CN117149996A

CN117149996A - Man-machine interface digital conversation mining method and AI system for artificial intelligence application

Info

Publication number: CN117149996A
Application number: CN202311207119.1A
Authority: CN
Inventors: 李�浩; 张姗姗; 陆冬良
Original assignee: Wuhan Tengxiangyue Technology Co ltd
Current assignee: Li Hao
Priority date: 2023-09-19
Filing date: 2023-09-19
Publication date: 2023-12-01

Abstract

The invention relates to the technical fields of artificial intelligence, digitization and big data mining, and provides a man-machine interface digital conversation mining method and an AI system for artificial intelligence application. According to the invention, the AI big data mining system can limit the similarity of the first thermodynamic relationship network and the second thermodynamic relationship network, so that the debugged session text analysis model obtained by debugging can improve the capturing precision of text blocks in HMI digital session text; in addition, when the basic session text mining network is debugged, the distributed data of each text block of the authenticated question-answer event in the authenticated HMI digital session text is not needed to be annotated, the acquisition limit of the authenticated HMI digital session text is broken, and the annotation processing of the text block distribution of the authenticated HMI digital session text can be reduced, so that the timeliness of the interactive intention mining of the HMI digital session text is improved.

Description

Man-machine interface digital conversation mining method and AI system for artificial intelligence application

Technical Field

The invention relates to the technical fields of artificial intelligence, digitization and big data mining, in particular to a man-machine interface digital conversation mining method and an AI system for artificial intelligence application.

Background

Artificial intelligence applications (Applications of artificial intelligence) are wide ranging and include, but are not limited to, computer science, digital finance, online customer service, digital virtual space, telecommunications interactions, smart cities, and the like. Along with the continuous maturity and the deep penetration of the artificial intelligence technology, the intelligent degree is higher and higher, and the artificial intelligence technology can be applied to most human-computer interaction scenes at present, so that the production efficiency is greatly released.

Human-machine interaction (Human Machine Interaction, HMI) is a technology that studies the interaction relationship between a system and a user. The human-machine interaction system may be a variety of machines, as well as computerized systems and software. Human-machine interaction interfaces generally refer to portions that are visible to a user. The user communicates with the system through a man-machine interaction interface and performs operation.

At present, the big data mining value of the man-machine interaction session is not ignored, and how to realize efficient mining analysis of the man-machine interaction session is a technical difficulty at present.

Disclosure of Invention

The invention provides at least a man-machine interface digital conversation mining method and an AI system for artificial intelligence application.

The technical scheme of the invention is realized by at least partial embodiments as follows.

An artificial intelligence application-oriented human-machine interface digital session mining method is applied to an AI big data mining system, and the method comprises the following steps:

processing the authenticated HMI digital conversation text by using a basic conversation text mining network to obtain a first interaction intention distinguishing report and a first thermodynamic relation network corresponding to the authenticated HMI digital conversation text;

processing the HMI conversation adjustment text subjected to conversation operation adjustment by utilizing the basic conversation text mining network to obtain a second interaction intention discrimination report and a second thermodynamic relation network corresponding to the HMI conversation adjustment text; the HMI session adjustment text is obtained by performing speaking adjustment on the authenticated HMI digital session text;

and improving the configuration variables of the basic conversation text mining network by combining the first thermodynamic relation network, the second thermodynamic relation network, the first interaction intention distinguishing report and the second interaction intention distinguishing report to obtain a debugged conversation text analysis model.

In some examples, the processing the authenticated HMI digital conversation text with the base conversation text mining network to obtain a first interaction intent discrimination report and a first thermodynamic relationship network corresponding to the authenticated HMI digital conversation text includes:

Acquiring an authenticated HMI digital conversation text, generating a first interaction intention distinguishing report corresponding to the authenticated HMI digital conversation text through a basic conversation text mining network, and generating a first thermodynamic relation network according to the first interaction intention distinguishing report and the authenticated semantic description of the authenticated HMI digital conversation text; the first interaction intention judging report is determined by a target query vector corresponding to an authenticated question-answer event in the authenticated HMI digital session text, and the first thermodynamic relationship network is used for representing the distributed data of text blocks of the authenticated question-answer event in the authenticated HMI digital session text;

processing the HMI conversation adjustment text subjected to conversation operation adjustment by using the basic conversation text mining network to obtain a second interaction intention discrimination report and a second thermodynamic relation network corresponding to the HMI conversation adjustment text; the HMI session adjustment text is obtained by performing a speaking adjustment on the authenticated HMI digital session text, comprising:

performing speaking adjustment on the authenticated HMI digital session text to obtain an HMI session adjustment text, generating a second interaction intention distinguishing report corresponding to the HMI session adjustment text through the basic session text mining network, and generating a second thermal relationship network according to the second interaction intention distinguishing report and the reconstruction semantic description of the HMI session adjustment text; the second interaction intention judging report is determined by a reconstruction query vector corresponding to an authenticated question-answer event in the HMI session adjustment text, and the second thermodynamic relationship network is used for representing the distribution data of text blocks of the authenticated question-answer event in the HMI session adjustment text;

The modifying the configuration variables of the basic conversation text mining network by combining the first thermodynamic relation network, the second thermodynamic relation network, the first interactive intention discrimination report and the second interactive intention discrimination report to obtain a debugged conversation text analysis model comprises the following steps:

determining similarity cost information of the basic conversation text mining network according to the first thermodynamic relation network and the second thermodynamic relation network, and determining intention judging cost information of the basic conversation text mining network according to the first interaction intention judging report, the second interaction intention judging report and text block intention words carried by the authenticated HMI digital conversation text;

combining the similarity cost information and the intention discrimination cost information, improving configuration variables of the basic session text mining network, and generating a debugged session text analysis model; the debugged session text analysis model is used for mining local intention and distribution reports of question and answer events corresponding to text blocks in the HMI digital session text to be processed.

In some examples, the generating, by the base session text mining network, a first interaction intent discrimination report corresponding to the authenticated HMI digital session text, generating a first thermodynamic relationship network according to the first interaction intent discrimination report and an authenticated semantic description of the authenticated HMI digital session text, includes:

Loading the authenticated HMI digital session text to the basic session text mining network, and acquiring a target query vector corresponding to the authenticated question-answer event in the authenticated HMI digital session text according to the basic session text mining network;

processing the target query vector according to a judging component in the basic session text mining network to obtain a first interaction intention judging report corresponding to the authenticated HMI digital session text;

acquiring authenticated semantic descriptions for the authenticated HMI digital conversation text generated by a target semantic coding component in the basic conversation text mining network, and weighting the first interaction intention discrimination report and the authenticated semantic descriptions to obtain an initial thermodynamic relationship network corresponding to the authenticated HMI digital conversation text;

and expanding the initial thermodynamic relation network to obtain a first thermodynamic relation network with the same visual scale as the authenticated HMI digital session text.

In some examples, the obtaining, according to the base session text mining network, a target query vector corresponding to the authenticated question-answer event in the authenticated HMI digital session text includes:

In the basic session text mining network, acquiring a scene inquiry vector corresponding to the authenticated question-answer event in the authenticated HMI digital session text, and generating a scene interaction intention discrimination report corresponding to the scene inquiry vector through a discrimination component in the basic session text mining network;

weighting the scene interaction intention discrimination report and the authenticated semantic description to obtain scene feature clusters corresponding to the authenticated HMI digital session text, and carrying out disassembly processing on the authenticated HMI digital session text according to the scene feature clusters to obtain X session text paragraphs; x is a positive integer;

sequentially loading the X session text paragraphs into the basic session text mining network, and acquiring paragraph query vectors corresponding to the X session text paragraphs in the basic session text mining network;

and vector aggregation is carried out on the scene query vector and paragraph query vectors corresponding to the X conversation text paragraphs, so that target query vectors corresponding to the authenticated question-answer events in the authenticated HMI digital conversation text are obtained.

