CN116501741A

CN116501741A - AI Chat bot-based structured database storage analysis method and software product

Info

Publication number: CN116501741A
Application number: CN202310500881.2A
Authority: CN
Inventors: 张洋; 杨阳
Original assignee: Handan Yaohui Network Technology Co ltd
Current assignee: Handan Yaohui Network Technology Co ltd
Priority date: 2023-05-06
Filing date: 2023-05-06
Publication date: 2023-07-28

Abstract

The embodiment of the invention relates to the technical fields of big data, artificial intelligence and databases, and provides a structured database storage analysis method and a software product based on an AI Chat bot. The big data decision analysis server can judge whether the storage condition reaches the standard or not by utilizing the AI deep decision tree network to store the AI Chat man-machine interaction text. The method has low overall constraint, can be applied to AI Chat man-machine interaction texts of different types and scenes, improves the storage analysis timeliness aiming at the structured database, and can ensure the order of data stored in the structured database.

Description

AI Chat bot-based structured database storage analysis method and software product

Technical Field

The invention relates to the technical fields of big data, artificial intelligence and databases, in particular to a structured database storage analysis method and a software product based on an AI Chat bot.

Background

Chat robots (chat bots), also known as alfibrates, are computer programs that talk via conversation or text. Along with the development of artificial intelligence (Artificial Intelligence), the intelligent degree of the chat robot is higher and higher, and the application scene of the current AI chat robot can relate to the fields of intelligent customer service, virtual robots, digital service, big data retrieval, information consultation and the like.

With the continuous maturation of leading edge technologies such as big data, cloud computing, blockchain, digitization and the like, the AI chat robots are more and more tightly combined with the fields, and the conversation data generated by the AI chat robots also have high analysis value, so that how to store the conversation data with high quality is important.

For the storage of the above-mentioned dialog data, structured storage is not lost as a preference. The structured storage method applies the principle of a tree file system to a single file, so that the single file can also contain subdirectories like the file system, the subdirectories can also contain subdirectories with deeper layers, and each directory can contain a plurality of files, thereby realizing the purpose that the content which is originally needed to be stored by a plurality of files is stored into one file according to the tree structure and the layers. The structured energy storage can intuitively reflect the characteristic connection of the dialogue data, and is convenient for targeted calling, big data mining analysis and the like in the later period. However, when the structured database performs structured storage on the session data, it is difficult to accurately and efficiently determine whether the related session data is suitable for structured storage.

Disclosure of Invention

The invention provides a structured database storage analysis method and a software product based on an AI Chat bot, which can utilize an AI depth decision tree network to judge whether storage conditions of an AI Chat man-machine interaction text to be stored reach standards. The method and the device have the advantages that the overall constraint is not strong, the method and the device can be applied to AI Chat man-machine interaction texts of different types and scenes, the storage analysis timeliness aiming at the structured database is improved, the order of data stored in the structured database can be ensured, and the following technical scheme is adopted for achieving the technical purposes.

The first aspect is a structured database storage analysis method based on AI Chat bot, applied to a big data decision analysis server, the method comprising:

acquiring an AI Chat man-machine interaction text to be stored, which corresponds to a structured storage task to be analyzed;

determining the input information corresponding to at least one AI algorithm component respectively; the input information of each AI algorithm component comprises at least one of the to-be-stored AI Chat man-machine interaction text and a text semantic aggregation vector generated by an upstream AI algorithm component before the corresponding AI algorithm component;

respectively carrying out text semantic mining on corresponding incoming information by utilizing each AI algorithm component to obtain scene-level text semantics and paragraph-level text semantics, and carrying out text semantic aggregation processing on the scene-level text semantics and the paragraph-level text semantics to obtain corresponding text semantic aggregation vectors;

and determining storage condition standard judging information corresponding to the structured storage task to be analyzed according to the text semantic aggregation vector respectively generated by each AI algorithm component.

In some optional examples, the acquiring the AI Chat human-computer interaction text to be stored corresponding to the structured storage task to be analyzed includes:

Acquiring a basic AI Chat man-machine interaction text, and determining an AI man-machine dialogue text set of a structured storage task to be analyzed in the basic AI Chat man-machine interaction text;

updating the scale of the AI man-machine conversation text set to obtain a target conversation text set;

and extracting the target dialogue text set to obtain the AI Chat man-machine interaction text to be stored, which is matched with the structured storage task to be analyzed.

In some optional examples, the determining the at least one AI algorithm component respectively corresponds to the incoming information includes:

for the first AI algorithm component in at least one AI algorithm component, taking the AI Chat man-machine interaction text to be stored as the incoming information of the first AI algorithm component;

for each of the at least one AI algorithm component except the first AI algorithm component, the text semantic aggregate vector generated by the upstream AI algorithm component preceding the corresponding AI algorithm component is used as the incoming information for the corresponding AI algorithm component.

In some optional examples, the text semantic mining is performed on the corresponding incoming information by using each AI algorithm component to obtain scene-level text semantics and paragraph-level text semantics, and the text semantic aggregation processing is performed on the scene-level text semantics and the paragraph-level text semantics to obtain corresponding text semantic aggregation vectors, which includes:

Performing text semantic mining on the to-be-stored AI Chat man-machine interaction text by using a first AI algorithm component in a plurality of AI algorithm components to obtain a first scene level text semantic and a first paragraph level text semantic, and performing text semantic aggregation processing on the first scene level text semantic and the first paragraph level text semantic to obtain a text semantic aggregation vector generated by the first AI algorithm component;

for each AI algorithm component except the first AI algorithm component, text semantic mining is carried out on text semantic aggregation vectors generated by upstream AI algorithm components before the corresponding AI algorithm component to obtain second scene level text semantics and second paragraph level text semantics, and text semantic aggregation processing is carried out on the second scene level text semantics and the second paragraph level text semantics to obtain text semantic aggregation vectors generated by the corresponding AI algorithm components.

In some optional examples, the text semantic mining of the corresponding incoming information by using each AI algorithm component to obtain scene-level text semantics and paragraph-level text semantics includes:

for each AI algorithm component in at least one AI algorithm component, extracting scene-level original semantics and paragraph-level original semantics in the corresponding incoming information by using the current AI algorithm component;

Performing first downsampling processing on the paragraph level original semantics by using the current AI algorithm component to obtain paragraph level downsampling semantics in the paragraph level original semantics, and obtaining corresponding paragraph level semantic differences according to semantic distances between the paragraph level original semantics and the paragraph level downsampling semantics;

and respectively carrying out semantic detail derivation processing on the scene-level original semantics and the paragraph-level semantic differences by utilizing the current AI algorithm component to obtain corresponding scene-level text semantics and paragraph-level text semantics.

In some optional examples, the performing semantic detail derivation processing on the scene-level original semantics and the paragraph-level semantic differences to obtain corresponding scene-level text semantics and paragraph-level text semantics includes:

extracting scene-level dialogue features from the paragraph-level semantic differences, and adding the scene-level dialogue features to the scene-level original semantics to obtain scene-level text semantics;

and extracting paragraph level attribute features from the scene level original semantics, and adding the paragraph level attribute features to the paragraph level semantic differences to obtain paragraph level text semantics.

In some optional examples, the performing text semantic aggregation processing on the scene-level text semantics and the paragraph-level text semantics to obtain corresponding text semantic aggregation vectors includes:

determining semantic enhancement factors respectively corresponding to the scene level text semantics and the paragraph level text semantics;

and carrying out text semantic aggregation processing on the scene layer text semantics and the paragraph layer text semantics based on semantic enhancement factors respectively corresponding to the scene layer text semantics and the paragraph layer text semantics to obtain corresponding text semantic aggregation vectors.

In some optional examples, the determining the semantic enhancement factors respectively corresponding to the scene-level text semantics and the paragraph-level text semantics includes:

windowing filtering operation is respectively carried out on the scene-level text semantics and the paragraph-level text semantics to obtain scene-level semantic filtering characteristics and paragraph-level semantic filtering characteristics, and splicing operation is carried out on the scene-level semantic filtering characteristics and the paragraph-level semantic filtering characteristics to obtain corresponding semantic splicing results;

and determining semantic enhancement factors respectively corresponding to the scene level text semantics and the paragraph level text semantics according to the semantic splicing result.

In some optional examples, determining the semantic enhancement factors corresponding to the scene-level text semantics and the paragraph-level text semantics respectively according to the semantic stitching result includes:

determining a first characteristic relation list and a first standardized algorithm layer corresponding to the scene level text semantics, and determining a second characteristic relation list and a second standardized algorithm layer corresponding to the paragraph level text semantics;

performing second downsampling processing on the semantic stitching result to obtain semantic stitching downsampling characteristics;

performing first feature mapping on the semantic stitching downsampling features by using the first feature relation list and a first standardized algorithm layer to obtain semantic enhancement factors corresponding to the scene layer text semantics;

and performing second feature mapping on the semantic stitching downsampled features by using the second feature relation list and a second standardized algorithm layer to obtain semantic enhancement factors corresponding to the paragraph level text semantics.

