CN113609266B

CN113609266B - Resource processing method and device

Info

Publication number: CN113609266B
Application number: CN202110780305.9A
Authority: CN
Inventors: 施晨; 胡于响; 张增明; 邵亮; 姜飞俊
Original assignee: Alibaba Innovation Co
Current assignee: Alibaba Innovation Co
Filing date: 2021-07-09
Publication date: 2024-07-16
Anticipated expiration: 2041-07-09

Abstract

The embodiment of the specification provides a resource processing method and a device, wherein the resource processing method comprises the following steps: receiving a resource acquisition request carrying a resource to be acquired, determining historical preference information of a user for each resource category under the condition that the resource to be acquired corresponds to at least two resource categories, acquiring the current resource category contained in a target dialogue round in the current dialogue of the user, and determining the target resource category of the resource to be acquired based on the resource acquisition request, the historical preference information and the current resource category.

Description

Resource processing method and device

Technical Field

The embodiment of the specification relates to the technical field of computers, in particular to a resource processing method. One or more embodiments of the present specification relate to a resource processing apparatus, a computing device, and a computer-readable storage medium.

Background

With the development of technology, more and more artificial intelligence dialogue devices are developed, and users can make the artificial intelligence dialogue devices play video and audio or weather broadcast for the users in a dialogue interaction or action interaction mode with the artificial intelligence dialogue devices.

However, with the continuous expansion of service items that can be provided for users in the artificial intelligent dialogue device, the access of third party applications or related skills in the artificial intelligent dialogue device is more and more, and often, the same type of applications or skills are also gradually increased in the same artificial intelligent device. Many times, multiple applications or skills may be accommodated for the same request by the user. Therefore, how to provide corresponding services for users by selecting applications or skills that are more accurate and more suitable for the needs of the users according to the requests of the users, so as to improve the use experience of the users, is a problem that needs to be solved.

Disclosure of Invention

In view of this, the present embodiments provide a resource processing method. One or more embodiments of the present specification are also directed to a resource processing apparatus, a computing device, and a computer-readable storage medium, which address the technical deficiencies of the prior art.

According to a first aspect of embodiments of the present disclosure, there is provided a resource processing method, including:

receiving a resource acquisition request carrying a resource to be acquired, and determining historical preference information of a user for each resource category under the condition that the resource to be acquired corresponds to at least two resource categories;

Acquiring a current resource category contained in a target dialogue round in a current dialogue of a user;

And determining the target resource category of the resource to be acquired based on the resource acquisition request, the historical preference information and the current resource category.

Optionally, the determining, based on the resource acquisition request, the historical preference information and the current resource category, the target resource category of the resource to be acquired includes:

the resource acquisition request, the historical preference information and the current resource category are input into a prediction model to carry out probability prediction, and a probability prediction result of negative feedback submitted by the user to each resource category is generated;

And determining the target resource category of the resource to be acquired according to the probability prediction result.

Optionally, the performing probability prediction on the resource acquisition request, the historical preference information and the current resource category input prediction model, generating a probability prediction result that the user submits negative feedback to each resource category includes:

Taking the resource acquisition request, the historical preference information and the current resource category as an input set, and performing coding processing by a vector coding module of an input prediction model to generate a coding vector of the input set;

and inputting the coding vector into a probability prediction module of the prediction model to carry out probability prediction, and generating a probability prediction result of negative feedback submitted by the user to each resource category.

Optionally, the encoding processing is performed by the vector encoding module of the input prediction model with the resource acquisition request, the historical preference information and the current resource category as an input set, and the generating an encoded vector of the input set includes:

the resource acquisition request is input into a first vector coding module of a prediction model to be coded, and a first coding vector is generated;

inputting the history preference information into a second vector coding module of a prediction model to carry out coding processing to generate a second coding vector;

inputting the current resource category and each resource category into a third vector coding module of a prediction model for coding processing to generate a corresponding third coding vector;

And the first code vector, the second code vector and the third code vector are used as the code vectors of the resource acquisition request, the historical preference information and the current resource category.

Optionally, the encoding processing is performed by the third vector encoding module that inputs the current resource category and each resource category into the prediction model, and a corresponding third encoding vector is generated, including:

respectively inputting the current resource category as an input set to a third vector coding module of the prediction model for coding processing to generate a first sub-coding vector corresponding to each current resource category;

taking each resource category as an input set together to input a third vector coding module of a prediction model for coding processing, and generating a second sub-coding vector corresponding to the input set;

inputting the first sub-coding vector and the second sub-coding vector into an attention calculating module for attention calculation to generate a corresponding third sub-coding vector;

Inputting the second sub-coding vector into a full convolution network and an average pooling layer of a prediction model for processing to generate a corresponding fourth sub-coding vector;

And the third sub-coding vector and the fourth sub-coding vector are used as the current resource category and the third coding vector corresponding to each resource category.

Optionally, the probability prediction module for inputting the coding vector into the prediction model performs probability prediction, and generating a probability prediction result of submitting negative feedback to each resource category by the user includes:

The first coding vector, the second coding vector and the third coding vector are input into an attention calculation submodule in a probability prediction module to carry out attention calculation, and a corresponding attention calculation result is generated;

performing multi-task learning based on the attention calculation result, and generating an initial probability prediction result of submitting negative feedback to each task dimension by a user under each resource category;

And determining a probability prediction result of negative feedback submitted by the user to each resource category according to the initial probability prediction result.

Optionally, the performing multi-task learning based on the attention calculation result, generating an initial probability prediction result that the user submits negative feedback to each task dimension under each resource category, including:

Inputting the attention calculation result into a full convolution network in the probability prediction module to perform dimension reduction processing, and generating dimension reduction processing results corresponding to each task dimension under each resource category;

And processing the dimension reduction processing result by using an activation function to generate an initial probability prediction result of negative feedback submitted by a user to each task dimension under each resource category.

Optionally, the resource processing method further includes:

Inquiring a common sense information coding vector with a mapping relation with the resource to be obtained in a pre-established common sense information mapping table;

And determining the target resource category of the resource to be acquired based on the resource acquisition request, the historical preference information, the current resource category and the common sense information encoding vector.

