval vector in the target neural network to form corresponding intermediate data retrieval features, wherein the target service data retrieval vector is the second service data retrieval vector or the aggregate service data retrieval vector; that is, the first service data retrieval vector and the second service data retrieval vector may be processed in the target neural network to form corresponding intermediate data retrieval features; the first service data retrieval vector and the aggregate service data retrieval vector can be processed in the target neural network to form corresponding intermediate data retrieval characteristics;
and according to the difference between the intermediate data retrieval feature and the first service data retrieval feature, optimizing network parameters of a plurality of text data analysis networks in the target neural network to form an optimized target neural network, wherein the optimized target neural network can be used for executing the characteristic reduction of the first service data retrieval vector according to the aggregate service data retrieval vector in the step S140, and outputting a target fusion data retrieval feature.
That is, the target neural network is further connected with a plurality of text data analysis networks in sequence on the basis of the text data recovery network, so that the target neural network can be trained before executing the step S140, only the plurality of text data analysis networks can be updated in the training process, parameters of the text data recovery network are fixed, and the text data recovery network can be the text data recovery network after training based on the difference between the initial fusion data retrieval feature and the first service data retrieval feature determined in the step S120, so that training is not needed.
Optionally, in some embodiments, the text data reduction network includes a first number of interference data suppression units and feature reduction units that are sequentially connected, where the first number is greater than or equal to 2, the target neural network is formed by setting a text data analysis network for each of the interference data suppression units on the basis of the text data reduction network, each of the interference data suppression units and the corresponding text data analysis network form a corresponding joint processing unit, and a sequential connection relationship between a plurality of the joint processing units is consistent with a sequential connection relationship between a plurality of the interference data suppression units. That is, the target neural network may include a first number of joint processing units and a feature recovery unit that are sequentially connected, based on which the steps of processing the first service data retrieval vector and the target service data retrieval vector in the target neural network to form corresponding intermediate data retrieval features may include:
The method comprises the steps of carrying out joint interference suppression processing on loading data of an ith joint processing unit by utilizing an ith joint processing unit in a first number of joint processing units which are connected in sequence, loading the ith joint interference suppression data output by the ith joint processing unit to the jth joint processing unit for joint interference suppression processing, and outputting the jth joint interference suppression data of the jth joint processing unit, namely, loading the output of the former joint processing unit to the latter joint processing unit for further joint interference suppression processing;
performing feature restoration on the joint interference suppression data output by the last joint processing unit, outputting corresponding intermediate data retrieval features, and realizing by using the feature restoration unit;
when i is equal to 1, the loading data of the i-th joint processing unit (i.e. the 1 st joint processing unit) is the first service data retrieval vector and the target service data retrieval vector, and when i is greater than or equal to 2 and less than the first quantity, the loading data of the i-th joint processing unit (i.e. the joint processing units other than the 1 st joint processing unit) is the joint interference suppression data output by the previous joint processing unit and the target service data retrieval vector.
That is, in the foregoing embodiments, the second service data search vector or the aggregate service data search vector is continuously subjected to association mining, so as to achieve fusion of association information, thereby reducing the influence of interference data.
Optionally, in some embodiments, the ith joint processing unit includes a plurality of feature extraction units (such as Feature Extraction Network), a plurality of reconstruction units (Reconstruction Network), and an ith text data analysis network corresponding to the ith interference data suppression unit (the interference data suppression unit and the text data analysis network are a functional overview, and specific functions are as follows), based on this, the step of using the ith joint processing unit of the first number of successively connected joint processing units to perform joint interference suppression processing on the loaded data of the ith joint processing unit, and loading the ith joint interference suppression data output by the ith joint processing unit to the jth joint processing unit to perform joint interference suppression processing, and outputting the jth joint interference suppression data of the jth joint processing unit may include:
Using the ith text data analysis network to perform association mining on the first service data retrieval vector and the target service data retrieval vector, and outputting a corresponding association mining vector, namely, the association mining vector can represent semantic features of the first service data retrieval vector and semantic features of the target service data retrieval vector;
performing feature extraction, such as convolution operation, on the target service data retrieval vector and the associated mining vector by using the feature extraction unit, and outputting a corresponding feature extraction result vector;
reconstructing the feature extraction result vector by using the reconstruction unit, for example, deconvolution operation, and outputting corresponding ith joint interference suppression data;
and loading the ith joint interference suppression data output by the ith joint processing unit to the jth joint processing unit to perform joint interference suppression processing, and outputting the jth joint interference suppression data of the jth joint processing unit, namely, performing further joint interference suppression processing on the joint interference suppression data output by the previous joint processing unit by utilizing the latter joint processing unit.