In some optional examples, the AI Chat bot-based structured database storage analysis method is performed by a deep decision tree network that is debugged using a network debug thread, the network debug thread comprising:

Acquiring a depth decision tree network to be debugged, an AI Chat man-machine interaction debugging text and debugging comments corresponding to the AI Chat man-machine interaction debugging text; the depth decision tree network comprises at least one AI algorithm component to be debugged;

determining debugging incoming data corresponding to at least one AI algorithm component to be debugged respectively;

the debugging incoming data of the current AI algorithm component to be debugged in the at least one AI algorithm component to be debugged comprises at least one of the AI Chat man-machine interaction debugging text and a text semantic aggregation deduction vector generated by the previous AI algorithm component to be debugged;

respectively carrying out text semantic mining on corresponding debugging input data by utilizing each AI algorithm component to obtain scene-level text deduction semantics and paragraph-level text deduction semantics, and carrying out text semantic aggregation processing on the scene-level text deduction semantics and the paragraph-level text deduction semantics to obtain corresponding text semantic aggregation deduction vectors;

determining a deduction analysis result of storage condition deduction analysis on the AI Chat man-machine interaction debugging text according to text semantic aggregation deduction vectors respectively generated by each AI algorithm component to be debugged;

And according to the deduction analysis result and the debugging annotation, debugging the deep decision tree network until the deep decision tree network meets the debugging completion requirement, and obtaining the debugged deep decision tree network.

In some optional examples, the deduction analysis results include a text structure distribution deduction result and a storage condition deduction analysis identity; the determining a deduction analysis result of storage condition deduction analysis on the AI Chat man-machine interaction debugging text according to the text semantic aggregation deduction vectors respectively generated by each AI algorithm component to be debugged comprises the following steps:

using text structure distribution generating components in the depth decision tree network to perform text structure distribution generating processing on the text semantic aggregation deduction vectors respectively generated by each AI algorithm component to be debugged, so as to obtain corresponding text structure distribution deduction results;

and carrying out storage condition deduction analysis on the text semantic aggregation deduction vectors respectively generated by each AI algorithm component to be debugged by using a storage condition analysis component in the depth decision tree network to obtain corresponding storage condition deduction analysis identifiers.

In some optional examples, the debug annotations include a target text structure distribution and a storage condition authentication annotation; and according to the deduction analysis result and the debugging annotation, debugging the depth decision tree network until the depth decision tree network meets the debugging completion requirement, wherein the method comprises the following steps:

Determining a first comparison result between the text structure distribution deduction result and the target text structure distribution, and determining a second comparison result between the storage condition deduction analysis identification and the storage condition authentication annotation;

and debugging the depth decision tree network by using the first comparison result and the second comparison result until the depth decision tree network meets the debugging completion requirement.

The second aspect is a big data decision analysis server comprising a memory and a processor; the memory is coupled to the processor; the memory is used for storing computer program codes, and the computer program codes comprise computer instructions; wherein the computer instructions, when executed by the processor, cause the big data decision analysis server to perform the method of the first aspect.

A third aspect is a software product for implementing an AI Chat bot-based structured database storage analysis method, comprising a computer program/instruction, wherein the computer program/instruction, when executed, implements the method of performing the first aspect.

A fourth aspect is a computer readable storage medium having stored thereon a computer program which, when run, performs the method of the first aspect.

According to the structured database storage analysis method and the software product based on the AI Chat bot, which are provided by the embodiment of the invention, by utilizing the text semantic aggregation vector generated by the AI Chat man-machine interaction text to be stored and the upstream AI algorithm components, the respectively corresponding incoming information of each AI algorithm component can be determined based on the text semantic aggregation vector generated by the AI Chat man-machine interaction text to be stored and the upstream AI algorithm components, so that each AI algorithm component can carry out text semantic mining and text semantic aggregation processing on the respectively corresponding incoming information to obtain the corresponding text semantic aggregation vector; and (3) utilizing the text semantic aggregation vectors generated by each AI algorithm component respectively to combine the text semantic aggregation vectors to obtain the storage condition standard judging information corresponding to the structured storage task to be analyzed.

In view of the fact that the scene-level text semantics and the paragraph-level text semantics are combined to judge whether the storage condition of the to-be-stored AI Chat human-computer interaction text meets the standard, compared with the fact that the disturbance condition is utilized to determine the storage condition meeting the standard judging information, the embodiment of the invention can still obtain accurate and reliable storage condition meeting the standard judging information when the to-be-stored AI Chat human-computer interaction text does not meet the standard of a disturbance condition possibly met by the distinction between the to-be-stored AI Chat human-computer interaction text and the standard AI Chat human-computer interaction text.

In addition, since the scene level text semantics and the paragraph level text semantics are not excessively limited, the constraint of the overall scheme is not too strong, and the method can be applied to AI Chat man-machine interaction texts of different types and scenes, so that the deep decision tree network can be used for accurately judging whether storage conditions of the AI Chat man-machine interaction texts in different states reach standards. According to the embodiment of the invention, the distribution form of the AI Chat man-machine interaction text is not required, the storage condition standard identification can be rapidly and efficiently carried out on the AI Chat man-machine interaction text, the storage analysis timeliness aiming at the structured database is improved, whether the AI Chat man-machine interaction text to be stored can realize structured storage or not can be accurately and reliably reflected by storage condition standard identification information, and the order of data stored in the structured database is ensured.

Drawings

Fig. 1 is a flow chart of a structured database storage analysis method based on AI Chat bot according to an embodiment of the present invention.

Detailed Description

Hereinafter, the terms "first," "second," and "third," etc. are used for descriptive purposes only and are not to be construed as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defining "a first", "a second", or "a third", etc., may explicitly or implicitly include one or more such feature.

Fig. 1 shows a flow chart of a method for storing and analyzing a structured database based on an AI Chat bot according to an embodiment of the present invention, where the method for storing and analyzing a structured database based on an AI Chat bot may be implemented by a big data decision analysis server, and the big data decision analysis server may include a memory and a processor; the memory is coupled to the processor; the memory is for storing computer program code, the computer program code comprising computer instructions.

Wherein the computer instructions, when executed by the processor, cause the big data decision analysis server to perform STEP202-STEP208.

STEP202 obtains the human-computer interaction text of the AI Chat to be stored corresponding to the structured storage task to be analyzed.

In the embodiment of the invention, the structured storage task to be analyzed can be sent to the big data decision analysis server by the database system corresponding to the structured database, and before the database system prepares to store the to-be-stored AI Chat man-machine interaction text corresponding to the structured storage task to be analyzed, whether the to-be-stored AI Chat man-machine interaction text can be converted into a tree structure or not needs to be judged in advance, namely whether the storage condition of the to-be-stored AI Chat man-machine interaction text meets the standard or not. In view of the fact that the structured database is focused on data storage and management, judging whether the storage condition of the AI Chat man-machine interaction text to be stored meets the standard or not can be achieved through a big data decision analysis server, so that task diversion and division cooperation can be achieved, and the operation efficiency of different systems/servers is improved.

For example, the AI Chat man-machine interaction text to be stored may be an interaction text record between the digital client and the AI Chat robot, such as, but not limited to, e-commerce dialogs, digital twin-related dialogs, meta-space services dialogs, data search dialogs, question-answer dialogs, digital office dialogs, distance education dialogs, intelligent medical care consultation dialogs, and the like.

For example, when the storage condition is required to reach the standard, the big data decision analysis server can acquire the human-computer interaction text of the AI Chat to be stored. The storage condition standard reaching judging information of the to-be-stored AI Chat man-machine interaction text can include that a to-be-analyzed structured storage task in the to-be-stored AI Chat man-machine interaction text meets the structured storage condition and that the to-be-analyzed structured storage task in the to-be-stored AI Chat man-machine interaction text does not meet the structured storage condition. When the storage condition standard judging information characterizes that a to-be-analyzed structured storage task in the to-be-stored AI Chat man-machine interaction text meets the structured storage condition, the to-be-stored AI Chat man-machine interaction text can be considered to be subjected to tree structured processing and stored; when the storage condition standard judging information characterizes that the to-be-analyzed structured storage task in the to-be-stored AI Chat man-machine interaction text does not meet the structured storage condition, the to-be-stored AI Chat man-machine interaction text can not be considered to be subjected to tree structured processing, so that the storage requirement of a database system is not met.

In some examples, obtaining the AI Chat human-machine interaction text to be stored corresponding to the structured storage task to be analyzed includes: acquiring a basic AI Chat man-machine interaction text, and determining an AI man-machine dialogue text set of a structured storage task to be analyzed in the basic AI Chat man-machine interaction text; updating the scale of the AI man-machine conversation text set to obtain a target conversation text set; and extracting the target dialogue text set to obtain the AI Chat man-machine interaction text to be stored, which is matched with the structured storage task to be analyzed.

The big data decision analysis server can obtain the basic AI Chat man-machine interaction text, and determine an AI man-machine conversation text set of the structured storage task to be analyzed in the basic AI Chat man-machine interaction text by utilizing the text capturing module, so as to obtain a text window for marking the AI man-machine conversation text set. The text capturing module can be self-contained in the big data decision analysis server, or can be obtained by calling other neural network models. Further, in order to ensure the analysis integrity of the structured storage task to be analyzed, the big data decision analysis server can also extend the scale of the text window so as to update the scale of the AI man-machine conversation text set, obtain a corresponding target conversation text set, extract the target conversation text set and obtain the AI Chat man-machine interaction text to be stored, which is matched with the structured storage task to be analyzed. The scale update can be understood as ensuring the maximization of the information quantity of the AI man-machine dialogue text set as much as possible, thereby providing a reliable basis for the structural analysis.

Therefore, the AI man-machine interaction text set is updated on a large scale, so that the AI Chat man-machine interaction text to be stored after the updating on a large scale can contain the AI man-machine interaction text set with as much content as possible, and the accuracy of the result of judging whether the storage condition meets the standard based on the complete AI man-machine interaction text set is ensured.

STEP204, determining at least one AI algorithm component to respectively correspond to the incoming information; the incoming information for each AI algorithm component includes at least one of an AI Chat man-machine interaction text to be stored, and a text semantic aggregation vector generated by an upstream AI algorithm component preceding the corresponding AI algorithm component.