Optionally, the resource processing method further includes:

extracting the resources to be acquired corresponding to the target resource category, and sending the resources to be acquired corresponding to the target resource category to the user to respond to the resource acquisition request.

Optionally, the resource processing method further includes:

Receiving feedback information submitted by the user for the resource to be acquired;

And under the condition that negative feedback is submitted by the user for the resources to be acquired corresponding to the target resource category according to the feedback information, adjusting the target resource category according to the historical preference information and the current resource category, and returning the resources to be acquired corresponding to the adjustment result to the user.

According to a second aspect of embodiments of the present specification, there is provided another resource processing method, including:

Receiving a voice instruction carrying an information resource acquisition request, and determining historical preference information of a user for each resource category under the condition that the information resource to be acquired in the information resource acquisition request corresponds to at least two resource categories;

acquiring a current resource category contained in a target dialogue round in a current dialogue of the user;

determining a target resource category of the information resource to be acquired based on the voice instruction, the historical preference information and the current resource category;

and extracting information resources to be acquired corresponding to the target resource category, and sending the information resources to the user so as to respond to the voice instruction.

According to a third aspect of embodiments of the present specification, there is provided a resource processing apparatus comprising:

The receiving module is configured to receive a resource acquisition request carrying a resource to be acquired, and determine historical preference information of a user for each resource category under the condition that the resource to be acquired corresponds to at least two resource categories;

The acquisition module is configured to acquire the current resource category contained in the target dialogue round in the current dialogue of the user;

And the determining module is configured to determine a target resource category of the resource to be acquired based on the resource acquisition request, the historical preference information and the current resource category.

According to a fourth aspect of embodiments of the present specification, there is provided a computing device comprising:

A memory and a processor;

The memory is for storing computer-executable instructions, and the processor is for executing the computer-executable instructions:

According to a fifth aspect of embodiments of the present description, there is provided a computer-readable storage medium storing computer-executable instructions which, when executed by a processor, implement the steps of the resource processing method.

According to one embodiment of the specification, a resource obtaining request carrying a resource to be obtained is received, under the condition that the resource to be obtained corresponds to at least two resource categories, historical preference information of a user for each resource category is determined, a current resource category contained in a target dialogue round in a current dialogue of the user is obtained, and the target resource category of the resource to be obtained is determined based on the resource obtaining request, the historical preference information and the current resource category.

After receiving a resource acquisition request of a user, the embodiment of the specification can provide sorting results of different resource categories for the user in a personalized way according to historical preference information of the user on each resource category and the resource category contained in a target dialogue round in the current dialogue of the user, and the sorting results are combined with the resources to be acquired in the resource acquisition request, so that the resources to be acquired of the target resource category which is more matched with the requirements of the user are selected according to the sorting results, thereby being beneficial to improving the accuracy of the resource acquisition results and improving the service experience of the user.

Drawings

FIG. 1 is a process flow diagram of a resource processing method provided by one embodiment of the present disclosure;

FIG. 2 is a schematic diagram of a resource handling process provided by one embodiment of the present disclosure;

FIG. 3 is a process flow diagram of a resource processing method according to one embodiment of the present disclosure;

FIG. 4 is a schematic diagram of a resource processing device according to one embodiment of the present disclosure;

FIG. 5 is a process flow diagram of another resource processing method provided by one embodiment of the present disclosure;

FIG. 6 is a schematic diagram of another resource processing device provided in one embodiment of the present disclosure;

FIG. 7 is a block diagram of a computing device provided in one embodiment of the present description.

Detailed Description

In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present description. This description may be embodied in many other forms than described herein and similarly generalized by those skilled in the art to whom this disclosure pertains without departing from the spirit of the disclosure and, therefore, this disclosure is not limited by the specific implementations disclosed below.

The terminology used in the one or more embodiments of the specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of the specification. As used in this specification, one or more embodiments and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used in one or more embodiments of the present specification refers to and encompasses any or all possible combinations of one or more of the associated listed items.

It should be understood that, although the terms first, second, etc. may be used in one or more embodiments of this specification to describe various information, these information should not be limited by these terms. These terms are only used to distinguish one type of information from another. For example, a first may also be referred to as a second, and similarly, a second may also be referred to as a first, without departing from the scope of one or more embodiments of the present description. The word "if" as used herein may be interpreted as "at … …" or "at … …" or "in response to a determination" depending on the context.

First, terms related to one or more embodiments of the present specification will be explained.

Negative feedback: when the user communicates with the artificial intelligence dialogue device, negative feedback behavior is generated due to dissatisfaction with the reply of the device.

Negative feedback rate: in a particular set of user requests, the number of requests comprising negative feedback actions of the user is proportional to the total number of requests.

The achievement rate is as follows: in a particular set of user requests, the number of requests for which user needs are successfully fulfilled is proportional to the total number of requests.

In the present specification, a resource processing method is provided, and the present specification relates to a resource processing apparatus, a computing device, and a computer-readable storage medium, which are described in detail in the following embodiments one by one.

After receiving the request of the user, the artificial intelligent dialogue device can perform flow distribution for the user to call related applications to provide a server for the user, but the current flow distribution system is relatively static, the distribution sequencing result of the request of the same sentence of users is fixed, the request is not changed along with different users or scenes, and the online user experience is not directly used as an algorithm to optimize a target adjustment strategy, so that the requirements of a new generation of intelligent assistant cannot be met. The new generation of intelligent flow distribution ordering model directly takes user experience as an online optimization target and has the characteristics of individuation, self-adaption and scene. Under the condition that the same sentence of user requests can be accepted in a plurality of fields or skills, how to develop a set of sequencing models which refer to different user preferences and different context dialogue scenes and adaptively adjust strategies according to user experience with time is a key for further improving the user experience and promoting development of skill ecology and quality.

Fig. 1 shows a process flow diagram of a resource processing method according to one embodiment of the present disclosure, including steps 102 to 106.