Optionally, in some embodiments, the ith text data analysis network includes a third number of focus mining units connected sequentially, where the third number is greater than or equal to 2, based on which, the step of performing association mining on the first service data retrieval vector and the target service data retrieval vector by using the ith text data analysis network, and outputting a corresponding association mining vector may include:
loading the x-th focus mining data of the x-th focus mining unit in the i-th text data analysis network to the y-th focus mining unit for focus feature mining, outputting the y-th focus mining data of the y-th focus mining unit in the corresponding i-th text data analysis network, wherein y=x+1, that is, using the latter focus mining unit for further focus feature mining on the focus mining data output by the former focus mining unit; wherein, the focusing characteristic digging process of the focusing digging unit can refer to the related description;
marking the focus mining data output by each focus mining unit as an associated mining vector; when x is equal to 1, the loading data of the xth focus mining unit (i.e. the first focus mining unit) is the joint interference suppression data output by the previous joint processing unit, and when x is greater than or equal to 2 and smaller than the third number, the loading data of the xth focus mining unit (i.e. the focus mining unit behind the first focus mining unit) is the focus mining data output by the previous focus mining unit in the ith text data analysis network.
Optionally, in some embodiments, the feature extraction unit may include a third number of successively connected focus mining units, based on which the step of performing feature extraction on the target service data retrieval vector and the associated mining vector by using the feature extraction unit, and outputting a corresponding feature extraction result vector may include:
utilizing an x-th focusing and mining unit in the feature extraction unit to perform focusing feature mining on the loading data of the x-th focusing and mining unit and the target service data retrieval vector, and outputting a corresponding focusing service data retrieval vector, as described above;
aggregating, such as superposing or splicing, the focus service data retrieval vector and the x focus mining data output by the x focus mining unit in the i text data analysis network to form the x focus mining data of the x focus mining unit in the feature extraction unit;
loading the x-th focus mining data of the x-th focus mining unit in the feature extraction units to the y-th focus mining unit in the feature extraction units to mine the y-th focus mining data of the y-th focus mining unit in the feature extraction units, as described above;
Marking the x-th focus mining data output by the x-th focus mining unit in the feature extraction unit, so that the x-th focus mining data is marked as a feature extraction result vector;
when x is equal to 1, the loading data of the xth focusing and excavating unit (i.e. the 1 st focusing and excavating unit) is the combined interference suppression data output by the previous combined processing unit, and when x is greater than or equal to 2 and smaller than the third quantity, the loading data of the xth focusing and excavating unit (i.e. each focusing and excavating unit behind the 1 st focusing and excavating unit) is the focusing and excavating data output by the previous focusing and excavating unit, so that cascade focusing and excavating are realized, and the excavating precision is ensured.
For example, performing focus feature mining on the input of the 1 st focus mining unit and the target service data retrieval vector through the 1 st focus mining unit of the feature extraction unit to obtain a focus service data retrieval vector; carrying out fusion processing, such as superposition, on the focus service data retrieval vector and the 1 st focus mining data output by the 1 st focus mining unit in the i text data analysis network to obtain the 1 st focus mining data of the 1 st focus mining unit in the feature extraction unit; loading the 1 st focus mining data of the 1 st focus mining unit in the feature extraction units to the 2 nd focus mining unit in the feature extraction units to obtain the 2 nd focus mining data of the 2 nd focus mining unit in the feature extraction units; taking the 3 rd focus mining data output by the 3 rd focus mining unit in the feature extraction unit as a feature extraction result vector, namely, the focus service data retrieval vector output by each focus mining unit of the feature extraction unit of the i-th joint processing unit is fused with the output of the corresponding focus mining unit in the i-th text data analysis network, and loading the fusion result to the next focus mining unit of the feature extraction unit for processing.