For example, the deep decision tree network can be utilized to judge whether the storage condition of the structured storage task to be analyzed in the AI Chat man-machine interaction text to be stored meets the standard. The deep decision tree network can be an initial artificial intelligent neural network with storage condition standard judging performance for completing debugging. The deep decision tree network may illustratively include at least one AI algorithm component, each AI algorithm component processing a respective incoming message to obtain a text semantic aggregate vector generated by each AI algorithm component, respectively. The input information of the current AI algorithm component in the at least one AI algorithm component comprises at least one of the AI Chat man-machine interaction text to be stored and a text semantic aggregation vector generated by an upstream AI algorithm component. Further, AI algorithm components may also be understood as processing units, network elements, network layers, or the like. The upstream AI algorithm component can be understood as the immediately preceding algorithm component of the current algorithm component. The text semantic aggregation vector is obtained by fusing text semantic features.

In some examples, determining the at least one AI algorithm component respectively corresponds to the incoming information includes: for the first AI algorithm component in the at least one AI algorithm component, taking the AI Chat man-machine interaction text to be stored as the incoming information of the first AI algorithm component; for each of the at least one AI algorithm component except the first AI algorithm component, the text semantic aggregate vector generated by the upstream AI algorithm component preceding the corresponding AI algorithm component is used as the incoming information for the corresponding AI algorithm component.

For example, each AI algorithm component in the deep decision tree network may be cascaded to form an AI algorithm component chain, and the big data decision analysis server uses the AI Chat man-machine interaction text to be stored as the incoming information of the first AI algorithm component in the AI algorithm component chain, that is, uses the AI Chat man-machine interaction text to be stored as the incoming information of the first AI algorithm component in the at least one AI algorithm component chain; for each AI algorithm component in the chain of AI algorithm components, except for the first AI algorithm component, the text semantic aggregate vector generated by the corresponding upstream AI algorithm component is used as the incoming information of the current AI algorithm component, in other words, the text semantic aggregate vector generated by the corresponding upstream AI algorithm component is used as the incoming information of the current AI algorithm component except for the first AI algorithm component in the at least one AI algorithm component. Wherein an upstream AI algorithm component can be understood as one AI algorithm component that is neighbor to and preceding the current AI algorithm component.

For example, the deep decision tree network may include three AI algorithm components, i.e., AI algorithm component Algorithm component1, AI algorithm component Algorithm component2, and AI algorithm component Algorithm component3, where the input information of AI algorithm component Algorithm component is the text semantic aggregation vector generated by AI algorithm component Algorithm component1, the input information of AI algorithm component Algorithm component is the text semantic aggregation vector generated by AI algorithm component Algorithm component2, and the input information of AI algorithm component Algorithm component is the text semantic aggregation vector generated by AI algorithm component Algorithm component.

In the embodiment of the invention, the corresponding incoming information can be processed by each AI algorithm component by determining the incoming information respectively corresponding to each AI algorithm component, so that the output information respectively generated by each AI algorithm component is obtained.

STEP206, respectively performing text semantic mining on the corresponding incoming information by utilizing each AI algorithm component to obtain scene level text semantics and paragraph level text semantics, and performing text semantic aggregation processing on the scene level text semantics and the paragraph level text semantics to obtain corresponding text semantic aggregation vectors.

The AI algorithm component comprises a text semantic mining sub-network and a text semantic aggregation sub-network. The text semantic mining subnetwork can be understood as a network architecture used for extracting scene-level text semantics and paragraph-level text semantics in the incoming information; a text semantic aggregation sub-network may be understood as a network architecture to aggregate scene-level text semantics and paragraph-level text semantics. Both the text-semantic mining sub-network and the text-semantic aggregation sub-network may be a federated architecture, for example, the text-semantic mining sub-network may be formed from a moving average sub-network, a pooled sub-network, and a sampled sub-network. The text semantic aggregation subnetwork may be formed from a moving average subnetwork, a second downsampling subnetwork, a normalized subnetwork, and a fully connected subnetwork.

Further, scene level text semantics may be understood as features to reflect the overall text corresponding to the structured storage task to be analyzed. Paragraph level text semantics can be understood as features to reflect the local text of the structured storage task to be analyzed. The scene-level text semantics can be, for example, features of a dialogue field, features of a dialogue summary, and the like, and the paragraph-level text semantics can be user interest text features, matching features of an AI response text, additional query features, AI question-answer service feedback features, and the like.

For example, the big data decision analysis server may perform text semantic mining processing (text feature extraction operation) on the corresponding incoming information by using each AI algorithm component, for example, for a current AI algorithm component in at least one AI algorithm component, the big data decision analysis server may extract scene-level text semantics and paragraph-level text semantics in the incoming information by using a text semantic mining subnet in the current AI algorithm component. Further, the big data decision analysis server can also utilize each AI algorithm component to respectively conduct text semantic aggregation processing on the scene level text semantics and the paragraph level text semantics obtained through respective mining so as to obtain corresponding text semantic aggregation vectors. For example, for the current AI algorithm component in at least one AI algorithm component, the big data decision analysis server can utilize the text semantic aggregation sub-network in the current AI algorithm component to perform text semantic aggregation processing on scene-level text semantics and paragraph-level text semantics extracted by the text semantic mining sub-network in the current AI algorithm component, so as to obtain a text semantic aggregation vector generated by the current AI algorithm component.

When only one AI algorithm component is included in the deep decision tree network, STEP202 and STEP204 can be implemented sequentially; when more than one AI algorithm component is included in the deep decision tree network, STEP202 and STEP204 may be alternatively implemented, for example, the incoming information of the first AI algorithm component may be determined first, and the incoming information may be processed by the first AI algorithm component to obtain a text semantic aggregation vector generated by the first AI algorithm component, then the incoming information of the subsequent AI algorithm component may be determined, and the incoming information may be processed by the subsequent AI algorithm component to obtain the text semantic aggregation vector generated by the subsequent AI algorithm component. And repeating the steps until the text semantic aggregation vector generated by the terminal AI algorithm component is obtained.

Under some examples, text semantic mining is performed on corresponding incoming information by using each AI algorithm component to obtain scene-level text semantics and paragraph-level text semantics, and text semantic aggregation processing is performed on the scene-level text semantics and the paragraph-level text semantics to obtain corresponding text semantic aggregation vectors, including: performing text semantic mining on the human-computer interaction text to be stored AI Chat by using a first AI algorithm component in a plurality of AI algorithm components to obtain a first scene level text semantic and a first paragraph level text semantic, and performing text semantic aggregation processing on the first scene level text semantic and the first paragraph level text semantic to obtain a text semantic aggregation vector generated by the first AI algorithm component; for each AI algorithm component except the first AI algorithm component, text semantic mining is carried out on text semantic aggregation vectors generated by upstream AI algorithm components before the corresponding AI algorithm component to obtain second scene level text semantics and second paragraph level text semantics, and text semantic aggregation processing is carried out on the second scene level text semantics and the second paragraph level text semantics to obtain text semantic aggregation vectors generated by the corresponding AI algorithm components.

For example, because the incoming information of the first AI algorithm component is the AI Chat man-machine interaction text to be stored, the big data decision analysis server performs text semantic mining processing on the AI Chat man-machine interaction text to be stored by using the first AI algorithm component to obtain first scene layer text semantics and first paragraph layer text semantics, performs text semantic aggregation processing on the first scene layer text semantics and the first paragraph layer text semantics to obtain a text semantic aggregation vector generated by the first AI algorithm component, for example, the first AI algorithm component performs weight reinforcement aggregation processing on the first scene layer text semantics and the first paragraph layer text semantics to obtain a corresponding text semantic aggregation vector. Further, the depth decision tree network takes the text semantic aggregation vector generated by the first AI algorithm component as the incoming information of the second AI algorithm component, and further performs text semantic mining and text semantic aggregation processing on the incoming information by using the second AI algorithm component to obtain the text semantic aggregation vector generated by the second AI algorithm component. And repeating the steps until the text semantic aggregation vector generated by the terminal AI algorithm component is obtained.

Under some possible examples, a big data decision analysis server acquires an AI Chat man-machine interaction text to be stored, which corresponds to a structured storage task to be analyzed, and performs text semantic mining on the AI Chat man-machine interaction text to be stored by utilizing a first AI algorithm component in a plurality of AI algorithm components to obtain a first scene level text semantic and a first paragraph level text semantic, and performs text semantic aggregation processing on the first scene level text semantic and the first paragraph level text semantic to obtain a text semantic aggregation vector generated by the first AI algorithm component; for each AI algorithm component except the first unit, text semantic mining is carried out on text semantic aggregation vectors generated by upstream AI algorithm components before the corresponding AI algorithm component to obtain second scene level text semantics and second paragraph level text semantics, and text semantic aggregation processing is carried out on the second scene level text semantics and the second paragraph level text semantics to obtain text semantic aggregation vectors generated by the corresponding AI algorithm components; and determining storage condition standard judging information corresponding to the structured storage task to be analyzed according to the text semantic aggregation vector respectively generated by each AI algorithm component.

In the embodiment of the invention, the detail expression of the text semantic aggregation vector generated by each AI algorithm component in the AI algorithm component chain can be gradually optimized by configuring a plurality of target AI algorithm components and taking the generation result of the upstream AI algorithm component as the input information of the current AI algorithm component so as to improve the quality of the text semantic aggregation vector.

STEP208 determines storage condition standard reaching discrimination information corresponding to the structured storage task to be analyzed according to the text semantic aggregation vector generated by each AI algorithm component respectively.