Step 102, receiving a resource acquisition request carrying a resource to be acquired, and determining historical preference information of a user for each resource category under the condition that the resource to be acquired corresponds to at least two resource categories.

The resource processing method provided in the embodiments of the present disclosure is applied to clients, including, but not limited to, clients such as large audio/video playing devices, game machines, desktop computers, smart phones, tablet computers, MP3 (Moving Picture Experts Group Audio Layer III, dynamic video expert compression standard audio layer 3) players, MP4 (Moving Picture Experts Group Audio Layer IV, dynamic video expert compression standard audio layer 4) players, laptop computers, electronic book readers, and other display terminals.

The user can send a resource acquisition request to the client, and after the client receives the resource acquisition request of the user, the client can provide the resource of the corresponding category for the user according to the resource category of the resource to be acquired in the resource acquisition request. However, in the case that the resource to be acquired has multiple resource categories, the client needs to determine the historical preference information of the user for each resource category, so as to select a resource of one category from the multiple resource categories according to the historical preference information, and return the resource to the user.

Specifically, the resources to be acquired, that is, the resources requested to be acquired by the user, include, but are not limited to, video resources, audio resources, text resources, and the like; the resource category, i.e. the category of the resource to be acquired, includes but is not limited to video category, audio category, text category, etc.

The client receives a resource acquisition request of the resource to be acquired, and can determine the historical preference information of the user for each resource category under the condition that the resource to be acquired corresponds to at least two resource categories.

Taking the resource acquisition request as "song M1 played by singer S1" as an example, since the resource category of song M1 includes two categories, i.e., audio (song) and video (music clip), in this case, the client needs to determine the historical preference information of the user for each resource category in the client to determine which category of resources is provided for the user.

In practical applications, the historical preference information of the user for each resource category includes, but is not limited to, the number of resource acquisition requests sent to the client by the user for 3 months/1 month/7 days/1 day, the negative feedback information submitted by the user to the client, the number of times each resource category is called for 3 months/1 month/7 days/1 day, the negative feedback information of each resource category, the number of resource acquisition requests submitted by the user under the resource category for 3 months/1 month/7 days/1 day, the negative feedback information submitted by the user for the resource category for 3 months/1 month/7 days/1 day, the first 10 resource categories preferred by the user for 3 months/7 days, the number of times the user submitted the resource acquisition request under the resource categories, the negative feedback information submitted by the user, and the like.

Step 104, obtaining the current resource category contained in the target dialogue round in the current dialogue of the user.

Specifically, if the user performs a session interaction with the client, that is, the user sends a resource acquisition request to the client by performing a session with the client, the current session, that is, the session carrying the resource acquisition request, may include other sessions except the session carrying the resource acquisition request, and may include multiple session rounds; one dialog turn is a turn of dialog performed by the user and the client; the target session may be 1 to 3 session cycles before the session in which the session carrying the resource acquisition request is located.

In addition, the current resource category, i.e. the resource category included in the target dialog turn, characterizes a current resource category that is mentioned in the process of dialog interaction with the client before submitting the resource acquisition request.

Corresponding to the foregoing history preference information, the current resource category is equivalent to real-time information of the user, and the embodiment of the present disclosure combines the history preference information and the real-time information to determine the target resource category of the resource to be acquired, which is beneficial to improving accuracy of the determination result.

And step 106, determining the target resource category of the resource to be acquired based on the resource acquisition request, the historical preference information and the current resource category.

Specifically, after the historical preference information of the user for each resource category is determined and the current resource category included in the target dialog turn in the current dialog of the user is acquired, the target resource category of the resource to be acquired can be determined based on the resource acquisition request, the historical preference information and the current resource category.

In the implementation, the determining the target resource category of the resource to be acquired based on the resource acquisition request, the historical preference information and the current resource category may be implemented in the following manner:

Specifically, after the historical preference information of the user for each resource category is determined and the current resource category included in the target dialog turn in the current dialog of the user is acquired, the resource acquisition request, the historical preference information and the current resource category can be input into a prediction model, the probability of submitting negative feedback to each resource category by the user is predicted by the prediction model, a corresponding prediction result is generated, and then the resource category with lower probability of submitting negative feedback by the user can be selected as the target resource category of the resource to be acquired according to the prediction result.

In specific implementation, probability prediction is performed on the resource acquisition request, the historical preference information and the current resource category input prediction model, and a probability prediction result of the user submitting negative feedback to each resource category is generated, which includes:

Further, taking the resource acquisition request, the historical preference information and the current resource category as an input set, and performing encoding processing by a vector encoding module of an input prediction model to generate an encoding vector of the input set, wherein the encoding method comprises the following steps:

Further, the encoding process is performed on the current resource category and the third vector encoding module of the prediction model input by each resource category, so as to generate a corresponding third encoding vector, which includes:

Specifically, the prediction model comprises a vector coding module and a probability prediction module, wherein the vector coding module is used for coding the resource acquisition request, the historical preference information and the current resource category; the probability prediction module is used for carrying out probability prediction on each resource category according to the coding result output by the vector coding module so as to obtain a probability prediction result of submitting negative feedback to each resource category by a user.

In addition, since there is a difference in the resource acquisition request, the history preference information, and the information contained in the current resource category, for example: the resource acquisition request and the current resource category contain some semantic information, and the history preference information contains some numerical information, so that for different input information, the prediction model adopts different vector coding modules to code the information so as to obtain more accurate coding results.

In the embodiment of the present disclosure, a first vector encoding module in a prediction model is used to encode a resource acquisition request to generate a first encoded vector, where the first vector encoding module may use an encoder of a transform model; a third vector coding module of the prediction model is adopted to code the current resource category and each resource category to generate a third coding vector, and the third vector coding module can also adopt an encoder of a transform model; and a second vector coding module adopting a prediction model is used for coding the history preference information to generate a second coding vector, wherein the second vector coding module can be realized by adopting a stat coding layer.

After the first code vector, the second code vector and the third code vector are obtained by coding, the first code vector, the second code vector and the third code vector can be used as code vectors of a resource acquisition request, history preference information and a current resource category.