In some examples, the base session text-mining network includes Y gradient focus sub-networks, each gradient focus sub-network including at least one semantic coding component, Y being a positive integer; the step of obtaining a scene query vector corresponding to the authenticated question-answer event in the authenticated HMI digital session text in the basic session text mining network comprises the following steps:

Acquiring the incoming information of a ith gradient focus subnet in the Y gradient focus subnets; when u is 1, the incoming information of the ith gradient focusing subnet is the authenticated HMI digital session text, and u is a positive integer smaller than Y;

according to at least one semantic coding component in the ith gradient focusing sub-network, carrying out sliding window coding on the incoming information of the ith gradient focusing sub-network to obtain initial semantic description;

integrating the initial semantic description with the incoming information of the (u) th gradient focus subnet to obtain a semantic description combination result of the (u) th gradient focus subnet, and taking the semantic description combination result of the (u) th gradient focus subnet as the incoming information of the (u+1) th gradient focus subnet; the u-th gradient focus subnet is cascaded with the u+1-th gradient focus subnet;

and determining the semantic description combination result of the Y-th gradient focus subnet as a scene query vector corresponding to the authenticated question-answer event in the authenticated HMI digital session text.

In some examples, the number of scene interrogation vectors is Z, Z being a positive integer; the generating, by the discriminating component in the basic session text mining network, a scene interaction intention discriminating report corresponding to the scene query vector includes:

Determining target feature variables corresponding to Z scene query vectors respectively, and integrating the target feature variables corresponding to the Z scene query vectors into scene text knowledge;

converting the scene text knowledge into scene text knowledge to be distinguished according to a feature mapping component in the basic session text mining network;

and loading the scene text knowledge to be discriminated into a discriminating component in the basic conversation text mining network, and generating a scene interaction intention discriminating report corresponding to the scene text knowledge to be discriminated through the discriminating component in the basic conversation text mining network.

In some examples, the determining similarity cost information for the underlying conversational text-mining network from the first and second thermal relationship networks includes:

performing the speaking adjustment on the second thermodynamic relationship network to obtain an adjusted thermodynamic relationship network;

and carrying out similarity analysis on the first thermodynamic relation network and the adjusted thermodynamic relation network, and determining similarity cost information of the basic session text mining network.

In some examples, the determining intent discrimination cost information of the underlying conversation text mining network according to the first interactive intent discrimination report, the second interactive intent discrimination report, and text block intent words carried by the authenticated HMI digital conversation text includes:

Acquiring a first evaluation offset between the first interaction intention discrimination report and text block intention words carried by the authenticated HMI digital conversation text, and determining training cost information of the basic conversation text mining network according to the first evaluation offset;

acquiring a second evaluation offset between the second interaction intention discrimination report and the text block intention word, and determining adjustment cost information of the basic session text mining network according to the second evaluation offset;

and determining the intention discrimination cost information of the basic session text mining network according to the training cost information and the adjustment cost information.

In some examples, the modifying the configuration variables of the base session text-mining network in combination with the similarity cost information and the intent discrimination cost information to generate a debugged session text parsing model includes:

determining comprehensive network cost information corresponding to the basic session text mining network according to the similarity cost information and the intention discrimination cost information;

and improving the configuration variables of the basic session text mining network by carrying out convergence processing on the comprehensive network cost information, and determining the basic session text mining network containing the improved configuration variables as a debugged session text analysis model.

In some examples, the method further comprises:

acquiring an HMI digital session text to be processed, acquiring a question-answer event query vector corresponding to a target question-answer event in the HMI digital session text to be processed through a debugged session text analysis model, and identifying a question-answer event local intention discrimination report corresponding to the question-answer event query vector; the local intention judgment report of the question-answer event is used for indicating a local intention label of the question-answer event corresponding to a text block of the target question-answer event;

generating a local feature cluster of the question-answer event according to the local intention discrimination report of the question-answer event and the semantic description of the question-answer event of the digital session text of the HMI to be processed;

acquiring a mean value characteristic variable corresponding to the local characteristic cluster of the question-answer event, determining a distribution report of text blocks in the target question-answer event in the digital session text of the HMI to be processed according to the mean value characteristic variable, and determining an inquiry data prediction report corresponding to the target question-answer event in the digital session text of the HMI to be processed according to the local intention label of the question-answer event and the distribution report.

In some examples, the obtaining, by the debugged session text parsing model, a query vector of a question-answer event corresponding to a target question-answer event in a digital session text of the HMI to be processed includes:

Loading the HMI digital session text to be processed into the debugged session text analysis model, acquiring scene question-answer event knowledge corresponding to the target question-answer event in the HMI digital session text to be processed in the debugged session text analysis model, and generating a scene question-answer event interaction intention judgment report corresponding to the scene question-answer event knowledge according to a judgment component in the debugged session text analysis model;

acquiring a question-answer event semantic description aiming at the to-be-processed HMI digital session text and generated by a target semantic coding component in the debugged session text analysis model, and weighting the scene question-answer event interaction intention discrimination report and the question-answer event semantic description to obtain a scene question-answer event knowledge cluster corresponding to the to-be-processed HMI digital session text;

according to the scene question-answer event knowledge clusters, decomposing the HMI digital session text to be processed to obtain X question-answer event text paragraphs, and according to the debugged session text analysis model, obtaining question-answer event paragraph semantics corresponding to the X question-answer event text paragraphs respectively; wherein X is a positive integer;

And aggregating the scene question-answer event knowledge and the question-answer event link vectors corresponding to the X question-answer event text paragraphs into the question-answer event query vector.

In some examples, the method further comprises:

when the query data prediction report is matched with target query data in a big data recommendation server, determining that a pushing pairing result of the to-be-processed HMI digital session text in the big data recommendation server is a trigger pushing condition, and performing personalized pushing strategy customization for digital session user equipment corresponding to the to-be-processed HMI digital session text.

An AI big data mining system, comprising: a processor, a memory, and a network interface; the processor is connected with the memory and the network interface; the network interface is for providing data communication functions, the memory is for storing program code, and the processor is for invoking the program code to perform the above-described method.

A computer readable storage medium having stored thereon a computer program which, when run, performs an artificial intelligence application oriented human interface digital session mining method.

A computer program product comprising a computer program or computer executable instructions that when executed by a processor implement a human interface digital session mining method for artificial intelligence applications.

According to one embodiment of the invention, a target query vector in an authenticated HMI digital session text can be extracted through a basic session text mining network, a first interaction intention discrimination report of the target query vector is obtained by carrying out interaction intention discrimination on the target query vector, and a first thermodynamic relationship network is generated by combining the first interaction intention discrimination report and authenticated semantic description of the authenticated HMI digital session text; based on the above, the authenticated HMI digital conversation text can be subjected to speaking operation adjustment to obtain an HMI conversation adjustment text, a reconstructed query vector in the HMI conversation adjustment text is extracted through a basic conversation text mining network, and a second thermodynamic relationship network is generated according to a second interaction intention discrimination report of the reconstructed query vector and a reconstructed semantic description of the HMI conversation adjustment text; the similarity limitation (namely similarity cost information) can be further carried out on the first thermal relationship network and the second thermal relationship network, so that the debugged session text analysis model obtained through debugging can improve the capturing precision of text blocks in the HMI digital session text; in addition, when the basic session text mining network is debugged, the distributed data of each text block of the authenticated question-answer event in the authenticated HMI digital session text is not needed to be annotated, the acquisition limit of the authenticated HMI digital session text is broken, and the annotation processing of the text block distribution of the authenticated HMI digital session text can be reduced, so that the timeliness of the interactive intention mining of the HMI digital session text is improved.

It is to be understood that both the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention as claimed.

Drawings

In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings that are necessary for the embodiments to be used are briefly described below, the drawings being incorporated in and forming a part of the description, these drawings showing embodiments according to the present invention and together with the description serve to illustrate the technical solutions of the present invention. It is to be understood that the following drawings illustrate only certain embodiments of the invention and are therefore not to be considered limiting of its scope, for the person of ordinary skill in the art may admit to other equally relevant drawings without inventive effort.

Fig. 1 shows a schematic diagram of an AI big data mining system according to an embodiment of the present invention.