In the embodiment of the invention, the deep decision tree network can also comprise a storage condition analysis component, and the storage condition analysis component can be utilized to output storage condition standard reaching discrimination information. For example, when the storage condition analysis component outputs a set discrimination value, the AI Chat human-computer interaction text to be stored can be considered to be subjected to tree-like structuring processing and stored, and when the storage condition analysis component outputs a non-set discrimination value, the AI Chat human-computer interaction text to be stored can be considered to be not subjected to tree-like structuring processing, so that the storage requirement of the database system is not met.

When the text semantic aggregation vector generated by each AI algorithm component is obtained, the depth decision tree network can load the text semantic aggregation vector generated by each AI algorithm component into a storage condition analysis component, and the storage condition analysis component is utilized to output a discriminant variable for judging whether the storage condition of the to-be-stored AI Chat man-machine interaction text meets the standard. Further, the depth decision tree network judges whether the judging variable reaches a set judging variable, if so, the AI Chat interactive text to be stored can be considered to be subjected to tree-like structural processing and stored, and if the AI Chat interactive text to be stored is smaller than the set judging variable, the AI Chat interactive text to be stored can not be considered to be subjected to tree-like structural processing, so that the storage requirement of the database system is not met.

In view of the fact that each AI algorithm component except the first AI algorithm component takes the generation result of the upstream AI algorithm component as the input information of the current AI algorithm component, the generation result of the upstream AI algorithm component is further processed by the current AI algorithm component to obtain a text semantic aggregation vector which is as accurate and reliable as possible, and in view of the fact, the feature orders of the text semantic aggregation vectors respectively generated by the AI algorithm components are different, so that storage condition standard reaching discrimination information obtained by combining the text semantic aggregation vectors with the different feature orders is as accurate and reliable as possible.

Through the above, by using the text semantic aggregation vector generated by the AI Chat man-machine interaction text to be stored and the upstream AI algorithm components, the input information respectively corresponding to each AI algorithm component can be determined based on the text semantic aggregation vector generated by the AI Chat man-machine interaction text to be stored and the upstream AI algorithm components, so that each AI algorithm component can perform text semantic mining and text semantic aggregation processing on the input information respectively corresponding to each AI algorithm component to obtain a corresponding text semantic aggregation vector; and (3) utilizing the text semantic aggregation vectors generated by each AI algorithm component respectively to combine the text semantic aggregation vectors to obtain the storage condition standard judging information corresponding to the structured storage task to be analyzed. In view of the fact that the scene-level text semantics and the paragraph-level text semantics are combined to judge whether the storage condition of the to-be-stored AI Chat human-computer interaction text meets the standard, compared with the fact that the disturbance condition is utilized to determine the storage condition meeting the standard judging information, the embodiment of the invention can still obtain accurate and reliable storage condition meeting the standard judging information when the to-be-stored AI Chat human-computer interaction text does not meet the standard of a disturbance condition possibly met by the distinction between the to-be-stored AI Chat human-computer interaction text and the standard AI Chat human-computer interaction text.

In some examples, text semantic mining is performed on corresponding incoming information by using each AI algorithm component to obtain scene-level text semantics and paragraph-level text semantics, including: for each AI algorithm component in at least one AI algorithm component, extracting scene-level original semantics and paragraph-level original semantics in the corresponding incoming information by using the current AI algorithm component; performing first downsampling processing on the original semantics of the paragraph level by using the current AI algorithm component to obtain downsampled semantics of the paragraph level in the original semantics of the paragraph level, and obtaining corresponding semantic differences of the paragraph level according to semantic distances between the original semantics of the paragraph level and the downsampled semantics of the paragraph level; and respectively carrying out semantic detail derivation processing on scene-level original semantics and paragraph-level semantic differences by using a current AI algorithm component to obtain corresponding scene-level text semantics and paragraph-level text semantics.

For example, in order to obtain the storage condition standard reaching discrimination information which is as accurate and reliable as possible, the text semantic mining sub-network in each AI algorithm component can mine the multidimensional semantic vector in the corresponding incoming information to obtain scene level text semantics and paragraph level text semantics.

For example, the text-semantic mining sub-network in the AI algorithm component may include a scene text-semantic mining branch and a paragraph text-semantic mining branch, and when the AI algorithm component obtains the incoming information, the incoming information may be loaded into the scene text-semantic mining branch and the paragraph text-semantic mining branch, so that the scene text-semantic mining branch performs a windowed filtering operation on the incoming information to obtain scene-level original semantics, and the paragraph text-semantic mining branch performs a windowed filtering operation on the incoming information to obtain paragraph-level original semantics. Where the original semantics can be understood as the initial features and the windowed filtering operation can be understood as the convolution process.

In order to further extract feature content in the incoming information to obtain paragraph level text semantics as good as possible, the paragraph text semantic mining branches in the text semantic mining sub-network may further perform first downsampling (local pooling) on paragraph level original semantics to obtain paragraph level downsampling semantics in the paragraph level original semantics, determine a semantic distance (feature difference) between the paragraph level original semantics and the paragraph level downsampling semantics, and obtain paragraph level text semantics as good as possible based on the semantic distance, i.e., obtain paragraph level semantic differences (local feature distances).

Because the scene-level original semantics can comprise characteristic content in the incoming information, and the paragraph-level semantic differences can also comprise integral semantic details in the incoming information, the text semantic mining sub-network can respectively conduct semantic detail derivation processing on the scene-level original semantics and the paragraph-level semantic differences to obtain scene-level text semantics and paragraph-level text semantics. In some possible examples, the feature size (vector size) of the scene-level original semantics is different from the feature size (vector size) of the paragraph-level original semantics.

In some possible examples, the AI algorithm component may perform subtraction processing on the paragraph-level original semantics and the paragraph-level downsampled semantics, thereby quickly obtaining corresponding paragraph-level semantic differences, and improving the timeliness of determination of the paragraph-level semantic differences.

In some possible examples, the AI algorithm component may include a text-semantic mining subnet to extract scene-level text semantics and paragraph-level text semantics in the incoming information. The text semantic mining sub-network can comprise scene text semantic mining branches and paragraph text semantic mining branches, scene-level original semantics in the incoming information can be extracted by utilizing the scene text semantic mining branches, and semantic detail derivation processing (semantic feature supplementing processing) is carried out on the scene-level original semantics to obtain scene-level text semantics; the paragraph text semantic mining branch is utilized to extract paragraph level original semantics in the incoming information, first downsampling processing is carried out on the paragraph level original semantics to obtain paragraph level downsampling semantics, and according to semantic distances between the paragraph level original semantics and the paragraph level downsampling semantics, paragraph level semantic differences are obtained, semantic detail derivation processing is carried out on the paragraph level semantic differences to obtain paragraph level text semantics. Since the paragraph text semantic mining branch in the AI algorithm component is used for extracting the paragraph level text semantic in the incoming information, which is mainly aimed at the feature content in the incoming information, and the scene text semantic mining branch is used for extracting the scene level text semantic in the incoming information, it is seen that in the paragraph text semantic mining branch, the paragraph level original semantic minus the paragraph level downsampling semantic can obtain better feature content. The scene text semantic mining branch does not need to extract feature content, so the scene text semantic mining branch does not need to implement the definite capture of the semantic distance.

Therefore, by determining the semantic distance between the original semantic of the paragraph level and the downsampled semantic of the paragraph level, the paragraph level semantic difference comprising the best possible characteristic content can be obtained, so that the paragraph level text semantic obtained based on the paragraph level semantic difference can be accurate and reliable as far as possible. In addition, the paragraph-level downsampling semantics are cleaned in the paragraph-level original semantics, so that the cleaning of noise characteristics can be realized, analysis errors are reduced, and the reusability of the deep decision tree network is improved.

Under some examples, performing semantic detail derivation processing on scene-level original semantics and paragraph-level semantic differences respectively to obtain corresponding scene-level text semantics and paragraph-level text semantics, including: extracting scene-level dialogue features from paragraph-level semantic differences, and adding the scene-level dialogue features to scene-level original semantics to obtain scene-level text semantics; and extracting paragraph-level attribute features from the scene-level original semantics, and adding the paragraph-level attribute features to paragraph-level semantic differences to obtain paragraph-level text semantics.

For example, the scene-level dialog feature may be included in the scene-level semantic difference, and correspondingly, the scene-level attribute feature may also be included in the scene-level original semantic, so, in order to obtain the paragraph-level text semantic including more semantic details, the AI algorithm component may further extract the paragraph-level attribute feature from the scene-level original semantic and add the paragraph-level attribute feature to the paragraph-level semantic difference, to obtain the paragraph-level text semantic. And in order to obtain scene-level text semantics including more overall features, the current AI algorithm component may extract scene-level dialog features from the paragraph-level semantic differences and add the scene-level dialog features to the scene-level original semantics to obtain scene-level text semantics. The scene-level dialogue features reflect the overall text of the structured storage task to be analyzed, and the paragraph-level attribute features reflect the information of text paragraphs of the structured storage task to be analyzed.

In some possible examples, the AI algorithm component may utilize a scene text semantic mining branch to perform a windowed filtering operation on the scene-level original semantic to obtain a first windowed filtering result, and perform a windowed filtering operation and feature expansion on the paragraph-level semantic difference to obtain a scene-level dialog feature. Further, the scene text semantic mining branch can aggregate the scene-level dialogue features and the first windowed filtering result to add the scene-level dialogue features to the scene-level original semantics to obtain scene-level text semantics. For example, the scene text semantic mining utilizes the arrangement of scene-level dialogue features and the first windowed filtering result to obtain corresponding scene-level text semantics.

The AI algorithm component can utilize paragraph text semantic mining branches to conduct windowing filtering operation on paragraph level semantic differences to obtain second windowing filtering results, conduct windowing filtering operation on scene level original semantics to obtain paragraph level attribute features, and conduct aggregation processing on the paragraph level attribute features and the second windowing filtering results to add the paragraph level attribute features to the paragraph level semantic differences to obtain paragraph detail features.