A schematic diagram of a resource processing procedure provided in the embodiment of the present disclosure is shown in fig. 2. After the resource acquisition request, the historical preference information and the current resource category are acquired, the resource acquisition request, the historical preference information and the current resource category are input into a prediction model for processing, and the prediction model comprises a vector coding module and a probability prediction module, so that after the resource acquisition request, the historical preference information and the current resource category are input into the prediction model, the vector coding module firstly codes the resource acquisition request, the historical preference information and the current resource category to generate a corresponding coding vector, and then the probability prediction module predicts the probability of submitting negative feedback to each resource category by a user based on a coding result so as to screen the target resource category of the resource to be acquired according to the prediction result.

As shown in fig. 2, the input of the prediction model includes a request text (Query), a resource category (Item), a current resource category (User real-time information, user real-time behavior) and historical preference information corresponding to the resource acquisition request, including User profile (User data), item profile (resource category data), userItem profile (resource category data of the User), and User preference (User preference).

The item is all resource categories contained in the client; user real-time behavior is User real-time information, specifically, in the current dialogue, the category of the resource contained in 1 to 3 rounds of dialogue before the dialogue round containing the resource acquisition request; user profile, i.e. the number of resource acquisition requests, submitted negative feedback information, etc. of the User over 3 months/1 month/7 days/1 day; item profile, i.e., the number of times an Item was invoked 3 months/1 month/7 days/1 day, negative feedback information submitted by a user for the Item, etc.; UI profile, namely the number of resource acquisition requests submitted by the user in the past 3 months/1 month/7 days/1 day, submitted negative feedback information and the like under the item; user preference, i.e., the first 10 resource categories that the User prefers over 3 months/7 days, as well as the number of resource acquisition requests submitted by the User under these resource categories, negative feedback information submitted, etc.

After inputting the information into a prediction model, a text coding layer Query Encoding Layer in a vector coding module codes the Query to generate a first coding vector; encoding the User profile, the Item profile, the UI profile and the User preference by a data encoding layer Stat Encoding Layer to generate a second encoding vector; the resource category coding layer Item Encoding Layer codes each resource category contained in the User real-time behavior to generate a first sub-coding vector corresponding to each current resource category; encoding the Item by a resource category encoding layer Item Encoding Layer to generate a corresponding second sub-encoding vector; then, the attention calculating module (text attention layer Context Attention) performs attention calculation on the first subcode vector and the second subcode vector to generate a corresponding third subcode vector; inputting the second sub-coded vector into a full convolution network (FCN Layer) for processing, and inputting the processing result into an average pooling Layer (Avg Pooling) for pooling processing to generate a fourth sub-coded vector; the third sub-coding vector and the fourth sub-coding vector can be used together as the current resource category and the third coding vector corresponding to each resource category.

Or after the average pooling layer performs pooling processing on the result output by the full convolution network, the generated fourth sub-coding vector is input into the drop out layer to be processed, and a fifth sub-coding vector is generated, wherein the third sub-coding vector and the fifth sub-coding vector can be jointly used as the current resource category and the third coding vector corresponding to each resource category.

In addition, the coding vector is input into a probability prediction module of the prediction model to carry out probability prediction, and a probability prediction result of negative feedback submitted by the user to each resource category is generated, which can be specifically realized by the following modes:

Further, performing multi-task learning based on the attention calculation result, generating an initial probability prediction result of negative feedback submitted by the user to each task dimension under each resource category, including:

Specifically, in the embodiment of the present disclosure, a schematic diagram of a probability prediction module in a prediction model is shown in fig. 2, after a vector encoding module encodes a resource acquisition request, history preference information, and a current resource category to generate a first encoding vector, a second encoding vector, and a third encoding vector, the first encoding vector, the second encoding vector, and the third encoding vector may be input into the probability prediction module, so that an Attention calculation sub-module (Target Attention layer Attention) in the probability prediction module performs Attention calculation on the first encoding vector, the second encoding vector, and the third encoding vector to generate corresponding Attention calculation results, and then performs multi-task learning based on the Attention calculation results to generate a probability prediction result that a user submits negative feedback to each resource category.

Because the embodiment of the specification needs to predict the probability of submitting negative feedback to different resource categories by the user, in practical application, whether the user submits negative feedback to the resource categories is mainly influenced by 8 different task dimension factors, such as changing intention, feedback error given to the user by the client, and the like; therefore, in order to ensure the accuracy of the probability prediction result, after the attention calculation result is generated, multi-task learning can be performed based on the attention calculation result, each task dimension is used as a learning task, the negative feedback probabilities of different task dimensions are respectively predicted, and the probability that the total user submits negative feedback to each resource category is determined according to the prediction result.

The specific multi-task learning process is that sub-coding vectors corresponding to different task dimensions are extracted from the attention calculation result, and then the different sub-coding vectors are respectively input into a full convolution network (FCN Layer) in a probability prediction module to carry out dimension reduction processing, so that dimension reduction processing results corresponding to the task dimensions under each resource category are generated; then, the dimension reduction processing results are processed by using an activation function 1 and an activation function 2, and an initial probability prediction result of negative feedback submitted by a user to each task dimension under each resource category is generated; wherein, the activation function 1 may be a linear unit function Leaky Relu with leakage correction, and the activation function 2 may be a Sigmoid function; and finally, calculating the probability prediction result of the user submitting the total negative feedback to each resource category according to the weights occupied by different task dimensions and the probability prediction results corresponding to different task dimensions.

The accuracy of the probability prediction result is improved by multi-task learning.

In addition, after receiving the resource acquisition request of the user, the common sense information encoding vector having a mapping relationship with the resource to be acquired may be queried in a pre-established common sense information mapping table, and the target resource category of the resource to be acquired may be determined based on the resource acquisition request, the history preference information, the current resource category and the common sense information encoding vector.