Fig. 2 shows a flowchart of a man-machine interface digital session mining method for artificial intelligence application according to an embodiment of the present invention.

Detailed Description

For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of the invention generally described and illustrated herein may be arranged and designed in a wide variety of different configurations.

Thus, the following detailed description of the embodiments of the invention is not intended to limit the scope of the invention, as claimed, but is merely representative of selected embodiments of the invention. Based on the embodiments of the present invention, one of ordinary skill in the art would obtain all other embodiments without undue burden, all falling within the scope of the present invention.

Fig. 1 shows a schematic diagram of an AI big data mining system provided by an embodiment of the present invention, the AI big data mining system 100 includes a processor 110, a memory 120, and a network interface 130. The processor 110 is connected to the memory 120 and the network interface 130. Further, the network interface 130 is configured to provide data communication functions, the memory 120 is configured to store program codes, and the processor 110 is configured to invoke the program codes to perform an artificial intelligence application oriented human-machine interface digital session mining method.

Human-machine interaction (Human Machine Interaction, HMI) is a technology that studies the interaction relationship between a system and a user. The human-machine interaction system may be a variety of machines, as well as computerized systems and software. Human-machine interaction interfaces generally refer to portions that are visible to a user. The user communicates with the system through a man-machine interaction interface and performs operation. The design of the man-machine interaction interface is to include the user's understanding of the system.

HMI digital conversation text is a digital service conversation record generated by a technical man-machine interaction technology, and the digital conversation can relate to different service fields such as electronic commerce, meta universe, virtual reality, network security, blockchain, intelligent power, cloud game, software APP, supply chain, distributed edge computing, intelligent community and the like. Further, the HMI digital session text may be a textual interaction record of the user terminal and the human-machine interaction system, including, but not limited to, an operational interaction record, a dialogue interaction record, etc. between the user terminal and the human-machine interaction system.

The overall design idea of the embodiment of the invention is as follows: processing the authenticated HMI digital conversation text by using a basic conversation text mining network to obtain a first interaction intention distinguishing report and a first thermodynamic relation network corresponding to the authenticated HMI digital conversation text; processing the HMI conversation adjustment text subjected to conversation operation adjustment by utilizing the basic conversation text mining network to obtain a second interaction intention discrimination report and a second thermodynamic relation network corresponding to the HMI conversation adjustment text; the HMI session adjustment text is obtained by performing speaking adjustment on the authenticated HMI digital session text; and improving the configuration variables of the basic conversation text mining network by combining the first thermodynamic relation network, the second thermodynamic relation network, the first interaction intention distinguishing report and the second interaction intention distinguishing report to obtain a debugged conversation text analysis model.

Therefore, the overall design thought focuses on training and debugging of a basic conversation text mining network, namely, the obtained debugged conversation text analysis model is ensured to be suitable for HMI digital conversation texts expressed by different conversations, so that accuracy and reliability of interactive intention mining on the HMI digital conversation texts are improved.

In addition, the whole design thought can be realized through the technical scheme described in the steps 01-04.

Fig. 2 is a flow chart illustrating an artificial intelligence application-oriented digital session mining method, which may be implemented in accordance with an embodiment of the present invention, and may be implemented by the AI big data mining system 100 shown in fig. 1, and the artificial intelligence application-oriented digital session mining method includes steps 01-04.

And step 01, acquiring an authenticated HMI digital conversation text, outputting a first interaction intention distinguishing report corresponding to the authenticated HMI digital conversation text through a basic conversation text mining network, and generating a first thermal relationship network according to the first interaction intention distinguishing report and the authenticated semantic description of the authenticated HMI digital conversation text.

The first interaction intention distinguishing report is determined through a target inquiry vector corresponding to an authenticated question-answer event in the authenticated HMI digital session text, and the first thermodynamic relation network is used for representing the distribution data of text blocks of the authenticated question-answer event in the authenticated HMI digital session text.

For example, the AI big data mining system may obtain authenticated HMI digital session text (HMI digital session text sample) for debugging the underlying session text mining network, where the authenticated HMI digital session text may be an e-commerce session text or a network security session text, the number of authenticated HMI digital session texts may be plural, each authenticated HMI digital session text may be an HMI digital session text containing an authenticated question-answer event (sample question-answer event), and each authenticated HMI digital session text may carry text block intent words that may be used to represent interaction intent corresponding to each text block of the authenticated question-answer event included in the authenticated HMI digital session text. Wherein the authenticated question-answer event contained in the authenticated HMI digital session text may include, but is not limited to: e-commerce flow consultation question-answer events and software APP upgrading question-answer events; the text blocks of the authenticated question-answering event may refer to the e-commerce process consulting the question-answering event or the software APP upgrading the text units of the question-answering event (including but not limited to a combination of words, sentences and paragraphs), different kinds of authenticated question-answering events may correspond to different text blocks, and each kind of authenticated question-answering event may correspond to a corresponding number of text blocks. For example, when the authenticated HMI digital session text is an HMI digital session text of an e-commerce process consultation question-answer event, each text block of the authenticated question-answer event included in the authenticated HMI digital session text may include a plurality of text units of e-commerce initiation flow, cross-border e-commerce consultation, forbidden operation area prompt, network fraud protection, etc., and combining these text blocks may describe a complete interaction state of the e-commerce process consultation question-answer event in the authenticated HMI digital session text; text block intent words carried by the authenticated HMI digital session text may include online operational preferences, business wind preferences, rights maintenance preferences, and the like. In addition, the text block intention word can reflect the interaction tendency or interest of different text blocks and can be used as the analysis basis for the subsequent big data pushing.

The AI big data mining system can obtain a general (initialized) HMI digital conversation text analysis model, namely a basic conversation text mining network, and in the process of debugging the basic conversation text mining network, one authenticated HMI digital conversation text can be input each time, and a group of authenticated HMI digital conversation texts can also be input. For any one of the plurality of authenticated HMI digital conversation texts, the AI big data mining system can load the authenticated HMI digital conversation text into a basic conversation text mining network, and can acquire a target query vector corresponding to an authenticated question-answer event in the authenticated HMI digital conversation text through the basic conversation text mining network, wherein the target query vector can be used for describing the interaction progress state of the authenticated question-answer event in the authenticated HMI digital conversation text; the target query vector is processed through a judging component (which can be understood as a classifying component) of the basic session text mining network, and a first interaction intention judging report corresponding to the target query vector can be obtained; further, a first thermal relationship network (which may be understood as a text block distribution record) may be generated according to the first interaction intention discrimination report and the authenticated semantic description (which may be understood as a semantic feature sample) corresponding to the authenticated HMI digital session text, where the first thermal relationship network may be used to represent distribution data (which may be understood as a relative location area) of each text block of the authenticated question-answer event in the authenticated HMI digital session text, respectively. The authenticated semantic description may refer to a semantic description for authenticated HMI digital session text generated by a target semantic encoding component in the base session text-mining network, where the target semantic encoding component may refer to a last semantic encoding component in the base session text-mining network.

The basic session text mining network may be one or a combination of a plurality of BERT, transformer, RNN, resNet, denseNet or LSTM models. The basic session text mining network can also be a neural network flexibly configured according to actual requirements, and the designed basic session text mining network can be used for mining multi-dimensional query vectors (namely, the target query vectors can comprise query vectors with different dimensions), and the multi-dimensional query vectors can be used for representing authenticated question-answering events with different priorities, so that the performance of the basic session text mining network is better; the invention does not limit the variety of the basic session text mining network.