With reference to fig. 3, the embodiment of the invention also provides a service data retrieval device combined with the platform database, which can be applied to the artificial intelligence system. The service data retrieval device combined with the platform database can comprise:
the search feature mining module is used for carrying out feature mining on the first service data search feature, outputting a first service data search vector, carrying out feature mining on the second service data search feature, and outputting a second service data search vector, wherein the first service data search feature and the second service data search feature belong to search features with two different data dimensions in a target service data search request;
the retrieval vector adjustment module is used for restoring the first service data retrieval vector according to the characteristics of the second service data retrieval vector, outputting initial fusion data retrieval characteristics, adjusting the second service data retrieval vector according to the differences between the initial fusion data retrieval characteristics and the first service data retrieval characteristics and outputting corresponding adjustment service data retrieval vectors;
the search vector aggregation module is used for aggregating the second service data search vector and the adjustment service data search vector and outputting an aggregate service data search vector;
And the service data retrieval module is used for carrying out feature restoration according to the aggregate service data retrieval vector on the first service data retrieval vector, outputting target fusion data retrieval features, and carrying out service data retrieval in a target platform database based on the target fusion data retrieval features to obtain target service data corresponding to the target service data retrieval request.
In summary, the service data retrieval method and the artificial intelligence system combined with the platform database provided by the invention can output the first service data retrieval vector and the second service data retrieval vector based on the first service data retrieval feature and the second service data retrieval feature; outputting initial fusion data retrieval characteristics based on the first service data retrieval vector and the second service data retrieval vector, and outputting adjustment service data retrieval vectors according to the initial fusion data retrieval characteristics, the first service data retrieval characteristics and the second service data retrieval vector; aggregating the second service data retrieval vector and adjusting the service data retrieval vector to output an aggregated service data retrieval vector; outputting target fusion data retrieval features based on the first service data retrieval vector and the aggregate service data retrieval vector, and performing service data retrieval based on the target fusion data retrieval features. Based on the foregoing, before service data retrieval, the retrieval features of two different data dimensions in the target service data retrieval request are fully fused, so that the obtained target fusion data retrieval feature can effectively represent the target service data retrieval request, and therefore, the reliability of service data retrieval based on the target fusion data retrieval feature can be ensured, and the problem of relatively low retrieval reliability in the prior art is solved.
The above description is only of the preferred embodiments of the present invention and is not intended to limit the present invention, but various modifications and variations can be made to the present invention by those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (7)
1. A service data retrieval method combined with a platform database is characterized by comprising the following steps:
performing feature mining on the first service data retrieval feature, outputting a first service data retrieval vector, performing feature mining on the second service data retrieval feature, and outputting a second service data retrieval vector, wherein the first service data retrieval feature and the second service data retrieval feature belong to retrieval features of two different data dimensions in a target service data retrieval request;
the first service data retrieval vector is subjected to feature restoration according to the second service data retrieval vector, initial fusion data retrieval features are output, and the second service data retrieval vector is adjusted according to the difference between the initial fusion data retrieval features and the first service data retrieval features, so that a corresponding adjustment service data retrieval vector is output;
Aggregating the second service data retrieval vector and the adjustment service data retrieval vector to output an aggregated service data retrieval vector;
performing feature restoration according to the aggregate service data retrieval vector on the first service data retrieval vector, outputting target fusion data retrieval features, and performing service data retrieval in a target platform database based on the target fusion data retrieval features to obtain target service data corresponding to the target service data retrieval request;
the first service data retrieval feature, the initial fusion data retrieval feature and the target fusion data retrieval feature have the same data dimension and belong to first dimension data, the second service data retrieval feature belongs to second dimension data, the first dimension data comprises text dimension data, and the second dimension data comprises image dimension data or voice dimension data;
the step of performing feature mining on the first service data retrieval feature, outputting a first service data retrieval vector, and performing feature mining on the second service data retrieval feature, outputting a second service data retrieval vector, includes:
Generating interference data features, wherein the interference data features have the same data dimension as the first service data retrieval features;
feature combination is carried out on the first service data retrieval feature and the interference data feature to form a corresponding combined service data retrieval feature;
feature mining is carried out on the combined service data retrieval features, and a first service data retrieval vector corresponding to the first service data retrieval features is output;
performing feature mining on the second service data retrieval features, and outputting second service data retrieval vectors corresponding to the second service data retrieval features;
the feature restoration of the first service data retrieval vector according to the second service data retrieval vector is performed by using a text data restoration network, wherein the text data restoration network comprises a first number of interference data suppression units and feature restoration units which are connected in sequence, and the first number is more than or equal to 2;
the step of restoring the first service data retrieval vector according to the features of the second service data retrieval vector, outputting an initial fusion data retrieval feature, adjusting the second service data retrieval vector according to the difference between the initial fusion data retrieval feature and the first service data retrieval feature, and outputting a corresponding adjusted service data retrieval vector, includes:
Using the ith interference data suppression unit in the first number of successively connected interference data suppression units to perform interference suppression processing on the loading data of the ith interference data suppression unit, loading the ith interference suppression data output by the ith interference data suppression unit to the jth interference data suppression unit to perform interference suppression processing, and outputting the jth interference suppression data corresponding to the jth interference data suppression unit, wherein j=i+1; when i is equal to 1, the loading data of the ith interference data suppression unit is the first service data retrieval vector and the second service data retrieval vector, and when i is greater than or equal to or less than the first quantity, the loading data of the ith interference data suppression unit is the interference suppression data output by the previous interference data suppression unit of the ith interference data suppression unit and the second service data retrieval vector;
the feature reduction unit is utilized to perform feature reduction on the interference suppression data output by the last interference data suppression unit, and corresponding initial fusion data retrieval features are output;
and adjusting the second service data retrieval vector according to the difference between the initial fusion data retrieval feature and the first service data retrieval feature, and outputting a corresponding adjustment service data retrieval vector.