Therefore, scene-level text semantics containing more text of the whole structured storage task to be analyzed can be obtained by adding scene-level dialogue features in the paragraph-level semantic differences into scene-level original semantics; the paragraph level text semantics containing more paragraph level attribute features of the structured storage task to be analyzed can be obtained by adding the paragraph level attribute features in the scene level original semantics to the paragraph level semantic differences; by using the scene layer text semantics containing more to-be-analyzed structured storage task overall texts and the paragraph layer text semantics containing more to-be-analyzed structured storage task paragraph level attribute features, the storage condition standard reaching discrimination information determined based on the scene layer text semantics and the paragraph layer text semantics can be as accurate and reliable as possible.

Under some examples, performing text semantic aggregation processing on scene-level text semantics and paragraph-level text semantics to obtain corresponding text semantic aggregation vectors, including: determining semantic enhancement factors respectively corresponding to scene level text semantics and paragraph level text semantics; and carrying out text semantic aggregation processing on the scene-level text semantics and the paragraph-level text semantics based on semantic enhancement factors respectively corresponding to the scene-level text semantics and the paragraph-level text semantics to obtain corresponding text semantic aggregation vectors.

For each AI algorithm component in at least one AI algorithm component, text semantic aggregation sub-networks in each AI algorithm component can be utilized to perform text semantic aggregation processing on corresponding scene-level text semantics and paragraph-level text semantics to obtain corresponding text semantic aggregation vectors.

When the text semantic mining sub-network in the AI algorithm component extracts the scene level text semantic and the paragraph level text semantic in the incoming information, the text semantic aggregation sub-network in the AI algorithm component can determine semantic enhancement factors respectively corresponding to the scene level text semantic and the paragraph level text semantic, and perform text semantic aggregation processing on the scene level text semantic and the paragraph level text semantic based on the semantic enhancement factors respectively corresponding to the scene level text semantic and the paragraph level text semantic to obtain corresponding text semantic aggregation vectors. For example, the text semantic aggregation sub-network in the AI algorithm component can perform weight enhancement aggregation processing on scene-level text semantics and paragraph-level text semantics to obtain text semantic aggregation vectors.

Therefore, the text semantic aggregation sub-network can adaptively refine high-quality multidimensional data through differentiated semantic enhancement factor configuration, so that the text semantic aggregation vector obtained based on the high-quality multidimensional data can improve the representation performance of text structure semantics.

In some examples, determining semantic enhancement factors to which scene-level text semantics and paragraph-level text semantics respectively correspond includes: windowing filtering operation is respectively carried out on the scene level text semantics and the paragraph level text semantics to obtain scene level semantic filtering characteristics and paragraph level semantic filtering characteristics, and splicing operation is carried out on the scene level semantic filtering characteristics and the paragraph level semantic filtering characteristics to obtain corresponding semantic splicing results (which can be understood as aggregation characteristics); and determining semantic enhancement factors respectively corresponding to the scene-level text semantics and the paragraph-level text semantics according to the semantic splicing result.

For example, when the AI algorithm component needs to determine the semantic enhancement factors corresponding to the scene-level text semantics and the paragraph-level text semantics respectively, the text semantic aggregation subnetwork in the AI algorithm component may perform windowed filtering operation on the scene-level text semantics and the paragraph-level text semantics respectively to obtain scene-level semantic filtering features and paragraph-level semantic filtering features, and fuse the scene-level semantic filtering features and the paragraph-level semantic filtering features to obtain a basis for determining the semantic enhancement factors.

Further, the text semantic aggregation sub-network determines semantic enhancement factors corresponding to the scene-level text semantics and the paragraph-level text semantics respectively according to the semantic splicing result. And then, the determined semantic enhancement factors can be utilized to select and obtain details with higher generalization scores from the paragraph level text semantics and the scene level text semantics, so that the interference of noise characteristics is avoided, and the reusability of the deep decision tree network is improved.

In some examples, determining semantic enhancement factors corresponding to the scene-level text semantics and the paragraph-level text semantics respectively according to the semantic stitching results includes: determining a first characteristic relation list and a first standardized algorithm layer corresponding to the scene level text semantics, and determining a second characteristic relation list and a second standardized algorithm layer corresponding to the paragraph level text semantics; performing second downsampling processing on the semantic splicing result to obtain semantic splicing downsampling characteristics; performing first feature mapping on the semantic stitching downsampling features by using a first feature relation list and a first standardized algorithm layer to obtain semantic enhancement factors corresponding to scene layer text semantics; and performing second feature mapping on the semantic splicing downsampling features by using a second feature relation list and a second normalization algorithm layer to obtain semantic enhancement factors corresponding to the text semantics of the paragraph level.

The feature relation list can be understood as a feature matrix, and the normalization process can be understood as a normalization process. For example, a text semantic aggregation subnet in the AI algorithm component can determine a first list of feature relationships and a first normalization algorithm layer corresponding to scene-level text semantics and a second list of feature relationships and a second normalization algorithm layer corresponding to paragraph-level text semantics. The first feature relation list, the first standardized algorithm layer, the second feature relation list and the second standardized algorithm layer can be obtained by utilizing a deep decision tree network for debugging, for example, the deep decision tree network can be obtained by utilizing the last cycle for debugging, so that the first feature relation list, the first standardized algorithm layer, the second feature relation list and the second standardized algorithm layer are obtained. Wherein the first feature relation list and the second feature relation list may be one fully connected subnet, respectively, as an example. The first normalization algorithm layer and the second normalization algorithm layer may specifically be one normalization subnet respectively.

Further, when the semantic stitching result is obtained, the text semantic aggregation sub-network may perform a second downsampling process (global average pooling process) on the semantic stitching result, so as to obtain a semantic stitching downsampling feature. For example, the text semantic aggregation sub-network can perform an averaging process on the whole semantic vector relation network corresponding to the semantic splicing result, so as to obtain the semantic splicing downsampling feature. The text semantic aggregation sub-network loads the semantic splicing downsampling feature to the full-connection sub-network corresponding to the first feature relation list, and loads data generated by the full-connection sub-network corresponding to the first feature relation list to the standardized sub-network corresponding to the first standardized algorithm layer, so that the first feature mapping is carried out on the semantic splicing downsampling feature by utilizing the first feature relation list and the first standardized algorithm layer, and a semantic enhancement factor corresponding to the scene-level text semantic is obtained. Correspondingly, the text semantic aggregation sub-network loads the semantic splicing downsampling feature to the full-connection sub-network corresponding to the second feature relation list, and loads data generated by the full-connection sub-network corresponding to the second feature relation list to the standardized sub-network corresponding to the second standardized algorithm layer, so that the second feature mapping is carried out on the semantic splicing downsampling feature by utilizing the second feature relation list and the second standardized algorithm layer, and a semantic enhancement factor corresponding to the paragraph level text semantic is obtained.

Therefore, the corresponding semantic enhancement factors can be determined by carrying out full-connection subnet processing and standardization processing on the semantic splicing downsampling characteristics, so that the determining timeliness of the semantic enhancement factors is improved.

Under some examples, when only one AI algorithm component is included in the deep decision tree network, the big data decision analysis server obtains the AI Chat man-machine interaction text to be stored, takes the AI Chat man-machine interaction text to be stored as the incoming information of the AI algorithm component, performs text semantic mining on the incoming information by using the AI algorithm component to obtain scene level text semantics and paragraph level text semantics, and performs text semantic aggregation processing on the scene level text semantics and the paragraph level text semantics to obtain a text semantic aggregation vector. Further, the depth decision tree network loads the text semantic aggregation vector generated by the AI algorithm component into the storage condition analysis component, and the storage condition analysis component is utilized to output storage condition standard judging information for judging whether the storage condition of the to-be-stored AI Chat man-machine interaction text is standard.

In some examples, when the deep decision tree network includes a plurality of AI algorithm components, the big data decision analysis server obtains the AI Chat man-machine interaction text to be stored, takes the AI Chat man-machine interaction text to be stored as the incoming information of the first AI algorithm component, and performs text semantic mining and text semantic aggregation processing on the incoming information by using the first AI algorithm component to obtain a text semantic aggregation vector generated by the first AI algorithm component. The depth decision tree network takes the text semantic aggregation vector generated by the first AI algorithm component as the incoming information of the second AI algorithm component, so that the second AI algorithm component performs text semantic mining and text semantic aggregation processing on the incoming information to obtain the text semantic aggregation vector generated by the second AI algorithm component, and the process is repeated until the text semantic aggregation vectors respectively generated by the AI algorithm components are obtained, so that the storage condition analysis component in the depth decision tree network can obtain the storage condition standard-reaching discrimination information based on the text semantic aggregation vectors respectively generated by the AI algorithm components.

The prior art is that a certain disturbance condition possibly met by the distinction between the to-be-stored AI Chat man-machine interaction text and the standard AI Chat man-machine interaction text is assumed, then an AI algorithm is utilized to determine a semantic vector relation network of unstructured elements, and storage condition standard judgment is carried out on the to-be-stored AI Chat man-machine interaction text based on the semantic vector relation network of the unstructured elements. However, the idea uses part of information (the information of unstructured elements) in the to-be-stored AI Chat man-machine interaction text to judge whether the storage condition meets the standard, and ignores the information of the structured elements in the to-be-stored AI Chat man-machine interaction text, so that when the to-be-stored AI Chat man-machine interaction text does not meet the standard of a certain disturbance condition possibly met by the difference between the to-be-stored AI Chat man-machine interaction text and the standard AI Chat man-machine interaction text, the judgment of the storage condition meeting the standard is difficult to accurately realize. The method can synthesize text information of different scales, so that the storage condition standard judging performance of the depth decision tree network on the to-be-stored AI Chat human-computer interactive texts of different forms is obviously improved, and the accuracy and the reliability of preamble judgment of structured storage are improved.