The AI big data mining system obtains a target query vector corresponding to an authenticated question-answer event in an authenticated HMI digital session text according to a basic session text mining network, and after determining a first interaction intention discrimination report corresponding to the target query vector, obtains an authenticated semantic description for the authenticated HMI digital session text generated by a target semantic coding component in the basic session text mining network, weights (such as multiplication) the first interaction intention discrimination report and the authenticated semantic description, and obtains an initial thermodynamic relation network corresponding to the authenticated HMI digital session text; the target semantic coding component may refer to a last semantic coding component in the basic session text mining network, and because the dimension of the authenticated semantic description generated by the last semantic coding component is smaller than the dimension of the authenticated HMI digital session text, that is, the dimension of the initial thermodynamic relationship network is smaller than the dimension of the authenticated HMI digital session text, based on this, the initial thermodynamic relationship network may be subjected to expansion processing (upsampling) to obtain a first thermodynamic relationship network having the same visualization scale as the authenticated HMI digital session text. The first interactive intention distinguishing report may include quantification possibility that text blocks of authenticated question-answer events belong to each interactive intention, where the quantification possibility in the first interactive intention distinguishing report may be considered as a confidence level corresponding to the authenticated semantic description generated by the target semantic coding component, and the authenticated semantic description is weighted by combining with the first interactive intention distinguishing report, so as to process a text set focused by the basic session text mining network. The first thermodynamic relationship network may also refer to, for example, HMI digital session text obtained by fusing the result of the expansion process with authenticated HMI digital session text.

And 02, performing speaking adjustment on the authenticated HMI digital conversation text to obtain an HMI conversation adjustment text, outputting a second interaction intention distinguishing report corresponding to the HMI conversation adjustment text through a basic conversation text mining network, and generating a second thermal relationship network according to the second interaction intention distinguishing report and the reconstructed semantic description of the HMI conversation adjustment text.

The second interaction intention judging report is determined by a reconstructed query vector corresponding to the authenticated question-answer event in the HMI session adjustment text, and the second thermodynamic relation network is used for representing the distribution data of the text blocks of the authenticated question-answer event in the HMI session adjustment text.

Illustratively, the AI big data mining system may perform a speaking adjustment (such as a process of repeating dialogue content) on the authenticated HMI digital session text to obtain HMI session adjustment texts corresponding to the authenticated HMI digital session text, where the number of HMI session adjustment texts may be at least one, and different HMI session adjustment texts may be obtained by performing different speaking adjustments on the authenticated HMI digital session text, where the speaking adjustment may include an integration of expression mode modification, vocabulary modification, semantic-based potential content mining augmentation, and the like, and the kind of the speaking adjustment in the present invention is not limited.

For at least one HMI session adjustment text obtained by speaking operation adjustment, the AI big data mining system can sequentially load the at least one HMI session adjustment text into the basic session text mining network, a reconstructed query vector (updated or adjusted query vector) corresponding to an authenticated question-answer event in the HMI session adjustment text can be obtained through the basic session text mining network, the reconstructed query vector is processed through a judging component of the basic session text mining network, a second interaction intention judging report corresponding to the HMI session adjustment text can be obtained, and then a second thermal relationship network is generated by combining the second interaction intention judging report and the reconstructed semantic description of the HMI session adjustment text. In other words, the processing thought of the AI big data mining system for adjusting the text of the HMI session by using the basic session text mining network is the same as that of the authenticated HMI digital session text. When the AI big data mining system adopts a speaking adjustment idea to perform speaking adjustment on the authenticated HMI digital conversation text, an HMI conversation adjustment text corresponding to the authenticated HMI digital conversation text can be obtained, and a second thermodynamic relationship network corresponding to the HMI conversation adjustment text can be generated through a basic conversation text mining network; when the AI big data mining system adopts a plurality of conversation adjustment ideas to carry out conversation adjustment on the authenticated HMI digital conversation text, a plurality of HMI conversation adjustment texts corresponding to the authenticated HMI digital conversation text can be obtained, and a second thermodynamic relation network respectively corresponding to the plurality of HMI conversation adjustment texts can be generated through a basic conversation text mining network, namely one HMI conversation adjustment text can correspond to one second thermodynamic relation network.

For example, after the AI big data mining system obtains the authenticated HMI digital session text sample_a, the text content corresponding to the authenticated question-answer event in the authenticated HMI digital session text sample_a may be subjected to repeated processing, so as to obtain an HMI session adjustment text_b corresponding to the authenticated HMI digital session text sample_a.

Or, by carrying out semantic-based potential content mining augmentation processing on the authenticated HMI digital session text sample_a, the HMI session adjustment text adjustment text_c corresponding to the authenticated HMI digital session text sample_a can be obtained.

Or, content sequence adjustment processing is carried out on the authenticated HMI digital session text sample_a, so that an HMI session adjustment text adjustment_d corresponding to the authenticated HMI digital session text sample_a can be obtained.

Or, by carrying out keyword highlighting processing on the authenticated HMI digital session text sample_a, the HMI session adjustment text adjustment text_e corresponding to the authenticated HMI digital session text sample_a can be obtained. The AI big data mining system can sequentially load the HMI session adjustment text adjustment text_b, the HMI session adjustment text adjustment text_c, the HMI session adjustment text adjustment text_d and the HMI session adjustment text adjustment text_e into a basic session text mining network, and the basic session text mining network can be used for respectively corresponding second thermodynamic relation networks of the HMI session adjustment texts.

Step 03, determining similarity cost information of a basic conversation text mining network according to the first thermodynamic relation network and the second thermodynamic relation network, and determining intention discrimination cost information of the basic conversation text mining network according to the first interaction intention discrimination report, the second interaction intention discrimination report and text block intention words carried by the authenticated HMI digital conversation text.

For example, the AI big data mining system may determine the commonality cost information (similarity loss) of the underlying conversational text mining network according to the first thermodynamic relationship network corresponding to the authenticated HMI digital conversational text and the second thermodynamic relationship network corresponding to the voice-operated adjusted HMI conversational adjustment text. The determining idea of the commonality cost information may include: the AI big data mining system can implement the same conversation adjustment on the second thermodynamic relationship network to obtain an adjusted thermodynamic relationship network, and further can perform similarity analysis (such as similarity constraint) on the first thermodynamic relationship network and the adjusted thermodynamic relationship network to determine the commonality cost information of the basic conversation text mining network. The authenticated question-answer event in the authenticated HMI digital conversation text and the authenticated question-answer event in each HMI conversation adjustment text are the same question-answer event, and it is expected that the first thermodynamic relation network corresponding to the authenticated HMI digital conversation text and the adjusted thermodynamic relation network should have the same distribution data, so when the basic conversation text mining network is debugged, similarity analysis can be performed for the first thermodynamic relation network and the adjusted thermodynamic relation network, and the basic conversation text mining network can master the distribution data of different text blocks to improve the capturing precision of the basic conversation text mining network.

In view of the fact that the authenticated HMI digital conversation text carries text block intent words corresponding to the authenticated question-answer event, whether the authenticated HMI digital conversation text and the speech-adjusted HMI conversation adjustment text are the same, the authenticated question-answer event and the correct interaction intent of the text block in the HMI conversation adjustment text are the same, and the correct interaction intent of the text block is the text block intent word carried by the authenticated HMI digital conversation text. The AI big data mining system can acquire a first quantized difference between the first interactive intention discrimination report and the text block intention word, training cost information of the basic conversation text mining network can be determined according to the first quantized difference, and the first quantized difference can refer to a characteristic distance between the first interactive intention discrimination report and the text block intention word; similarly, a second quantization difference between the second interaction intention judging report and the text block intention word can be obtained, the adjustment cost information of the basic session text mining network can be determined according to the second quantization difference, and the second quantization difference can be the characteristic distance between the second interaction intention judging report and the text block intention word; and determining intention discrimination cost information of the basic session text mining network according to the training cost information and the adjustment cost information. For example, the training cost information cost1 and the adjustment cost information cost2 may be summed, and the summed result (cost 1+cost 2) is used as the intention discrimination cost information of the basic session text mining network; or the weight q1 can be set for the training cost information cost1, the weight q2 can be set for the adjustment cost information cost2, and (q1+q2×cost 2) can be used as the intention discrimination cost information of the basic session text mining network; the invention does not limit the integration mode of the training cost information and the adjustment cost information.

Step 04, improving configuration variables of the basic session text mining network by combining similarity cost information and intention discrimination cost information to generate a debugged session text analysis model; the debugged session text analysis model is used for mining local intention and distribution reports of the question-answer event corresponding to the text blocks in the digital session text of the HMI to be processed.