Specifically, as shown in fig. 2, in addition to inputting the resource acquisition request, the history preference information and the current resource category into the prediction model, the common sense information encoding vector may be also input into the prediction model, where the difference is that the information encoding module in the prediction model needs to encode the resource acquisition request, the history preference information and the current resource category, and does not need to encode the common sense information encoding vector; the vector encoding module encodes the resource acquisition request, the history preference information and the current resource category to generate a first encoding vector, a second encoding vector and a third encoding vector, and then inputs the first encoding vector, the second encoding vector, the third encoding vector and the common sense information encoding vector into the probability prediction module for processing.

In practical application, the common sense information encoding vector can be obtained by querying a pre-established common sense information mapping table, and the common sense information mapping table can be updated by a user. Since the common sense information in the embodiment of the present specification refers to the category 41, such as "Person", "Artist", "Organization", etc., and typically three granularity, namely, the domain, the intention, and the value of the information slot, are involved in one resource acquisition request, the client may determine the value of the information slot in the resource acquisition request after acquiring the resource acquisition request, and then query the pre-established common sense information mapping table for the common sense information encoding vector corresponding to the value of the information slot, and the common sense information encoding vector is 41 dimensions, and the value of each dimension is a score of one common sense information, such as 0.2, 0.6, or 1.0, and this score is used to characterize the confidence of each common sense information.

Taking the resource obtaining request as an example of "playing the song M1 of the singer S1", the value of the information slot may be extracted here to be the singer S1 or the song M1, and then the common sense information encoding vector corresponding to the singer S1 or the song M1 may be queried in the pre-established common sense information mapping table, and the target resource category of the resource to be obtained may be determined based on the resource obtaining request, the history preference information, the current resource category and the common sense information encoding vector.

Finally, after determining the target resource category of the resource to be acquired, the resource to be acquired corresponding to the target resource category can be extracted, and the resource to be acquired corresponding to the target resource category is sent to the user to respond to the resource acquisition request. For example, if the target resource category is audio, a song is played for the user; if the target asset class is a music clip, then a music clip (MV) of song M1 is played for the user.

In this embodiment of the present disclosure, after sending a resource to be acquired corresponding to a target resource category to a user, the user may submit feedback for the resource to be acquired, and may also adjust the resource category of the resource to be acquired for the user if the feedback submitted by the user is negative feedback, which may be specifically implemented by:

Specifically, under the condition that feedback information of a user aiming at the resource to be acquired corresponding to the target resource category is received, whether the user submits negative feedback or not can be determined according to the feedback information, if the user submits the negative feedback, the resource to be acquired of the target resource category does not meet the requirement of the user, therefore, the resource category meeting the requirement of the user can be redetermined according to the historical preference information of the user on the resource category and the current resource category and combined with the feedback information, namely, the target resource category is adjusted, and the resource to be acquired corresponding to the adjustment result is returned to the user.

For example, when the target resource type is audio, a song is played for the user through the player a, but when negative feedback of "change one player" submitted by the user is received during the playing process of the song, the player B for playing the song for the user can be redetermined according to the feedback information in combination with the historical preference information of the user and the current resource type, and the song is continuously played for the user through the player B.

The embodiment of the specification provides an intelligent flow distribution sequencing system which directly takes negative behavioral feedback of a user as an online optimization target and has individuation, self-adaption and scene. The system can reduce negative feedback of online users, improve the user demand achievement rate, and individually give the sequencing results according to users with different preferences, give the sequencing results in different context dialogue histories in a scene mode, and adaptively adjust the sequencing strategy according to user experience over time. In addition, the item list under each resource acquisition request of each user is input in the scheme, and the item list has pointwise property (pointwise), so that a large amount of training data which is easy to collect is trained; according to the scheme, the Query and the Item are respectively encoded, and the semantic information of the Query can be directly referred in downstream sequencing, so that error conduction caused by analysis errors of an upstream natural language understanding module is avoided; according to the scheme, a randomly generated negative sampling sample is particularly added in the item coding module, so that semantic mismatching of item dimensions is avoided to a certain extent.

The resource processing method provided in the present specification is further described below by taking an application of the resource processing method in an actual scenario as an example with reference to fig. 3. Fig. 3 is a flowchart of a processing procedure of a resource processing method according to an embodiment of the present disclosure, where specific steps include steps 302 to 332.

Step 302, receiving a resource acquisition request carrying a resource to be acquired, and determining historical preference information of a user for each resource category under the condition that the resource to be acquired corresponds to at least two resource categories.

Step 304, obtaining a current resource category included in the target dialogue round in the current dialogue of the user.

And 306, performing coding processing on the first vector coding module of the resource acquisition request input prediction model to generate a first coding vector.

And 308, inputting the history preference information into a second vector coding module of the prediction model to carry out coding processing, and generating a second coding vector.

And step 310, respectively taking the current resource category as an input set to be input into a third vector coding module of the prediction model for coding processing, and generating a first sub-coding vector corresponding to each current resource category.

And 312, taking each resource category as an input set to input a third vector coding module of the prediction model to carry out coding processing, and generating a second sub-coding vector corresponding to the input set.

In step 314, the first sub-encoded vector and the second sub-encoded vector are input to an attention calculation module for performing attention calculation, so as to generate a corresponding third sub-encoded vector.

Step 316, inputting the second sub-coding vector into the full convolution network and the average pooling layer of the prediction model for processing, and generating a corresponding fourth sub-coding vector.

And step 318, the third sub-coding vector and the fourth sub-coding vector are used together as the current resource category and the third coding vector corresponding to each resource category.

Step 320, using the first code vector, the second code vector, and the third code vector together as the code vectors of the resource acquisition request, the history preference information, and the current resource category.

Step 322, query the common sense information encoding vector having a mapping relationship with the resource to be obtained in the pre-established common sense information mapping table.

Step 324, inputting the first encoding vector, the second encoding vector, the third encoding vector and the common sense information encoding vector into an attention calculating submodule in the probability prediction module for performing attention calculation, and generating a corresponding attention calculating result.

And 326, inputting the attention calculation result into a full convolution network in the probability prediction module to perform dimension reduction processing, and generating dimension reduction processing results corresponding to each task dimension under each resource category.