In some examples, the AI Chat bot-based structured database storage analysis method is performed by a deep decision tree network, the deep decision tree network being adapted with a network adapted concept, the network adapted concept comprising: acquiring a depth decision tree network to be debugged, an AI Chat man-machine interaction debugging text and debugging comments corresponding to the AI Chat man-machine interaction debugging text; the deep decision tree network comprises at least one AI algorithm component to be debugged; determining debugging incoming data corresponding to at least one AI algorithm component to be debugged respectively; the debugging incoming data of the current AI algorithm component to be debugged in the at least one AI algorithm component to be debugged comprises at least one of an AI Chat man-machine interaction debugging text and a text semantic aggregation deduction vector generated by the previous AI algorithm component to be debugged; respectively carrying out text semantic mining on corresponding debugging input data by utilizing each AI algorithm component to obtain scene-level text deduction semantics and paragraph-level text deduction semantics, and carrying out text semantic aggregation processing on the scene-level text deduction semantics and the paragraph-level text deduction semantics to obtain corresponding text semantic aggregation deduction vectors; according to the text semantic aggregation deduction vector generated by each AI algorithm component to be debugged, determining a deduction analysis result of performing storage condition deduction analysis on the AI Chat man-machine interaction debugging text; and according to the deduction analysis result and the debugging annotation, debugging the deep decision tree network until the deep decision tree network meets the debugging completion requirement, and obtaining the debugged deep decision tree network.

The above-mentioned debug text can be understood as training text, the debug comment is usually understood as a sample tag, and further, the text semantic aggregation deduction vector can be understood as a prediction result of the text semantic aggregation vector.

For example, before the storage condition standard is judged for the to-be-stored AI Chat man-machine interaction text by using the deep decision tree network, the deep decision tree network is debugged. The big data decision analysis server can acquire sufficient AI Chat man-machine interaction debugging texts and debugging comments corresponding to the AI Chat man-machine interaction debugging texts, so that the big data decision analysis server can carry out cyclic debugging on the depth decision tree network to be debugged based on the acquired AI Chat man-machine interaction debugging texts and corresponding debugging comments. In a single debugging process of cyclic debugging, the big data decision analysis server takes the AI Chat man-machine interaction debugging text as debugging input data of a first AI algorithm component to be debugged in at least one AI algorithm component to be debugged, so that the first AI algorithm component to be debugged can conduct text semantic mining on the debugging input data to obtain scene-level text deduction semantics and paragraph-level text deduction semantics, and conduct text semantic aggregation processing on the scene-level text deduction semantics and the paragraph-level text deduction semantics to obtain a text semantic aggregation deduction vector generated by the first AI algorithm component to be debugged. Further, when the depth decision tree network to be debugged comprises a plurality of AI algorithm components, the depth decision tree network can take the text semantic aggregation deduction vector generated by the first AI algorithm component to be debugged as the debugging incoming data of the second AI algorithm component to enable the second AI algorithm component to be debugged to perform text semantic mining and text semantic aggregation processing on the debugging incoming data so as to obtain the text semantic aggregation deduction vector generated by the second AI algorithm component to be debugged. Further, the depth decision tree network uses the text semantic aggregation deduction vector generated by the second AI algorithm module to be debugged as the debugging input data of the third AI algorithm module to be debugged, and the process is repeated until the text semantic aggregation deduction vector generated by the terminal AI algorithm module to be debugged is obtained. The process of outputting the text semantic aggregation deduction vector by the AI algorithm component to be debugged can refer to the process of outputting the text semantic aggregation vector by the AI algorithm component.

Further, the depth decision tree network aggregates deduction vectors according to text semantics respectively generated by each AI algorithm component to be debugged, determines deduction analysis results of storage condition deduction analysis on the AI Chat man-machine interaction debugging text, and debugs the depth decision tree network according to differences between the deduction analysis results (prediction results of whether storage conditions reach standards or not) and corresponding debugging comments. When the set round of cyclic debugging is carried out or the result generated by the deep decision tree network meets the debugging completion condition, the big data decision analysis server can be considered to be in accordance with the debugging completion requirement, and the deep decision tree network after the debugging is completed is obtained by stopping the debugging of the deep decision tree network.

In the embodiment of the invention, the depth decision tree network to be debugged is debugged, so that the storage condition standard reaching judgment result generated based on the debugged depth decision tree network can be as accurate as possible.

In some examples, the deduction analysis results include a text structure distribution deduction result and a storage condition deduction analysis identity; according to the text semantic aggregation deduction vector generated by each AI algorithm component to be debugged, determining a deduction analysis result of performing storage condition deduction analysis on the AI Chat man-machine interaction debugging text, wherein the deduction analysis result comprises the following steps: using a text structure distribution generating component in the depth decision tree network to perform text structure distribution generating processing on the text semantic aggregation deduction vectors respectively generated by each AI algorithm component to be debugged, so as to obtain a corresponding text structure distribution deduction result; and carrying out storage condition deduction analysis on the text semantic aggregation deduction vectors respectively generated by each AI algorithm component to be debugged by using a storage condition analysis component in the depth decision tree network to obtain corresponding storage condition deduction analysis identifiers.

The text structure distribution deduction result can be understood as text structure distribution which is generated by the depth decision tree network to be debugged and is used for reflecting text logic architecture data of the structured storage task to be analyzed in the AI Chat man-machine interaction debugging text, for example, the text structure distribution deduction result can be specifically the text structure distribution of the structured storage task item to be analyzed in the AI Chat man-machine interaction debugging text. The storage condition deduction analysis identifier may be understood as an annotation generated by the deep decision tree network to be debugged, wherein the annotation is used for reflecting whether the debug text meets the storage condition, for example, the storage condition deduction analysis identifier may be "the AI Chat man-machine interaction debug text can be subjected to tree structure processing and stored", and may be "the AI Chat man-machine interaction debug text cannot be subjected to tree structure processing and thus cannot be subjected to structure storage".

For example, a text structure distribution generation component and a storage condition analysis component can be included in a deep decision tree network to be debugged. When the text semantic aggregation deduction vectors respectively generated by the AI algorithm components are obtained, the storage condition analysis component can load the text semantic aggregation deduction vectors respectively generated by the AI algorithm components to the text structure distribution generation component and to the storage condition analysis component, so that the text structure distribution generation component can obtain a text structure distribution deduction result corresponding to the AI Chat man-machine interaction debugging text based on the text semantic aggregation deduction vectors respectively generated by the AI algorithm components, and the storage condition analysis component can obtain a storage condition deduction analysis identification corresponding to the AI Chat man-machine interaction debugging text based on the text semantic aggregation deduction vectors respectively generated by the AI algorithm components. The storage condition analysis component can be, for example, a classifier. In some possible examples, the text structure distribution generation component may be used only to debug the deep decision tree network, without involving the process of discriminating between AI Chat human-machine interactive text to be stored.

In some possible examples, the deep decision tree network to be debugged may output corresponding text semantic aggregation deduction vectors by using each AI algorithm component to be debugged, so that the text structure distribution generating component and the storage condition analysis component may obtain corresponding text structure distribution deduction results and storage condition deduction analysis identifications based on each text semantic aggregation deduction vector.

Therefore, the text structure distribution deduction result and the storage condition deduction analysis identifier are obtained, and the to-be-debugged depth decision tree network can be debugged based on the text structure distribution deduction result and the storage condition deduction analysis identifier.

In some examples, the debug annotations include a target text structure distribution and a storage condition authentication annotation; according to the deduction analysis result and the debugging annotation, the deep decision tree network is debugged until meeting the debugging completion requirement, and the method comprises the following steps: determining a first comparison result between the text structure distribution deduction result and the target text structure distribution, and determining a second comparison result between the storage condition deduction analysis identifier and the storage condition authentication annotation; and debugging the deep decision tree network by using the first comparison result and the second comparison result until the deep decision tree network meets the debugging completion requirement.

The target text structure distribution can be understood as standard text structure distribution, and the storage condition authentication annotation can be understood as a correct storage condition standard judging result. When the AI Chat man-machine interaction debugging text is obtained, annotation processing can be carried out on the AI Chat man-machine interaction debugging text, and corresponding target text structure distribution and storage condition authentication annotation can be obtained.

For example, the big data decision analysis server may determine a first comparison result between the text structure distribution deduction result and the corresponding target text structure distribution, and determine a second comparison result between the storage condition deduction analysis identifier and the storage condition authentication annotation, generate a first training cost function using the first comparison result, generate a second training cost function using the second comparison result, and debug the deep decision tree network using the generated first training cost function and second training cost function until meeting a debug completion requirement. The training cost function can be understood as a loss function.

Therefore, by utilizing the generated text structure distribution deduction result and the storage condition deduction analysis mark, the depth decision tree network can be debugged by combining a first comparison result between the text structure distribution deduction result and the target text structure distribution and a second comparison result between the storage condition deduction analysis mark and the storage condition authentication annotation, so that the debugged depth decision tree network can be as accurate as possible.

Under another design concept, the structured database storage analysis method based on the AI Chat bot comprises the following contents.

STEP302, acquiring a basic AI Chat man-machine interaction text, and determining an AI man-machine dialogue text set of a structured storage task to be analyzed in the basic AI Chat man-machine interaction text; updating the scale of the AI man-machine conversation text set to obtain a target conversation text set; and extracting a target dialogue text set to obtain the to-be-stored AI Chat man-machine interaction text matched with the to-be-analyzed structured storage task.