2. The method for searching service data in combination with a platform database according to claim 1, wherein the ith interference data suppression unit comprises a second number of successively connected focusing and mining units, and the second number is greater than or equal to 2;
the step of using the ith interference data suppression unit in the first number of interference data suppression units connected in sequence to perform interference suppression processing on the loading data of the ith interference data suppression unit, and loading the ith interference suppression data output by the ith interference data suppression unit to the jth interference data suppression unit to perform interference suppression processing, and outputting the jth interference suppression data corresponding to the jth interference data suppression unit includes:
utilizing an a-th focusing excavation unit in the i-th interference data suppression unit to perform focusing characteristic excavation on the loading data of the a-th focusing excavation unit and the second service data retrieval vector, outputting a corresponding focusing service data retrieval vector, and marking the corresponding focusing service data retrieval vector as a-th focusing excavation data of the a-th focusing excavation unit in the i-th interference data suppression unit;
loading the a-th focus mining data of the a-th focus mining unit in the i-th interference data suppression unit to the b-th focus mining unit to perform focus feature mining, outputting the b-th focus mining data of the b-th focus mining unit in the i-th interference data suppression unit, wherein when b is equal to 1, the loading data of the a-th focus mining unit is interference suppression data output by a previous interference data suppression unit of the i-th interference data suppression unit, and when b is greater than or equal to 2 and less than the second quantity, the loading data of the a-th focus mining unit is focus mining data output by a previous focus mining unit;
Marking the focus mining data output by the last focus mining unit in the ith interference data suppression unit, so that the focus mining data is marked as the ith interference suppression data;
and loading the ith interference suppression data output by the ith interference data suppression unit to the jth interference data suppression unit to perform interference suppression processing, and outputting the jth interference suppression data corresponding to the jth interference data suppression unit.
3. The method for searching service data in combination with a platform database according to claim 2, wherein the step of utilizing an a-th focus mining unit in the i-th interference data suppression unit to perform focus feature mining on the loading data of the a-th focus mining unit and the second service data search vector, outputting a corresponding focus service data search vector, and marking the corresponding focus service data search vector as a-th focus mining data of an a-th focus mining unit in the i-th interference data suppression unit comprises the steps of:
carrying out first linear conversion of an attention mechanism on the loading data of the a-th focusing excavation unit, and outputting a corresponding first linear conversion result;
performing a second linear conversion of the attention mechanism on the second service data retrieval vector, and outputting a corresponding second linear conversion result;
Performing third linear conversion of an attention mechanism on the second service data retrieval vector, and outputting a corresponding third linear conversion result;
based on the first linear conversion result and the second linear conversion result, determining similarity parameters between the loading data of the a-th focusing mining unit and the second service data retrieval vector;
and weighting the third linear conversion result based on the similarity parameter, outputting a corresponding focus service data retrieval vector, and marking the focus service data retrieval vector as the a focus mining data of the a focus mining unit in the i interference data suppression unit.