And step 328, processing the dimension reduction processing result by using an activation function, and generating an initial probability prediction result of submitting negative feedback to each task dimension by a user under each resource category.

And step 330, determining a probability prediction result of negative feedback submitted by the user to each resource category according to the initial probability prediction result.

And step 332, determining the target resource category of the resource to be acquired according to the probability prediction result.

The present embodiments customize the coding structure for different features for the input features described above. Specifically, for the Query feature and the Item feature, encoding is performed by using a text encoder based on a transducer respectively; for User real-time behavior, firstly, splitting the User real-time behavior into independent items, then adding position codes, and passing through dropout after a full connection layer; furthermore, to capture the semantic associations between these historical and current items, we have particularly added text-based attention mechanisms between the two.

After all feature codes are finished, the embodiment of the specification specifically introduces a layer of attention mechanism module to perform weighted combination of the features. The method comprises the steps of firstly splicing all the features together to be used as target features, and comparing an original feature list with the target features based on an attention mechanism to obtain the weight of each feature to the target features. And finally, weighting each feature based on the weight to obtain a final processed feature vector.

In addition, in the embodiment of the specification, multi-task learning of each sub-dimension (such as playing interruption, repeated text sending and the like) aiming at the overall user negative feedback and negative feedback is adopted, the prediction of the overall negative feedback is taken as a main task, and the predictions of other dimensions are taken as auxiliary tasks. Considering that the importance of each subtask is different in different scenes (for example, in an singing order scene, the importance of playing interruption is significantly higher than that in other scenes), such a multi-task learning mechanism facilitates that we flexibly adjust strategies for scenes.

Corresponding to the above method embodiments, the present disclosure further provides an embodiment of a resource processing device, and fig. 4 shows a schematic diagram of a resource processing device provided in one embodiment of the present disclosure. As shown in fig. 4, the apparatus includes:

A receiving module 402, configured to receive a resource acquisition request carrying a resource to be acquired, and determine historical preference information of a user for each resource category if it is determined that the resource to be acquired corresponds to at least two resource categories;

An obtaining module 404 configured to obtain a current resource category included in a target dialog turn in a current dialog of the user;

a determining module 406 is configured to determine a target resource category of the resource to be acquired based on the resource acquisition request, the historical preference information and the current resource category.

Optionally, the determining module 406 includes:

The prediction sub-module is configured to carry out probability prediction on the resource acquisition request, the historical preference information and the current resource category input prediction model, and generate a probability prediction result of negative feedback submitted by the user to each resource category;

and the determining submodule is configured to determine the target resource category of the resource to be acquired according to the probability prediction result.

Optionally, the prediction submodule includes:

the encoding unit is configured to take the resource acquisition request, the historical preference information and the current resource category as an input set, and a vector encoding module of an input prediction model performs encoding processing to generate an encoding vector of the input set;

and the prediction unit is configured to input the coding vector into a probability prediction module of the prediction model to carry out probability prediction and generate a probability prediction result of the user submitting negative feedback to each resource category.

Optionally, the encoding unit includes:

A first coding subunit configured to perform coding processing on a first vector coding module of the resource acquisition request input prediction model to generate a first coding vector;

A second encoding subunit configured to input the history preference information into a second vector encoding module of the prediction model for encoding processing, and generate a second encoded vector;

The third coding subunit is configured to code the current resource category and a third vector coding module of each resource category input prediction model to generate a corresponding third coding vector;

A processing subunit configured to use the first encoding vector, the second encoding vector, and the third encoding vector together as encoding vectors for the resource acquisition request, the history preference information, and the current resource category.

Optionally, the third encoding subunit is further configured to:

Optionally, the prediction unit includes:

A calculation subunit configured to input the first coding vector, the second coding vector and the third coding vector into an attention calculation submodule in the probability prediction module to perform attention calculation and generate a corresponding attention calculation result;

The generating subunit is configured to perform multi-task learning based on the attention calculation result and generate an initial probability prediction result of submitting negative feedback to each task dimension by a user under each resource category;

A determining subunit configured to determine a probability prediction result of the user submitting negative feedback to each resource category according to the initial probability prediction result.

Optionally, the generating subunit is further configured to:

Optionally, the resource processing device further includes a query module configured to:

Optionally, the resource processing device further includes:

and the extraction module is configured to extract the resources to be acquired corresponding to the target resource category and send the resources to be acquired corresponding to the target resource category to the user so as to respond to the resource acquisition request.

Optionally, the resource processing device further includes an adjustment module configured to:

The above is a schematic scheme of a resource processing device of the present embodiment. It should be noted that, the technical solution of the resource processing device and the technical solution of the resource processing method belong to the same conception, and details of the technical solution of the resource processing device, which are not described in detail, can be referred to the description of the technical solution of the resource processing method.

Fig. 5 shows a process flow diagram of another resource processing method provided in accordance with one embodiment of the present disclosure, including steps 502 through 506.

Step 502, receiving a voice command carrying an information resource obtaining request, and determining historical preference information of a user for each resource category under the condition that the information resource to be obtained in the information resource obtaining request corresponds to at least two resource categories.

Step 504, obtaining a current resource category included in the target dialog turn in the current dialog of the user.

Step 506, determining a target resource category of the information resource to be acquired based on the voice command, the historical preference information and the current resource category.

And step 508, extracting information resources to be acquired corresponding to the target resource category and sending the information resources to the user so as to respond to the voice instruction.

Specifically, the resource processing method provided in the embodiments of the present disclosure is applied to a client, where the client may be an intelligent robot for man-machine interaction, and a user may perform voice interaction with the client by sending a voice command, that is, send a resource acquisition request to the client by sending a voice command, and after receiving the voice command carrying an information resource acquisition request of the user, the client may provide a resource of a corresponding class for the user according to a resource class of an information resource to be acquired in the information resource acquisition request. However, in the case that two or more resource categories exist in the information resource to be acquired, the client needs to determine the historical preference information of the user for each resource category, acquire the current resource category included in the current dialogue of the user in a target dialogue round, select the information resource to be acquired in one target resource category from the two or more resource categories according to the historical preference information, the current resource category and the information resource acquisition request, and send the information resource to the user to respond to the voice command.