STEP304, for the first AI algorithm component of the at least one AI algorithm component, uses the AI Chat man-machine interaction text to be stored as the incoming information of the first AI algorithm component.

STEP306, for each AI algorithm component of the at least one AI algorithm component except the first AI algorithm component, uses the text semantic aggregate vector generated by the upstream AI algorithm component preceding the corresponding AI algorithm component as the incoming information of the corresponding AI algorithm component, where the upstream AI algorithm component may be understood as an AI algorithm component that is neighboring to and preceding the current AI algorithm component.

STEP308, AI algorithm components include text semantic mining sub-networks, for each AI algorithm component in at least one AI algorithm component, the text semantic mining sub-network in the current AI algorithm component is utilized to extract scene-level original semantics and paragraph-level original semantics in the corresponding incoming information, and the paragraph-level original semantics are subjected to first downsampling processing to obtain paragraph-level downsampling semantics in the paragraph-level original semantics, and corresponding paragraph-level semantic differences are obtained according to semantic distances between the paragraph-level original semantics and the paragraph-level downsampling semantics.

STEP310, extracting scene-level dialogue features from paragraph-level semantic differences by utilizing text semantic mining subnets in the current AI algorithm components, and adding the scene-level dialogue features to scene-level original semantics to obtain scene-level text semantics; and extracting paragraph-level attribute features from the scene-level original semantics, and adding the paragraph-level attribute features to paragraph-level semantic differences to obtain paragraph-level text semantics.

STEP312, AI algorithm component still includes text semantic aggregation subnetwork, utilize text semantic aggregation subnetwork to carry out window filtering operation to scene level text semantic and paragraph level text semantic respectively, obtain scene level semantic filtering characteristic and paragraph level semantic filtering characteristic, and splice operation to scene level semantic filtering characteristic and paragraph level semantic filtering characteristic, obtain the semantic concatenation result.

STEP314, determining a first feature relation list and a first standardized algorithm layer corresponding to the scene level text semantics by using the text semantic aggregation subnet, and determining a second feature relation list and a second standardized algorithm layer corresponding to the paragraph level text semantics, where, for example, the first feature relation list and the second feature relation list may be a fully connected subnet, and the first standardized algorithm layer and the second standardized algorithm layer may be a standardized subnet, respectively.

STEP316, the semantic splicing result is subjected to second downsampling treatment by using the text semantic aggregation sub-network, so that semantic splicing downsampling characteristics are obtained; for example, the text semantic aggregation sub-network can perform an averaging process on the whole semantic vector relation network corresponding to the semantic splicing result, so as to obtain the semantic splicing downsampling feature.

STEP318, utilizing a text semantic aggregation subnet, and performing first feature mapping on the semantic splicing downsampling features based on a first feature relation list and a first standardized algorithm layer to obtain semantic enhancement factors corresponding to scene-level text semantics; and performing second feature mapping on the semantic splicing downsampling features by using a second feature relation list and a second normalization algorithm layer to obtain semantic enhancement factors corresponding to the text semantics of the paragraph level.

STEP320 utilizes the text semantic aggregation sub-network, and performs text semantic aggregation processing on the scene-level text semantics and the paragraph-level text semantics based on semantic enhancement factors respectively corresponding to the scene-level text semantics and the paragraph-level text semantics, so as to obtain corresponding text semantic aggregation vectors.

STEP322, according to the text semantic aggregation vector generated by each AI algorithm component, determines the storage condition standard reaching judgment information corresponding to the structured storage task to be analyzed, for example, the text semantic aggregation vector generated by each AI algorithm component is loaded into the storage condition analysis component, and the storage condition analysis component is utilized to output the judgment variable for judging the storage condition standard reaching of the AI Chat man-machine interaction text to be stored.

In the structured database storage analysis method based on the AI Chat bot, the text semantic aggregation vector generated by the AI Chat man-machine interaction text to be stored and the upstream AI algorithm components is obtained, and the incoming information respectively corresponding to each AI algorithm component can be determined based on the text semantic aggregation vector generated by the AI Chat man-machine interaction text to be stored and the upstream AI algorithm components, so that each AI algorithm component can carry out text semantic mining and text semantic aggregation processing on the incoming information respectively corresponding to each AI algorithm component to obtain the corresponding text semantic aggregation vector; and combining the text semantic aggregation vectors by utilizing the text semantic aggregation vectors generated by each AI algorithm component, so as to obtain storage condition standard judging information corresponding to the structured storage task to be analyzed. In view of the fact that the scene-level text semantics and the paragraph-level text semantics are combined to judge whether the storage condition of the to-be-stored AI Chat human-computer interaction text meets the standard, compared with the fact that the disturbance condition is utilized to determine the storage condition meeting the standard judging information, the embodiment of the invention can still obtain accurate and reliable storage condition meeting the standard judging information when the to-be-stored AI Chat human-computer interaction text does not meet the standard of a disturbance condition possibly met by the distinction between the to-be-stored AI Chat human-computer interaction text and the standard AI Chat human-computer interaction text.

STEP402, obtaining a depth decision tree network to be debugged, an AI Chat man-machine interaction debugging text and a debugging annotation corresponding to the AI Chat man-machine interaction debugging text; the deep decision tree network includes at least one AI algorithm component to be debugged.

STEP404, determining at least one debugging incoming data corresponding to the AI algorithm components to be debugged respectively, for example, the debugging incoming data corresponding to the first AI algorithm component to be debugged is an AI Chat man-machine interaction debugging text, and the debugging incoming data corresponding to the other AI algorithm components except the first AI algorithm component to be debugged is a text semantic aggregation deduction vector; the debugging incoming data of the current AI algorithm component to be debugged in the at least one AI algorithm component to be debugged comprises at least one of AI Chat man-machine interaction debugging text and text semantic aggregation deduction vectors generated by the previous AI algorithm component to be debugged.

STEP406 performs text semantic mining on the corresponding debugging input data by utilizing each AI algorithm component to obtain scene-level text deduction semantics and paragraph-level text deduction semantics, and performs text semantic aggregation processing on the scene-level text deduction semantics and the paragraph-level text deduction semantics to obtain corresponding text semantic aggregation deduction vectors.

STEP408, using the text structure distribution generating component in the deep decision tree network, performs text structure distribution generating processing on the text semantic aggregation deduction vectors generated by each to-be-debugged AI algorithm component respectively to obtain a corresponding text structure distribution deduction result, for example, obtain the text structure distribution of the to-be-analyzed structured storage task item in the AI Chat man-machine interaction debugging text.

STEP410 performs storage condition deduction analysis on the text semantic aggregation deduction vectors respectively generated by each AI algorithm component to be debugged by using a storage condition analysis component in the deep decision tree network to obtain corresponding storage condition deduction analysis identifiers, for example, a discrimination result of "AI Chat man-machine interaction debugging text can be subjected to tree structure processing and stored" is obtained, or a discrimination result of "AI Chat man-machine interaction debugging text cannot be subjected to tree structure processing and therefore cannot be subjected to structural storage" is obtained.

STEP412, determining a first comparison result between the text structure distribution deduction result and the target text structure distribution, for example, subtracting the text structure distribution deduction result from the target text structure distribution to obtain a first comparison result, and determining a second comparison result between the storage condition deduction analysis identifier and the storage condition authentication annotation; and debugging the deep decision tree network by using the first comparison result and the second comparison result until the debugging completion requirement is met, for example, until the preset cycle debugging times are met.

STEP414, acquiring an AI Chat man-machine interaction text to be stored, which corresponds to the structured storage task to be analyzed; determining the input information corresponding to at least one AI algorithm component respectively; the incoming information for each AI algorithm component includes at least one of an AI Chat man-machine interaction text to be stored, and a text semantic aggregation vector generated by an upstream AI algorithm component preceding the corresponding AI algorithm component.

STEP416 performs text semantic mining on the corresponding incoming information by using each AI algorithm component to obtain scene-level text semantics and paragraph-level text semantics, and performs text semantic aggregation processing on the scene-level text semantics and the paragraph-level text semantics to obtain corresponding text semantic aggregation vectors.

STEP418 determines storage condition standard reaching discrimination information corresponding to the structured storage task to be analyzed according to the text semantic aggregation vector generated by each AI algorithm component respectively.

Based on the foregoing, in some independent embodiments, after determining the storage condition criterion discrimination information corresponding to the to-be-analyzed structured storage task according to the text semantic aggregation vector generated by each AI algorithm component, the method further includes: responding to a structured storage auxiliary request aiming at the storage condition standard judging information as the storage condition standard, and performing tree structured conversion on the AI Chat man-machine interaction text to be stored to obtain a structured AI Chat text; and sending the structured AI Chat text to a database system.

Therefore, the tree-structured conversion task of the database system can be shared, so that the processing pressure of the database system is released, and the data storage and management efficiency of the database system is improved.

Based on the foregoing, in some independent embodiments, after obtaining the structured AI Chat text, the method further comprises: caching the structured AI Chat text; responding to a call request aiming at the structured AI Chat text, carrying out data security detection on a call server to obtain a detection result, and issuing the structured AI Chat text to the call server on the premise that the detection result characterizes that the call server has no potential safety hazard.

Therefore, after the big data decision analysis server converts the structured AI Chat text to obtain the structured AI Chat text, the structured AI Chat text can be cached, and in the cache time, if a call request for the structured AI Chat text is received, the security detection of the call server can be realized, so that the structured AI Chat text is issued to the call server on the premise that the call server has no potential safety hazard. Therefore, the database system can be prevented from repeatedly processing related call requests in the process of storing the structured AI Chat text, and the calculation power of the big data decision analysis server is stronger than that of the database system, so that accurate safety analysis before calling can be realized, and the safety of the structured AI Chat text in the calling process is improved.