4. The business data retrieval method in combination with the platform database according to claim 1, wherein the feature restoration of the first business data retrieval vector according to the second business data retrieval vector is performed using a text data restoration network, and the feature restoration of the first business data retrieval vector according to the aggregate business data retrieval vector is performed using a target neural network including the text data restoration network and a plurality of text data analysis networks;
Before the step of restoring the first service data retrieval vector according to the feature of the aggregate service data retrieval vector, outputting a target fusion data retrieval feature, and performing service data retrieval in a target platform database based on the target fusion data retrieval feature to obtain target service data corresponding to the target service data retrieval request, the service data retrieval method of the combined platform database further comprises:
processing the first service data retrieval vector and a target service data retrieval vector in the target neural network to form corresponding intermediate data retrieval features, wherein the target service data retrieval vector is the second service data retrieval vector or the aggregate service data retrieval vector;
and optimizing network parameters of a plurality of text data analysis networks in the target neural network according to the difference between the intermediate data retrieval feature and the first service data retrieval feature to form an optimized target neural network.
5. The business data retrieval method combining the platform database according to claim 4, wherein the text data reduction network comprises a first number of interference data suppression units and feature reduction units which are sequentially connected, the first number is greater than or equal to 2, the target neural network is formed by setting a text data analysis network for each interference data suppression unit on the basis of the text data reduction network, each interference data suppression unit and the corresponding text data analysis network form a corresponding combined processing unit, and the sequential connection relationship among a plurality of combined processing units is consistent with the sequential connection relationship among a plurality of interference data suppression units;
The step of processing the first service data retrieval vector and the target service data retrieval vector in the target neural network to form corresponding intermediate data retrieval features comprises the following steps:
the method comprises the steps of carrying out joint interference suppression processing on loading data of an ith joint processing unit by utilizing an ith joint processing unit in a first number of joint processing units which are connected in sequence, loading the ith joint interference suppression data output by the ith joint processing unit to the jth joint processing unit for joint interference suppression processing, and outputting the jth joint interference suppression data of the jth joint processing unit;
performing feature restoration on the joint interference suppression data output by the last joint processing unit, and outputting corresponding intermediate data retrieval features;
and when i is equal to 1, the loading data of the ith joint processing unit is the first service data retrieval vector and the target service data retrieval vector, and when i is more than or equal to 2 and less than the first quantity, the loading data of the ith joint processing unit is the joint interference suppression data output by the previous joint processing unit and the target service data retrieval vector.
6. The service data retrieval method in combination with a platform database according to any one of claims 1 to 5, wherein the step of aggregating the second service data retrieval vector and the adjustment service data retrieval vector to output an aggregated service data retrieval vector comprises:
performing vector decimation according to a first size on the second service data retrieval vector to form a corresponding first size decimated vector, wherein the first size decimated vector comprises vector local data of a first size with a front distribution coordinate in the second service data retrieval vector;
performing vector decimation according to a second size on the adjustment service data retrieval vector to form a corresponding second-size decimated vector, wherein the second-size decimated vector comprises vector local data of a second size with a front distribution coordinate in the adjustment service data retrieval vector;
and cascading the second size decimation vector to the tail end position of the first size decimation vector to form a corresponding aggregate service data retrieval vector.
7. An artificial intelligence system, comprising a processor and a memory, the memory for storing a computer program, the processor for executing the computer program to implement the method of business data retrieval incorporating a platform database according to any one of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311243360.XA CN116991919B (en) | 2023-09-26 | 2023-09-26 | Service data retrieval method combined with platform database and artificial intelligent system |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202311243360.XA CN116991919B (en) | 2023-09-26 | 2023-09-26 | Service data retrieval method combined with platform database and artificial intelligent system |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116991919A CN116991919A (en) | 2023-11-03 |