In practical applications, the information resource to be acquired may be a multimedia resource, such as an audio or video resource. The user can instruct the intelligent robot to play audio or video resources by sending a voice command, and after receiving the voice command, if two or more players available for playing audio or video exist in the intelligent robot, the intelligent robot can select a target player meeting the user requirement from the two or more players according to the historical preference information of the user on each player and the players mentioned in the process of dialogue interaction between the user and the client before submitting the voice command, and play the audio or video by utilizing the target player to respond to the voice command.

Or the information resource to be acquired may be an information-type resource such as news information. The user can instruct the intelligent robot to display news information by sending a voice command, and after receiving the voice command, if two or more applications available for displaying news information exist, the intelligent robot can select a target application meeting the user requirement from the two or more applications according to the historical preference information of the user for each application and the applications mentioned in the process of dialogue interaction between the user and the client before submitting the voice command, and display the news information by utilizing the target application to respond to the voice command.

Or the information resource to be obtained can also be commodity to be traded, such as lipstick, household appliance and the like. The user can instruct the intelligent robot to recommend commodities to the user in a mode of sending a voice command, after receiving the voice command, if two or more trading platforms which can be used for commodity trade exist in the intelligent robot, the intelligent robot can select a target trading platform meeting the user requirement from the two or more trading platforms according to historical preference information of the user on each trading platform and the trading platform mentioned in the process of dialogue interaction between the user and the client before submitting the voice command, and the commodity recommendation is performed by utilizing the target trading platform to respond to the voice command; in addition, after receiving the voice command, if only one transaction platform for commodity transaction exists, but multiple types or brands of commodities to be transacted may exist in the platform, the intelligent robot may select a target commodity type or commodity brand meeting the user demand from multiple commodity types or brands according to historical preference information of the user for each type or brand and commodity types or brands mentioned in the process of dialogue interaction between the user and the client before submitting the voice command, and conduct commodity recommendation in response to the voice command.

In addition, the user can interact with the client through a mode of sending voice instructions, and can send corresponding instructions through a mode of inputting words, after the client obtains the words input by the user, the information resources to be obtained of the user can be determined through a mode of word recognition and semantic recognition, so that the information resources to be obtained of the target resource type are sent to the user to respond to the instructions, and a specific implementation process is similar to a processing process of the voice instructions of the user and is not repeated.

After receiving the voice command carrying the information resource obtaining request, the embodiment of the specification can provide the sorting results of different resource categories for the user in a personalized way according to the historical preference information of the user on each resource category and the resource category contained in the target dialogue round in the current dialogue of the user, and the information resource to be obtained in the information resource obtaining request is combined, so that the information resource to be obtained of the target resource category which is more suitable for the requirement of the user is selected according to the sorting results, thereby being beneficial to improving the accuracy of the information resource obtaining result and the service experience of the user.

The above is a schematic scheme of another resource processing method of the present embodiment. It should be noted that, the technical solution of the resource processing method and the technical solution of the resource processing method described above belong to the same concept, and details of the technical solution of the resource processing method which are not described in detail may be referred to the description of the technical solution of the resource processing method described above.

Corresponding to the above method embodiments, the present disclosure further provides an embodiment of a resource processing device, and fig. 6 shows a schematic diagram of another resource processing device provided in one embodiment of the present disclosure. As shown in fig. 6, the apparatus includes:

The instruction receiving module 602 is configured to receive a voice instruction carrying an information resource obtaining request, and determine historical preference information of a user for each resource category under the condition that it is determined that an information resource to be obtained in the information resource obtaining request corresponds to at least two resource categories;

a resource category obtaining module 604 configured to obtain a current resource category included in a target dialog turn in a current dialog of the user;

A resource category determination module 606 configured to determine a target resource category of the information resource to be acquired based on the voice instruction, the history preference information, and the current resource category;

and a response module 608, configured to extract information resources to be acquired corresponding to the target resource category and send the information resources to the user, so as to respond to the voice instruction.

The above is another illustrative version of the resource processing device of the present embodiment. It should be noted that, the technical solution of the resource processing device and the technical solution of the other resource processing method belong to the same concept, and details of the technical solution of the resource processing device, which are not described in detail, can be referred to the description of the technical solution of the other resource processing method.

Fig. 7 illustrates a block diagram of a computing device 700 provided in accordance with one embodiment of the present description. The components of computing device 700 include, but are not limited to, memory 710 and processor 720. Processor 720 is coupled to memory 710 via bus 730, and database 750 is used to store data.

Computing device 700 also includes access device 740, access device 740 enabling computing device 700 to communicate via one or more networks 760. Examples of such networks include the Public Switched Telephone Network (PSTN), a Local Area Network (LAN), a Wide Area Network (WAN), a Personal Area Network (PAN), or a combination of communication networks such as the internet. The access device 740 may include one or more of any type of network interface, wired or wireless (e.g., a Network Interface Card (NIC)), such as an IEEE802.11 Wireless Local Area Network (WLAN) wireless interface, a worldwide interoperability for microwave access (Wi-MAX) interface, an ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a bluetooth interface, a Near Field Communication (NFC) interface, and so forth.

In one embodiment of the present description, the above-described components of computing device 700, as well as other components not shown in FIG. 7, may also be connected to each other, such as by a bus. It should be understood that the block diagram of the computing device illustrated in FIG. 7 is for exemplary purposes only and is not intended to limit the scope of the present description. Those skilled in the art may add or replace other components as desired.

Computing device 700 may be any type of stationary or mobile computing device including a mobile computer or mobile computing device (e.g., tablet, personal digital assistant, laptop, notebook, netbook, etc.), mobile phone (e.g., smart phone), wearable computing device (e.g., smart watch, smart glasses, etc.), or other type of mobile device, or a stationary computing device such as a desktop computer or PC. Computing device 700 may also be a mobile or stationary server.