Based on the foregoing, in some independent embodiments, the performing data security detection on the call server to obtain a detection result includes: acquiring a detection response linear array and an interactive feedback linear array of the calling server, wherein the detection response linear array reflects detection response data in the calling server, and the interactive feedback linear array reflects interactive feedback data in the calling server; for the detection response linear array, acquiring a first safety evaluation linear array based on a part of linear arrays associated with the detection response linear array in the interactive feedback linear array, wherein the first safety evaluation linear array reflects the detection response linear array fused with the interactive feedback linear array; based on a part of linear arrays associated with the interactive feedback linear array in the detection response linear array, the interactive feedback linear array is used for acquiring a second safety evaluation linear array, and the second safety evaluation linear array reflects the interactive feedback linear array fused with the detection response linear array; fusing the first security evaluation linear array and the second security evaluation linear array to obtain a linkage security evaluation linear array; and determining a security detection tag corresponding to the calling server based on the linkage security assessment linear array.

Therefore, the security detection label corresponding to the calling server can be accurately obtained by comprehensively analyzing the detection response data and the interactive feedback data.

Based on the foregoing, in some independent embodiments, the obtaining the detection response linear array and the interaction feedback linear array of the call server includes: extracting a detection response trend linear array of the detection response data, and fusing the detection response trend linear array with a detection response state linear array to obtain the detection response linear array; and extracting an interactive feedback view point linear array of the interactive feedback data, and fusing the interactive feedback view point linear array and the interactive feedback state linear array to obtain the interactive feedback linear array.

Based on the foregoing, in some independent embodiments, the detection reply data includes reply item data and reply content data; the detection response linear array comprises a response item linear array and a response content linear array; the extracting the detection response trend linear array of the detection response data comprises the following steps: extracting a response item trend linear array of the response item data and a response content trend linear array of the response content data; the fusing the detection response trend linear array with the detection response state linear array to obtain the detection response linear array comprises the following steps: fusing the response item trend linear array with the response item state linear array to obtain the response item linear array; and fusing the response content trend linear array with the response content state linear array to obtain the response content linear array.

Based on the foregoing, in some independent embodiments, the first security assessment linear array includes a response item security assessment linear array and a response content security assessment linear array, and the obtaining, for the detection response linear array, based on a part of the linear arrays associated with the detection response linear array in the interactive feedback linear array, the first security assessment linear array includes: for the response item linear array, based on the response content linear array and a part of linear arrays respectively associated with the response item linear array in the interactive feedback linear array, acquiring a response item security evaluation linear array, wherein the response item security evaluation linear array reflects the response item linear array after the response content linear array and the interactive feedback linear array are fused; and for the response content linear array, acquiring the response content security assessment linear array based on partial linear arrays respectively associated with the response content linear array in the response item linear array and the interactive feedback linear array, wherein the response content security assessment linear array reflects the response content linear array after the response item linear array and the interactive feedback linear array are fused.

Based on the foregoing, in some independent embodiments, the obtaining the response item security assessment linear array based on the response content linear array and the partial linear array associated with the response item linear array in the interactive feedback linear array respectively includes: based on a part of linear arrays associated with the response item linear array in the response content linear array, acquiring a first response item security evaluation linear array; based on a part of linear arrays associated with the response item linear array in the interactive feedback linear array, acquiring a second response item security evaluation linear array; splicing the first response item security assessment linear array and the second response item security assessment linear array to obtain a third response item security assessment linear array; and processing the third response item security evaluation linear array to obtain the response item security evaluation linear array.

Based on the foregoing, in some independent embodiments, the obtaining the response content security assessment linear array based on the response item linear array and the partial linear array associated with the response content linear array in the interactive feedback linear array respectively includes: based on a part of linear arrays associated with the response content linear array in the response item linear array, acquiring a first response content security evaluation linear array; based on a part of linear arrays associated with the response content linear array in the interactive feedback linear array, acquiring a second response content security assessment linear array; splicing the first response content security assessment linear array and the second response content security assessment linear array to obtain a third response content security assessment linear array; and processing the third response content security evaluation linear array to obtain the response content security evaluation linear array.

Based on the foregoing, in some independent embodiments, the obtaining, for the interactive feedback linear array, a second security assessment linear array based on a portion of the detection response linear array associated with the interactive feedback linear array includes: and for the interactive feedback linear array, acquiring the second security evaluation linear array based on the response item linear array and the partial linear array respectively associated with the interactive feedback linear array in the response content linear array.

Based on the foregoing, in some independent embodiments, the obtaining the second security assessment linear array based on the response item linear array and the response content linear array, which are respectively associated with the interactive feedback linear array, includes: based on a part of linear arrays associated with the interactive feedback linear array in the response item linear array, a first interactive feedback security evaluation linear array is obtained; based on a part of linear arrays associated with the interactive feedback linear array in the response content linear array, a second interactive feedback security evaluation linear array is obtained; splicing the first interactive feedback security assessment linear array and the second interactive feedback security assessment linear array to obtain a third interactive feedback security assessment linear array; and processing the third interactive feedback security assessment linear array to obtain the second security assessment linear array.

Based on the foregoing, in some independent embodiments, the determining, based on the linked security assessment linear array, the security detection tag corresponding to the call server includes: predicting the linkage safety evaluation linear array to obtain a safety prediction linear array; performing feature mapping on the safe prediction linear array to obtain a plurality of trusted coefficients of the calling server belonging to a plurality of candidate labels; and selecting the candidate label corresponding to the maximum trusted coefficient as the security detection label corresponding to the calling server.

The embodiment of the invention also provides a software product for realizing the AI Chat bot-based structured database storage analysis method, which comprises a computer program/instruction, wherein the computer program/instruction realizes the method when being executed.

The embodiment of the invention also provides a computer readable storage medium, on which a computer program is stored, which when run performs the above method.

In summary, the big data decision analysis server can utilize the AI deep decision tree network to judge whether the storage condition reaches the standard or not for the AI Chat man-machine interaction text to be stored. The scheme has weak overall constraint, can be applied to AI Chat man-machine interaction texts of different types and scenes, improves the storage analysis timeliness aiming at the structured database, and can ensure the order of data stored in the structured database

The foregoing is only a specific embodiment of the present invention. Variations and alternatives will occur to those skilled in the art based on the detailed description provided herein and are intended to be included within the scope of the invention.

Claims

1. A structured database storage analysis method based on AI Chat bot, characterized by being applied to a big data decision analysis server, the method comprising:

2. The method of claim 1, wherein the obtaining the AI Chat human-machine interaction text to be stored corresponding to the structured storage task to be analyzed comprises:

3. The method of claim 1, wherein determining the respective corresponding incoming information for the at least one AI algorithm component comprises:

4. The method of claim 3, wherein the performing text semantic mining on the corresponding incoming information by using each AI algorithm component to obtain scene-level text semantics and paragraph-level text semantics, and performing text semantic aggregation processing on the scene-level text semantics and the paragraph-level text semantics to obtain corresponding text semantic aggregation vectors, includes:

5. The method of claim 1, wherein the text semantic mining of the corresponding incoming information with each AI algorithm component to obtain scene-level text semantics and paragraph-level text semantics, respectively, comprises:

6. The method of claim 5, wherein the performing semantic detail derivation processing on the scene-level original semantics and the paragraph-level semantic differences, respectively, to obtain corresponding scene-level text semantics and paragraph-level text semantics, comprises:

7. The method of claim 1, wherein performing text semantic aggregation processing on the scene-level text semantics and the paragraph-level text semantics to obtain corresponding text semantic aggregation vectors comprises:

based on semantic enhancement factors respectively corresponding to the scene-level text semantics and the paragraph-level text semantics, performing text semantic aggregation processing on the scene-level text semantics and the paragraph-level text semantics to obtain corresponding text semantic aggregation vectors;

the determining the semantic enhancement factors respectively corresponding to the scene level text semantics and the paragraph level text semantics comprises the following steps:

Determining semantic enhancement factors respectively corresponding to the scene level text semantics and the paragraph level text semantics according to the semantic splicing result;

the determining, according to the semantic splicing result, semantic enhancement factors corresponding to the scene level text semantics and the paragraph level text semantics respectively includes:

8. The method of claim 1, wherein the AI Chat bot-based structured database storage analysis method is performed by a deep decision tree network that is debugged using a network debug thread comprising:

According to the deduction analysis result and the debugging annotation, debugging the depth decision tree network until the depth decision tree network meets the debugging completion requirement, and obtaining a depth decision tree network with the completion of debugging;

the deduction analysis result comprises a text structure distribution deduction result and a storage condition deduction analysis mark; the determining a deduction analysis result of storage condition deduction analysis on the AI Chat man-machine interaction debugging text according to the text semantic aggregation deduction vectors respectively generated by each AI algorithm component to be debugged comprises the following steps:

using a storage condition analysis component in the depth decision tree network to carry out storage condition deduction analysis on the text semantic aggregation deduction vectors respectively generated by each AI algorithm component to be debugged, so as to obtain corresponding storage condition deduction analysis identifiers;

the debugging comments comprise target text structure distribution and storage condition authentication comments; and according to the deduction analysis result and the debugging annotation, debugging the depth decision tree network until the depth decision tree network meets the debugging completion requirement, wherein the method comprises the following steps:

9. A software product for implementing an AI Chat bot-based structured database storage analysis method, comprising a computer program/instruction, wherein the computer program/instruction, when executed, implements performing the method of one or more of claims 1-8.

10. A computer readable storage medium, characterized in that it has stored thereon a computer program which, when run, is a method according to one or more of claims 1-8.