| CN116991919B true CN116991919B (en) | 2023-12-08 |
Family
ID=88528709
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202311243360.XA Active CN116991919B (en) | 2023-09-26 | 2023-09-26 | Service data retrieval method combined with platform database and artificial intelligent system |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116991919B (en) |
Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2006155556A (en) * | 2004-10-27 | 2006-06-15 | Hitachi Software Eng Co Ltd | Text mining method and text mining server |
| CN1967536A (en) * | 2006-11-16 | 2007-05-23 | 华中科技大学 | Region based multiple features Integration and multiple-stage feedback latent semantic image retrieval method |
| CN112884005A (en) * | 2021-01-21 | 2021-06-01 | 汉唐信通(北京)科技有限公司 | Image retrieval method and device based on SPTAG and convolutional neural network |
| CN113657450A (en) * | 2021-07-16 | 2021-11-16 | 中国人民解放军陆军炮兵防空兵学院 | Attention mechanism-based land battlefield image-text cross-modal retrieval method and system |
| CN113971222A (en) * | 2021-10-28 | 2022-01-25 | 重庆紫光华山智安科技有限公司 | Multi-mode composite coding image retrieval method and system |
| CN114048340A (en) * | 2021-11-15 | 2022-02-15 | 电子科技大学 | Hierarchical fusion combined query image retrieval method |
| CN114416733A (en) * | 2021-12-29 | 2022-04-29 | 中国电信股份有限公司 | Data retrieval processing method and device, electronic equipment and storage medium |
| CN115033670A (en) * | 2022-06-02 | 2022-09-09 | 西安电子科技大学 | Cross-modal image-text retrieval method with multi-granularity feature fusion |
| CN116522011A (en) * | 2023-05-16 | 2023-08-01 | 深圳九星互动科技有限公司 | Big data-based pushing method and pushing system |
-
2023
- 2023-09-26 CN CN202311243360.XA patent/CN116991919B/en active Active
Patent Citations (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2006155556A (en) * | 2004-10-27 | 2006-06-15 | Hitachi Software Eng Co Ltd | Text mining method and text mining server |
| CN1967536A (en) * | 2006-11-16 | 2007-05-23 | 华中科技大学 | Region based multiple features Integration and multiple-stage feedback latent semantic image retrieval method |
| CN112884005A (en) * | 2021-01-21 | 2021-06-01 | 汉唐信通(北京)科技有限公司 | Image retrieval method and device based on SPTAG and convolutional neural network |
| CN113657450A (en) * | 2021-07-16 | 2021-11-16 | 中国人民解放军陆军炮兵防空兵学院 | Attention mechanism-based land battlefield image-text cross-modal retrieval method and system |
| CN113971222A (en) * | 2021-10-28 | 2022-01-25 | 重庆紫光华山智安科技有限公司 | Multi-mode composite coding image retrieval method and system |
| CN114048340A (en) * | 2021-11-15 | 2022-02-15 | 电子科技大学 | Hierarchical fusion combined query image retrieval method |
| CN114416733A (en) * | 2021-12-29 | 2022-04-29 | 中国电信股份有限公司 | Data retrieval processing method and device, electronic equipment and storage medium |
| CN115033670A (en) * | 2022-06-02 | 2022-09-09 | 西安电子科技大学 | Cross-modal image-text retrieval method with multi-granularity feature fusion |
| CN116522011A (en) * | 2023-05-16 | 2023-08-01 | 深圳九星互动科技有限公司 | Big data-based pushing method and pushing system |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116991919A (en) | 2023-11-03 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111625635B (en) | Question-answering processing method, device, equipment and storage medium | |
| US20220180202A1 (en) | Text processing model training method, and text processing method and apparatus | |
| CN108334487B (en) | Missing semantic information completion method and device, computer equipment and storage medium | |
| CA3047353C (en) | Learning document embeddings with convolutional neural network architectures | |
| CN113762322B (en) | Video classification method, device and equipment based on multi-modal representation and storage medium | |
| CN111914085B (en) | Text fine granularity emotion classification method, system, device and storage medium | |
| CN111222305A (en) | Information structuring method and device | |
| EP4336378A1 (en) | Data processing method and related device | |
| JP2022050379A (en) | Semantic retrieval method, apparatus, electronic device, storage medium, and computer program product | |
| CN112417855A (en) | Text intention recognition method and device and related equipment | |
| CN108959388B (en) | Information generation method and device | |
| CN113239169A (en) | Artificial intelligence-based answer generation method, device, equipment and storage medium | |
| CN112925898B (en) | Question-answering method and device based on artificial intelligence, server and storage medium | |
| CN112836502A (en) | Implicit causal relationship extraction method for events in financial field | |
| CN116975350A (en) | Image-text retrieval method, device, equipment and storage medium | |
| CN113486659B (en) | Text matching method, device, computer equipment and storage medium | |
| CN117635275B (en) | Intelligent electronic commerce operation commodity management platform and method based on big data | |
| US20240037939A1 (en) | Contrastive captioning for image groups | |
| CN109857843A (en) | Exchange method and system based on document | |
| US20210271705A1 (en) | Generating accurate and natural captions for figures | |
| CN116991919B (en) | Service data retrieval method combined with platform database and artificial intelligent system | |
| CN116957043A (en) | Model quantization method, device, equipment and medium | |
| EP4322066A1 (en) | Method and apparatus for generating training data | |
| CN113010717B (en) | Image verse description generation method, device and equipment | |
| CN115080039A (en) | Front-end code generation method, device, computer equipment, storage medium and product |
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 |