Wherein the memory 710 is configured to store computer-executable instructions and the processor 720 is configured to execute the computer-executable instructions as follows:

The foregoing is a schematic illustration of a computing device of this embodiment. It should be noted that, the technical solution of the computing device and the technical solution of the resource processing method belong to the same concept, and details of the technical solution of the computing device, which are not described in detail, can be referred to the description of the technical solution of the resource processing method.

An embodiment of the present specification also provides a computer-readable storage medium storing computer instructions that, when executed by a processor, are configured to implement the steps of the resource processing method.

The above is an exemplary version of a computer-readable storage medium of the present embodiment. It should be noted that, the technical solution of the storage medium and the technical solution of the resource processing method belong to the same concept, and details of the technical solution of the storage medium which are not described in detail can be referred to the description of the technical solution of the resource processing method.

The foregoing describes specific embodiments of the present disclosure. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

The computer instructions include computer program code that may be in source code form, object code form, executable file or some intermediate form, etc. The computer readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a U disk, a removable hard disk, a magnetic disk, an optical disk, a computer Memory, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), an electrical carrier signal, a telecommunications signal, a software distribution medium, and so forth. It should be noted that the computer readable medium contains content that can be appropriately scaled according to the requirements of jurisdictions in which such content is subject to legislation and patent practice, such as in certain jurisdictions in which such content is subject to legislation and patent practice, the computer readable medium does not include electrical carrier signals and telecommunication signals.

It should be noted that, for simplicity of description, the foregoing method embodiments are all expressed as a series of combinations of actions, but it should be understood by those skilled in the art that the embodiments are not limited by the order of actions described, as some steps may be performed in other order or simultaneously according to the embodiments of the present disclosure. Further, those skilled in the art will appreciate that the embodiments described in the specification are all preferred embodiments, and that the acts and modules referred to are not necessarily all required for the embodiments described in the specification.

In the foregoing embodiments, the descriptions of the embodiments are emphasized, and for parts of one embodiment that are not described in detail, reference may be made to the related descriptions of other embodiments.

The preferred embodiments of the present specification disclosed above are merely used to help clarify the present specification. Alternative embodiments are not intended to be exhaustive or to limit the invention to the precise form disclosed. Obviously, many modifications and variations are possible in light of the teaching of the embodiments. The embodiments were chosen and described in order to best explain the principles of the embodiments and the practical application, to thereby enable others skilled in the art to best understand and utilize the invention. This specification is to be limited only by the claims and the full scope and equivalents thereof.

Claims

1. A resource processing method, comprising:

Inputting the resource acquisition request and the history preference information into a vector coding module of a prediction model to carry out coding processing to generate a first coding vector and a second coding vector;

Taking each resource category as an input set to input a third vector coding module of a prediction model to carry out coding processing, and generating a second sub-coding vector corresponding to the input set;

inputting the second sub-coding vector into a full convolution network and an average pooling layer of the prediction model for processing to generate a corresponding fourth sub-coding vector;

the third sub-coding vector and the fourth sub-coding vector are used together as the current resource category and the third coding vector corresponding to each resource category;

The first code vector, the second code vector and the third code vector are used as the code vectors of the resource acquisition request, the history preference information and the current resource category together;

Inputting the coding vector into a probability prediction module of the prediction model to carry out probability prediction, and generating a probability prediction result of submitting negative feedback to each resource category by the user;

2. The resource processing method according to claim 1, wherein the vector encoding module that inputs the resource acquisition request and the history preference information into the prediction model performs encoding processing to generate a first encoded vector and a second encoded vector, comprising:

and inputting the history preference information into a second vector coding module of the prediction model to carry out coding processing to generate a second coding vector.

3. The resource processing method according to claim 1, wherein the probability prediction module for inputting the coding vector into the prediction model performs probability prediction, and generating a probability prediction result of the user submitting negative feedback to each resource category includes:

4. The resource processing method according to claim 3, wherein the performing the multi-task learning based on the attention calculation result generates an initial probability prediction result that the user submits negative feedback to each task dimension under each resource category, including:

5. The resource processing method according to claim 1, further comprising:

6. The resource processing method according to claim 1, further comprising:

7. The resource processing method according to claim 6, further comprising:

8. A resource processing method, comprising:

Inputting the voice instruction and the history preference information into a vector coding module of a prediction model to carry out coding processing to generate a first coding vector and a second coding vector;

Determining a target resource category of the information resource to be acquired according to the probability prediction result;

9. A resource processing apparatus, comprising:

The determining module is configured to input the resource acquisition request and the history preference information into a vector coding module of a prediction model for coding processing to generate a first coding vector and a second coding vector; respectively inputting the current resource category as an input set to a third vector coding module of the prediction model for coding processing to generate a first sub-coding vector corresponding to each current resource category; taking each resource category as an input set to input a third vector coding module of a prediction model to carry out coding processing, and generating a second sub-coding vector corresponding to the input set; inputting the first sub-coding vector and the second sub-coding vector into an attention calculating module for attention calculation to generate a corresponding third sub-coding vector; inputting the second sub-coding vector into a full convolution network and an average pooling layer of the prediction model for processing to generate a corresponding fourth sub-coding vector;

The third sub-coding vector and the fourth sub-coding vector are used together as the current resource category and the third coding vector corresponding to each resource category; the first code vector, the second code vector and the third code vector are used as the code vectors of the resource acquisition request, the history preference information and the current resource category together; inputting the coding vector into a probability prediction module of the prediction model to carry out probability prediction, and generating a probability prediction result of submitting negative feedback to each resource category by the user; and determining the target resource category of the resource to be acquired according to the probability prediction result.

10. A computing device, comprising:

A memory and a processor;

The memory is configured to store computer executable instructions, and the processor is configured to implement the steps of the resource processing method of any of claims 1 to 8 when the computer executable instructions are executed.

11. A computer readable storage medium storing computer instructions which, when executed by a processor, implement the steps of the resource processing method of any of claims 1 to 8.