By way of example, the AI big data mining system may combine the similarity cost information and the intent discrimination cost information to improve configuration variables of the basic session text mining network, and determine the debugged basic session text mining network as a debugged session text parsing model.

The AI big data mining system may determine comprehensive network cost information corresponding to the basic session text mining network according to the similarity cost information and the intention discrimination cost information, where the comprehensive network cost information may refer to a comprehensive result of the similarity cost information and the intention discrimination cost information, or may refer to a result obtained by multiplying the similarity cost information and the intention discrimination cost information by respective corresponding weights and then summing the multiplied result. By performing convergence processing on the comprehensive network cost information of the basic session text-mining network (minimizing the comprehensive network cost information), configuration variables of the basic session text-mining network can be improved, the basic session text-mining network containing the improved configuration variables is determined as a debugged session text-parsing model, and the debugged session text-parsing model can refer to the debugged basic session text-mining network. In other words, by converging comprehensive network cost information, continuously debugging configuration variables of the basic session text mining network, namely repeatedly improving the configuration variables of the basic session text mining network, when the debugging turn of the basic session text mining network reaches a preset cycle number threshold value or the debugging of the basic session text mining network tends to be stable, the configuration variables under the process can be recorded, and the basic session text mining network containing the configuration variables is determined as a debugged session text analysis model; the debugged session text parsing model may be used to mine local intent and distribution reports of questioning and answering events corresponding to text blocks in the pending HMI digital session text, which may be used to determine predictive reports of questioning and answering data corresponding to target questioning and answering events in the pending HMI digital session text (which may also be understood as local event interest and distribution location areas corresponding to text blocks). Thus, the interactive interest mining analysis of the HMI digital session text can be realized, and a decision basis is provided for subsequent pushing processing.

For example, after acquiring the authenticated HMI digital session text e_text1, the AI big data mining system may load the authenticated HMI digital session text e_text1 into a transform algorithm (the transform algorithm herein may be considered as the basic session text mining network), obtain a target query vector required vector through the transform algorithm, and process the target query vector required vector to obtain a first interaction intention discrimination report1 corresponding to the authenticated HMI digital session text e_text1; the first interaction intention discrimination report1 and the authenticated semantic description generated by the last semantic coding component in the transform algorithm are weighted to obtain an Activity map1 (namely the first thermal relationship network), the number of the Activity map1 can be at least one, and the Activity map1 is the same as the number of text block interaction intention of authenticated question-answer events contained in the authenticated HMI digital conversation text e_text1, namely the Activity map1 can comprise a thermal relationship network corresponding to each text block interaction intention in the authenticated HMI digital conversation text e_text 1. The AI big data mining system can determine training cost information loss information of a transform algorithm according to a first quantization difference between a text block intent word information words carried by the first interactive intent discrimination report intent report1 and the authenticated HMI digital conversation text e_text 1.

Further, the AI big data mining system can adjust the authenticated question and answer event in the authenticated HMI digital session text e_text1 to obtain an HMI session adjustment text repeated text, further can load the HMI session adjustment text repeated text into a Transformer algorithm, can obtain a reconstruction query vector reconsitution vector through the Transformer algorithm, and can obtain a second interaction intention discrimination report entry report2 corresponding to the authenticated HMI digital session text e_text1 by processing the reconstruction query vector reconsitution vector; and weighting the reconstructed semantic description generated by the last semantic coding component in the second interaction intention discrimination report2 and the transform algorithm to obtain an Activity map2 (namely the second thermodynamic relationship network), wherein the Activity map2 can comprise thermodynamic relationship networks corresponding to interaction intention of each text block in the HMI session adjustment text repeated text. The AI big data mining system may determine the adjustment cost information reconsitution loss of the transform algorithm according to the second quantization difference between the second interactive intent discrimination report intent report2 and the text block intent word intent words.

The AI big data mining system can also analyze the Similarity for the Activity map1 and the Activity map2, determine the adjustment cost information Similarity of the transform algorithm, determine the comprehensive network cost information of the transform algorithm through training the cost information loss information, the adjustment cost information reconsitution loss and the adjustment cost information Similarity, and continuously debug the configuration variables of the transform algorithm until the debugging turn reaches a preset cycle number threshold (or the debugging tends to be stable) through convergence processing of the comprehensive network cost information, so as to obtain the debugged session text analysis model after the completion of the debugging.

In the embodiment of the invention, a target query vector in an authenticated HMI digital session text is extracted through a basic session text mining network, a first interaction intention discrimination report of the target query vector is obtained by carrying out interaction intention discrimination on the target query vector, and a first thermodynamic relationship network is generated by combining the first interaction intention discrimination report and authenticated semantic description of the authenticated HMI digital session text; based on the above, the authenticated HMI digital conversation text can be subjected to speaking operation adjustment to obtain an HMI conversation adjustment text, a reconstructed query vector in the HMI conversation adjustment text is extracted through a basic conversation text mining network, and a second thermodynamic relationship network is generated according to a second interaction intention discrimination report of the reconstructed query vector and a reconstructed semantic description of the HMI conversation adjustment text; the similarity limitation (namely similarity cost information) can be further carried out on the first thermal relationship network and the second thermal relationship network, so that the debugged session text analysis model obtained through debugging can improve the capturing precision of text blocks in the HMI digital session text; in addition, when the basic session text mining network is debugged, the distributed data of each text block of the authenticated question-answer event in the authenticated HMI digital session text is not needed to be annotated, the acquisition limit of the authenticated HMI digital session text is broken, and the annotation processing of the text block distribution of the authenticated HMI digital session text can be reduced, so that the timeliness of the interactive intention mining of the HMI digital session text is improved.

The man-machine interface digital conversation mining method facing the artificial intelligence application provided by the embodiment of the invention can comprise the following steps 10-18.

And step 10, acquiring an authenticated HMI digital session text, acquiring a scene inquiry vector corresponding to an authenticated question-answer event in the authenticated HMI digital session text in a basic session text mining network, and outputting a scene interaction intention discrimination report corresponding to the scene inquiry vector through a discrimination component in the basic session text mining network.

The AI big data mining system can load the authenticated HMI digital session text into a basic session text mining network after acquiring the authenticated HMI digital session text, and can obtain a scene inquiry vector corresponding to the authenticated question-answer event in the authenticated HMI digital session text through the basic session text mining network, wherein the scene inquiry vector can be used for describing the overall interaction progress state of the authenticated question-answer event in the authenticated HMI digital session text; and processing the scene inquiry vector through a judging component of the basic session text mining network to obtain a scene interaction intention judging report corresponding to the scene inquiry vector. The scene query vector can be understood as a global query vector, so that the scene interaction intention discrimination report can be understood as an interaction intention discrimination result of the whole layer.

Taking the example that the base session text-mining network is a transform algorithm, the base session text-mining network may include Y gradient focus sub-networks (residual networks), each of which may include at least one semantic coding component, where Y is a positive integer. The mining concept of the scene query vector may include: the AI big data mining system can acquire the incoming information of a u-th gradient focusing sub-network in the Y gradient focusing sub-networks, when u is 1, the incoming information of the u-th gradient focusing sub-network can be the authenticated HMI digital session text, and u can be a positive integer smaller than Y; illustratively, before the Y gradient focus sub-networks of the underlying session text-mining network, the underlying session text-mining network may further include at least one unshared semantic coding component (convolution component or convolution layer), and the 1 st semantic coding component (u is 1) may be a semantic description generated by passing the authenticated HMI digital session text through one or two unshared semantic coding components in the underlying session text-mining network.

Through at least one semantic coding component in the u-th gradient focus sub-network, sliding window coding (convolution operation) is carried out on the incoming information of the u-th gradient focus sub-network, initial semantic description can be obtained, further, the initial semantic description and the incoming information of the u-th initial gradient focus sub-network can be integrated (for example, the integration can be feature summation) to obtain a semantic description combination result of the u-th gradient focus sub-network, the semantic description combination result of the u-th gradient focus sub-network is used as the incoming information of the u+1-th gradient focus sub-network, and the semantic description combination result of the Y-th gradient focus sub-network is determined to be a scene query vector corresponding to an authenticated question-answer event in an authenticated HMI digital session text; wherein the u-th gradient focus subnet is cascaded with the u+1-th gradient focus subnet. For example, when the dimension of the initial semantic description is inconsistent with the dimension of the incoming information of the u-th initial gradient focus subnet, feature mapping may be performed on the incoming information of the u-th initial gradient focus subnet, so that the dimension of the feature mapping result is the same as the dimension of the initial semantic description, and further the feature mapping result and the initial semantic description may be summed to obtain a semantic description combination result of the u-th gradient focus subnet. In other words, the Y gradient focus sub-networks in the basic session text mining network are sequentially connected, and the semantic description combination result of the former gradient focus sub-network (such as the u-th gradient focus sub-network) can be used as the incoming information of the latter gradient focus sub-network (the u+1th gradient focus sub-network), and finally the semantic description combination result of the last gradient focus sub-network (the Y-th gradient focus sub-network) can be used as the scene query vector corresponding to the authenticated question-answer event in the authenticated HMI digital session text.

Illustratively, the number of scene query vectors is Z, which may be a positive integer, where Z may be considered the number of attention indicators of the scene query vectors; the AI big data mining system can determine target feature variables corresponding to the Z scene query vectors respectively, and integrate the target feature variables corresponding to the Z scene query vectors into scene text knowledge; and then the scene text knowledge can be converted into scene text knowledge to be distinguished according to the feature mapping component in the basic conversation text mining network, the scene text knowledge to be distinguished is loaded to the distinguishing component of the basic conversation text mining network, and the scene interaction intention distinguishing report corresponding to the scene text knowledge to be distinguished is output through the distinguishing component of the basic conversation text mining network. For example, the AI big data mining system may perform global-level average pooling processing on the Z target query vectors, map each scene query vector into a numerical value, that is, the Z target query vectors may be converted into a Z-dimensional scene text knowledge, and the scene text knowledge is loaded to the discrimination component after passing through the feature mapping component, so as to obtain the scene interaction intent discrimination report. The relative distribution characteristics in the authenticated HMI digital session text can be recorded through the global-level average pooling process, so that the capturing precision of the model is improved.

Step 12, weighting the scene interaction intention discrimination report and the authenticated semantic description to obtain scene feature clusters corresponding to the authenticated HMI digital session text, and carrying out disassembling processing on the authenticated HMI digital session text according to the scene feature clusters to obtain X session text paragraphs; x is a positive integer.

For example, the AI big data mining system may weight the scene interaction intent discrimination report and the authenticated semantic description of the authenticated HMI digital session text to obtain a scene feature cluster corresponding to the authenticated HMI digital session text, where the scene feature cluster may include a thermodynamic relationship network corresponding to each text block of the authenticated question-answer event in the authenticated HMI digital session text; the determination mode of the scene feature cluster can be combined with the determination mode of the first thermodynamic relation network in the step 01.

Furthermore, the AI big data mining system can take the thermodynamic relation network of each text block as a real reference of a capturing frame, and disassemble the authenticated HMI digital conversation text, so that X conversation text paragraphs can be obtained, wherein X can be a positive integer. In other words, the authenticated HMI digital session text can be segmented according to the scene feature clusters to obtain session text segments corresponding to each link.

And step 13, sequentially loading the X session text paragraphs into a basic session text mining network, and acquiring paragraph query vectors corresponding to the X session text paragraphs in the basic session text mining network.

The AI big data mining system may load the X session text paragraphs obtained by segmentation of the paragraphs into the basic session text mining network again, and may obtain semantic vectors with better detail characterization performance through the basic session text mining network, that is, paragraph query vectors corresponding to each session text paragraph respectively, where the paragraph query vectors may be used to express interaction progress states of links of authenticated question-answer events included in the authenticated HMI digital session text. The processing concept of the basic session text mining network for a single session text paragraph can be combined with the processing concept of the authenticated HMI digital session text in step 10.

And 14, vector aggregation is carried out on the scene query vector and paragraph query vectors corresponding to the X conversation text paragraphs, so that target query vectors corresponding to authenticated question-answer events in authenticated HMI digital conversation texts are obtained.

For example, the AI big data mining system may aggregate the scene query vector and the X paragraph query vectors grasped by the basic session text mining network, e.g., combine the scene query vector with the X paragraph query vectors to obtain a target query vector corresponding to the authenticated question-answer event in the authenticated HMI digital session text; the target query vector here may contain both a paragraph query vector for each link of an authenticated question-answer event and a scene query vector for an authenticated question-answer event. By considering text segmentation learning of the combined link attention in the basic session text mining network, the detail characterization performance of the target query vector can be enhanced, and further the capturing precision of the basic session text mining network can be improved.

Step 15, processing the target query vector according to a judging component in the basic session text mining network to obtain a first interaction intention judging report corresponding to the authenticated HMI digital session text; a first network of thermal relationships is generated based on the first interactive intent discrimination report and the authenticated semantic description of the authenticated HMI digital conversation text.

Step 16, performing speaking adjustment on the authenticated HMI digital conversation text to obtain an HMI conversation adjustment text, outputting a second interaction intention distinguishing report corresponding to the HMI conversation adjustment text through a basic conversation text mining network, and generating a second thermal relationship network according to the second interaction intention distinguishing report and the reconstructed semantic description of the HMI conversation adjustment text; the second interactive intention discrimination report is determined by a reconstructed query vector corresponding to the authenticated question-answer event in the HMI session adjustment text, and the second thermodynamic relationship network is used for representing the distribution data of the text blocks of the authenticated question-answer event in the HMI session adjustment text.

And step 17, determining similarity cost information of the basic conversation text mining network according to the first thermodynamic relation network and the second thermodynamic relation network, and determining intention discrimination cost information of the basic conversation text mining network according to the first interaction intention discrimination report, the second interaction intention discrimination report and text block intention words carried by the authenticated HMI digital conversation text.

Step 18, combining the similarity cost information and the intention discrimination cost information to improve the configuration variables of the basic session text mining network and generate a debugged session text analysis model; the debugged session text analysis model is used for mining local intention and distribution reports of the question-answer event corresponding to the text blocks in the digital session text of the HMI to be processed.

The design ideas of the steps 15 to 18 may be combined with the steps 01 to 04. In an exemplary embodiment of the present invention, the intent discrimination cost information of the basic session text mining network may include scene cost information in addition to training cost information and adjustment cost information; wherein the scene cost information may be determined by a third quantitative difference between the scene interaction intent discrimination report and the text block intent word.

The debugging of the basic session text mining network provided by the embodiment of the invention adds text segmentation learning concerned by the combined event link on the basis of the previous basic session text mining network. After the AI big data mining system obtains the scene query vector corresponding to the authenticated HMI digital conversation text e_text1 through a transducer algorithm, the scene query vector can be further processed by utilizing the global-level average pooling processing and feature mapping component, the processed result is judged to obtain a scene interaction intention judgment report, and the scene interaction intention judgment report is multiplied by the authenticated semantic description generated by the last semantic coding component to obtain a scene feature cluster. And then the authenticated HMI digital conversation text can be disassembled according to the scene feature cluster to obtain X conversation text paragraphs, the X conversation text paragraphs are sequentially loaded into a transducer algorithm, and paragraph query vectors corresponding to the X conversation text paragraphs respectively can be obtained through the transducer algorithm. Vector aggregation is carried out on the X paragraph query vectors and scene query vectors of authenticated HMI digital conversation text to obtain target query vectors, a first interaction intention discrimination report can be obtained by processing the target query vectors, the authenticated semantic description generated by the last semantic coding component in the first interaction intention discrimination report and a transducer algorithm is weighted, the characteristic HMI digital conversation text (namely the first thermodynamic relation network) can be obtained, the first thermodynamic relation network pays attention to an event link area of the authenticated HMI digital conversation text, and the subsequent processing thinking can be combined with the related content.

In the embodiment of the invention, a target query vector in an authenticated HMI digital session text is extracted through a basic session text mining network, a first interaction intention discrimination report of the target query vector is obtained by carrying out interaction intention discrimination on the target query vector, and a first thermodynamic relationship network is generated by combining the first interaction intention discrimination report and authenticated semantic description of the authenticated HMI digital session text; based on the above, the authenticated HMI digital conversation text can be subjected to speaking operation adjustment to obtain an HMI conversation adjustment text, a reconstructed query vector in the HMI conversation adjustment text is extracted through a basic conversation text mining network, and a second thermodynamic relationship network is generated according to a second interaction intention discrimination report of the reconstructed query vector and a reconstructed semantic description of the HMI conversation adjustment text; the similarity limitation (namely similarity cost information) can be further carried out on the first thermal relationship network and the second thermal relationship network, so that the debugged session text analysis model obtained through debugging can improve the capturing precision of text blocks in the HMI digital session text; in addition, when the basic session text mining network is debugged, the distributed data of each text block of the authenticated question-answer event in the authenticated HMI digital session text is not needed to be annotated, the acquisition limit of the authenticated HMI digital session text is broken, and the annotation processing of the text block distribution of the authenticated HMI digital session text can be reduced, so that the timeliness of mining the interaction intention of the HMI digital session text is improved; the text segmentation learning which is concerned by combining the event links is considered in the basic session text mining network, so that semantic vectors with better detail representation performance can be mastered, a first thermodynamic relation network which is concerned with the event link area is further obtained, and the capturing precision of the basic session text mining network can be further improved.

The man-machine interface digital conversation mining method for artificial intelligence application provided by the embodiment of the invention can comprise the following steps 21-23.

Step 21, acquiring an HMI digital session text to be processed, acquiring a question-answer event query vector corresponding to a target question-answer event in the HMI digital session text to be processed through a debugged session text analysis model, and identifying a question-answer event local intention discrimination report corresponding to the question-answer event query vector; the local intention discrimination report of the question-answer event is used for expressing the local intention label of the question-answer event corresponding to the text block of the target question-answer event.

Illustratively, after the above-mentioned basic session text mining network completes the debugging, the debugged basic session text mining network may be referred to as a debugged session text parsing model. The AI big data mining system may obtain a pending HMI digital session text, where the pending HMI digital session text may include a target question-answer event that needs to make an interactive process state prediction, where the target question-answer event may include, but is not limited to: e-commerce flow consults a question-answer event, a software APP upgrades a question-answer event, and so on. The method comprises the steps that an HMI digital session text to be processed is loaded into a debugged session text analysis model after debugging is completed, a question-answer event query vector corresponding to a target question-answer event in the HMI digital session text to be processed can be obtained through the debugged session text analysis model, a question-answer event local intention discrimination report corresponding to the question-answer event query vector can be output through a discrimination component of the debugged session text analysis model, and the question-answer event local intention discrimination report can be used for expressing a question-answer event local intention label corresponding to a text block (such as an E-commerce flow consultation question-answer event text unit) of the target question-answer event. The question-answer event query vector may be a scene question-answer event knowledge for the target question-answer event extracted through the debugged session text parsing model, or may be an aggregate knowledge vector between a scene question-answer event knowledge corresponding to the target question-answer event and a local question-answer event feature. When the question-answer event query vector is scene question-answer event knowledge corresponding to a target question-answer event in the HMI digital session text to be processed, the method shows that text segmentation learning focused by combining event links is not considered in the process of carrying out semantic vector mining on the HMI digital session text to be processed by utilizing a debugged session text analysis model; when the question-answer event query vector is an aggregate knowledge vector between scene question-answer event knowledge and local question-answer event characteristics corresponding to a target question-answer event in the HMI digital session text to be processed, the text segmentation learning concerned by combining event links is considered in the process of carrying out semantic vector mining on the HMI digital session text to be processed by utilizing the debugged session text analysis model.

For example, if text segmentation learning focused by combining event links is considered in the process of carrying out semantic vector mining on an HMI digital conversation text to be processed by using a debugged conversation text analysis model, loading the HMI digital conversation text to be processed into the debugged conversation text analysis model by the AI big data mining system, acquiring scene question-answer event knowledge corresponding to a target question-answer event in the HMI digital conversation text to be processed in the debugged conversation text analysis model, and outputting scene question-answer event interaction intention discrimination reports corresponding to the scene question-answer event knowledge according to a discrimination component in the debugged conversation text analysis model; acquiring a question-answer event semantic description aiming at an HMI digital session text to be processed and generated by a target semantic coding component in a debugged session text analysis model, and weighting a scene question-answer event interaction intention discrimination report and the question-answer event semantic description to obtain a scene question-answer event knowledge cluster corresponding to the HMI digital session text to be processed; according to the scene question-answer event knowledge clusters, decomposing the HMI digital session text to be processed to obtain X question-answer event text paragraphs, and according to the debugged session text analysis model, obtaining the question-answer event paragraph semantics corresponding to the X question-answer event text paragraphs respectively; x is a positive integer; and aggregating the scene question-answer event knowledge and the question-answer event link vectors corresponding to the X question-answer event text paragraphs into question-answer event query vectors. The mining concept of the question-answer event query vector may be combined with the mining concept of the target query vector in steps 10 to 14.

In view of the capability of the debugged session text parsing model to master the corresponding distributed data, the process of the HMI session adjustment text after the session adjustment is unnecessary when the debugged session text parsing model is applied. In other words, when the debugged session text parsing model is used, the processing such as the speaking adjustment is not necessary to be considered.

And step 22, generating a local feature cluster of the question-answer event according to the local intention discrimination report of the question-answer event and the semantic description of the question-answer event of the digital session text of the HMI to be processed.

For example, after obtaining the local intention discrimination report of the question-answer event, the AI big data mining system may multiply the local intention discrimination report of the question-answer event with the semantic description of the question-answer event (semantic feature of the question-answer event) of the HMI digital session text to be processed, to generate a local feature cluster of the question-answer event, which is similar to the first thermodynamic relationship network in the above corresponding embodiment.

Step 23, acquiring a mean value characteristic variable corresponding to the local characteristic cluster of the question-answer event, determining a distribution report of text blocks in the target question-answer event in the to-be-processed HMI digital session text according to the mean value characteristic variable, and determining an inquiry data prediction report corresponding to the target question-answer event in the to-be-processed HMI digital session text according to the local intention label and the distribution report of the question-answer event.

The AI big data mining system can take a mean feature variable from the local feature cluster of the question-answer event, determine the mean feature variable as a distribution report of text blocks in the target question-answer event in the text of the to-be-processed HMI digital session, and determine a question-answer event interest tag of the target question-answer event in the text of the to-be-processed HMI digital session according to the local intention tag and the distribution report, wherein the question-answer event interest tag can be used as a query data prediction report corresponding to the target question-answer event in the text of the to-be-processed HMI digital session.

In some examples, the HMI digital session text to be processed may include at least one target question-answer event, after acquiring the HMI digital session text to be processed, the AI big data mining system may first identify, through the debugged session text parsing model, the target question-answer event included in the HMI digital session text to be processed, determine a text set corresponding to the single target question-answer event in the HMI digital session text to be processed, and further perform interactive process state estimation on the text set corresponding to the single target question-answer event, that is, perform semantic vector mining on the text set corresponding to the single target question-answer event, identify all text blocks (such as all electronic commerce process consultation question-answer event text units) included in the single target question-answer event, and perform local intention labels and distribution reports of the question-answer event corresponding to each text block, and may combine all text blocks identified according to the answer event local intention labels and the distribution reports of each text block, so as to obtain a question-answer event interest label corresponding to the single target question-answer event, where the question-answer event interest label may be used to represent the predicted data of the single target question-answer event.

Under some design ideas which can be independently implemented, when the query data prediction report is matched with target query data in the big data recommendation server, determining that a pushing pairing result of the to-be-processed HMI digital session text in the big data recommendation server is a trigger pushing condition, and performing personalized pushing strategy customization for digital session user equipment corresponding to the to-be-processed HMI digital session text. Therefore, targeted pushing auxiliary processing can be performed based on the interaction interest/interaction intention corresponding to the query data prediction report, and the execution efficiency of the data pushing task is improved.

In the embodiment of the invention, the interaction process state is speculated on the HMI digital session text to be processed through the debugged session text analysis model, so that the capturing precision of text blocks in the HMI digital session text can be improved; in addition, text segmentation learning focused by combining the event links is considered in the debugged session text analysis model, so that semantic vectors with better detail characterization performance can be mastered, question-answering event local feature clusters focused on the event link area can be further obtained, and the capturing precision can be further improved.

Under some design ideas which can be implemented independently, the personalized push strategy customization for the digital session user equipment corresponding to the to-be-processed HMI digital session text comprises the following steps: obtaining a push service response description and a push preference portrait description of the digital session user device, the push service response description being used to characterize push service response data in the digital session user device, the push preference portrait description being used to characterize push preference portrait data in the digital session user device, and at least a portion of the push preference portrait data being determined based on the query data prediction report; obtaining a first push matching feature item for the push service response description based on a local description vector associated with the push service response description in the push preference portrait description, wherein the first push matching feature item reflects the push service response description spliced with the push preference portrait description; obtaining a second push matching feature item for the push preference portrait description based on a local description vector associated with the push preference portrait description in the push service response description, wherein the second push matching feature item reflects the push preference portrait description spliced with the push service response description; combining the first pushing matching feature item and the second pushing matching feature item to obtain a combined pushing matching feature item; based on the joint push matching feature item, determining personalized push element features corresponding to the digital session user equipment; and generating a target push strategy of the digital session user equipment based on the personalized push element characteristics.

It can be seen that, by respectively extracting the push service response description and the push preference portrait description for the push service response data and the push preference portrait data of the digital session user equipment, and utilizing the joint push matching feature items thereof, the personalized push element features corresponding to the digital session user equipment are determined, and simultaneously suggestions of the push service response dimension and the push preference portrait dimension for the personalized push element features are considered, and the extracted joint push matching feature items are not combinations of the push service response description and the push preference portrait description, and can reflect deeper detail features, so that the determination progress of the personalized push element features is obviously improved, and the generation quality of a target push strategy is ensured.

Based on the same or similar technical ideas described above, the embodiments of the present invention further provide a computer-readable storage medium having a computer program stored thereon, the computer program executing an artificial intelligence application oriented human-machine interface digital session mining method at runtime.

Based on the same or similar technical ideas described above, the embodiments of the present invention further provide a computer program product, which includes a computer program or a computer executable instruction, where the computer program or the computer executable instruction implements a man-machine interface digital session mining method facing an artificial intelligence application when executed by a processor.

The embodiment of the invention provides a man-machine interface digital conversation mining method and an AI system for artificial intelligence application, which can extract a target inquiry vector in an authenticated HMI digital conversation text through a basic conversation text mining network, obtain a first interaction intention discrimination report of the target inquiry vector by carrying out interaction intention discrimination on the target inquiry vector, and generate a first thermal relation network by combining the first interaction intention discrimination report and authenticated semantic description of the authenticated HMI digital conversation text; based on the above, the authenticated HMI digital conversation text can be subjected to speaking operation adjustment to obtain an HMI conversation adjustment text, a reconstructed query vector in the HMI conversation adjustment text is extracted through a basic conversation text mining network, and a second thermodynamic relationship network is generated according to a second interaction intention discrimination report of the reconstructed query vector and a reconstructed semantic description of the HMI conversation adjustment text; the similarity limitation (namely similarity cost information) can be further carried out on the first thermal relationship network and the second thermal relationship network, so that the debugged session text analysis model obtained through debugging can improve the capturing precision of text blocks in the HMI digital session text; in addition, when the basic session text mining network is debugged, the distributed data of each text block of the authenticated question-answer event in the authenticated HMI digital session text is not needed to be annotated, the acquisition limit of the authenticated HMI digital session text is broken, and the annotation processing of the text block distribution of the authenticated HMI digital session text can be reduced, so that the timeliness of the interactive intention mining of the HMI digital session text is improved.

Those of ordinary skill in the art will appreciate that the elements and algorithm steps described in connection with the embodiments disclosed herein may be embodied in electronic hardware, in computer software, or in a combination of the two, and that the elements and steps of the examples have been generally described in terms of function in the foregoing description to clearly illustrate the interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the solution. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

The foregoing disclosure is merely illustrative of the presently preferred embodiments of the present invention, and it is to be understood that the scope of the invention is not limited thereto, but is intended to cover modifications as fall within the scope of the present invention.

Claims

1. The man-machine interface digital conversation mining method facing artificial intelligence application is characterized by being applied to an AI big data mining system, and comprises the following steps:

2. The method of claim 1, wherein processing the authenticated HMI digital conversation text using the base conversation text mining network to obtain a first interaction intent discrimination report and a first thermodynamic relationship network corresponding to the authenticated HMI digital conversation text, comprises:

3. The method of claim 2, wherein generating a first interaction intent discrimination report corresponding to the authenticated HMI digital session text through a base session text mining network, generating a first thermodynamic relationship network based on the first interaction intent discrimination report and the authenticated semantic description of the authenticated HMI digital session text, comprises:

expanding the initial thermodynamic relation network to obtain a first thermodynamic relation network with the same visual scale as the authenticated HMI digital session text;

the obtaining, according to the basic session text mining network, a target query vector corresponding to the authenticated question-answer event in the authenticated HMI digital session text includes:

vector aggregation is carried out on the scene query vector and paragraph query vectors corresponding to the X conversation text paragraphs, so that target query vectors corresponding to the authenticated question-answer events in the authenticated HMI digital conversation text are obtained;

the basic session text mining network comprises Y gradient focus sub-networks, each gradient focus sub-network comprises at least one semantic coding component, and Y is a positive integer; the step of obtaining a scene query vector corresponding to the authenticated question-answer event in the authenticated HMI digital session text in the basic session text mining network comprises the following steps:

determining a semantic description combination result of a Y-th gradient focus subnet as a scene query vector corresponding to the authenticated question-answer event in the authenticated HMI digital session text;

the number of the scene inquiry vectors is Z, and Z is a positive integer; the generating, by the discriminating component in the basic session text mining network, a scene interaction intention discriminating report corresponding to the scene query vector includes:

4. The method of claim 2, wherein said determining similarity cost information for the underlying conversational text-mining network from the first and second networks of thermodynamic relationships comprises:

5. The method of claim 2, wherein determining intent discrimination cost information for the underlying conversation text mining network based on the first interactive intent discrimination report, the second interactive intent discrimination report, and text block intent words carried by the authenticated HMI digital conversation text comprises:

6. The method of claim 2, wherein the modifying the configuration variables of the underlying conversational text-mining network to generate the debugged conversational text parsing model, in combination with the similarity cost information and the intent discrimination cost information, comprises:

7. The method of claim 2, wherein the method further comprises:

acquiring a mean value characteristic variable corresponding to the local characteristic cluster of the question-answer event, determining a distribution report of text blocks in the target question-answer event in the digital session text of the HMI to be processed according to the mean value characteristic variable, and determining an inquiry data prediction report corresponding to the target question-answer event in the digital session text of the HMI to be processed according to the local intention label of the question-answer event and the distribution report;

the acquiring the question-answer event query vector corresponding to the target question-answer event in the digital session text of the HMI to be processed through the debugged session text analysis model comprises the following steps:

8. An AI big data mining system, comprising: a processor, a memory, and a network interface; the processor is connected with the memory and the network interface; the network interface is for providing data communication functionality, the memory is for storing program code, and the processor is for invoking the program code to perform the artificial intelligence application oriented human-machine interface digital session mining method of any of claims 1-7.

9. A computer readable storage medium, having stored thereon a computer program, which when run performs the artificial intelligence application oriented human-machine interface digital session mining method of any of claims 1-7.

10. A computer program product comprising a computer program or computer-executable instructions which, when executed by a processor, implement the artificial intelligence application oriented human-machine interface digital session mining method of any of claims 1-7.